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 "amdgpu_atomfirmware.h" 33 #include "imu_v12_0.h" 34 #include "soc24.h" 35 #include "nvd.h" 36 37 #include "gc/gc_12_0_0_offset.h" 38 #include "gc/gc_12_0_0_sh_mask.h" 39 #include "soc24_enum.h" 40 #include "ivsrcid/gfx/irqsrcs_gfx_11_0_0.h" 41 42 #include "soc15.h" 43 #include "soc15d.h" 44 #include "clearstate_gfx12.h" 45 #include "v12_structs.h" 46 #include "gfx_v12_0.h" 47 #include "nbif_v6_3_1.h" 48 #include "mes_v12_0.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 MODULE_FIRMWARE("amdgpu/gc_12_0_0_pfp.bin"); 56 MODULE_FIRMWARE("amdgpu/gc_12_0_0_me.bin"); 57 MODULE_FIRMWARE("amdgpu/gc_12_0_0_mec.bin"); 58 MODULE_FIRMWARE("amdgpu/gc_12_0_0_rlc.bin"); 59 MODULE_FIRMWARE("amdgpu/gc_12_0_0_toc.bin"); 60 MODULE_FIRMWARE("amdgpu/gc_12_0_1_pfp.bin"); 61 MODULE_FIRMWARE("amdgpu/gc_12_0_1_me.bin"); 62 MODULE_FIRMWARE("amdgpu/gc_12_0_1_mec.bin"); 63 MODULE_FIRMWARE("amdgpu/gc_12_0_1_rlc.bin"); 64 MODULE_FIRMWARE("amdgpu/gc_12_0_1_toc.bin"); 65 66 static const struct amdgpu_hwip_reg_entry gc_reg_list_12_0[] = { 67 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS), 68 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS2), 69 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS3), 70 SOC15_REG_ENTRY_STR(GC, 0, regCP_STALLED_STAT1), 71 SOC15_REG_ENTRY_STR(GC, 0, regCP_STALLED_STAT2), 72 SOC15_REG_ENTRY_STR(GC, 0, regCP_STALLED_STAT3), 73 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPC_STALLED_STAT1), 74 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPF_STALLED_STAT1), 75 SOC15_REG_ENTRY_STR(GC, 0, regCP_BUSY_STAT), 76 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPC_BUSY_STAT), 77 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPF_BUSY_STAT), 78 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPC_BUSY_STAT2), 79 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPF_BUSY_STAT2), 80 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPF_STATUS), 81 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_ERROR), 82 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HPD_STATUS0), 83 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_BASE), 84 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_RPTR), 85 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_WPTR), 86 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB0_BASE), 87 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB0_RPTR), 88 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB0_WPTR), 89 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_CMD_BUFSZ), 90 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB2_CMD_BUFSZ), 91 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BASE_LO), 92 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BASE_HI), 93 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BUFSZ), 94 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB2_BASE_LO), 95 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB2_BASE_HI), 96 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB2_BUFSZ), 97 SOC15_REG_ENTRY_STR(GC, 0, regCPF_UTCL1_STATUS), 98 SOC15_REG_ENTRY_STR(GC, 0, regCPC_UTCL1_STATUS), 99 SOC15_REG_ENTRY_STR(GC, 0, regCPG_UTCL1_STATUS), 100 SOC15_REG_ENTRY_STR(GC, 0, regIA_UTCL1_STATUS), 101 SOC15_REG_ENTRY_STR(GC, 0, regIA_UTCL1_STATUS_2), 102 SOC15_REG_ENTRY_STR(GC, 0, regPA_CL_CNTL_STATUS), 103 SOC15_REG_ENTRY_STR(GC, 0, regRMI_UTCL1_STATUS), 104 SOC15_REG_ENTRY_STR(GC, 0, regSQC_CACHES), 105 SOC15_REG_ENTRY_STR(GC, 0, regSQG_STATUS), 106 SOC15_REG_ENTRY_STR(GC, 0, regWD_UTCL1_STATUS), 107 SOC15_REG_ENTRY_STR(GC, 0, regGCVM_L2_PROTECTION_FAULT_CNTL), 108 SOC15_REG_ENTRY_STR(GC, 0, regGCVM_L2_PROTECTION_FAULT_STATUS_LO32), 109 SOC15_REG_ENTRY_STR(GC, 0, regGCVM_L2_PROTECTION_FAULT_STATUS_HI32), 110 SOC15_REG_ENTRY_STR(GC, 0, regCP_DEBUG), 111 SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_CNTL), 112 SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_CNTL), 113 SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_INSTR_PNTR), 114 SOC15_REG_ENTRY_STR(GC, 0, regCP_ME_INSTR_PNTR), 115 SOC15_REG_ENTRY_STR(GC, 0, regCP_PFP_INSTR_PNTR), 116 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPC_STATUS), 117 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_RS64_INSTR_PNTR0), 118 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_RS64_INSTR_PNTR1), 119 SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_RS64_INSTR_PNTR), 120 121 /* cp header registers */ 122 SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_ME1_HEADER_DUMP), 123 SOC15_REG_ENTRY_STR(GC, 0, regCP_PFP_HEADER_DUMP), 124 SOC15_REG_ENTRY_STR(GC, 0, regCP_ME_HEADER_DUMP), 125 SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_HEADER_DUMP), 126 /* SE status registers */ 127 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS_SE0), 128 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS_SE1), 129 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS_SE2), 130 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS_SE3) 131 }; 132 133 static const struct amdgpu_hwip_reg_entry gc_cp_reg_list_12[] = { 134 /* compute registers */ 135 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_VMID), 136 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PERSISTENT_STATE), 137 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PIPE_PRIORITY), 138 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_QUEUE_PRIORITY), 139 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_QUANTUM), 140 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_BASE), 141 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_BASE_HI), 142 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_RPTR), 143 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR), 144 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR_HI), 145 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL), 146 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_CONTROL), 147 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_IB_BASE_ADDR), 148 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_IB_BASE_ADDR_HI), 149 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_IB_RPTR), 150 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_IB_CONTROL), 151 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_DEQUEUE_REQUEST), 152 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_BASE_ADDR), 153 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_BASE_ADDR_HI), 154 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_CONTROL), 155 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_RPTR), 156 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_WPTR), 157 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_EVENTS), 158 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CTX_SAVE_BASE_ADDR_LO), 159 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CTX_SAVE_BASE_ADDR_HI), 160 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CTX_SAVE_CONTROL), 161 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CNTL_STACK_OFFSET), 162 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CNTL_STACK_SIZE), 163 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_WG_STATE_OFFSET), 164 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CTX_SAVE_SIZE), 165 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_GDS_RESOURCE_STATE), 166 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_ERROR), 167 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_WPTR_MEM), 168 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_WPTR_LO), 169 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_WPTR_HI), 170 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_SUSPEND_CNTL_STACK_OFFSET), 171 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_SUSPEND_CNTL_STACK_DW_CNT), 172 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_SUSPEND_WG_STATE_OFFSET), 173 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_DEQUEUE_STATUS) 174 }; 175 176 static const struct amdgpu_hwip_reg_entry gc_gfx_queue_reg_list_12[] = { 177 /* gfx queue registers */ 178 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_ACTIVE), 179 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_VMID), 180 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_QUEUE_PRIORITY), 181 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_QUANTUM), 182 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_BASE), 183 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_BASE_HI), 184 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_OFFSET), 185 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_CNTL), 186 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_CSMD_RPTR), 187 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_WPTR), 188 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_WPTR_HI), 189 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_DEQUEUE_REQUEST), 190 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_MAPPED), 191 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_QUE_MGR_CONTROL), 192 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_HQ_CONTROL0), 193 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_HQ_STATUS0), 194 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_MQD_BASE_ADDR), 195 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_MQD_BASE_ADDR_HI), 196 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_WPTR_POLL_ADDR_LO), 197 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_WPTR_POLL_ADDR_HI), 198 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_RPTR), 199 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BASE_LO), 200 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BASE_HI), 201 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_CMD_BUFSZ), 202 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BUFSZ) 203 }; 204 205 static const struct soc15_reg_golden golden_settings_gc_12_0[] = { 206 SOC15_REG_GOLDEN_VALUE(GC, 0, regDB_MEM_CONFIG, 0x0000000f, 0x0000000f), 207 SOC15_REG_GOLDEN_VALUE(GC, 0, regCB_HW_CONTROL_1, 0x03000000, 0x03000000), 208 SOC15_REG_GOLDEN_VALUE(GC, 0, regGL2C_CTRL5, 0x00000070, 0x00000020) 209 }; 210 211 #define DEFAULT_SH_MEM_CONFIG \ 212 ((SH_MEM_ADDRESS_MODE_64 << SH_MEM_CONFIG__ADDRESS_MODE__SHIFT) | \ 213 (SH_MEM_ALIGNMENT_MODE_UNALIGNED << SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT) | \ 214 (3 << SH_MEM_CONFIG__INITIAL_INST_PREFETCH__SHIFT)) 215 216 static void gfx_v12_0_disable_gpa_mode(struct amdgpu_device *adev); 217 static void gfx_v12_0_set_ring_funcs(struct amdgpu_device *adev); 218 static void gfx_v12_0_set_irq_funcs(struct amdgpu_device *adev); 219 static void gfx_v12_0_set_rlc_funcs(struct amdgpu_device *adev); 220 static void gfx_v12_0_set_mqd_funcs(struct amdgpu_device *adev); 221 static void gfx_v12_0_set_imu_funcs(struct amdgpu_device *adev); 222 static int gfx_v12_0_get_cu_info(struct amdgpu_device *adev, 223 struct amdgpu_cu_info *cu_info); 224 static uint64_t gfx_v12_0_get_gpu_clock_counter(struct amdgpu_device *adev); 225 static void gfx_v12_0_select_se_sh(struct amdgpu_device *adev, u32 se_num, 226 u32 sh_num, u32 instance, int xcc_id); 227 static u32 gfx_v12_0_get_wgp_active_bitmap_per_sh(struct amdgpu_device *adev); 228 229 static void gfx_v12_0_ring_emit_frame_cntl(struct amdgpu_ring *ring, bool start, bool secure); 230 static void gfx_v12_0_ring_emit_wreg(struct amdgpu_ring *ring, uint32_t reg, 231 uint32_t val); 232 static int gfx_v12_0_wait_for_rlc_autoload_complete(struct amdgpu_device *adev); 233 static void gfx_v12_0_ring_invalidate_tlbs(struct amdgpu_ring *ring, 234 uint16_t pasid, uint32_t flush_type, 235 bool all_hub, uint8_t dst_sel); 236 static void gfx_v12_0_set_safe_mode(struct amdgpu_device *adev, int xcc_id); 237 static void gfx_v12_0_unset_safe_mode(struct amdgpu_device *adev, int xcc_id); 238 static void gfx_v12_0_update_perf_clk(struct amdgpu_device *adev, 239 bool enable); 240 241 static void gfx_v12_0_kiq_set_resources(struct amdgpu_ring *kiq_ring, 242 uint64_t queue_mask) 243 { 244 amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_SET_RESOURCES, 6)); 245 amdgpu_ring_write(kiq_ring, PACKET3_SET_RESOURCES_VMID_MASK(0) | 246 PACKET3_SET_RESOURCES_QUEUE_TYPE(0)); /* vmid_mask:0 queue_type:0 (KIQ) */ 247 amdgpu_ring_write(kiq_ring, lower_32_bits(queue_mask)); /* queue mask lo */ 248 amdgpu_ring_write(kiq_ring, upper_32_bits(queue_mask)); /* queue mask hi */ 249 amdgpu_ring_write(kiq_ring, 0); /* gws mask lo */ 250 amdgpu_ring_write(kiq_ring, 0); /* gws mask hi */ 251 amdgpu_ring_write(kiq_ring, 0); /* oac mask */ 252 amdgpu_ring_write(kiq_ring, 0); 253 } 254 255 static void gfx_v12_0_kiq_map_queues(struct amdgpu_ring *kiq_ring, 256 struct amdgpu_ring *ring) 257 { 258 uint64_t mqd_addr = amdgpu_bo_gpu_offset(ring->mqd_obj); 259 uint64_t wptr_addr = ring->wptr_gpu_addr; 260 uint32_t me = 0, eng_sel = 0; 261 262 switch (ring->funcs->type) { 263 case AMDGPU_RING_TYPE_COMPUTE: 264 me = 1; 265 eng_sel = 0; 266 break; 267 case AMDGPU_RING_TYPE_GFX: 268 me = 0; 269 eng_sel = 4; 270 break; 271 case AMDGPU_RING_TYPE_MES: 272 me = 2; 273 eng_sel = 5; 274 break; 275 default: 276 WARN_ON(1); 277 } 278 279 amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_MAP_QUEUES, 5)); 280 /* Q_sel:0, vmid:0, vidmem: 1, engine:0, num_Q:1*/ 281 amdgpu_ring_write(kiq_ring, /* Q_sel: 0, vmid: 0, engine: 0, num_Q: 1 */ 282 PACKET3_MAP_QUEUES_QUEUE_SEL(0) | /* Queue_Sel */ 283 PACKET3_MAP_QUEUES_VMID(0) | /* VMID */ 284 PACKET3_MAP_QUEUES_QUEUE(ring->queue) | 285 PACKET3_MAP_QUEUES_PIPE(ring->pipe) | 286 PACKET3_MAP_QUEUES_ME((me)) | 287 PACKET3_MAP_QUEUES_QUEUE_TYPE(0) | /*queue_type: normal compute queue */ 288 PACKET3_MAP_QUEUES_ALLOC_FORMAT(0) | /* alloc format: all_on_one_pipe */ 289 PACKET3_MAP_QUEUES_ENGINE_SEL(eng_sel) | 290 PACKET3_MAP_QUEUES_NUM_QUEUES(1)); /* num_queues: must be 1 */ 291 amdgpu_ring_write(kiq_ring, PACKET3_MAP_QUEUES_DOORBELL_OFFSET(ring->doorbell_index)); 292 amdgpu_ring_write(kiq_ring, lower_32_bits(mqd_addr)); 293 amdgpu_ring_write(kiq_ring, upper_32_bits(mqd_addr)); 294 amdgpu_ring_write(kiq_ring, lower_32_bits(wptr_addr)); 295 amdgpu_ring_write(kiq_ring, upper_32_bits(wptr_addr)); 296 } 297 298 static void gfx_v12_0_kiq_unmap_queues(struct amdgpu_ring *kiq_ring, 299 struct amdgpu_ring *ring, 300 enum amdgpu_unmap_queues_action action, 301 u64 gpu_addr, u64 seq) 302 { 303 struct amdgpu_device *adev = kiq_ring->adev; 304 uint32_t eng_sel = ring->funcs->type == AMDGPU_RING_TYPE_GFX ? 4 : 0; 305 306 if (adev->enable_mes && !adev->gfx.kiq[0].ring.sched.ready) { 307 amdgpu_mes_unmap_legacy_queue(adev, ring, action, gpu_addr, seq); 308 return; 309 } 310 311 amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_UNMAP_QUEUES, 4)); 312 amdgpu_ring_write(kiq_ring, /* Q_sel: 0, vmid: 0, engine: 0, num_Q: 1 */ 313 PACKET3_UNMAP_QUEUES_ACTION(action) | 314 PACKET3_UNMAP_QUEUES_QUEUE_SEL(0) | 315 PACKET3_UNMAP_QUEUES_ENGINE_SEL(eng_sel) | 316 PACKET3_UNMAP_QUEUES_NUM_QUEUES(1)); 317 amdgpu_ring_write(kiq_ring, 318 PACKET3_UNMAP_QUEUES_DOORBELL_OFFSET0(ring->doorbell_index)); 319 320 if (action == PREEMPT_QUEUES_NO_UNMAP) { 321 amdgpu_ring_write(kiq_ring, lower_32_bits(gpu_addr)); 322 amdgpu_ring_write(kiq_ring, upper_32_bits(gpu_addr)); 323 amdgpu_ring_write(kiq_ring, seq); 324 } else { 325 amdgpu_ring_write(kiq_ring, 0); 326 amdgpu_ring_write(kiq_ring, 0); 327 amdgpu_ring_write(kiq_ring, 0); 328 } 329 } 330 331 static void gfx_v12_0_kiq_query_status(struct amdgpu_ring *kiq_ring, 332 struct amdgpu_ring *ring, 333 u64 addr, u64 seq) 334 { 335 uint32_t eng_sel = ring->funcs->type == AMDGPU_RING_TYPE_GFX ? 4 : 0; 336 337 amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_QUERY_STATUS, 5)); 338 amdgpu_ring_write(kiq_ring, 339 PACKET3_QUERY_STATUS_CONTEXT_ID(0) | 340 PACKET3_QUERY_STATUS_INTERRUPT_SEL(0) | 341 PACKET3_QUERY_STATUS_COMMAND(2)); 342 amdgpu_ring_write(kiq_ring, /* Q_sel: 0, vmid: 0, engine: 0, num_Q: 1 */ 343 PACKET3_QUERY_STATUS_DOORBELL_OFFSET(ring->doorbell_index) | 344 PACKET3_QUERY_STATUS_ENG_SEL(eng_sel)); 345 amdgpu_ring_write(kiq_ring, lower_32_bits(addr)); 346 amdgpu_ring_write(kiq_ring, upper_32_bits(addr)); 347 amdgpu_ring_write(kiq_ring, lower_32_bits(seq)); 348 amdgpu_ring_write(kiq_ring, upper_32_bits(seq)); 349 } 350 351 static void gfx_v12_0_kiq_invalidate_tlbs(struct amdgpu_ring *kiq_ring, 352 uint16_t pasid, 353 uint32_t flush_type, 354 bool all_hub) 355 { 356 gfx_v12_0_ring_invalidate_tlbs(kiq_ring, pasid, flush_type, all_hub, 1); 357 } 358 359 static const struct kiq_pm4_funcs gfx_v12_0_kiq_pm4_funcs = { 360 .kiq_set_resources = gfx_v12_0_kiq_set_resources, 361 .kiq_map_queues = gfx_v12_0_kiq_map_queues, 362 .kiq_unmap_queues = gfx_v12_0_kiq_unmap_queues, 363 .kiq_query_status = gfx_v12_0_kiq_query_status, 364 .kiq_invalidate_tlbs = gfx_v12_0_kiq_invalidate_tlbs, 365 .set_resources_size = 8, 366 .map_queues_size = 7, 367 .unmap_queues_size = 6, 368 .query_status_size = 7, 369 .invalidate_tlbs_size = 2, 370 }; 371 372 static void gfx_v12_0_set_kiq_pm4_funcs(struct amdgpu_device *adev) 373 { 374 adev->gfx.kiq[0].pmf = &gfx_v12_0_kiq_pm4_funcs; 375 } 376 377 static void gfx_v12_0_wait_reg_mem(struct amdgpu_ring *ring, int eng_sel, 378 int mem_space, int opt, uint32_t addr0, 379 uint32_t addr1, uint32_t ref, 380 uint32_t mask, uint32_t inv) 381 { 382 amdgpu_ring_write(ring, PACKET3(PACKET3_WAIT_REG_MEM, 5)); 383 amdgpu_ring_write(ring, 384 /* memory (1) or register (0) */ 385 (WAIT_REG_MEM_MEM_SPACE(mem_space) | 386 WAIT_REG_MEM_OPERATION(opt) | /* wait */ 387 WAIT_REG_MEM_FUNCTION(3) | /* equal */ 388 WAIT_REG_MEM_ENGINE(eng_sel))); 389 390 if (mem_space) 391 BUG_ON(addr0 & 0x3); /* Dword align */ 392 amdgpu_ring_write(ring, addr0); 393 amdgpu_ring_write(ring, addr1); 394 amdgpu_ring_write(ring, ref); 395 amdgpu_ring_write(ring, mask); 396 amdgpu_ring_write(ring, inv); /* poll interval */ 397 } 398 399 static int gfx_v12_0_ring_test_ring(struct amdgpu_ring *ring) 400 { 401 struct amdgpu_device *adev = ring->adev; 402 uint32_t scratch = SOC15_REG_OFFSET(GC, 0, regSCRATCH_REG0); 403 uint32_t tmp = 0; 404 unsigned i; 405 int r; 406 407 WREG32(scratch, 0xCAFEDEAD); 408 r = amdgpu_ring_alloc(ring, 5); 409 if (r) { 410 dev_err(adev->dev, 411 "amdgpu: cp failed to lock ring %d (%d).\n", 412 ring->idx, r); 413 return r; 414 } 415 416 if (ring->funcs->type == AMDGPU_RING_TYPE_KIQ) { 417 gfx_v12_0_ring_emit_wreg(ring, scratch, 0xDEADBEEF); 418 } else { 419 amdgpu_ring_write(ring, PACKET3(PACKET3_SET_UCONFIG_REG, 1)); 420 amdgpu_ring_write(ring, scratch - 421 PACKET3_SET_UCONFIG_REG_START); 422 amdgpu_ring_write(ring, 0xDEADBEEF); 423 } 424 amdgpu_ring_commit(ring); 425 426 for (i = 0; i < adev->usec_timeout; i++) { 427 tmp = RREG32(scratch); 428 if (tmp == 0xDEADBEEF) 429 break; 430 if (amdgpu_emu_mode == 1) 431 msleep(1); 432 else 433 udelay(1); 434 } 435 436 if (i >= adev->usec_timeout) 437 r = -ETIMEDOUT; 438 return r; 439 } 440 441 static int gfx_v12_0_ring_test_ib(struct amdgpu_ring *ring, long timeout) 442 { 443 struct amdgpu_device *adev = ring->adev; 444 struct amdgpu_ib ib; 445 struct dma_fence *f = NULL; 446 unsigned index; 447 uint64_t gpu_addr; 448 volatile uint32_t *cpu_ptr; 449 long r; 450 451 /* MES KIQ fw hasn't indirect buffer support for now */ 452 if (adev->enable_mes_kiq && 453 ring->funcs->type == AMDGPU_RING_TYPE_KIQ) 454 return 0; 455 456 memset(&ib, 0, sizeof(ib)); 457 458 if (ring->is_mes_queue) { 459 uint32_t padding, offset; 460 461 offset = amdgpu_mes_ctx_get_offs(ring, AMDGPU_MES_CTX_IB_OFFS); 462 padding = amdgpu_mes_ctx_get_offs(ring, 463 AMDGPU_MES_CTX_PADDING_OFFS); 464 465 ib.gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset); 466 ib.ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset); 467 468 gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, padding); 469 cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, padding); 470 *cpu_ptr = cpu_to_le32(0xCAFEDEAD); 471 } else { 472 r = amdgpu_device_wb_get(adev, &index); 473 if (r) 474 return r; 475 476 gpu_addr = adev->wb.gpu_addr + (index * 4); 477 adev->wb.wb[index] = cpu_to_le32(0xCAFEDEAD); 478 cpu_ptr = &adev->wb.wb[index]; 479 480 r = amdgpu_ib_get(adev, NULL, 16, AMDGPU_IB_POOL_DIRECT, &ib); 481 if (r) { 482 dev_err(adev->dev, "amdgpu: failed to get ib (%ld).\n", r); 483 goto err1; 484 } 485 } 486 487 ib.ptr[0] = PACKET3(PACKET3_WRITE_DATA, 3); 488 ib.ptr[1] = WRITE_DATA_DST_SEL(5) | WR_CONFIRM; 489 ib.ptr[2] = lower_32_bits(gpu_addr); 490 ib.ptr[3] = upper_32_bits(gpu_addr); 491 ib.ptr[4] = 0xDEADBEEF; 492 ib.length_dw = 5; 493 494 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f); 495 if (r) 496 goto err2; 497 498 r = dma_fence_wait_timeout(f, false, timeout); 499 if (r == 0) { 500 r = -ETIMEDOUT; 501 goto err2; 502 } else if (r < 0) { 503 goto err2; 504 } 505 506 if (le32_to_cpu(*cpu_ptr) == 0xDEADBEEF) 507 r = 0; 508 else 509 r = -EINVAL; 510 err2: 511 if (!ring->is_mes_queue) 512 amdgpu_ib_free(adev, &ib, NULL); 513 dma_fence_put(f); 514 err1: 515 if (!ring->is_mes_queue) 516 amdgpu_device_wb_free(adev, index); 517 return r; 518 } 519 520 static void gfx_v12_0_free_microcode(struct amdgpu_device *adev) 521 { 522 amdgpu_ucode_release(&adev->gfx.pfp_fw); 523 amdgpu_ucode_release(&adev->gfx.me_fw); 524 amdgpu_ucode_release(&adev->gfx.rlc_fw); 525 amdgpu_ucode_release(&adev->gfx.mec_fw); 526 527 kfree(adev->gfx.rlc.register_list_format); 528 } 529 530 static int gfx_v12_0_init_toc_microcode(struct amdgpu_device *adev, const char *ucode_prefix) 531 { 532 const struct psp_firmware_header_v1_0 *toc_hdr; 533 int err = 0; 534 535 err = amdgpu_ucode_request(adev, &adev->psp.toc_fw, 536 "amdgpu/%s_toc.bin", ucode_prefix); 537 if (err) 538 goto out; 539 540 toc_hdr = (const struct psp_firmware_header_v1_0 *)adev->psp.toc_fw->data; 541 adev->psp.toc.fw_version = le32_to_cpu(toc_hdr->header.ucode_version); 542 adev->psp.toc.feature_version = le32_to_cpu(toc_hdr->sos.fw_version); 543 adev->psp.toc.size_bytes = le32_to_cpu(toc_hdr->header.ucode_size_bytes); 544 adev->psp.toc.start_addr = (uint8_t *)toc_hdr + 545 le32_to_cpu(toc_hdr->header.ucode_array_offset_bytes); 546 return 0; 547 out: 548 amdgpu_ucode_release(&adev->psp.toc_fw); 549 return err; 550 } 551 552 static int gfx_v12_0_init_microcode(struct amdgpu_device *adev) 553 { 554 char ucode_prefix[15]; 555 int err; 556 const struct rlc_firmware_header_v2_0 *rlc_hdr; 557 uint16_t version_major; 558 uint16_t version_minor; 559 560 DRM_DEBUG("\n"); 561 562 amdgpu_ucode_ip_version_decode(adev, GC_HWIP, ucode_prefix, sizeof(ucode_prefix)); 563 564 err = amdgpu_ucode_request(adev, &adev->gfx.pfp_fw, 565 "amdgpu/%s_pfp.bin", ucode_prefix); 566 if (err) 567 goto out; 568 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_PFP); 569 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_PFP_P0_STACK); 570 571 err = amdgpu_ucode_request(adev, &adev->gfx.me_fw, 572 "amdgpu/%s_me.bin", ucode_prefix); 573 if (err) 574 goto out; 575 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_ME); 576 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_ME_P0_STACK); 577 578 if (!amdgpu_sriov_vf(adev)) { 579 err = amdgpu_ucode_request(adev, &adev->gfx.rlc_fw, 580 "amdgpu/%s_rlc.bin", ucode_prefix); 581 if (err) 582 goto out; 583 rlc_hdr = (const struct rlc_firmware_header_v2_0 *)adev->gfx.rlc_fw->data; 584 version_major = le16_to_cpu(rlc_hdr->header.header_version_major); 585 version_minor = le16_to_cpu(rlc_hdr->header.header_version_minor); 586 err = amdgpu_gfx_rlc_init_microcode(adev, version_major, version_minor); 587 if (err) 588 goto out; 589 } 590 591 err = amdgpu_ucode_request(adev, &adev->gfx.mec_fw, 592 "amdgpu/%s_mec.bin", ucode_prefix); 593 if (err) 594 goto out; 595 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_MEC); 596 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_MEC_P0_STACK); 597 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_MEC_P1_STACK); 598 599 if (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) 600 err = gfx_v12_0_init_toc_microcode(adev, ucode_prefix); 601 602 /* only one MEC for gfx 12 */ 603 adev->gfx.mec2_fw = NULL; 604 605 if (adev->gfx.imu.funcs) { 606 if (adev->gfx.imu.funcs->init_microcode) { 607 err = adev->gfx.imu.funcs->init_microcode(adev); 608 if (err) 609 dev_err(adev->dev, "Failed to load imu firmware!\n"); 610 } 611 } 612 613 out: 614 if (err) { 615 amdgpu_ucode_release(&adev->gfx.pfp_fw); 616 amdgpu_ucode_release(&adev->gfx.me_fw); 617 amdgpu_ucode_release(&adev->gfx.rlc_fw); 618 amdgpu_ucode_release(&adev->gfx.mec_fw); 619 } 620 621 return err; 622 } 623 624 static u32 gfx_v12_0_get_csb_size(struct amdgpu_device *adev) 625 { 626 u32 count = 0; 627 const struct cs_section_def *sect = NULL; 628 const struct cs_extent_def *ext = NULL; 629 630 count += 1; 631 632 for (sect = gfx12_cs_data; sect->section != NULL; ++sect) { 633 if (sect->id == SECT_CONTEXT) { 634 for (ext = sect->section; ext->extent != NULL; ++ext) 635 count += 2 + ext->reg_count; 636 } else 637 return 0; 638 } 639 640 return count; 641 } 642 643 static void gfx_v12_0_get_csb_buffer(struct amdgpu_device *adev, 644 volatile u32 *buffer) 645 { 646 u32 count = 0, clustercount = 0, i; 647 const struct cs_section_def *sect = NULL; 648 const struct cs_extent_def *ext = NULL; 649 650 if (adev->gfx.rlc.cs_data == NULL) 651 return; 652 if (buffer == NULL) 653 return; 654 655 count += 1; 656 657 for (sect = adev->gfx.rlc.cs_data; sect->section != NULL; ++sect) { 658 if (sect->id == SECT_CONTEXT) { 659 for (ext = sect->section; ext->extent != NULL; ++ext) { 660 clustercount++; 661 buffer[count++] = ext->reg_count; 662 buffer[count++] = ext->reg_index; 663 664 for (i = 0; i < ext->reg_count; i++) 665 buffer[count++] = cpu_to_le32(ext->extent[i]); 666 } 667 } else 668 return; 669 } 670 671 buffer[0] = clustercount; 672 } 673 674 static void gfx_v12_0_rlc_fini(struct amdgpu_device *adev) 675 { 676 /* clear state block */ 677 amdgpu_bo_free_kernel(&adev->gfx.rlc.clear_state_obj, 678 &adev->gfx.rlc.clear_state_gpu_addr, 679 (void **)&adev->gfx.rlc.cs_ptr); 680 681 /* jump table block */ 682 amdgpu_bo_free_kernel(&adev->gfx.rlc.cp_table_obj, 683 &adev->gfx.rlc.cp_table_gpu_addr, 684 (void **)&adev->gfx.rlc.cp_table_ptr); 685 } 686 687 static void gfx_v12_0_init_rlcg_reg_access_ctrl(struct amdgpu_device *adev) 688 { 689 struct amdgpu_rlcg_reg_access_ctrl *reg_access_ctrl; 690 691 reg_access_ctrl = &adev->gfx.rlc.reg_access_ctrl[0]; 692 reg_access_ctrl->scratch_reg0 = SOC15_REG_OFFSET(GC, 0, regSCRATCH_REG0); 693 reg_access_ctrl->scratch_reg1 = SOC15_REG_OFFSET(GC, 0, regSCRATCH_REG1); 694 reg_access_ctrl->scratch_reg2 = SOC15_REG_OFFSET(GC, 0, regSCRATCH_REG2); 695 reg_access_ctrl->scratch_reg3 = SOC15_REG_OFFSET(GC, 0, regSCRATCH_REG3); 696 reg_access_ctrl->grbm_cntl = SOC15_REG_OFFSET(GC, 0, regGRBM_GFX_CNTL); 697 reg_access_ctrl->grbm_idx = SOC15_REG_OFFSET(GC, 0, regGRBM_GFX_INDEX); 698 reg_access_ctrl->spare_int = SOC15_REG_OFFSET(GC, 0, regRLC_SPARE_INT_0); 699 adev->gfx.rlc.rlcg_reg_access_supported = true; 700 } 701 702 static int gfx_v12_0_rlc_init(struct amdgpu_device *adev) 703 { 704 const struct cs_section_def *cs_data; 705 int r; 706 707 adev->gfx.rlc.cs_data = gfx12_cs_data; 708 709 cs_data = adev->gfx.rlc.cs_data; 710 711 if (cs_data) { 712 /* init clear state block */ 713 r = amdgpu_gfx_rlc_init_csb(adev); 714 if (r) 715 return r; 716 } 717 718 /* init spm vmid with 0xf */ 719 if (adev->gfx.rlc.funcs->update_spm_vmid) 720 adev->gfx.rlc.funcs->update_spm_vmid(adev, NULL, 0xf); 721 722 return 0; 723 } 724 725 static void gfx_v12_0_mec_fini(struct amdgpu_device *adev) 726 { 727 amdgpu_bo_free_kernel(&adev->gfx.mec.hpd_eop_obj, NULL, NULL); 728 amdgpu_bo_free_kernel(&adev->gfx.mec.mec_fw_obj, NULL, NULL); 729 amdgpu_bo_free_kernel(&adev->gfx.mec.mec_fw_data_obj, NULL, NULL); 730 } 731 732 static void gfx_v12_0_me_init(struct amdgpu_device *adev) 733 { 734 bitmap_zero(adev->gfx.me.queue_bitmap, AMDGPU_MAX_GFX_QUEUES); 735 736 amdgpu_gfx_graphics_queue_acquire(adev); 737 } 738 739 static int gfx_v12_0_mec_init(struct amdgpu_device *adev) 740 { 741 int r; 742 u32 *hpd; 743 size_t mec_hpd_size; 744 745 bitmap_zero(adev->gfx.mec_bitmap[0].queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES); 746 747 /* take ownership of the relevant compute queues */ 748 amdgpu_gfx_compute_queue_acquire(adev); 749 mec_hpd_size = adev->gfx.num_compute_rings * GFX12_MEC_HPD_SIZE; 750 751 if (mec_hpd_size) { 752 r = amdgpu_bo_create_reserved(adev, mec_hpd_size, PAGE_SIZE, 753 AMDGPU_GEM_DOMAIN_GTT, 754 &adev->gfx.mec.hpd_eop_obj, 755 &adev->gfx.mec.hpd_eop_gpu_addr, 756 (void **)&hpd); 757 if (r) { 758 dev_warn(adev->dev, "(%d) create HDP EOP bo failed\n", r); 759 gfx_v12_0_mec_fini(adev); 760 return r; 761 } 762 763 memset(hpd, 0, mec_hpd_size); 764 765 amdgpu_bo_kunmap(adev->gfx.mec.hpd_eop_obj); 766 amdgpu_bo_unreserve(adev->gfx.mec.hpd_eop_obj); 767 } 768 769 return 0; 770 } 771 772 static uint32_t wave_read_ind(struct amdgpu_device *adev, uint32_t wave, uint32_t address) 773 { 774 WREG32_SOC15(GC, 0, regSQ_IND_INDEX, 775 (wave << SQ_IND_INDEX__WAVE_ID__SHIFT) | 776 (address << SQ_IND_INDEX__INDEX__SHIFT)); 777 return RREG32_SOC15(GC, 0, regSQ_IND_DATA); 778 } 779 780 static void wave_read_regs(struct amdgpu_device *adev, uint32_t wave, 781 uint32_t thread, uint32_t regno, 782 uint32_t num, uint32_t *out) 783 { 784 WREG32_SOC15(GC, 0, regSQ_IND_INDEX, 785 (wave << SQ_IND_INDEX__WAVE_ID__SHIFT) | 786 (regno << SQ_IND_INDEX__INDEX__SHIFT) | 787 (thread << SQ_IND_INDEX__WORKITEM_ID__SHIFT) | 788 (SQ_IND_INDEX__AUTO_INCR_MASK)); 789 while (num--) 790 *(out++) = RREG32_SOC15(GC, 0, regSQ_IND_DATA); 791 } 792 793 static void gfx_v12_0_read_wave_data(struct amdgpu_device *adev, 794 uint32_t xcc_id, 795 uint32_t simd, uint32_t wave, 796 uint32_t *dst, int *no_fields) 797 { 798 /* in gfx12 the SIMD_ID is specified as part of the INSTANCE 799 * field when performing a select_se_sh so it should be 800 * zero here */ 801 WARN_ON(simd != 0); 802 803 /* type 4 wave data */ 804 dst[(*no_fields)++] = 4; 805 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_STATUS); 806 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_PC_LO); 807 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_PC_HI); 808 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_EXEC_LO); 809 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_EXEC_HI); 810 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_HW_ID1); 811 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_HW_ID2); 812 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_GPR_ALLOC); 813 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_LDS_ALLOC); 814 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_IB_STS); 815 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_IB_STS2); 816 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_IB_DBG1); 817 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_M0); 818 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_MODE); 819 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_STATE_PRIV); 820 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_EXCP_FLAG_PRIV); 821 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_EXCP_FLAG_USER); 822 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_TRAP_CTRL); 823 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_ACTIVE); 824 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_VALID_AND_IDLE); 825 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_DVGPR_ALLOC_LO); 826 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_DVGPR_ALLOC_HI); 827 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_SCHED_MODE); 828 } 829 830 static void gfx_v12_0_read_wave_sgprs(struct amdgpu_device *adev, 831 uint32_t xcc_id, uint32_t simd, 832 uint32_t wave, uint32_t start, 833 uint32_t size, uint32_t *dst) 834 { 835 WARN_ON(simd != 0); 836 837 wave_read_regs( 838 adev, wave, 0, start + SQIND_WAVE_SGPRS_OFFSET, size, 839 dst); 840 } 841 842 static void gfx_v12_0_read_wave_vgprs(struct amdgpu_device *adev, 843 uint32_t xcc_id, uint32_t simd, 844 uint32_t wave, uint32_t thread, 845 uint32_t start, uint32_t size, 846 uint32_t *dst) 847 { 848 wave_read_regs( 849 adev, wave, thread, 850 start + SQIND_WAVE_VGPRS_OFFSET, size, dst); 851 } 852 853 static void gfx_v12_0_select_me_pipe_q(struct amdgpu_device *adev, 854 u32 me, u32 pipe, u32 q, u32 vm, u32 xcc_id) 855 { 856 soc24_grbm_select(adev, me, pipe, q, vm); 857 } 858 859 static const struct amdgpu_gfx_funcs gfx_v12_0_gfx_funcs = { 860 .get_gpu_clock_counter = &gfx_v12_0_get_gpu_clock_counter, 861 .select_se_sh = &gfx_v12_0_select_se_sh, 862 .read_wave_data = &gfx_v12_0_read_wave_data, 863 .read_wave_sgprs = &gfx_v12_0_read_wave_sgprs, 864 .read_wave_vgprs = &gfx_v12_0_read_wave_vgprs, 865 .select_me_pipe_q = &gfx_v12_0_select_me_pipe_q, 866 .update_perfmon_mgcg = &gfx_v12_0_update_perf_clk, 867 }; 868 869 static int gfx_v12_0_gpu_early_init(struct amdgpu_device *adev) 870 { 871 872 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 873 case IP_VERSION(12, 0, 0): 874 case IP_VERSION(12, 0, 1): 875 adev->gfx.config.max_hw_contexts = 8; 876 adev->gfx.config.sc_prim_fifo_size_frontend = 0x20; 877 adev->gfx.config.sc_prim_fifo_size_backend = 0x100; 878 adev->gfx.config.sc_hiz_tile_fifo_size = 0; 879 adev->gfx.config.sc_earlyz_tile_fifo_size = 0x4C0; 880 break; 881 default: 882 BUG(); 883 break; 884 } 885 886 return 0; 887 } 888 889 static int gfx_v12_0_gfx_ring_init(struct amdgpu_device *adev, int ring_id, 890 int me, int pipe, int queue) 891 { 892 int r; 893 struct amdgpu_ring *ring; 894 unsigned int irq_type; 895 896 ring = &adev->gfx.gfx_ring[ring_id]; 897 898 ring->me = me; 899 ring->pipe = pipe; 900 ring->queue = queue; 901 902 ring->ring_obj = NULL; 903 ring->use_doorbell = true; 904 905 if (!ring_id) 906 ring->doorbell_index = adev->doorbell_index.gfx_ring0 << 1; 907 else 908 ring->doorbell_index = adev->doorbell_index.gfx_ring1 << 1; 909 ring->vm_hub = AMDGPU_GFXHUB(0); 910 sprintf(ring->name, "gfx_%d.%d.%d", ring->me, ring->pipe, ring->queue); 911 912 irq_type = AMDGPU_CP_IRQ_GFX_ME0_PIPE0_EOP + ring->pipe; 913 r = amdgpu_ring_init(adev, ring, 1024, &adev->gfx.eop_irq, irq_type, 914 AMDGPU_RING_PRIO_DEFAULT, NULL); 915 if (r) 916 return r; 917 return 0; 918 } 919 920 static int gfx_v12_0_compute_ring_init(struct amdgpu_device *adev, int ring_id, 921 int mec, int pipe, int queue) 922 { 923 int r; 924 unsigned irq_type; 925 struct amdgpu_ring *ring; 926 unsigned int hw_prio; 927 928 ring = &adev->gfx.compute_ring[ring_id]; 929 930 /* mec0 is me1 */ 931 ring->me = mec + 1; 932 ring->pipe = pipe; 933 ring->queue = queue; 934 935 ring->ring_obj = NULL; 936 ring->use_doorbell = true; 937 ring->doorbell_index = (adev->doorbell_index.mec_ring0 + ring_id) << 1; 938 ring->eop_gpu_addr = adev->gfx.mec.hpd_eop_gpu_addr 939 + (ring_id * GFX12_MEC_HPD_SIZE); 940 ring->vm_hub = AMDGPU_GFXHUB(0); 941 sprintf(ring->name, "comp_%d.%d.%d", ring->me, ring->pipe, ring->queue); 942 943 irq_type = AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE0_EOP 944 + ((ring->me - 1) * adev->gfx.mec.num_pipe_per_mec) 945 + ring->pipe; 946 hw_prio = amdgpu_gfx_is_high_priority_compute_queue(adev, ring) ? 947 AMDGPU_GFX_PIPE_PRIO_HIGH : AMDGPU_GFX_PIPE_PRIO_NORMAL; 948 /* type-2 packets are deprecated on MEC, use type-3 instead */ 949 r = amdgpu_ring_init(adev, ring, 1024, &adev->gfx.eop_irq, irq_type, 950 hw_prio, NULL); 951 if (r) 952 return r; 953 954 return 0; 955 } 956 957 static struct { 958 SOC24_FIRMWARE_ID id; 959 unsigned int offset; 960 unsigned int size; 961 unsigned int size_x16; 962 } rlc_autoload_info[SOC24_FIRMWARE_ID_MAX]; 963 964 #define RLC_TOC_OFFSET_DWUNIT 8 965 #define RLC_SIZE_MULTIPLE 1024 966 #define RLC_TOC_UMF_SIZE_inM 23ULL 967 #define RLC_TOC_FORMAT_API 165ULL 968 969 static void gfx_v12_0_parse_rlc_toc(struct amdgpu_device *adev, void *rlc_toc) 970 { 971 RLC_TABLE_OF_CONTENT_V2 *ucode = rlc_toc; 972 973 while (ucode && (ucode->id > SOC24_FIRMWARE_ID_INVALID)) { 974 rlc_autoload_info[ucode->id].id = ucode->id; 975 rlc_autoload_info[ucode->id].offset = 976 ucode->offset * RLC_TOC_OFFSET_DWUNIT * 4; 977 rlc_autoload_info[ucode->id].size = 978 ucode->size_x16 ? ucode->size * RLC_SIZE_MULTIPLE * 4 : 979 ucode->size * 4; 980 ucode++; 981 } 982 } 983 984 static uint32_t gfx_v12_0_calc_toc_total_size(struct amdgpu_device *adev) 985 { 986 uint32_t total_size = 0; 987 SOC24_FIRMWARE_ID id; 988 989 gfx_v12_0_parse_rlc_toc(adev, adev->psp.toc.start_addr); 990 991 for (id = SOC24_FIRMWARE_ID_RLC_G_UCODE; id < SOC24_FIRMWARE_ID_MAX; id++) 992 total_size += rlc_autoload_info[id].size; 993 994 /* In case the offset in rlc toc ucode is aligned */ 995 if (total_size < rlc_autoload_info[SOC24_FIRMWARE_ID_MAX-1].offset) 996 total_size = rlc_autoload_info[SOC24_FIRMWARE_ID_MAX-1].offset + 997 rlc_autoload_info[SOC24_FIRMWARE_ID_MAX-1].size; 998 if (total_size < (RLC_TOC_UMF_SIZE_inM << 20)) 999 total_size = RLC_TOC_UMF_SIZE_inM << 20; 1000 1001 return total_size; 1002 } 1003 1004 static int gfx_v12_0_rlc_autoload_buffer_init(struct amdgpu_device *adev) 1005 { 1006 int r; 1007 uint32_t total_size; 1008 1009 total_size = gfx_v12_0_calc_toc_total_size(adev); 1010 1011 r = amdgpu_bo_create_reserved(adev, total_size, 64 * 1024, 1012 AMDGPU_GEM_DOMAIN_VRAM, 1013 &adev->gfx.rlc.rlc_autoload_bo, 1014 &adev->gfx.rlc.rlc_autoload_gpu_addr, 1015 (void **)&adev->gfx.rlc.rlc_autoload_ptr); 1016 1017 if (r) { 1018 dev_err(adev->dev, "(%d) failed to create fw autoload bo\n", r); 1019 return r; 1020 } 1021 1022 return 0; 1023 } 1024 1025 static void gfx_v12_0_rlc_backdoor_autoload_copy_ucode(struct amdgpu_device *adev, 1026 SOC24_FIRMWARE_ID id, 1027 const void *fw_data, 1028 uint32_t fw_size) 1029 { 1030 uint32_t toc_offset; 1031 uint32_t toc_fw_size; 1032 char *ptr = adev->gfx.rlc.rlc_autoload_ptr; 1033 1034 if (id <= SOC24_FIRMWARE_ID_INVALID || id >= SOC24_FIRMWARE_ID_MAX) 1035 return; 1036 1037 toc_offset = rlc_autoload_info[id].offset; 1038 toc_fw_size = rlc_autoload_info[id].size; 1039 1040 if (fw_size == 0) 1041 fw_size = toc_fw_size; 1042 1043 if (fw_size > toc_fw_size) 1044 fw_size = toc_fw_size; 1045 1046 memcpy(ptr + toc_offset, fw_data, fw_size); 1047 1048 if (fw_size < toc_fw_size) 1049 memset(ptr + toc_offset + fw_size, 0, toc_fw_size - fw_size); 1050 } 1051 1052 static void 1053 gfx_v12_0_rlc_backdoor_autoload_copy_toc_ucode(struct amdgpu_device *adev) 1054 { 1055 void *data; 1056 uint32_t size; 1057 uint32_t *toc_ptr; 1058 1059 data = adev->psp.toc.start_addr; 1060 size = rlc_autoload_info[SOC24_FIRMWARE_ID_RLC_TOC].size; 1061 1062 toc_ptr = (uint32_t *)data + size / 4 - 2; 1063 *toc_ptr = (RLC_TOC_FORMAT_API << 24) | 0x1; 1064 1065 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RLC_TOC, 1066 data, size); 1067 } 1068 1069 static void 1070 gfx_v12_0_rlc_backdoor_autoload_copy_gfx_ucode(struct amdgpu_device *adev) 1071 { 1072 const __le32 *fw_data; 1073 uint32_t fw_size; 1074 const struct gfx_firmware_header_v2_0 *cpv2_hdr; 1075 const struct rlc_firmware_header_v2_0 *rlc_hdr; 1076 const struct rlc_firmware_header_v2_1 *rlcv21_hdr; 1077 const struct rlc_firmware_header_v2_2 *rlcv22_hdr; 1078 uint16_t version_major, version_minor; 1079 1080 /* pfp ucode */ 1081 cpv2_hdr = (const struct gfx_firmware_header_v2_0 *) 1082 adev->gfx.pfp_fw->data; 1083 /* instruction */ 1084 fw_data = (const __le32 *)(adev->gfx.pfp_fw->data + 1085 le32_to_cpu(cpv2_hdr->ucode_offset_bytes)); 1086 fw_size = le32_to_cpu(cpv2_hdr->ucode_size_bytes); 1087 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_PFP, 1088 fw_data, fw_size); 1089 /* data */ 1090 fw_data = (const __le32 *)(adev->gfx.pfp_fw->data + 1091 le32_to_cpu(cpv2_hdr->data_offset_bytes)); 1092 fw_size = le32_to_cpu(cpv2_hdr->data_size_bytes); 1093 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_PFP_P0_STACK, 1094 fw_data, fw_size); 1095 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_PFP_P1_STACK, 1096 fw_data, fw_size); 1097 /* me ucode */ 1098 cpv2_hdr = (const struct gfx_firmware_header_v2_0 *) 1099 adev->gfx.me_fw->data; 1100 /* instruction */ 1101 fw_data = (const __le32 *)(adev->gfx.me_fw->data + 1102 le32_to_cpu(cpv2_hdr->ucode_offset_bytes)); 1103 fw_size = le32_to_cpu(cpv2_hdr->ucode_size_bytes); 1104 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_ME, 1105 fw_data, fw_size); 1106 /* data */ 1107 fw_data = (const __le32 *)(adev->gfx.me_fw->data + 1108 le32_to_cpu(cpv2_hdr->data_offset_bytes)); 1109 fw_size = le32_to_cpu(cpv2_hdr->data_size_bytes); 1110 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_ME_P0_STACK, 1111 fw_data, fw_size); 1112 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_ME_P1_STACK, 1113 fw_data, fw_size); 1114 /* mec ucode */ 1115 cpv2_hdr = (const struct gfx_firmware_header_v2_0 *) 1116 adev->gfx.mec_fw->data; 1117 /* instruction */ 1118 fw_data = (const __le32 *) (adev->gfx.mec_fw->data + 1119 le32_to_cpu(cpv2_hdr->ucode_offset_bytes)); 1120 fw_size = le32_to_cpu(cpv2_hdr->ucode_size_bytes); 1121 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_MEC, 1122 fw_data, fw_size); 1123 /* data */ 1124 fw_data = (const __le32 *) (adev->gfx.mec_fw->data + 1125 le32_to_cpu(cpv2_hdr->data_offset_bytes)); 1126 fw_size = le32_to_cpu(cpv2_hdr->data_size_bytes); 1127 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_MEC_P0_STACK, 1128 fw_data, fw_size); 1129 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_MEC_P1_STACK, 1130 fw_data, fw_size); 1131 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_MEC_P2_STACK, 1132 fw_data, fw_size); 1133 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RS64_MEC_P3_STACK, 1134 fw_data, fw_size); 1135 1136 /* rlc ucode */ 1137 rlc_hdr = (const struct rlc_firmware_header_v2_0 *) 1138 adev->gfx.rlc_fw->data; 1139 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data + 1140 le32_to_cpu(rlc_hdr->header.ucode_array_offset_bytes)); 1141 fw_size = le32_to_cpu(rlc_hdr->header.ucode_size_bytes); 1142 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RLC_G_UCODE, 1143 fw_data, fw_size); 1144 1145 version_major = le16_to_cpu(rlc_hdr->header.header_version_major); 1146 version_minor = le16_to_cpu(rlc_hdr->header.header_version_minor); 1147 if (version_major == 2) { 1148 if (version_minor >= 1) { 1149 rlcv21_hdr = (const struct rlc_firmware_header_v2_1 *)adev->gfx.rlc_fw->data; 1150 1151 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data + 1152 le32_to_cpu(rlcv21_hdr->save_restore_list_gpm_offset_bytes)); 1153 fw_size = le32_to_cpu(rlcv21_hdr->save_restore_list_gpm_size_bytes); 1154 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RLCG_SCRATCH, 1155 fw_data, fw_size); 1156 1157 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data + 1158 le32_to_cpu(rlcv21_hdr->save_restore_list_srm_offset_bytes)); 1159 fw_size = le32_to_cpu(rlcv21_hdr->save_restore_list_srm_size_bytes); 1160 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RLC_SRM_ARAM, 1161 fw_data, fw_size); 1162 } 1163 if (version_minor >= 2) { 1164 rlcv22_hdr = (const struct rlc_firmware_header_v2_2 *)adev->gfx.rlc_fw->data; 1165 1166 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data + 1167 le32_to_cpu(rlcv22_hdr->rlc_iram_ucode_offset_bytes)); 1168 fw_size = le32_to_cpu(rlcv22_hdr->rlc_iram_ucode_size_bytes); 1169 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RLX6_UCODE, 1170 fw_data, fw_size); 1171 1172 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data + 1173 le32_to_cpu(rlcv22_hdr->rlc_dram_ucode_offset_bytes)); 1174 fw_size = le32_to_cpu(rlcv22_hdr->rlc_dram_ucode_size_bytes); 1175 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_RLX6_DRAM_BOOT, 1176 fw_data, fw_size); 1177 } 1178 } 1179 } 1180 1181 static void 1182 gfx_v12_0_rlc_backdoor_autoload_copy_sdma_ucode(struct amdgpu_device *adev) 1183 { 1184 const __le32 *fw_data; 1185 uint32_t fw_size; 1186 const struct sdma_firmware_header_v3_0 *sdma_hdr; 1187 1188 sdma_hdr = (const struct sdma_firmware_header_v3_0 *) 1189 adev->sdma.instance[0].fw->data; 1190 fw_data = (const __le32 *) (adev->sdma.instance[0].fw->data + 1191 le32_to_cpu(sdma_hdr->ucode_offset_bytes)); 1192 fw_size = le32_to_cpu(sdma_hdr->ucode_size_bytes); 1193 1194 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, SOC24_FIRMWARE_ID_SDMA_UCODE_TH0, 1195 fw_data, fw_size); 1196 } 1197 1198 static void 1199 gfx_v12_0_rlc_backdoor_autoload_copy_mes_ucode(struct amdgpu_device *adev) 1200 { 1201 const __le32 *fw_data; 1202 unsigned fw_size; 1203 const struct mes_firmware_header_v1_0 *mes_hdr; 1204 int pipe, ucode_id, data_id; 1205 1206 for (pipe = 0; pipe < 2; pipe++) { 1207 if (pipe == 0) { 1208 ucode_id = SOC24_FIRMWARE_ID_RS64_MES_P0; 1209 data_id = SOC24_FIRMWARE_ID_RS64_MES_P0_STACK; 1210 } else { 1211 ucode_id = SOC24_FIRMWARE_ID_RS64_MES_P1; 1212 data_id = SOC24_FIRMWARE_ID_RS64_MES_P1_STACK; 1213 } 1214 1215 mes_hdr = (const struct mes_firmware_header_v1_0 *) 1216 adev->mes.fw[pipe]->data; 1217 1218 fw_data = (const __le32 *)(adev->mes.fw[pipe]->data + 1219 le32_to_cpu(mes_hdr->mes_ucode_offset_bytes)); 1220 fw_size = le32_to_cpu(mes_hdr->mes_ucode_size_bytes); 1221 1222 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, ucode_id, fw_data, fw_size); 1223 1224 fw_data = (const __le32 *)(adev->mes.fw[pipe]->data + 1225 le32_to_cpu(mes_hdr->mes_ucode_data_offset_bytes)); 1226 fw_size = le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes); 1227 1228 gfx_v12_0_rlc_backdoor_autoload_copy_ucode(adev, data_id, fw_data, fw_size); 1229 } 1230 } 1231 1232 static int gfx_v12_0_rlc_backdoor_autoload_enable(struct amdgpu_device *adev) 1233 { 1234 uint32_t rlc_g_offset, rlc_g_size; 1235 uint64_t gpu_addr; 1236 uint32_t data; 1237 1238 /* RLC autoload sequence 2: copy ucode */ 1239 gfx_v12_0_rlc_backdoor_autoload_copy_sdma_ucode(adev); 1240 gfx_v12_0_rlc_backdoor_autoload_copy_gfx_ucode(adev); 1241 gfx_v12_0_rlc_backdoor_autoload_copy_mes_ucode(adev); 1242 gfx_v12_0_rlc_backdoor_autoload_copy_toc_ucode(adev); 1243 1244 rlc_g_offset = rlc_autoload_info[SOC24_FIRMWARE_ID_RLC_G_UCODE].offset; 1245 rlc_g_size = rlc_autoload_info[SOC24_FIRMWARE_ID_RLC_G_UCODE].size; 1246 gpu_addr = adev->gfx.rlc.rlc_autoload_gpu_addr + rlc_g_offset - adev->gmc.vram_start; 1247 1248 WREG32_SOC15(GC, 0, regGFX_IMU_RLC_BOOTLOADER_ADDR_HI, upper_32_bits(gpu_addr)); 1249 WREG32_SOC15(GC, 0, regGFX_IMU_RLC_BOOTLOADER_ADDR_LO, lower_32_bits(gpu_addr)); 1250 1251 WREG32_SOC15(GC, 0, regGFX_IMU_RLC_BOOTLOADER_SIZE, rlc_g_size); 1252 1253 if (adev->gfx.imu.funcs && (amdgpu_dpm > 0)) { 1254 /* RLC autoload sequence 3: load IMU fw */ 1255 if (adev->gfx.imu.funcs->load_microcode) 1256 adev->gfx.imu.funcs->load_microcode(adev); 1257 /* RLC autoload sequence 4 init IMU fw */ 1258 if (adev->gfx.imu.funcs->setup_imu) 1259 adev->gfx.imu.funcs->setup_imu(adev); 1260 if (adev->gfx.imu.funcs->start_imu) 1261 adev->gfx.imu.funcs->start_imu(adev); 1262 1263 /* RLC autoload sequence 5 disable gpa mode */ 1264 gfx_v12_0_disable_gpa_mode(adev); 1265 } else { 1266 /* unhalt rlc to start autoload without imu */ 1267 data = RREG32_SOC15(GC, 0, regRLC_GPM_THREAD_ENABLE); 1268 data = REG_SET_FIELD(data, RLC_GPM_THREAD_ENABLE, THREAD0_ENABLE, 1); 1269 data = REG_SET_FIELD(data, RLC_GPM_THREAD_ENABLE, THREAD1_ENABLE, 1); 1270 WREG32_SOC15(GC, 0, regRLC_GPM_THREAD_ENABLE, data); 1271 WREG32_SOC15(GC, 0, regRLC_CNTL, RLC_CNTL__RLC_ENABLE_F32_MASK); 1272 } 1273 1274 return 0; 1275 } 1276 1277 static void gfx_v12_0_alloc_ip_dump(struct amdgpu_device *adev) 1278 { 1279 uint32_t reg_count = ARRAY_SIZE(gc_reg_list_12_0); 1280 uint32_t *ptr; 1281 uint32_t inst; 1282 1283 ptr = kcalloc(reg_count, sizeof(uint32_t), GFP_KERNEL); 1284 if (ptr == NULL) { 1285 DRM_ERROR("Failed to allocate memory for GFX IP Dump\n"); 1286 adev->gfx.ip_dump_core = NULL; 1287 } else { 1288 adev->gfx.ip_dump_core = ptr; 1289 } 1290 1291 /* Allocate memory for compute queue registers for all the instances */ 1292 reg_count = ARRAY_SIZE(gc_cp_reg_list_12); 1293 inst = adev->gfx.mec.num_mec * adev->gfx.mec.num_pipe_per_mec * 1294 adev->gfx.mec.num_queue_per_pipe; 1295 1296 ptr = kcalloc(reg_count * inst, sizeof(uint32_t), GFP_KERNEL); 1297 if (ptr == NULL) { 1298 DRM_ERROR("Failed to allocate memory for Compute Queues IP Dump\n"); 1299 adev->gfx.ip_dump_compute_queues = NULL; 1300 } else { 1301 adev->gfx.ip_dump_compute_queues = ptr; 1302 } 1303 1304 /* Allocate memory for gfx queue registers for all the instances */ 1305 reg_count = ARRAY_SIZE(gc_gfx_queue_reg_list_12); 1306 inst = adev->gfx.me.num_me * adev->gfx.me.num_pipe_per_me * 1307 adev->gfx.me.num_queue_per_pipe; 1308 1309 ptr = kcalloc(reg_count * inst, sizeof(uint32_t), GFP_KERNEL); 1310 if (ptr == NULL) { 1311 DRM_ERROR("Failed to allocate memory for GFX Queues IP Dump\n"); 1312 adev->gfx.ip_dump_gfx_queues = NULL; 1313 } else { 1314 adev->gfx.ip_dump_gfx_queues = ptr; 1315 } 1316 } 1317 1318 static int gfx_v12_0_sw_init(void *handle) 1319 { 1320 int i, j, k, r, ring_id = 0; 1321 unsigned num_compute_rings; 1322 int xcc_id = 0; 1323 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1324 1325 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 1326 case IP_VERSION(12, 0, 0): 1327 case IP_VERSION(12, 0, 1): 1328 adev->gfx.me.num_me = 1; 1329 adev->gfx.me.num_pipe_per_me = 1; 1330 adev->gfx.me.num_queue_per_pipe = 1; 1331 adev->gfx.mec.num_mec = 2; 1332 adev->gfx.mec.num_pipe_per_mec = 2; 1333 adev->gfx.mec.num_queue_per_pipe = 4; 1334 break; 1335 default: 1336 adev->gfx.me.num_me = 1; 1337 adev->gfx.me.num_pipe_per_me = 1; 1338 adev->gfx.me.num_queue_per_pipe = 1; 1339 adev->gfx.mec.num_mec = 1; 1340 adev->gfx.mec.num_pipe_per_mec = 4; 1341 adev->gfx.mec.num_queue_per_pipe = 8; 1342 break; 1343 } 1344 1345 /* recalculate compute rings to use based on hardware configuration */ 1346 num_compute_rings = (adev->gfx.mec.num_pipe_per_mec * 1347 adev->gfx.mec.num_queue_per_pipe) / 2; 1348 adev->gfx.num_compute_rings = min(adev->gfx.num_compute_rings, 1349 num_compute_rings); 1350 1351 /* EOP Event */ 1352 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GRBM_CP, 1353 GFX_11_0_0__SRCID__CP_EOP_INTERRUPT, 1354 &adev->gfx.eop_irq); 1355 if (r) 1356 return r; 1357 1358 /* Bad opcode Event */ 1359 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GRBM_CP, 1360 GFX_11_0_0__SRCID__CP_BAD_OPCODE_ERROR, 1361 &adev->gfx.bad_op_irq); 1362 if (r) 1363 return r; 1364 1365 /* Privileged reg */ 1366 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GRBM_CP, 1367 GFX_11_0_0__SRCID__CP_PRIV_REG_FAULT, 1368 &adev->gfx.priv_reg_irq); 1369 if (r) 1370 return r; 1371 1372 /* Privileged inst */ 1373 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GRBM_CP, 1374 GFX_11_0_0__SRCID__CP_PRIV_INSTR_FAULT, 1375 &adev->gfx.priv_inst_irq); 1376 if (r) 1377 return r; 1378 1379 adev->gfx.gfx_current_status = AMDGPU_GFX_NORMAL_MODE; 1380 1381 gfx_v12_0_me_init(adev); 1382 1383 r = gfx_v12_0_rlc_init(adev); 1384 if (r) { 1385 dev_err(adev->dev, "Failed to init rlc BOs!\n"); 1386 return r; 1387 } 1388 1389 r = gfx_v12_0_mec_init(adev); 1390 if (r) { 1391 dev_err(adev->dev, "Failed to init MEC BOs!\n"); 1392 return r; 1393 } 1394 1395 /* set up the gfx ring */ 1396 for (i = 0; i < adev->gfx.me.num_me; i++) { 1397 for (j = 0; j < adev->gfx.me.num_queue_per_pipe; j++) { 1398 for (k = 0; k < adev->gfx.me.num_pipe_per_me; k++) { 1399 if (!amdgpu_gfx_is_me_queue_enabled(adev, i, k, j)) 1400 continue; 1401 1402 r = gfx_v12_0_gfx_ring_init(adev, ring_id, 1403 i, k, j); 1404 if (r) 1405 return r; 1406 ring_id++; 1407 } 1408 } 1409 } 1410 1411 ring_id = 0; 1412 /* set up the compute queues - allocate horizontally across pipes */ 1413 for (i = 0; i < adev->gfx.mec.num_mec; ++i) { 1414 for (j = 0; j < adev->gfx.mec.num_queue_per_pipe; j++) { 1415 for (k = 0; k < adev->gfx.mec.num_pipe_per_mec; k++) { 1416 if (!amdgpu_gfx_is_mec_queue_enabled(adev, 1417 0, i, k, j)) 1418 continue; 1419 1420 r = gfx_v12_0_compute_ring_init(adev, ring_id, 1421 i, k, j); 1422 if (r) 1423 return r; 1424 1425 ring_id++; 1426 } 1427 } 1428 } 1429 1430 if (!adev->enable_mes_kiq) { 1431 r = amdgpu_gfx_kiq_init(adev, GFX12_MEC_HPD_SIZE, 0); 1432 if (r) { 1433 dev_err(adev->dev, "Failed to init KIQ BOs!\n"); 1434 return r; 1435 } 1436 1437 r = amdgpu_gfx_kiq_init_ring(adev, xcc_id); 1438 if (r) 1439 return r; 1440 } 1441 1442 r = amdgpu_gfx_mqd_sw_init(adev, sizeof(struct v12_compute_mqd), 0); 1443 if (r) 1444 return r; 1445 1446 /* allocate visible FB for rlc auto-loading fw */ 1447 if (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) { 1448 r = gfx_v12_0_rlc_autoload_buffer_init(adev); 1449 if (r) 1450 return r; 1451 } 1452 1453 r = gfx_v12_0_gpu_early_init(adev); 1454 if (r) 1455 return r; 1456 1457 gfx_v12_0_alloc_ip_dump(adev); 1458 1459 return 0; 1460 } 1461 1462 static void gfx_v12_0_pfp_fini(struct amdgpu_device *adev) 1463 { 1464 amdgpu_bo_free_kernel(&adev->gfx.pfp.pfp_fw_obj, 1465 &adev->gfx.pfp.pfp_fw_gpu_addr, 1466 (void **)&adev->gfx.pfp.pfp_fw_ptr); 1467 1468 amdgpu_bo_free_kernel(&adev->gfx.pfp.pfp_fw_data_obj, 1469 &adev->gfx.pfp.pfp_fw_data_gpu_addr, 1470 (void **)&adev->gfx.pfp.pfp_fw_data_ptr); 1471 } 1472 1473 static void gfx_v12_0_me_fini(struct amdgpu_device *adev) 1474 { 1475 amdgpu_bo_free_kernel(&adev->gfx.me.me_fw_obj, 1476 &adev->gfx.me.me_fw_gpu_addr, 1477 (void **)&adev->gfx.me.me_fw_ptr); 1478 1479 amdgpu_bo_free_kernel(&adev->gfx.me.me_fw_data_obj, 1480 &adev->gfx.me.me_fw_data_gpu_addr, 1481 (void **)&adev->gfx.me.me_fw_data_ptr); 1482 } 1483 1484 static void gfx_v12_0_rlc_autoload_buffer_fini(struct amdgpu_device *adev) 1485 { 1486 amdgpu_bo_free_kernel(&adev->gfx.rlc.rlc_autoload_bo, 1487 &adev->gfx.rlc.rlc_autoload_gpu_addr, 1488 (void **)&adev->gfx.rlc.rlc_autoload_ptr); 1489 } 1490 1491 static int gfx_v12_0_sw_fini(void *handle) 1492 { 1493 int i; 1494 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 1495 1496 for (i = 0; i < adev->gfx.num_gfx_rings; i++) 1497 amdgpu_ring_fini(&adev->gfx.gfx_ring[i]); 1498 for (i = 0; i < adev->gfx.num_compute_rings; i++) 1499 amdgpu_ring_fini(&adev->gfx.compute_ring[i]); 1500 1501 amdgpu_gfx_mqd_sw_fini(adev, 0); 1502 1503 if (!adev->enable_mes_kiq) { 1504 amdgpu_gfx_kiq_free_ring(&adev->gfx.kiq[0].ring); 1505 amdgpu_gfx_kiq_fini(adev, 0); 1506 } 1507 1508 gfx_v12_0_pfp_fini(adev); 1509 gfx_v12_0_me_fini(adev); 1510 gfx_v12_0_rlc_fini(adev); 1511 gfx_v12_0_mec_fini(adev); 1512 1513 if (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) 1514 gfx_v12_0_rlc_autoload_buffer_fini(adev); 1515 1516 gfx_v12_0_free_microcode(adev); 1517 1518 kfree(adev->gfx.ip_dump_core); 1519 kfree(adev->gfx.ip_dump_compute_queues); 1520 kfree(adev->gfx.ip_dump_gfx_queues); 1521 1522 return 0; 1523 } 1524 1525 static void gfx_v12_0_select_se_sh(struct amdgpu_device *adev, u32 se_num, 1526 u32 sh_num, u32 instance, int xcc_id) 1527 { 1528 u32 data; 1529 1530 if (instance == 0xffffffff) 1531 data = REG_SET_FIELD(0, GRBM_GFX_INDEX, 1532 INSTANCE_BROADCAST_WRITES, 1); 1533 else 1534 data = REG_SET_FIELD(0, GRBM_GFX_INDEX, INSTANCE_INDEX, 1535 instance); 1536 1537 if (se_num == 0xffffffff) 1538 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SE_BROADCAST_WRITES, 1539 1); 1540 else 1541 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SE_INDEX, se_num); 1542 1543 if (sh_num == 0xffffffff) 1544 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SA_BROADCAST_WRITES, 1545 1); 1546 else 1547 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SA_INDEX, sh_num); 1548 1549 WREG32_SOC15(GC, 0, regGRBM_GFX_INDEX, data); 1550 } 1551 1552 static u32 gfx_v12_0_get_sa_active_bitmap(struct amdgpu_device *adev) 1553 { 1554 u32 gc_disabled_sa_mask, gc_user_disabled_sa_mask, sa_mask; 1555 1556 gc_disabled_sa_mask = RREG32_SOC15(GC, 0, regGRBM_CC_GC_SA_UNIT_DISABLE); 1557 gc_disabled_sa_mask = REG_GET_FIELD(gc_disabled_sa_mask, 1558 GRBM_CC_GC_SA_UNIT_DISABLE, 1559 SA_DISABLE); 1560 gc_user_disabled_sa_mask = RREG32_SOC15(GC, 0, regGRBM_GC_USER_SA_UNIT_DISABLE); 1561 gc_user_disabled_sa_mask = REG_GET_FIELD(gc_user_disabled_sa_mask, 1562 GRBM_GC_USER_SA_UNIT_DISABLE, 1563 SA_DISABLE); 1564 sa_mask = amdgpu_gfx_create_bitmask(adev->gfx.config.max_sh_per_se * 1565 adev->gfx.config.max_shader_engines); 1566 1567 return sa_mask & (~(gc_disabled_sa_mask | gc_user_disabled_sa_mask)); 1568 } 1569 1570 static u32 gfx_v12_0_get_rb_active_bitmap(struct amdgpu_device *adev) 1571 { 1572 u32 gc_disabled_rb_mask, gc_user_disabled_rb_mask; 1573 u32 rb_mask; 1574 1575 gc_disabled_rb_mask = RREG32_SOC15(GC, 0, regCC_RB_BACKEND_DISABLE); 1576 gc_disabled_rb_mask = REG_GET_FIELD(gc_disabled_rb_mask, 1577 CC_RB_BACKEND_DISABLE, 1578 BACKEND_DISABLE); 1579 gc_user_disabled_rb_mask = RREG32_SOC15(GC, 0, regGC_USER_RB_BACKEND_DISABLE); 1580 gc_user_disabled_rb_mask = REG_GET_FIELD(gc_user_disabled_rb_mask, 1581 GC_USER_RB_BACKEND_DISABLE, 1582 BACKEND_DISABLE); 1583 rb_mask = amdgpu_gfx_create_bitmask(adev->gfx.config.max_backends_per_se * 1584 adev->gfx.config.max_shader_engines); 1585 1586 return rb_mask & (~(gc_disabled_rb_mask | gc_user_disabled_rb_mask)); 1587 } 1588 1589 static void gfx_v12_0_setup_rb(struct amdgpu_device *adev) 1590 { 1591 u32 rb_bitmap_width_per_sa; 1592 u32 max_sa; 1593 u32 active_sa_bitmap; 1594 u32 global_active_rb_bitmap; 1595 u32 active_rb_bitmap = 0; 1596 u32 i; 1597 1598 /* query sa bitmap from SA_UNIT_DISABLE registers */ 1599 active_sa_bitmap = gfx_v12_0_get_sa_active_bitmap(adev); 1600 /* query rb bitmap from RB_BACKEND_DISABLE registers */ 1601 global_active_rb_bitmap = gfx_v12_0_get_rb_active_bitmap(adev); 1602 1603 /* generate active rb bitmap according to active sa bitmap */ 1604 max_sa = adev->gfx.config.max_shader_engines * 1605 adev->gfx.config.max_sh_per_se; 1606 rb_bitmap_width_per_sa = adev->gfx.config.max_backends_per_se / 1607 adev->gfx.config.max_sh_per_se; 1608 for (i = 0; i < max_sa; i++) { 1609 if (active_sa_bitmap & (1 << i)) 1610 active_rb_bitmap |= (0x3 << (i * rb_bitmap_width_per_sa)); 1611 } 1612 1613 active_rb_bitmap |= global_active_rb_bitmap; 1614 adev->gfx.config.backend_enable_mask = active_rb_bitmap; 1615 adev->gfx.config.num_rbs = hweight32(active_rb_bitmap); 1616 } 1617 1618 #define LDS_APP_BASE 0x1 1619 #define SCRATCH_APP_BASE 0x2 1620 1621 static void gfx_v12_0_init_compute_vmid(struct amdgpu_device *adev) 1622 { 1623 int i; 1624 uint32_t sh_mem_bases; 1625 uint32_t data; 1626 1627 /* 1628 * Configure apertures: 1629 * LDS: 0x60000000'00000000 - 0x60000001'00000000 (4GB) 1630 * Scratch: 0x60000001'00000000 - 0x60000002'00000000 (4GB) 1631 * GPUVM: 0x60010000'00000000 - 0x60020000'00000000 (1TB) 1632 */ 1633 sh_mem_bases = (LDS_APP_BASE << SH_MEM_BASES__SHARED_BASE__SHIFT) | 1634 SCRATCH_APP_BASE; 1635 1636 mutex_lock(&adev->srbm_mutex); 1637 for (i = adev->vm_manager.first_kfd_vmid; i < AMDGPU_NUM_VMID; i++) { 1638 soc24_grbm_select(adev, 0, 0, 0, i); 1639 /* CP and shaders */ 1640 WREG32_SOC15(GC, 0, regSH_MEM_CONFIG, DEFAULT_SH_MEM_CONFIG); 1641 WREG32_SOC15(GC, 0, regSH_MEM_BASES, sh_mem_bases); 1642 1643 /* Enable trap for each kfd vmid. */ 1644 data = RREG32_SOC15(GC, 0, regSPI_GDBG_PER_VMID_CNTL); 1645 data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, TRAP_EN, 1); 1646 WREG32_SOC15(GC, 0, regSPI_GDBG_PER_VMID_CNTL, data); 1647 } 1648 soc24_grbm_select(adev, 0, 0, 0, 0); 1649 mutex_unlock(&adev->srbm_mutex); 1650 } 1651 1652 static void gfx_v12_0_tcp_harvest(struct amdgpu_device *adev) 1653 { 1654 /* TODO: harvest feature to be added later. */ 1655 } 1656 1657 static void gfx_v12_0_get_tcc_info(struct amdgpu_device *adev) 1658 { 1659 } 1660 1661 static void gfx_v12_0_constants_init(struct amdgpu_device *adev) 1662 { 1663 u32 tmp; 1664 int i; 1665 1666 if (!amdgpu_sriov_vf(adev)) 1667 WREG32_FIELD15_PREREG(GC, 0, GRBM_CNTL, READ_TIMEOUT, 0xff); 1668 1669 gfx_v12_0_setup_rb(adev); 1670 gfx_v12_0_get_cu_info(adev, &adev->gfx.cu_info); 1671 gfx_v12_0_get_tcc_info(adev); 1672 adev->gfx.config.pa_sc_tile_steering_override = 0; 1673 1674 /* XXX SH_MEM regs */ 1675 /* where to put LDS, scratch, GPUVM in FSA64 space */ 1676 mutex_lock(&adev->srbm_mutex); 1677 for (i = 0; i < adev->vm_manager.id_mgr[AMDGPU_GFXHUB(0)].num_ids; i++) { 1678 soc24_grbm_select(adev, 0, 0, 0, i); 1679 /* CP and shaders */ 1680 WREG32_SOC15(GC, 0, regSH_MEM_CONFIG, DEFAULT_SH_MEM_CONFIG); 1681 if (i != 0) { 1682 tmp = REG_SET_FIELD(0, SH_MEM_BASES, PRIVATE_BASE, 1683 (adev->gmc.private_aperture_start >> 48)); 1684 tmp = REG_SET_FIELD(tmp, SH_MEM_BASES, SHARED_BASE, 1685 (adev->gmc.shared_aperture_start >> 48)); 1686 WREG32_SOC15(GC, 0, regSH_MEM_BASES, tmp); 1687 } 1688 } 1689 soc24_grbm_select(adev, 0, 0, 0, 0); 1690 1691 mutex_unlock(&adev->srbm_mutex); 1692 1693 gfx_v12_0_init_compute_vmid(adev); 1694 } 1695 1696 static u32 gfx_v12_0_get_cpg_int_cntl(struct amdgpu_device *adev, 1697 int me, int pipe) 1698 { 1699 if (me != 0) 1700 return 0; 1701 1702 switch (pipe) { 1703 case 0: 1704 return SOC15_REG_OFFSET(GC, 0, regCP_INT_CNTL_RING0); 1705 default: 1706 return 0; 1707 } 1708 } 1709 1710 static u32 gfx_v12_0_get_cpc_int_cntl(struct amdgpu_device *adev, 1711 int me, int pipe) 1712 { 1713 /* 1714 * amdgpu controls only the first MEC. That's why this function only 1715 * handles the setting of interrupts for this specific MEC. All other 1716 * pipes' interrupts are set by amdkfd. 1717 */ 1718 if (me != 1) 1719 return 0; 1720 1721 switch (pipe) { 1722 case 0: 1723 return SOC15_REG_OFFSET(GC, 0, regCP_ME1_PIPE0_INT_CNTL); 1724 case 1: 1725 return SOC15_REG_OFFSET(GC, 0, regCP_ME1_PIPE1_INT_CNTL); 1726 default: 1727 return 0; 1728 } 1729 } 1730 1731 static void gfx_v12_0_enable_gui_idle_interrupt(struct amdgpu_device *adev, 1732 bool enable) 1733 { 1734 u32 tmp, cp_int_cntl_reg; 1735 int i, j; 1736 1737 if (amdgpu_sriov_vf(adev)) 1738 return; 1739 1740 for (i = 0; i < adev->gfx.me.num_me; i++) { 1741 for (j = 0; j < adev->gfx.me.num_pipe_per_me; j++) { 1742 cp_int_cntl_reg = gfx_v12_0_get_cpg_int_cntl(adev, i, j); 1743 1744 if (cp_int_cntl_reg) { 1745 tmp = RREG32_SOC15_IP(GC, cp_int_cntl_reg); 1746 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL_RING0, CNTX_BUSY_INT_ENABLE, 1747 enable ? 1 : 0); 1748 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL_RING0, CNTX_EMPTY_INT_ENABLE, 1749 enable ? 1 : 0); 1750 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL_RING0, CMP_BUSY_INT_ENABLE, 1751 enable ? 1 : 0); 1752 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL_RING0, GFX_IDLE_INT_ENABLE, 1753 enable ? 1 : 0); 1754 WREG32_SOC15_IP(GC, cp_int_cntl_reg, tmp); 1755 } 1756 } 1757 } 1758 } 1759 1760 static int gfx_v12_0_init_csb(struct amdgpu_device *adev) 1761 { 1762 adev->gfx.rlc.funcs->get_csb_buffer(adev, adev->gfx.rlc.cs_ptr); 1763 1764 WREG32_SOC15(GC, 0, regRLC_CSIB_ADDR_HI, 1765 adev->gfx.rlc.clear_state_gpu_addr >> 32); 1766 WREG32_SOC15(GC, 0, regRLC_CSIB_ADDR_LO, 1767 adev->gfx.rlc.clear_state_gpu_addr & 0xfffffffc); 1768 WREG32_SOC15(GC, 0, regRLC_CSIB_LENGTH, adev->gfx.rlc.clear_state_size); 1769 1770 return 0; 1771 } 1772 1773 static void gfx_v12_0_rlc_stop(struct amdgpu_device *adev) 1774 { 1775 u32 tmp = RREG32_SOC15(GC, 0, regRLC_CNTL); 1776 1777 tmp = REG_SET_FIELD(tmp, RLC_CNTL, RLC_ENABLE_F32, 0); 1778 WREG32_SOC15(GC, 0, regRLC_CNTL, tmp); 1779 } 1780 1781 static void gfx_v12_0_rlc_reset(struct amdgpu_device *adev) 1782 { 1783 WREG32_FIELD15_PREREG(GC, 0, GRBM_SOFT_RESET, SOFT_RESET_RLC, 1); 1784 udelay(50); 1785 WREG32_FIELD15_PREREG(GC, 0, GRBM_SOFT_RESET, SOFT_RESET_RLC, 0); 1786 udelay(50); 1787 } 1788 1789 static void gfx_v12_0_rlc_smu_handshake_cntl(struct amdgpu_device *adev, 1790 bool enable) 1791 { 1792 uint32_t rlc_pg_cntl; 1793 1794 rlc_pg_cntl = RREG32_SOC15(GC, 0, regRLC_PG_CNTL); 1795 1796 if (!enable) { 1797 /* RLC_PG_CNTL[23] = 0 (default) 1798 * RLC will wait for handshake acks with SMU 1799 * GFXOFF will be enabled 1800 * RLC_PG_CNTL[23] = 1 1801 * RLC will not issue any message to SMU 1802 * hence no handshake between SMU & RLC 1803 * GFXOFF will be disabled 1804 */ 1805 rlc_pg_cntl |= RLC_PG_CNTL__SMU_HANDSHAKE_DISABLE_MASK; 1806 } else 1807 rlc_pg_cntl &= ~RLC_PG_CNTL__SMU_HANDSHAKE_DISABLE_MASK; 1808 WREG32_SOC15(GC, 0, regRLC_PG_CNTL, rlc_pg_cntl); 1809 } 1810 1811 static void gfx_v12_0_rlc_start(struct amdgpu_device *adev) 1812 { 1813 /* TODO: enable rlc & smu handshake until smu 1814 * and gfxoff feature works as expected */ 1815 if (!(amdgpu_pp_feature_mask & PP_GFXOFF_MASK)) 1816 gfx_v12_0_rlc_smu_handshake_cntl(adev, false); 1817 1818 WREG32_FIELD15_PREREG(GC, 0, RLC_CNTL, RLC_ENABLE_F32, 1); 1819 udelay(50); 1820 } 1821 1822 static void gfx_v12_0_rlc_enable_srm(struct amdgpu_device *adev) 1823 { 1824 uint32_t tmp; 1825 1826 /* enable Save Restore Machine */ 1827 tmp = RREG32(SOC15_REG_OFFSET(GC, 0, regRLC_SRM_CNTL)); 1828 tmp |= RLC_SRM_CNTL__AUTO_INCR_ADDR_MASK; 1829 tmp |= RLC_SRM_CNTL__SRM_ENABLE_MASK; 1830 WREG32(SOC15_REG_OFFSET(GC, 0, regRLC_SRM_CNTL), tmp); 1831 } 1832 1833 static void gfx_v12_0_load_rlcg_microcode(struct amdgpu_device *adev) 1834 { 1835 const struct rlc_firmware_header_v2_0 *hdr; 1836 const __le32 *fw_data; 1837 unsigned i, fw_size; 1838 1839 hdr = (const struct rlc_firmware_header_v2_0 *)adev->gfx.rlc_fw->data; 1840 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data + 1841 le32_to_cpu(hdr->header.ucode_array_offset_bytes)); 1842 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4; 1843 1844 WREG32_SOC15(GC, 0, regRLC_GPM_UCODE_ADDR, 1845 RLCG_UCODE_LOADING_START_ADDRESS); 1846 1847 for (i = 0; i < fw_size; i++) 1848 WREG32_SOC15(GC, 0, regRLC_GPM_UCODE_DATA, 1849 le32_to_cpup(fw_data++)); 1850 1851 WREG32_SOC15(GC, 0, regRLC_GPM_UCODE_ADDR, adev->gfx.rlc_fw_version); 1852 } 1853 1854 static void gfx_v12_0_load_rlc_iram_dram_microcode(struct amdgpu_device *adev) 1855 { 1856 const struct rlc_firmware_header_v2_2 *hdr; 1857 const __le32 *fw_data; 1858 unsigned i, fw_size; 1859 u32 tmp; 1860 1861 hdr = (const struct rlc_firmware_header_v2_2 *)adev->gfx.rlc_fw->data; 1862 1863 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data + 1864 le32_to_cpu(hdr->rlc_iram_ucode_offset_bytes)); 1865 fw_size = le32_to_cpu(hdr->rlc_iram_ucode_size_bytes) / 4; 1866 1867 WREG32_SOC15(GC, 0, regRLC_LX6_IRAM_ADDR, 0); 1868 1869 for (i = 0; i < fw_size; i++) { 1870 if ((amdgpu_emu_mode == 1) && (i % 100 == 99)) 1871 msleep(1); 1872 WREG32_SOC15(GC, 0, regRLC_LX6_IRAM_DATA, 1873 le32_to_cpup(fw_data++)); 1874 } 1875 1876 WREG32_SOC15(GC, 0, regRLC_LX6_IRAM_ADDR, adev->gfx.rlc_fw_version); 1877 1878 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data + 1879 le32_to_cpu(hdr->rlc_dram_ucode_offset_bytes)); 1880 fw_size = le32_to_cpu(hdr->rlc_dram_ucode_size_bytes) / 4; 1881 1882 WREG32_SOC15(GC, 0, regRLC_LX6_DRAM_ADDR, 0); 1883 for (i = 0; i < fw_size; i++) { 1884 if ((amdgpu_emu_mode == 1) && (i % 100 == 99)) 1885 msleep(1); 1886 WREG32_SOC15(GC, 0, regRLC_LX6_DRAM_DATA, 1887 le32_to_cpup(fw_data++)); 1888 } 1889 1890 WREG32_SOC15(GC, 0, regRLC_LX6_IRAM_ADDR, adev->gfx.rlc_fw_version); 1891 1892 tmp = RREG32_SOC15(GC, 0, regRLC_LX6_CNTL); 1893 tmp = REG_SET_FIELD(tmp, RLC_LX6_CNTL, PDEBUG_ENABLE, 1); 1894 tmp = REG_SET_FIELD(tmp, RLC_LX6_CNTL, BRESET, 0); 1895 WREG32_SOC15(GC, 0, regRLC_LX6_CNTL, tmp); 1896 } 1897 1898 static int gfx_v12_0_rlc_load_microcode(struct amdgpu_device *adev) 1899 { 1900 const struct rlc_firmware_header_v2_0 *hdr; 1901 uint16_t version_major; 1902 uint16_t version_minor; 1903 1904 if (!adev->gfx.rlc_fw) 1905 return -EINVAL; 1906 1907 hdr = (const struct rlc_firmware_header_v2_0 *)adev->gfx.rlc_fw->data; 1908 amdgpu_ucode_print_rlc_hdr(&hdr->header); 1909 1910 version_major = le16_to_cpu(hdr->header.header_version_major); 1911 version_minor = le16_to_cpu(hdr->header.header_version_minor); 1912 1913 if (version_major == 2) { 1914 gfx_v12_0_load_rlcg_microcode(adev); 1915 if (amdgpu_dpm == 1) { 1916 if (version_minor >= 2) 1917 gfx_v12_0_load_rlc_iram_dram_microcode(adev); 1918 } 1919 1920 return 0; 1921 } 1922 1923 return -EINVAL; 1924 } 1925 1926 static int gfx_v12_0_rlc_resume(struct amdgpu_device *adev) 1927 { 1928 int r; 1929 1930 if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) { 1931 gfx_v12_0_init_csb(adev); 1932 1933 if (!amdgpu_sriov_vf(adev)) /* enable RLC SRM */ 1934 gfx_v12_0_rlc_enable_srm(adev); 1935 } else { 1936 if (amdgpu_sriov_vf(adev)) { 1937 gfx_v12_0_init_csb(adev); 1938 return 0; 1939 } 1940 1941 adev->gfx.rlc.funcs->stop(adev); 1942 1943 /* disable CG */ 1944 WREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL, 0); 1945 1946 /* disable PG */ 1947 WREG32_SOC15(GC, 0, regRLC_PG_CNTL, 0); 1948 1949 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) { 1950 /* legacy rlc firmware loading */ 1951 r = gfx_v12_0_rlc_load_microcode(adev); 1952 if (r) 1953 return r; 1954 } 1955 1956 gfx_v12_0_init_csb(adev); 1957 1958 adev->gfx.rlc.funcs->start(adev); 1959 } 1960 1961 return 0; 1962 } 1963 1964 static void gfx_v12_0_config_gfx_rs64(struct amdgpu_device *adev) 1965 { 1966 const struct gfx_firmware_header_v2_0 *pfp_hdr; 1967 const struct gfx_firmware_header_v2_0 *me_hdr; 1968 const struct gfx_firmware_header_v2_0 *mec_hdr; 1969 uint32_t pipe_id, tmp; 1970 1971 mec_hdr = (const struct gfx_firmware_header_v2_0 *) 1972 adev->gfx.mec_fw->data; 1973 me_hdr = (const struct gfx_firmware_header_v2_0 *) 1974 adev->gfx.me_fw->data; 1975 pfp_hdr = (const struct gfx_firmware_header_v2_0 *) 1976 adev->gfx.pfp_fw->data; 1977 1978 /* config pfp program start addr */ 1979 for (pipe_id = 0; pipe_id < 2; pipe_id++) { 1980 soc24_grbm_select(adev, 0, pipe_id, 0, 0); 1981 WREG32_SOC15(GC, 0, regCP_PFP_PRGRM_CNTR_START, 1982 (pfp_hdr->ucode_start_addr_hi << 30) | 1983 (pfp_hdr->ucode_start_addr_lo >> 2)); 1984 WREG32_SOC15(GC, 0, regCP_PFP_PRGRM_CNTR_START_HI, 1985 pfp_hdr->ucode_start_addr_hi >> 2); 1986 } 1987 soc24_grbm_select(adev, 0, 0, 0, 0); 1988 1989 /* reset pfp pipe */ 1990 tmp = RREG32_SOC15(GC, 0, regCP_ME_CNTL); 1991 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, PFP_PIPE0_RESET, 1); 1992 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, PFP_PIPE1_RESET, 1); 1993 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 1994 1995 /* clear pfp pipe reset */ 1996 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, PFP_PIPE0_RESET, 0); 1997 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, PFP_PIPE1_RESET, 0); 1998 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 1999 2000 /* config me program start addr */ 2001 for (pipe_id = 0; pipe_id < 2; pipe_id++) { 2002 soc24_grbm_select(adev, 0, pipe_id, 0, 0); 2003 WREG32_SOC15(GC, 0, regCP_ME_PRGRM_CNTR_START, 2004 (me_hdr->ucode_start_addr_hi << 30) | 2005 (me_hdr->ucode_start_addr_lo >> 2)); 2006 WREG32_SOC15(GC, 0, regCP_ME_PRGRM_CNTR_START_HI, 2007 me_hdr->ucode_start_addr_hi>>2); 2008 } 2009 soc24_grbm_select(adev, 0, 0, 0, 0); 2010 2011 /* reset me pipe */ 2012 tmp = RREG32_SOC15(GC, 0, regCP_ME_CNTL); 2013 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, ME_PIPE0_RESET, 1); 2014 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, ME_PIPE1_RESET, 1); 2015 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 2016 2017 /* clear me pipe reset */ 2018 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, ME_PIPE0_RESET, 0); 2019 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, ME_PIPE1_RESET, 0); 2020 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 2021 2022 /* config mec program start addr */ 2023 for (pipe_id = 0; pipe_id < 4; pipe_id++) { 2024 soc24_grbm_select(adev, 1, pipe_id, 0, 0); 2025 WREG32_SOC15(GC, 0, regCP_MEC_RS64_PRGRM_CNTR_START, 2026 mec_hdr->ucode_start_addr_lo >> 2 | 2027 mec_hdr->ucode_start_addr_hi << 30); 2028 WREG32_SOC15(GC, 0, regCP_MEC_RS64_PRGRM_CNTR_START_HI, 2029 mec_hdr->ucode_start_addr_hi >> 2); 2030 } 2031 soc24_grbm_select(adev, 0, 0, 0, 0); 2032 2033 /* reset mec pipe */ 2034 tmp = RREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL); 2035 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE0_RESET, 1); 2036 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE1_RESET, 1); 2037 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE2_RESET, 1); 2038 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE3_RESET, 1); 2039 WREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL, tmp); 2040 2041 /* clear mec pipe reset */ 2042 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE0_RESET, 0); 2043 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE1_RESET, 0); 2044 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE2_RESET, 0); 2045 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE3_RESET, 0); 2046 WREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL, tmp); 2047 } 2048 2049 static void gfx_v12_0_set_pfp_ucode_start_addr(struct amdgpu_device *adev) 2050 { 2051 const struct gfx_firmware_header_v2_0 *cp_hdr; 2052 unsigned pipe_id, tmp; 2053 2054 cp_hdr = (const struct gfx_firmware_header_v2_0 *) 2055 adev->gfx.pfp_fw->data; 2056 mutex_lock(&adev->srbm_mutex); 2057 for (pipe_id = 0; pipe_id < adev->gfx.me.num_pipe_per_me; pipe_id++) { 2058 soc24_grbm_select(adev, 0, pipe_id, 0, 0); 2059 WREG32_SOC15(GC, 0, regCP_PFP_PRGRM_CNTR_START, 2060 (cp_hdr->ucode_start_addr_hi << 30) | 2061 (cp_hdr->ucode_start_addr_lo >> 2)); 2062 WREG32_SOC15(GC, 0, regCP_PFP_PRGRM_CNTR_START_HI, 2063 cp_hdr->ucode_start_addr_hi>>2); 2064 2065 /* 2066 * Program CP_ME_CNTL to reset given PIPE to take 2067 * effect of CP_PFP_PRGRM_CNTR_START. 2068 */ 2069 tmp = RREG32_SOC15(GC, 0, regCP_ME_CNTL); 2070 if (pipe_id == 0) 2071 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 2072 PFP_PIPE0_RESET, 1); 2073 else 2074 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 2075 PFP_PIPE1_RESET, 1); 2076 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 2077 2078 /* Clear pfp pipe0 reset bit. */ 2079 if (pipe_id == 0) 2080 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 2081 PFP_PIPE0_RESET, 0); 2082 else 2083 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 2084 PFP_PIPE1_RESET, 0); 2085 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 2086 } 2087 soc24_grbm_select(adev, 0, 0, 0, 0); 2088 mutex_unlock(&adev->srbm_mutex); 2089 } 2090 2091 static void gfx_v12_0_set_me_ucode_start_addr(struct amdgpu_device *adev) 2092 { 2093 const struct gfx_firmware_header_v2_0 *cp_hdr; 2094 unsigned pipe_id, tmp; 2095 2096 cp_hdr = (const struct gfx_firmware_header_v2_0 *) 2097 adev->gfx.me_fw->data; 2098 mutex_lock(&adev->srbm_mutex); 2099 for (pipe_id = 0; pipe_id < adev->gfx.me.num_pipe_per_me; pipe_id++) { 2100 soc24_grbm_select(adev, 0, pipe_id, 0, 0); 2101 WREG32_SOC15(GC, 0, regCP_ME_PRGRM_CNTR_START, 2102 (cp_hdr->ucode_start_addr_hi << 30) | 2103 (cp_hdr->ucode_start_addr_lo >> 2) ); 2104 WREG32_SOC15(GC, 0, regCP_ME_PRGRM_CNTR_START_HI, 2105 cp_hdr->ucode_start_addr_hi>>2); 2106 2107 /* 2108 * Program CP_ME_CNTL to reset given PIPE to take 2109 * effect of CP_ME_PRGRM_CNTR_START. 2110 */ 2111 tmp = RREG32_SOC15(GC, 0, regCP_ME_CNTL); 2112 if (pipe_id == 0) 2113 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 2114 ME_PIPE0_RESET, 1); 2115 else 2116 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 2117 ME_PIPE1_RESET, 1); 2118 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 2119 2120 /* Clear pfp pipe0 reset bit. */ 2121 if (pipe_id == 0) 2122 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 2123 ME_PIPE0_RESET, 0); 2124 else 2125 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 2126 ME_PIPE1_RESET, 0); 2127 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 2128 } 2129 soc24_grbm_select(adev, 0, 0, 0, 0); 2130 mutex_unlock(&adev->srbm_mutex); 2131 } 2132 2133 static void gfx_v12_0_set_mec_ucode_start_addr(struct amdgpu_device *adev) 2134 { 2135 const struct gfx_firmware_header_v2_0 *cp_hdr; 2136 unsigned pipe_id; 2137 2138 cp_hdr = (const struct gfx_firmware_header_v2_0 *) 2139 adev->gfx.mec_fw->data; 2140 mutex_lock(&adev->srbm_mutex); 2141 for (pipe_id = 0; pipe_id < adev->gfx.mec.num_pipe_per_mec; pipe_id++) { 2142 soc24_grbm_select(adev, 1, pipe_id, 0, 0); 2143 WREG32_SOC15(GC, 0, regCP_MEC_RS64_PRGRM_CNTR_START, 2144 cp_hdr->ucode_start_addr_lo >> 2 | 2145 cp_hdr->ucode_start_addr_hi << 30); 2146 WREG32_SOC15(GC, 0, regCP_MEC_RS64_PRGRM_CNTR_START_HI, 2147 cp_hdr->ucode_start_addr_hi >> 2); 2148 } 2149 soc24_grbm_select(adev, 0, 0, 0, 0); 2150 mutex_unlock(&adev->srbm_mutex); 2151 } 2152 2153 static int gfx_v12_0_wait_for_rlc_autoload_complete(struct amdgpu_device *adev) 2154 { 2155 uint32_t cp_status; 2156 uint32_t bootload_status; 2157 int i; 2158 2159 for (i = 0; i < adev->usec_timeout; i++) { 2160 cp_status = RREG32_SOC15(GC, 0, regCP_STAT); 2161 bootload_status = RREG32_SOC15(GC, 0, regRLC_RLCS_BOOTLOAD_STATUS); 2162 2163 if ((cp_status == 0) && 2164 (REG_GET_FIELD(bootload_status, 2165 RLC_RLCS_BOOTLOAD_STATUS, BOOTLOAD_COMPLETE) == 1)) { 2166 break; 2167 } 2168 udelay(1); 2169 if (amdgpu_emu_mode) 2170 msleep(10); 2171 } 2172 2173 if (i >= adev->usec_timeout) { 2174 dev_err(adev->dev, "rlc autoload: gc ucode autoload timeout\n"); 2175 return -ETIMEDOUT; 2176 } 2177 2178 if (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) { 2179 gfx_v12_0_set_pfp_ucode_start_addr(adev); 2180 gfx_v12_0_set_me_ucode_start_addr(adev); 2181 gfx_v12_0_set_mec_ucode_start_addr(adev); 2182 } 2183 2184 return 0; 2185 } 2186 2187 static int gfx_v12_0_cp_gfx_enable(struct amdgpu_device *adev, bool enable) 2188 { 2189 int i; 2190 u32 tmp = RREG32_SOC15(GC, 0, regCP_ME_CNTL); 2191 2192 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, ME_HALT, enable ? 0 : 1); 2193 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, PFP_HALT, enable ? 0 : 1); 2194 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 2195 2196 for (i = 0; i < adev->usec_timeout; i++) { 2197 if (RREG32_SOC15(GC, 0, regCP_STAT) == 0) 2198 break; 2199 udelay(1); 2200 } 2201 2202 if (i >= adev->usec_timeout) 2203 DRM_ERROR("failed to %s cp gfx\n", enable ? "unhalt" : "halt"); 2204 2205 return 0; 2206 } 2207 2208 static int gfx_v12_0_cp_gfx_load_pfp_microcode_rs64(struct amdgpu_device *adev) 2209 { 2210 int r; 2211 const struct gfx_firmware_header_v2_0 *pfp_hdr; 2212 const __le32 *fw_ucode, *fw_data; 2213 unsigned i, pipe_id, fw_ucode_size, fw_data_size; 2214 uint32_t tmp; 2215 uint32_t usec_timeout = 50000; /* wait for 50ms */ 2216 2217 pfp_hdr = (const struct gfx_firmware_header_v2_0 *) 2218 adev->gfx.pfp_fw->data; 2219 2220 amdgpu_ucode_print_gfx_hdr(&pfp_hdr->header); 2221 2222 /* instruction */ 2223 fw_ucode = (const __le32 *)(adev->gfx.pfp_fw->data + 2224 le32_to_cpu(pfp_hdr->ucode_offset_bytes)); 2225 fw_ucode_size = le32_to_cpu(pfp_hdr->ucode_size_bytes); 2226 /* data */ 2227 fw_data = (const __le32 *)(adev->gfx.pfp_fw->data + 2228 le32_to_cpu(pfp_hdr->data_offset_bytes)); 2229 fw_data_size = le32_to_cpu(pfp_hdr->data_size_bytes); 2230 2231 /* 64kb align */ 2232 r = amdgpu_bo_create_reserved(adev, fw_ucode_size, 2233 64 * 1024, AMDGPU_GEM_DOMAIN_VRAM, 2234 &adev->gfx.pfp.pfp_fw_obj, 2235 &adev->gfx.pfp.pfp_fw_gpu_addr, 2236 (void **)&adev->gfx.pfp.pfp_fw_ptr); 2237 if (r) { 2238 dev_err(adev->dev, "(%d) failed to create pfp ucode fw bo\n", r); 2239 gfx_v12_0_pfp_fini(adev); 2240 return r; 2241 } 2242 2243 r = amdgpu_bo_create_reserved(adev, fw_data_size, 2244 64 * 1024, AMDGPU_GEM_DOMAIN_VRAM, 2245 &adev->gfx.pfp.pfp_fw_data_obj, 2246 &adev->gfx.pfp.pfp_fw_data_gpu_addr, 2247 (void **)&adev->gfx.pfp.pfp_fw_data_ptr); 2248 if (r) { 2249 dev_err(adev->dev, "(%d) failed to create pfp data fw bo\n", r); 2250 gfx_v12_0_pfp_fini(adev); 2251 return r; 2252 } 2253 2254 memcpy(adev->gfx.pfp.pfp_fw_ptr, fw_ucode, fw_ucode_size); 2255 memcpy(adev->gfx.pfp.pfp_fw_data_ptr, fw_data, fw_data_size); 2256 2257 amdgpu_bo_kunmap(adev->gfx.pfp.pfp_fw_obj); 2258 amdgpu_bo_kunmap(adev->gfx.pfp.pfp_fw_data_obj); 2259 amdgpu_bo_unreserve(adev->gfx.pfp.pfp_fw_obj); 2260 amdgpu_bo_unreserve(adev->gfx.pfp.pfp_fw_data_obj); 2261 2262 if (amdgpu_emu_mode == 1) 2263 adev->hdp.funcs->flush_hdp(adev, NULL); 2264 2265 WREG32_SOC15(GC, 0, regCP_PFP_IC_BASE_LO, 2266 lower_32_bits(adev->gfx.pfp.pfp_fw_gpu_addr)); 2267 WREG32_SOC15(GC, 0, regCP_PFP_IC_BASE_HI, 2268 upper_32_bits(adev->gfx.pfp.pfp_fw_gpu_addr)); 2269 2270 tmp = RREG32_SOC15(GC, 0, regCP_PFP_IC_BASE_CNTL); 2271 tmp = REG_SET_FIELD(tmp, CP_PFP_IC_BASE_CNTL, VMID, 0); 2272 tmp = REG_SET_FIELD(tmp, CP_PFP_IC_BASE_CNTL, CACHE_POLICY, 0); 2273 tmp = REG_SET_FIELD(tmp, CP_PFP_IC_BASE_CNTL, EXE_DISABLE, 0); 2274 WREG32_SOC15(GC, 0, regCP_PFP_IC_BASE_CNTL, tmp); 2275 2276 /* 2277 * Programming any of the CP_PFP_IC_BASE registers 2278 * forces invalidation of the ME L1 I$. Wait for the 2279 * invalidation complete 2280 */ 2281 for (i = 0; i < usec_timeout; i++) { 2282 tmp = RREG32_SOC15(GC, 0, regCP_PFP_IC_OP_CNTL); 2283 if (1 == REG_GET_FIELD(tmp, CP_PFP_IC_OP_CNTL, 2284 INVALIDATE_CACHE_COMPLETE)) 2285 break; 2286 udelay(1); 2287 } 2288 2289 if (i >= usec_timeout) { 2290 dev_err(adev->dev, "failed to invalidate instruction cache\n"); 2291 return -EINVAL; 2292 } 2293 2294 /* Prime the L1 instruction caches */ 2295 tmp = RREG32_SOC15(GC, 0, regCP_PFP_IC_OP_CNTL); 2296 tmp = REG_SET_FIELD(tmp, CP_PFP_IC_OP_CNTL, PRIME_ICACHE, 1); 2297 WREG32_SOC15(GC, 0, regCP_PFP_IC_OP_CNTL, tmp); 2298 /* Waiting for cache primed*/ 2299 for (i = 0; i < usec_timeout; i++) { 2300 tmp = RREG32_SOC15(GC, 0, regCP_PFP_IC_OP_CNTL); 2301 if (1 == REG_GET_FIELD(tmp, CP_PFP_IC_OP_CNTL, 2302 ICACHE_PRIMED)) 2303 break; 2304 udelay(1); 2305 } 2306 2307 if (i >= usec_timeout) { 2308 dev_err(adev->dev, "failed to prime instruction cache\n"); 2309 return -EINVAL; 2310 } 2311 2312 mutex_lock(&adev->srbm_mutex); 2313 for (pipe_id = 0; pipe_id < adev->gfx.me.num_pipe_per_me; pipe_id++) { 2314 soc24_grbm_select(adev, 0, pipe_id, 0, 0); 2315 2316 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE0_LO, 2317 lower_32_bits(adev->gfx.pfp.pfp_fw_data_gpu_addr)); 2318 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE0_HI, 2319 upper_32_bits(adev->gfx.pfp.pfp_fw_data_gpu_addr)); 2320 } 2321 soc24_grbm_select(adev, 0, 0, 0, 0); 2322 mutex_unlock(&adev->srbm_mutex); 2323 2324 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE_CNTL); 2325 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_BASE_CNTL, VMID, 0); 2326 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_BASE_CNTL, CACHE_POLICY, 0); 2327 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE_CNTL, tmp); 2328 2329 /* Invalidate the data caches */ 2330 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL); 2331 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_OP_CNTL, INVALIDATE_DCACHE, 1); 2332 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL, tmp); 2333 2334 for (i = 0; i < usec_timeout; i++) { 2335 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL); 2336 if (1 == REG_GET_FIELD(tmp, CP_GFX_RS64_DC_OP_CNTL, 2337 INVALIDATE_DCACHE_COMPLETE)) 2338 break; 2339 udelay(1); 2340 } 2341 2342 if (i >= usec_timeout) { 2343 dev_err(adev->dev, "failed to invalidate RS64 data cache\n"); 2344 return -EINVAL; 2345 } 2346 2347 gfx_v12_0_set_pfp_ucode_start_addr(adev); 2348 2349 return 0; 2350 } 2351 2352 static int gfx_v12_0_cp_gfx_load_me_microcode_rs64(struct amdgpu_device *adev) 2353 { 2354 int r; 2355 const struct gfx_firmware_header_v2_0 *me_hdr; 2356 const __le32 *fw_ucode, *fw_data; 2357 unsigned i, pipe_id, fw_ucode_size, fw_data_size; 2358 uint32_t tmp; 2359 uint32_t usec_timeout = 50000; /* wait for 50ms */ 2360 2361 me_hdr = (const struct gfx_firmware_header_v2_0 *) 2362 adev->gfx.me_fw->data; 2363 2364 amdgpu_ucode_print_gfx_hdr(&me_hdr->header); 2365 2366 /* instruction */ 2367 fw_ucode = (const __le32 *)(adev->gfx.me_fw->data + 2368 le32_to_cpu(me_hdr->ucode_offset_bytes)); 2369 fw_ucode_size = le32_to_cpu(me_hdr->ucode_size_bytes); 2370 /* data */ 2371 fw_data = (const __le32 *)(adev->gfx.me_fw->data + 2372 le32_to_cpu(me_hdr->data_offset_bytes)); 2373 fw_data_size = le32_to_cpu(me_hdr->data_size_bytes); 2374 2375 /* 64kb align*/ 2376 r = amdgpu_bo_create_reserved(adev, fw_ucode_size, 2377 64 * 1024, AMDGPU_GEM_DOMAIN_VRAM, 2378 &adev->gfx.me.me_fw_obj, 2379 &adev->gfx.me.me_fw_gpu_addr, 2380 (void **)&adev->gfx.me.me_fw_ptr); 2381 if (r) { 2382 dev_err(adev->dev, "(%d) failed to create me ucode bo\n", r); 2383 gfx_v12_0_me_fini(adev); 2384 return r; 2385 } 2386 2387 r = amdgpu_bo_create_reserved(adev, fw_data_size, 2388 64 * 1024, AMDGPU_GEM_DOMAIN_VRAM, 2389 &adev->gfx.me.me_fw_data_obj, 2390 &adev->gfx.me.me_fw_data_gpu_addr, 2391 (void **)&adev->gfx.me.me_fw_data_ptr); 2392 if (r) { 2393 dev_err(adev->dev, "(%d) failed to create me data bo\n", r); 2394 gfx_v12_0_pfp_fini(adev); 2395 return r; 2396 } 2397 2398 memcpy(adev->gfx.me.me_fw_ptr, fw_ucode, fw_ucode_size); 2399 memcpy(adev->gfx.me.me_fw_data_ptr, fw_data, fw_data_size); 2400 2401 amdgpu_bo_kunmap(adev->gfx.me.me_fw_obj); 2402 amdgpu_bo_kunmap(adev->gfx.me.me_fw_data_obj); 2403 amdgpu_bo_unreserve(adev->gfx.me.me_fw_obj); 2404 amdgpu_bo_unreserve(adev->gfx.me.me_fw_data_obj); 2405 2406 if (amdgpu_emu_mode == 1) 2407 adev->hdp.funcs->flush_hdp(adev, NULL); 2408 2409 WREG32_SOC15(GC, 0, regCP_ME_IC_BASE_LO, 2410 lower_32_bits(adev->gfx.me.me_fw_gpu_addr)); 2411 WREG32_SOC15(GC, 0, regCP_ME_IC_BASE_HI, 2412 upper_32_bits(adev->gfx.me.me_fw_gpu_addr)); 2413 2414 tmp = RREG32_SOC15(GC, 0, regCP_ME_IC_BASE_CNTL); 2415 tmp = REG_SET_FIELD(tmp, CP_ME_IC_BASE_CNTL, VMID, 0); 2416 tmp = REG_SET_FIELD(tmp, CP_ME_IC_BASE_CNTL, CACHE_POLICY, 0); 2417 tmp = REG_SET_FIELD(tmp, CP_ME_IC_BASE_CNTL, EXE_DISABLE, 0); 2418 WREG32_SOC15(GC, 0, regCP_ME_IC_BASE_CNTL, tmp); 2419 2420 /* 2421 * Programming any of the CP_ME_IC_BASE registers 2422 * forces invalidation of the ME L1 I$. Wait for the 2423 * invalidation complete 2424 */ 2425 for (i = 0; i < usec_timeout; i++) { 2426 tmp = RREG32_SOC15(GC, 0, regCP_ME_IC_OP_CNTL); 2427 if (1 == REG_GET_FIELD(tmp, CP_ME_IC_OP_CNTL, 2428 INVALIDATE_CACHE_COMPLETE)) 2429 break; 2430 udelay(1); 2431 } 2432 2433 if (i >= usec_timeout) { 2434 dev_err(adev->dev, "failed to invalidate instruction cache\n"); 2435 return -EINVAL; 2436 } 2437 2438 /* Prime the instruction caches */ 2439 tmp = RREG32_SOC15(GC, 0, regCP_ME_IC_OP_CNTL); 2440 tmp = REG_SET_FIELD(tmp, CP_ME_IC_OP_CNTL, PRIME_ICACHE, 1); 2441 WREG32_SOC15(GC, 0, regCP_ME_IC_OP_CNTL, tmp); 2442 2443 /* Waiting for instruction cache primed*/ 2444 for (i = 0; i < usec_timeout; i++) { 2445 tmp = RREG32_SOC15(GC, 0, regCP_ME_IC_OP_CNTL); 2446 if (1 == REG_GET_FIELD(tmp, CP_ME_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_BASE1_LO, 2462 lower_32_bits(adev->gfx.me.me_fw_data_gpu_addr)); 2463 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE1_HI, 2464 upper_32_bits(adev->gfx.me.me_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_me_ucode_start_addr(adev); 2493 2494 return 0; 2495 } 2496 2497 static int gfx_v12_0_cp_gfx_load_microcode(struct amdgpu_device *adev) 2498 { 2499 int r; 2500 2501 if (!adev->gfx.me_fw || !adev->gfx.pfp_fw) 2502 return -EINVAL; 2503 2504 gfx_v12_0_cp_gfx_enable(adev, false); 2505 2506 r = gfx_v12_0_cp_gfx_load_pfp_microcode_rs64(adev); 2507 if (r) { 2508 dev_err(adev->dev, "(%d) failed to load pfp fw\n", r); 2509 return r; 2510 } 2511 2512 r = gfx_v12_0_cp_gfx_load_me_microcode_rs64(adev); 2513 if (r) { 2514 dev_err(adev->dev, "(%d) failed to load me fw\n", r); 2515 return r; 2516 } 2517 2518 return 0; 2519 } 2520 2521 static int gfx_v12_0_cp_gfx_start(struct amdgpu_device *adev) 2522 { 2523 /* init the CP */ 2524 WREG32_SOC15(GC, 0, regCP_MAX_CONTEXT, 2525 adev->gfx.config.max_hw_contexts - 1); 2526 WREG32_SOC15(GC, 0, regCP_DEVICE_ID, 1); 2527 2528 if (!amdgpu_async_gfx_ring) 2529 gfx_v12_0_cp_gfx_enable(adev, true); 2530 2531 return 0; 2532 } 2533 2534 static void gfx_v12_0_cp_gfx_switch_pipe(struct amdgpu_device *adev, 2535 CP_PIPE_ID pipe) 2536 { 2537 u32 tmp; 2538 2539 tmp = RREG32_SOC15(GC, 0, regGRBM_GFX_CNTL); 2540 tmp = REG_SET_FIELD(tmp, GRBM_GFX_CNTL, PIPEID, pipe); 2541 2542 WREG32_SOC15(GC, 0, regGRBM_GFX_CNTL, tmp); 2543 } 2544 2545 static void gfx_v12_0_cp_gfx_set_doorbell(struct amdgpu_device *adev, 2546 struct amdgpu_ring *ring) 2547 { 2548 u32 tmp; 2549 2550 tmp = RREG32_SOC15(GC, 0, regCP_RB_DOORBELL_CONTROL); 2551 if (ring->use_doorbell) { 2552 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL, 2553 DOORBELL_OFFSET, ring->doorbell_index); 2554 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL, 2555 DOORBELL_EN, 1); 2556 } else { 2557 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL, 2558 DOORBELL_EN, 0); 2559 } 2560 WREG32_SOC15(GC, 0, regCP_RB_DOORBELL_CONTROL, tmp); 2561 2562 tmp = REG_SET_FIELD(0, CP_RB_DOORBELL_RANGE_LOWER, 2563 DOORBELL_RANGE_LOWER, ring->doorbell_index); 2564 WREG32_SOC15(GC, 0, regCP_RB_DOORBELL_RANGE_LOWER, tmp); 2565 2566 WREG32_SOC15(GC, 0, regCP_RB_DOORBELL_RANGE_UPPER, 2567 CP_RB_DOORBELL_RANGE_UPPER__DOORBELL_RANGE_UPPER_MASK); 2568 } 2569 2570 static int gfx_v12_0_cp_gfx_resume(struct amdgpu_device *adev) 2571 { 2572 struct amdgpu_ring *ring; 2573 u32 tmp; 2574 u32 rb_bufsz; 2575 u64 rb_addr, rptr_addr, wptr_gpu_addr; 2576 u32 i; 2577 2578 /* Set the write pointer delay */ 2579 WREG32_SOC15(GC, 0, regCP_RB_WPTR_DELAY, 0); 2580 2581 /* set the RB to use vmid 0 */ 2582 WREG32_SOC15(GC, 0, regCP_RB_VMID, 0); 2583 2584 /* Init gfx ring 0 for pipe 0 */ 2585 mutex_lock(&adev->srbm_mutex); 2586 gfx_v12_0_cp_gfx_switch_pipe(adev, PIPE_ID0); 2587 2588 /* Set ring buffer size */ 2589 ring = &adev->gfx.gfx_ring[0]; 2590 rb_bufsz = order_base_2(ring->ring_size / 8); 2591 tmp = REG_SET_FIELD(0, CP_RB0_CNTL, RB_BUFSZ, rb_bufsz); 2592 tmp = REG_SET_FIELD(tmp, CP_RB0_CNTL, RB_BLKSZ, rb_bufsz - 2); 2593 WREG32_SOC15(GC, 0, regCP_RB0_CNTL, tmp); 2594 2595 /* Initialize the ring buffer's write pointers */ 2596 ring->wptr = 0; 2597 WREG32_SOC15(GC, 0, regCP_RB0_WPTR, lower_32_bits(ring->wptr)); 2598 WREG32_SOC15(GC, 0, regCP_RB0_WPTR_HI, upper_32_bits(ring->wptr)); 2599 2600 /* set the wb address wether it's enabled or not */ 2601 rptr_addr = ring->rptr_gpu_addr; 2602 WREG32_SOC15(GC, 0, regCP_RB0_RPTR_ADDR, lower_32_bits(rptr_addr)); 2603 WREG32_SOC15(GC, 0, regCP_RB0_RPTR_ADDR_HI, upper_32_bits(rptr_addr) & 2604 CP_RB_RPTR_ADDR_HI__RB_RPTR_ADDR_HI_MASK); 2605 2606 wptr_gpu_addr = ring->wptr_gpu_addr; 2607 WREG32_SOC15(GC, 0, regCP_RB_WPTR_POLL_ADDR_LO, 2608 lower_32_bits(wptr_gpu_addr)); 2609 WREG32_SOC15(GC, 0, regCP_RB_WPTR_POLL_ADDR_HI, 2610 upper_32_bits(wptr_gpu_addr)); 2611 2612 mdelay(1); 2613 WREG32_SOC15(GC, 0, regCP_RB0_CNTL, tmp); 2614 2615 rb_addr = ring->gpu_addr >> 8; 2616 WREG32_SOC15(GC, 0, regCP_RB0_BASE, rb_addr); 2617 WREG32_SOC15(GC, 0, regCP_RB0_BASE_HI, upper_32_bits(rb_addr)); 2618 2619 WREG32_SOC15(GC, 0, regCP_RB_ACTIVE, 1); 2620 2621 gfx_v12_0_cp_gfx_set_doorbell(adev, ring); 2622 mutex_unlock(&adev->srbm_mutex); 2623 2624 /* Switch to pipe 0 */ 2625 mutex_lock(&adev->srbm_mutex); 2626 gfx_v12_0_cp_gfx_switch_pipe(adev, PIPE_ID0); 2627 mutex_unlock(&adev->srbm_mutex); 2628 2629 /* start the ring */ 2630 gfx_v12_0_cp_gfx_start(adev); 2631 2632 for (i = 0; i < adev->gfx.num_gfx_rings; i++) { 2633 ring = &adev->gfx.gfx_ring[i]; 2634 ring->sched.ready = true; 2635 } 2636 2637 return 0; 2638 } 2639 2640 static void gfx_v12_0_cp_compute_enable(struct amdgpu_device *adev, bool enable) 2641 { 2642 u32 data; 2643 2644 data = RREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL); 2645 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_INVALIDATE_ICACHE, 2646 enable ? 0 : 1); 2647 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE0_RESET, 2648 enable ? 0 : 1); 2649 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE1_RESET, 2650 enable ? 0 : 1); 2651 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE2_RESET, 2652 enable ? 0 : 1); 2653 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE3_RESET, 2654 enable ? 0 : 1); 2655 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE0_ACTIVE, 2656 enable ? 1 : 0); 2657 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE1_ACTIVE, 2658 enable ? 1 : 0); 2659 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE2_ACTIVE, 2660 enable ? 1 : 0); 2661 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE3_ACTIVE, 2662 enable ? 1 : 0); 2663 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_HALT, 2664 enable ? 0 : 1); 2665 WREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL, data); 2666 2667 adev->gfx.kiq[0].ring.sched.ready = enable; 2668 2669 udelay(50); 2670 } 2671 2672 static int gfx_v12_0_cp_compute_load_microcode_rs64(struct amdgpu_device *adev) 2673 { 2674 const struct gfx_firmware_header_v2_0 *mec_hdr; 2675 const __le32 *fw_ucode, *fw_data; 2676 u32 tmp, fw_ucode_size, fw_data_size; 2677 u32 i, usec_timeout = 50000; /* Wait for 50 ms */ 2678 u32 *fw_ucode_ptr, *fw_data_ptr; 2679 int r; 2680 2681 if (!adev->gfx.mec_fw) 2682 return -EINVAL; 2683 2684 gfx_v12_0_cp_compute_enable(adev, false); 2685 2686 mec_hdr = (const struct gfx_firmware_header_v2_0 *)adev->gfx.mec_fw->data; 2687 amdgpu_ucode_print_gfx_hdr(&mec_hdr->header); 2688 2689 fw_ucode = (const __le32 *) (adev->gfx.mec_fw->data + 2690 le32_to_cpu(mec_hdr->ucode_offset_bytes)); 2691 fw_ucode_size = le32_to_cpu(mec_hdr->ucode_size_bytes); 2692 2693 fw_data = (const __le32 *) (adev->gfx.mec_fw->data + 2694 le32_to_cpu(mec_hdr->data_offset_bytes)); 2695 fw_data_size = le32_to_cpu(mec_hdr->data_size_bytes); 2696 2697 r = amdgpu_bo_create_reserved(adev, fw_ucode_size, 2698 64 * 1024, AMDGPU_GEM_DOMAIN_VRAM, 2699 &adev->gfx.mec.mec_fw_obj, 2700 &adev->gfx.mec.mec_fw_gpu_addr, 2701 (void **)&fw_ucode_ptr); 2702 if (r) { 2703 dev_err(adev->dev, "(%d) failed to create mec fw ucode bo\n", r); 2704 gfx_v12_0_mec_fini(adev); 2705 return r; 2706 } 2707 2708 r = amdgpu_bo_create_reserved(adev, 2709 ALIGN(fw_data_size, 64 * 1024) * 2710 adev->gfx.mec.num_pipe_per_mec, 2711 64 * 1024, AMDGPU_GEM_DOMAIN_VRAM, 2712 &adev->gfx.mec.mec_fw_data_obj, 2713 &adev->gfx.mec.mec_fw_data_gpu_addr, 2714 (void **)&fw_data_ptr); 2715 if (r) { 2716 dev_err(adev->dev, "(%d) failed to create mec fw ucode bo\n", r); 2717 gfx_v12_0_mec_fini(adev); 2718 return r; 2719 } 2720 2721 memcpy(fw_ucode_ptr, fw_ucode, fw_ucode_size); 2722 for (i = 0; i < adev->gfx.mec.num_pipe_per_mec; i++) { 2723 memcpy(fw_data_ptr + i * ALIGN(fw_data_size, 64 * 1024) / 4, fw_data, fw_data_size); 2724 } 2725 2726 amdgpu_bo_kunmap(adev->gfx.mec.mec_fw_obj); 2727 amdgpu_bo_kunmap(adev->gfx.mec.mec_fw_data_obj); 2728 amdgpu_bo_unreserve(adev->gfx.mec.mec_fw_obj); 2729 amdgpu_bo_unreserve(adev->gfx.mec.mec_fw_data_obj); 2730 2731 tmp = RREG32_SOC15(GC, 0, regCP_CPC_IC_BASE_CNTL); 2732 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_BASE_CNTL, VMID, 0); 2733 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_BASE_CNTL, EXE_DISABLE, 0); 2734 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_BASE_CNTL, CACHE_POLICY, 0); 2735 WREG32_SOC15(GC, 0, regCP_CPC_IC_BASE_CNTL, tmp); 2736 2737 tmp = RREG32_SOC15(GC, 0, regCP_MEC_DC_BASE_CNTL); 2738 tmp = REG_SET_FIELD(tmp, CP_MEC_DC_BASE_CNTL, VMID, 0); 2739 tmp = REG_SET_FIELD(tmp, CP_MEC_DC_BASE_CNTL, CACHE_POLICY, 0); 2740 WREG32_SOC15(GC, 0, regCP_MEC_DC_BASE_CNTL, tmp); 2741 2742 mutex_lock(&adev->srbm_mutex); 2743 for (i = 0; i < adev->gfx.mec.num_pipe_per_mec; i++) { 2744 soc24_grbm_select(adev, 1, i, 0, 0); 2745 2746 WREG32_SOC15(GC, 0, regCP_MEC_MDBASE_LO, 2747 lower_32_bits(adev->gfx.mec.mec_fw_data_gpu_addr + 2748 i * ALIGN(fw_data_size, 64 * 1024))); 2749 WREG32_SOC15(GC, 0, regCP_MEC_MDBASE_HI, 2750 upper_32_bits(adev->gfx.mec.mec_fw_data_gpu_addr + 2751 i * ALIGN(fw_data_size, 64 * 1024))); 2752 2753 WREG32_SOC15(GC, 0, regCP_CPC_IC_BASE_LO, 2754 lower_32_bits(adev->gfx.mec.mec_fw_gpu_addr)); 2755 WREG32_SOC15(GC, 0, regCP_CPC_IC_BASE_HI, 2756 upper_32_bits(adev->gfx.mec.mec_fw_gpu_addr)); 2757 } 2758 mutex_unlock(&adev->srbm_mutex); 2759 soc24_grbm_select(adev, 0, 0, 0, 0); 2760 2761 /* Trigger an invalidation of the L1 instruction caches */ 2762 tmp = RREG32_SOC15(GC, 0, regCP_MEC_DC_OP_CNTL); 2763 tmp = REG_SET_FIELD(tmp, CP_MEC_DC_OP_CNTL, INVALIDATE_DCACHE, 1); 2764 WREG32_SOC15(GC, 0, regCP_MEC_DC_OP_CNTL, tmp); 2765 2766 /* Wait for invalidation complete */ 2767 for (i = 0; i < usec_timeout; i++) { 2768 tmp = RREG32_SOC15(GC, 0, regCP_MEC_DC_OP_CNTL); 2769 if (1 == REG_GET_FIELD(tmp, CP_MEC_DC_OP_CNTL, 2770 INVALIDATE_DCACHE_COMPLETE)) 2771 break; 2772 udelay(1); 2773 } 2774 2775 if (i >= usec_timeout) { 2776 dev_err(adev->dev, "failed to invalidate instruction cache\n"); 2777 return -EINVAL; 2778 } 2779 2780 /* Trigger an invalidation of the L1 instruction caches */ 2781 tmp = RREG32_SOC15(GC, 0, regCP_CPC_IC_OP_CNTL); 2782 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_OP_CNTL, INVALIDATE_CACHE, 1); 2783 WREG32_SOC15(GC, 0, regCP_CPC_IC_OP_CNTL, tmp); 2784 2785 /* Wait for invalidation complete */ 2786 for (i = 0; i < usec_timeout; i++) { 2787 tmp = RREG32_SOC15(GC, 0, regCP_CPC_IC_OP_CNTL); 2788 if (1 == REG_GET_FIELD(tmp, CP_CPC_IC_OP_CNTL, 2789 INVALIDATE_CACHE_COMPLETE)) 2790 break; 2791 udelay(1); 2792 } 2793 2794 if (i >= usec_timeout) { 2795 dev_err(adev->dev, "failed to invalidate instruction cache\n"); 2796 return -EINVAL; 2797 } 2798 2799 gfx_v12_0_set_mec_ucode_start_addr(adev); 2800 2801 return 0; 2802 } 2803 2804 static void gfx_v12_0_kiq_setting(struct amdgpu_ring *ring) 2805 { 2806 uint32_t tmp; 2807 struct amdgpu_device *adev = ring->adev; 2808 2809 /* tell RLC which is KIQ queue */ 2810 tmp = RREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS); 2811 tmp &= 0xffffff00; 2812 tmp |= (ring->me << 5) | (ring->pipe << 3) | (ring->queue); 2813 WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp); 2814 tmp |= 0x80; 2815 WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp); 2816 } 2817 2818 static void gfx_v12_0_cp_set_doorbell_range(struct amdgpu_device *adev) 2819 { 2820 /* set graphics engine doorbell range */ 2821 WREG32_SOC15(GC, 0, regCP_RB_DOORBELL_RANGE_LOWER, 2822 (adev->doorbell_index.gfx_ring0 * 2) << 2); 2823 WREG32_SOC15(GC, 0, regCP_RB_DOORBELL_RANGE_UPPER, 2824 (adev->doorbell_index.gfx_userqueue_end * 2) << 2); 2825 2826 /* set compute engine doorbell range */ 2827 WREG32_SOC15(GC, 0, regCP_MEC_DOORBELL_RANGE_LOWER, 2828 (adev->doorbell_index.kiq * 2) << 2); 2829 WREG32_SOC15(GC, 0, regCP_MEC_DOORBELL_RANGE_UPPER, 2830 (adev->doorbell_index.userqueue_end * 2) << 2); 2831 } 2832 2833 static int gfx_v12_0_gfx_mqd_init(struct amdgpu_device *adev, void *m, 2834 struct amdgpu_mqd_prop *prop) 2835 { 2836 struct v12_gfx_mqd *mqd = m; 2837 uint64_t hqd_gpu_addr, wb_gpu_addr; 2838 uint32_t tmp; 2839 uint32_t rb_bufsz; 2840 2841 /* set up gfx hqd wptr */ 2842 mqd->cp_gfx_hqd_wptr = 0; 2843 mqd->cp_gfx_hqd_wptr_hi = 0; 2844 2845 /* set the pointer to the MQD */ 2846 mqd->cp_mqd_base_addr = prop->mqd_gpu_addr & 0xfffffffc; 2847 mqd->cp_mqd_base_addr_hi = upper_32_bits(prop->mqd_gpu_addr); 2848 2849 /* set up mqd control */ 2850 tmp = RREG32_SOC15(GC, 0, regCP_GFX_MQD_CONTROL); 2851 tmp = REG_SET_FIELD(tmp, CP_GFX_MQD_CONTROL, VMID, 0); 2852 tmp = REG_SET_FIELD(tmp, CP_GFX_MQD_CONTROL, PRIV_STATE, 1); 2853 tmp = REG_SET_FIELD(tmp, CP_GFX_MQD_CONTROL, CACHE_POLICY, 0); 2854 mqd->cp_gfx_mqd_control = tmp; 2855 2856 /* set up gfx_hqd_vimd with 0x0 to indicate the ring buffer's vmid */ 2857 tmp = RREG32_SOC15(GC, 0, regCP_GFX_HQD_VMID); 2858 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_VMID, VMID, 0); 2859 mqd->cp_gfx_hqd_vmid = 0; 2860 2861 /* set up default queue priority level 2862 * 0x0 = low priority, 0x1 = high priority */ 2863 tmp = RREG32_SOC15(GC, 0, regCP_GFX_HQD_QUEUE_PRIORITY); 2864 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_QUEUE_PRIORITY, PRIORITY_LEVEL, 0); 2865 mqd->cp_gfx_hqd_queue_priority = tmp; 2866 2867 /* set up time quantum */ 2868 tmp = RREG32_SOC15(GC, 0, regCP_GFX_HQD_QUANTUM); 2869 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_QUANTUM, QUANTUM_EN, 1); 2870 mqd->cp_gfx_hqd_quantum = tmp; 2871 2872 /* set up gfx hqd base. this is similar as CP_RB_BASE */ 2873 hqd_gpu_addr = prop->hqd_base_gpu_addr >> 8; 2874 mqd->cp_gfx_hqd_base = hqd_gpu_addr; 2875 mqd->cp_gfx_hqd_base_hi = upper_32_bits(hqd_gpu_addr); 2876 2877 /* set up hqd_rptr_addr/_hi, similar as CP_RB_RPTR */ 2878 wb_gpu_addr = prop->rptr_gpu_addr; 2879 mqd->cp_gfx_hqd_rptr_addr = wb_gpu_addr & 0xfffffffc; 2880 mqd->cp_gfx_hqd_rptr_addr_hi = 2881 upper_32_bits(wb_gpu_addr) & 0xffff; 2882 2883 /* set up rb_wptr_poll addr */ 2884 wb_gpu_addr = prop->wptr_gpu_addr; 2885 mqd->cp_rb_wptr_poll_addr_lo = wb_gpu_addr & 0xfffffffc; 2886 mqd->cp_rb_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr) & 0xffff; 2887 2888 /* set up the gfx_hqd_control, similar as CP_RB0_CNTL */ 2889 rb_bufsz = order_base_2(prop->queue_size / 4) - 1; 2890 tmp = RREG32_SOC15(GC, 0, regCP_GFX_HQD_CNTL); 2891 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_CNTL, RB_BUFSZ, rb_bufsz); 2892 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_CNTL, RB_BLKSZ, rb_bufsz - 2); 2893 #ifdef __BIG_ENDIAN 2894 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_CNTL, BUF_SWAP, 1); 2895 #endif 2896 mqd->cp_gfx_hqd_cntl = tmp; 2897 2898 /* set up cp_doorbell_control */ 2899 tmp = RREG32_SOC15(GC, 0, regCP_RB_DOORBELL_CONTROL); 2900 if (prop->use_doorbell) { 2901 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL, 2902 DOORBELL_OFFSET, prop->doorbell_index); 2903 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL, 2904 DOORBELL_EN, 1); 2905 } else 2906 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL, 2907 DOORBELL_EN, 0); 2908 mqd->cp_rb_doorbell_control = tmp; 2909 2910 /* reset read and write pointers, similar to CP_RB0_WPTR/_RPTR */ 2911 mqd->cp_gfx_hqd_rptr = RREG32_SOC15(GC, 0, regCP_GFX_HQD_RPTR); 2912 2913 /* active the queue */ 2914 mqd->cp_gfx_hqd_active = 1; 2915 2916 return 0; 2917 } 2918 2919 static int gfx_v12_0_gfx_init_queue(struct amdgpu_ring *ring) 2920 { 2921 struct amdgpu_device *adev = ring->adev; 2922 struct v12_gfx_mqd *mqd = ring->mqd_ptr; 2923 int mqd_idx = ring - &adev->gfx.gfx_ring[0]; 2924 2925 if (!amdgpu_in_reset(adev) && !adev->in_suspend) { 2926 memset((void *)mqd, 0, sizeof(*mqd)); 2927 mutex_lock(&adev->srbm_mutex); 2928 soc24_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0); 2929 amdgpu_ring_init_mqd(ring); 2930 soc24_grbm_select(adev, 0, 0, 0, 0); 2931 mutex_unlock(&adev->srbm_mutex); 2932 if (adev->gfx.me.mqd_backup[mqd_idx]) 2933 memcpy_fromio(adev->gfx.me.mqd_backup[mqd_idx], mqd, sizeof(*mqd)); 2934 } else { 2935 /* restore mqd with the backup copy */ 2936 if (adev->gfx.me.mqd_backup[mqd_idx]) 2937 memcpy_toio(mqd, adev->gfx.me.mqd_backup[mqd_idx], sizeof(*mqd)); 2938 /* reset the ring */ 2939 ring->wptr = 0; 2940 *ring->wptr_cpu_addr = 0; 2941 amdgpu_ring_clear_ring(ring); 2942 } 2943 2944 return 0; 2945 } 2946 2947 static int gfx_v12_0_cp_async_gfx_ring_resume(struct amdgpu_device *adev) 2948 { 2949 int r, i; 2950 struct amdgpu_ring *ring; 2951 2952 for (i = 0; i < adev->gfx.num_gfx_rings; i++) { 2953 ring = &adev->gfx.gfx_ring[i]; 2954 2955 r = amdgpu_bo_reserve(ring->mqd_obj, false); 2956 if (unlikely(r != 0)) 2957 goto done; 2958 2959 r = amdgpu_bo_kmap(ring->mqd_obj, (void **)&ring->mqd_ptr); 2960 if (!r) { 2961 r = gfx_v12_0_gfx_init_queue(ring); 2962 amdgpu_bo_kunmap(ring->mqd_obj); 2963 ring->mqd_ptr = NULL; 2964 } 2965 amdgpu_bo_unreserve(ring->mqd_obj); 2966 if (r) 2967 goto done; 2968 } 2969 2970 r = amdgpu_gfx_enable_kgq(adev, 0); 2971 if (r) 2972 goto done; 2973 2974 r = gfx_v12_0_cp_gfx_start(adev); 2975 if (r) 2976 goto done; 2977 2978 for (i = 0; i < adev->gfx.num_gfx_rings; i++) { 2979 ring = &adev->gfx.gfx_ring[i]; 2980 ring->sched.ready = true; 2981 } 2982 done: 2983 return r; 2984 } 2985 2986 static int gfx_v12_0_compute_mqd_init(struct amdgpu_device *adev, void *m, 2987 struct amdgpu_mqd_prop *prop) 2988 { 2989 struct v12_compute_mqd *mqd = m; 2990 uint64_t hqd_gpu_addr, wb_gpu_addr, eop_base_addr; 2991 uint32_t tmp; 2992 2993 mqd->header = 0xC0310800; 2994 mqd->compute_pipelinestat_enable = 0x00000001; 2995 mqd->compute_static_thread_mgmt_se0 = 0xffffffff; 2996 mqd->compute_static_thread_mgmt_se1 = 0xffffffff; 2997 mqd->compute_static_thread_mgmt_se2 = 0xffffffff; 2998 mqd->compute_static_thread_mgmt_se3 = 0xffffffff; 2999 mqd->compute_misc_reserved = 0x00000007; 3000 3001 eop_base_addr = prop->eop_gpu_addr >> 8; 3002 mqd->cp_hqd_eop_base_addr_lo = eop_base_addr; 3003 mqd->cp_hqd_eop_base_addr_hi = upper_32_bits(eop_base_addr); 3004 3005 /* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */ 3006 tmp = RREG32_SOC15(GC, 0, regCP_HQD_EOP_CONTROL); 3007 tmp = REG_SET_FIELD(tmp, CP_HQD_EOP_CONTROL, EOP_SIZE, 3008 (order_base_2(GFX12_MEC_HPD_SIZE / 4) - 1)); 3009 3010 mqd->cp_hqd_eop_control = tmp; 3011 3012 /* enable doorbell? */ 3013 tmp = RREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL); 3014 3015 if (prop->use_doorbell) { 3016 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 3017 DOORBELL_OFFSET, prop->doorbell_index); 3018 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 3019 DOORBELL_EN, 1); 3020 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 3021 DOORBELL_SOURCE, 0); 3022 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 3023 DOORBELL_HIT, 0); 3024 } else { 3025 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 3026 DOORBELL_EN, 0); 3027 } 3028 3029 mqd->cp_hqd_pq_doorbell_control = tmp; 3030 3031 /* disable the queue if it's active */ 3032 mqd->cp_hqd_dequeue_request = 0; 3033 mqd->cp_hqd_pq_rptr = 0; 3034 mqd->cp_hqd_pq_wptr_lo = 0; 3035 mqd->cp_hqd_pq_wptr_hi = 0; 3036 3037 /* set the pointer to the MQD */ 3038 mqd->cp_mqd_base_addr_lo = prop->mqd_gpu_addr & 0xfffffffc; 3039 mqd->cp_mqd_base_addr_hi = upper_32_bits(prop->mqd_gpu_addr); 3040 3041 /* set MQD vmid to 0 */ 3042 tmp = RREG32_SOC15(GC, 0, regCP_MQD_CONTROL); 3043 tmp = REG_SET_FIELD(tmp, CP_MQD_CONTROL, VMID, 0); 3044 mqd->cp_mqd_control = tmp; 3045 3046 /* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */ 3047 hqd_gpu_addr = prop->hqd_base_gpu_addr >> 8; 3048 mqd->cp_hqd_pq_base_lo = hqd_gpu_addr; 3049 mqd->cp_hqd_pq_base_hi = upper_32_bits(hqd_gpu_addr); 3050 3051 /* set up the HQD, this is similar to CP_RB0_CNTL */ 3052 tmp = RREG32_SOC15(GC, 0, regCP_HQD_PQ_CONTROL); 3053 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, QUEUE_SIZE, 3054 (order_base_2(prop->queue_size / 4) - 1)); 3055 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, RPTR_BLOCK_SIZE, 3056 (order_base_2(AMDGPU_GPU_PAGE_SIZE / 4) - 1)); 3057 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, UNORD_DISPATCH, 0); 3058 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, TUNNEL_DISPATCH, 0); 3059 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, PRIV_STATE, 1); 3060 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, KMD_QUEUE, 1); 3061 mqd->cp_hqd_pq_control = tmp; 3062 3063 /* set the wb address whether it's enabled or not */ 3064 wb_gpu_addr = prop->rptr_gpu_addr; 3065 mqd->cp_hqd_pq_rptr_report_addr_lo = wb_gpu_addr & 0xfffffffc; 3066 mqd->cp_hqd_pq_rptr_report_addr_hi = 3067 upper_32_bits(wb_gpu_addr) & 0xffff; 3068 3069 /* only used if CP_PQ_WPTR_POLL_CNTL.CP_PQ_WPTR_POLL_CNTL__EN_MASK=1 */ 3070 wb_gpu_addr = prop->wptr_gpu_addr; 3071 mqd->cp_hqd_pq_wptr_poll_addr_lo = wb_gpu_addr & 0xfffffffc; 3072 mqd->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr) & 0xffff; 3073 3074 tmp = 0; 3075 /* enable the doorbell if requested */ 3076 if (prop->use_doorbell) { 3077 tmp = RREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL); 3078 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 3079 DOORBELL_OFFSET, prop->doorbell_index); 3080 3081 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 3082 DOORBELL_EN, 1); 3083 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 3084 DOORBELL_SOURCE, 0); 3085 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 3086 DOORBELL_HIT, 0); 3087 } 3088 3089 mqd->cp_hqd_pq_doorbell_control = tmp; 3090 3091 /* reset read and write pointers, similar to CP_RB0_WPTR/_RPTR */ 3092 mqd->cp_hqd_pq_rptr = RREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR); 3093 3094 /* set the vmid for the queue */ 3095 mqd->cp_hqd_vmid = 0; 3096 3097 tmp = RREG32_SOC15(GC, 0, regCP_HQD_PERSISTENT_STATE); 3098 tmp = REG_SET_FIELD(tmp, CP_HQD_PERSISTENT_STATE, PRELOAD_SIZE, 0x55); 3099 mqd->cp_hqd_persistent_state = tmp; 3100 3101 /* set MIN_IB_AVAIL_SIZE */ 3102 tmp = RREG32_SOC15(GC, 0, regCP_HQD_IB_CONTROL); 3103 tmp = REG_SET_FIELD(tmp, CP_HQD_IB_CONTROL, MIN_IB_AVAIL_SIZE, 3); 3104 mqd->cp_hqd_ib_control = tmp; 3105 3106 /* set static priority for a compute queue/ring */ 3107 mqd->cp_hqd_pipe_priority = prop->hqd_pipe_priority; 3108 mqd->cp_hqd_queue_priority = prop->hqd_queue_priority; 3109 3110 mqd->cp_hqd_active = prop->hqd_active; 3111 3112 return 0; 3113 } 3114 3115 static int gfx_v12_0_kiq_init_register(struct amdgpu_ring *ring) 3116 { 3117 struct amdgpu_device *adev = ring->adev; 3118 struct v12_compute_mqd *mqd = ring->mqd_ptr; 3119 int j; 3120 3121 /* inactivate the queue */ 3122 if (amdgpu_sriov_vf(adev)) 3123 WREG32_SOC15(GC, 0, regCP_HQD_ACTIVE, 0); 3124 3125 /* disable wptr polling */ 3126 WREG32_FIELD15_PREREG(GC, 0, CP_PQ_WPTR_POLL_CNTL, EN, 0); 3127 3128 /* write the EOP addr */ 3129 WREG32_SOC15(GC, 0, regCP_HQD_EOP_BASE_ADDR, 3130 mqd->cp_hqd_eop_base_addr_lo); 3131 WREG32_SOC15(GC, 0, regCP_HQD_EOP_BASE_ADDR_HI, 3132 mqd->cp_hqd_eop_base_addr_hi); 3133 3134 /* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */ 3135 WREG32_SOC15(GC, 0, regCP_HQD_EOP_CONTROL, 3136 mqd->cp_hqd_eop_control); 3137 3138 /* enable doorbell? */ 3139 WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, 3140 mqd->cp_hqd_pq_doorbell_control); 3141 3142 /* disable the queue if it's active */ 3143 if (RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1) { 3144 WREG32_SOC15(GC, 0, regCP_HQD_DEQUEUE_REQUEST, 1); 3145 for (j = 0; j < adev->usec_timeout; j++) { 3146 if (!(RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1)) 3147 break; 3148 udelay(1); 3149 } 3150 WREG32_SOC15(GC, 0, regCP_HQD_DEQUEUE_REQUEST, 3151 mqd->cp_hqd_dequeue_request); 3152 WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR, 3153 mqd->cp_hqd_pq_rptr); 3154 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_LO, 3155 mqd->cp_hqd_pq_wptr_lo); 3156 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_HI, 3157 mqd->cp_hqd_pq_wptr_hi); 3158 } 3159 3160 /* set the pointer to the MQD */ 3161 WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR, 3162 mqd->cp_mqd_base_addr_lo); 3163 WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR_HI, 3164 mqd->cp_mqd_base_addr_hi); 3165 3166 /* set MQD vmid to 0 */ 3167 WREG32_SOC15(GC, 0, regCP_MQD_CONTROL, 3168 mqd->cp_mqd_control); 3169 3170 /* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */ 3171 WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE, 3172 mqd->cp_hqd_pq_base_lo); 3173 WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE_HI, 3174 mqd->cp_hqd_pq_base_hi); 3175 3176 /* set up the HQD, this is similar to CP_RB0_CNTL */ 3177 WREG32_SOC15(GC, 0, regCP_HQD_PQ_CONTROL, 3178 mqd->cp_hqd_pq_control); 3179 3180 /* set the wb address whether it's enabled or not */ 3181 WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR, 3182 mqd->cp_hqd_pq_rptr_report_addr_lo); 3183 WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR_HI, 3184 mqd->cp_hqd_pq_rptr_report_addr_hi); 3185 3186 /* only used if CP_PQ_WPTR_POLL_CNTL.CP_PQ_WPTR_POLL_CNTL__EN_MASK=1 */ 3187 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR, 3188 mqd->cp_hqd_pq_wptr_poll_addr_lo); 3189 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR_HI, 3190 mqd->cp_hqd_pq_wptr_poll_addr_hi); 3191 3192 /* enable the doorbell if requested */ 3193 if (ring->use_doorbell) { 3194 WREG32_SOC15(GC, 0, regCP_MEC_DOORBELL_RANGE_LOWER, 3195 (adev->doorbell_index.kiq * 2) << 2); 3196 WREG32_SOC15(GC, 0, regCP_MEC_DOORBELL_RANGE_UPPER, 3197 (adev->doorbell_index.userqueue_end * 2) << 2); 3198 } 3199 3200 WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, 3201 mqd->cp_hqd_pq_doorbell_control); 3202 3203 /* reset read and write pointers, similar to CP_RB0_WPTR/_RPTR */ 3204 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_LO, 3205 mqd->cp_hqd_pq_wptr_lo); 3206 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_HI, 3207 mqd->cp_hqd_pq_wptr_hi); 3208 3209 /* set the vmid for the queue */ 3210 WREG32_SOC15(GC, 0, regCP_HQD_VMID, mqd->cp_hqd_vmid); 3211 3212 WREG32_SOC15(GC, 0, regCP_HQD_PERSISTENT_STATE, 3213 mqd->cp_hqd_persistent_state); 3214 3215 /* activate the queue */ 3216 WREG32_SOC15(GC, 0, regCP_HQD_ACTIVE, 3217 mqd->cp_hqd_active); 3218 3219 if (ring->use_doorbell) 3220 WREG32_FIELD15_PREREG(GC, 0, CP_PQ_STATUS, DOORBELL_ENABLE, 1); 3221 3222 return 0; 3223 } 3224 3225 static int gfx_v12_0_kiq_init_queue(struct amdgpu_ring *ring) 3226 { 3227 struct amdgpu_device *adev = ring->adev; 3228 struct v12_compute_mqd *mqd = ring->mqd_ptr; 3229 int mqd_idx = AMDGPU_MAX_COMPUTE_RINGS; 3230 3231 gfx_v12_0_kiq_setting(ring); 3232 3233 if (amdgpu_in_reset(adev)) { /* for GPU_RESET case */ 3234 /* reset MQD to a clean status */ 3235 if (adev->gfx.mec.mqd_backup[mqd_idx]) 3236 memcpy(mqd, adev->gfx.mec.mqd_backup[mqd_idx], sizeof(*mqd)); 3237 3238 /* reset ring buffer */ 3239 ring->wptr = 0; 3240 amdgpu_ring_clear_ring(ring); 3241 3242 mutex_lock(&adev->srbm_mutex); 3243 soc24_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0); 3244 gfx_v12_0_kiq_init_register(ring); 3245 soc24_grbm_select(adev, 0, 0, 0, 0); 3246 mutex_unlock(&adev->srbm_mutex); 3247 } else { 3248 memset((void *)mqd, 0, sizeof(*mqd)); 3249 if (amdgpu_sriov_vf(adev) && adev->in_suspend) 3250 amdgpu_ring_clear_ring(ring); 3251 mutex_lock(&adev->srbm_mutex); 3252 soc24_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0); 3253 amdgpu_ring_init_mqd(ring); 3254 gfx_v12_0_kiq_init_register(ring); 3255 soc24_grbm_select(adev, 0, 0, 0, 0); 3256 mutex_unlock(&adev->srbm_mutex); 3257 3258 if (adev->gfx.mec.mqd_backup[mqd_idx]) 3259 memcpy(adev->gfx.mec.mqd_backup[mqd_idx], mqd, sizeof(*mqd)); 3260 } 3261 3262 return 0; 3263 } 3264 3265 static int gfx_v12_0_kcq_init_queue(struct amdgpu_ring *ring) 3266 { 3267 struct amdgpu_device *adev = ring->adev; 3268 struct v12_compute_mqd *mqd = ring->mqd_ptr; 3269 int mqd_idx = ring - &adev->gfx.compute_ring[0]; 3270 3271 if (!amdgpu_in_reset(adev) && !adev->in_suspend) { 3272 memset((void *)mqd, 0, sizeof(*mqd)); 3273 mutex_lock(&adev->srbm_mutex); 3274 soc24_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0); 3275 amdgpu_ring_init_mqd(ring); 3276 soc24_grbm_select(adev, 0, 0, 0, 0); 3277 mutex_unlock(&adev->srbm_mutex); 3278 3279 if (adev->gfx.mec.mqd_backup[mqd_idx]) 3280 memcpy_fromio(adev->gfx.mec.mqd_backup[mqd_idx], mqd, sizeof(*mqd)); 3281 } else { 3282 /* restore MQD to a clean status */ 3283 if (adev->gfx.mec.mqd_backup[mqd_idx]) 3284 memcpy_toio(mqd, adev->gfx.mec.mqd_backup[mqd_idx], sizeof(*mqd)); 3285 /* reset ring buffer */ 3286 ring->wptr = 0; 3287 atomic64_set((atomic64_t *)ring->wptr_cpu_addr, 0); 3288 amdgpu_ring_clear_ring(ring); 3289 } 3290 3291 return 0; 3292 } 3293 3294 static int gfx_v12_0_kiq_resume(struct amdgpu_device *adev) 3295 { 3296 struct amdgpu_ring *ring; 3297 int r; 3298 3299 ring = &adev->gfx.kiq[0].ring; 3300 3301 r = amdgpu_bo_reserve(ring->mqd_obj, false); 3302 if (unlikely(r != 0)) 3303 return r; 3304 3305 r = amdgpu_bo_kmap(ring->mqd_obj, (void **)&ring->mqd_ptr); 3306 if (unlikely(r != 0)) { 3307 amdgpu_bo_unreserve(ring->mqd_obj); 3308 return r; 3309 } 3310 3311 gfx_v12_0_kiq_init_queue(ring); 3312 amdgpu_bo_kunmap(ring->mqd_obj); 3313 ring->mqd_ptr = NULL; 3314 amdgpu_bo_unreserve(ring->mqd_obj); 3315 ring->sched.ready = true; 3316 return 0; 3317 } 3318 3319 static int gfx_v12_0_kcq_resume(struct amdgpu_device *adev) 3320 { 3321 struct amdgpu_ring *ring = NULL; 3322 int r = 0, i; 3323 3324 if (!amdgpu_async_gfx_ring) 3325 gfx_v12_0_cp_compute_enable(adev, true); 3326 3327 for (i = 0; i < adev->gfx.num_compute_rings; i++) { 3328 ring = &adev->gfx.compute_ring[i]; 3329 3330 r = amdgpu_bo_reserve(ring->mqd_obj, false); 3331 if (unlikely(r != 0)) 3332 goto done; 3333 r = amdgpu_bo_kmap(ring->mqd_obj, (void **)&ring->mqd_ptr); 3334 if (!r) { 3335 r = gfx_v12_0_kcq_init_queue(ring); 3336 amdgpu_bo_kunmap(ring->mqd_obj); 3337 ring->mqd_ptr = NULL; 3338 } 3339 amdgpu_bo_unreserve(ring->mqd_obj); 3340 if (r) 3341 goto done; 3342 } 3343 3344 r = amdgpu_gfx_enable_kcq(adev, 0); 3345 done: 3346 return r; 3347 } 3348 3349 static int gfx_v12_0_cp_resume(struct amdgpu_device *adev) 3350 { 3351 int r, i; 3352 struct amdgpu_ring *ring; 3353 3354 if (!(adev->flags & AMD_IS_APU)) 3355 gfx_v12_0_enable_gui_idle_interrupt(adev, false); 3356 3357 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) { 3358 /* legacy firmware loading */ 3359 r = gfx_v12_0_cp_gfx_load_microcode(adev); 3360 if (r) 3361 return r; 3362 3363 r = gfx_v12_0_cp_compute_load_microcode_rs64(adev); 3364 if (r) 3365 return r; 3366 } 3367 3368 gfx_v12_0_cp_set_doorbell_range(adev); 3369 3370 if (amdgpu_async_gfx_ring) { 3371 gfx_v12_0_cp_compute_enable(adev, true); 3372 gfx_v12_0_cp_gfx_enable(adev, true); 3373 } 3374 3375 if (adev->enable_mes_kiq && adev->mes.kiq_hw_init) 3376 r = amdgpu_mes_kiq_hw_init(adev); 3377 else 3378 r = gfx_v12_0_kiq_resume(adev); 3379 if (r) 3380 return r; 3381 3382 r = gfx_v12_0_kcq_resume(adev); 3383 if (r) 3384 return r; 3385 3386 if (!amdgpu_async_gfx_ring) { 3387 r = gfx_v12_0_cp_gfx_resume(adev); 3388 if (r) 3389 return r; 3390 } else { 3391 r = gfx_v12_0_cp_async_gfx_ring_resume(adev); 3392 if (r) 3393 return r; 3394 } 3395 3396 for (i = 0; i < adev->gfx.num_gfx_rings; i++) { 3397 ring = &adev->gfx.gfx_ring[i]; 3398 r = amdgpu_ring_test_helper(ring); 3399 if (r) 3400 return r; 3401 } 3402 3403 for (i = 0; i < adev->gfx.num_compute_rings; i++) { 3404 ring = &adev->gfx.compute_ring[i]; 3405 r = amdgpu_ring_test_helper(ring); 3406 if (r) 3407 return r; 3408 } 3409 3410 return 0; 3411 } 3412 3413 static void gfx_v12_0_cp_enable(struct amdgpu_device *adev, bool enable) 3414 { 3415 gfx_v12_0_cp_gfx_enable(adev, enable); 3416 gfx_v12_0_cp_compute_enable(adev, enable); 3417 } 3418 3419 static int gfx_v12_0_gfxhub_enable(struct amdgpu_device *adev) 3420 { 3421 int r; 3422 bool value; 3423 3424 r = adev->gfxhub.funcs->gart_enable(adev); 3425 if (r) 3426 return r; 3427 3428 adev->hdp.funcs->flush_hdp(adev, NULL); 3429 3430 value = (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) ? 3431 false : true; 3432 3433 adev->gfxhub.funcs->set_fault_enable_default(adev, value); 3434 /* TODO investigate why this and the hdp flush above is needed, 3435 * are we missing a flush somewhere else? */ 3436 adev->gmc.gmc_funcs->flush_gpu_tlb(adev, 0, AMDGPU_GFXHUB(0), 0); 3437 3438 return 0; 3439 } 3440 3441 static int get_gb_addr_config(struct amdgpu_device *adev) 3442 { 3443 u32 gb_addr_config; 3444 3445 gb_addr_config = RREG32_SOC15(GC, 0, regGB_ADDR_CONFIG); 3446 if (gb_addr_config == 0) 3447 return -EINVAL; 3448 3449 adev->gfx.config.gb_addr_config_fields.num_pkrs = 3450 1 << REG_GET_FIELD(gb_addr_config, GB_ADDR_CONFIG, NUM_PKRS); 3451 3452 adev->gfx.config.gb_addr_config = gb_addr_config; 3453 3454 adev->gfx.config.gb_addr_config_fields.num_pipes = 1 << 3455 REG_GET_FIELD(adev->gfx.config.gb_addr_config, 3456 GB_ADDR_CONFIG, NUM_PIPES); 3457 3458 adev->gfx.config.max_tile_pipes = 3459 adev->gfx.config.gb_addr_config_fields.num_pipes; 3460 3461 adev->gfx.config.gb_addr_config_fields.max_compress_frags = 1 << 3462 REG_GET_FIELD(adev->gfx.config.gb_addr_config, 3463 GB_ADDR_CONFIG, MAX_COMPRESSED_FRAGS); 3464 adev->gfx.config.gb_addr_config_fields.num_rb_per_se = 1 << 3465 REG_GET_FIELD(adev->gfx.config.gb_addr_config, 3466 GB_ADDR_CONFIG, NUM_RB_PER_SE); 3467 adev->gfx.config.gb_addr_config_fields.num_se = 1 << 3468 REG_GET_FIELD(adev->gfx.config.gb_addr_config, 3469 GB_ADDR_CONFIG, NUM_SHADER_ENGINES); 3470 adev->gfx.config.gb_addr_config_fields.pipe_interleave_size = 1 << (8 + 3471 REG_GET_FIELD(adev->gfx.config.gb_addr_config, 3472 GB_ADDR_CONFIG, PIPE_INTERLEAVE_SIZE)); 3473 3474 return 0; 3475 } 3476 3477 static void gfx_v12_0_disable_gpa_mode(struct amdgpu_device *adev) 3478 { 3479 uint32_t data; 3480 3481 data = RREG32_SOC15(GC, 0, regCPC_PSP_DEBUG); 3482 data |= CPC_PSP_DEBUG__GPA_OVERRIDE_MASK; 3483 WREG32_SOC15(GC, 0, regCPC_PSP_DEBUG, data); 3484 3485 data = RREG32_SOC15(GC, 0, regCPG_PSP_DEBUG); 3486 data |= CPG_PSP_DEBUG__GPA_OVERRIDE_MASK; 3487 WREG32_SOC15(GC, 0, regCPG_PSP_DEBUG, data); 3488 } 3489 3490 static void gfx_v12_0_init_golden_registers(struct amdgpu_device *adev) 3491 { 3492 if (amdgpu_sriov_vf(adev)) 3493 return; 3494 3495 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 3496 case IP_VERSION(12, 0, 0): 3497 case IP_VERSION(12, 0, 1): 3498 if (adev->rev_id == 0) 3499 soc15_program_register_sequence(adev, 3500 golden_settings_gc_12_0, 3501 (const u32)ARRAY_SIZE(golden_settings_gc_12_0)); 3502 break; 3503 default: 3504 break; 3505 } 3506 } 3507 3508 static int gfx_v12_0_hw_init(void *handle) 3509 { 3510 int r; 3511 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 3512 3513 if (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) { 3514 if (adev->gfx.imu.funcs && (amdgpu_dpm > 0)) { 3515 /* RLC autoload sequence 1: Program rlc ram */ 3516 if (adev->gfx.imu.funcs->program_rlc_ram) 3517 adev->gfx.imu.funcs->program_rlc_ram(adev); 3518 } 3519 /* rlc autoload firmware */ 3520 r = gfx_v12_0_rlc_backdoor_autoload_enable(adev); 3521 if (r) 3522 return r; 3523 } else { 3524 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) { 3525 if (adev->gfx.imu.funcs && (amdgpu_dpm > 0)) { 3526 if (adev->gfx.imu.funcs->load_microcode) 3527 adev->gfx.imu.funcs->load_microcode(adev); 3528 if (adev->gfx.imu.funcs->setup_imu) 3529 adev->gfx.imu.funcs->setup_imu(adev); 3530 if (adev->gfx.imu.funcs->start_imu) 3531 adev->gfx.imu.funcs->start_imu(adev); 3532 } 3533 3534 /* disable gpa mode in backdoor loading */ 3535 gfx_v12_0_disable_gpa_mode(adev); 3536 } 3537 } 3538 3539 if ((adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) || 3540 (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP)) { 3541 r = gfx_v12_0_wait_for_rlc_autoload_complete(adev); 3542 if (r) { 3543 dev_err(adev->dev, "(%d) failed to wait rlc autoload complete\n", r); 3544 return r; 3545 } 3546 } 3547 3548 if (!amdgpu_emu_mode) 3549 gfx_v12_0_init_golden_registers(adev); 3550 3551 adev->gfx.is_poweron = true; 3552 3553 if (get_gb_addr_config(adev)) 3554 DRM_WARN("Invalid gb_addr_config !\n"); 3555 3556 if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) 3557 gfx_v12_0_config_gfx_rs64(adev); 3558 3559 r = gfx_v12_0_gfxhub_enable(adev); 3560 if (r) 3561 return r; 3562 3563 if ((adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT || 3564 adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) && 3565 (amdgpu_dpm == 1)) { 3566 /** 3567 * For gfx 12, rlc firmware loading relies on smu firmware is 3568 * loaded firstly, so in direct type, it has to load smc ucode 3569 * here before rlc. 3570 */ 3571 r = amdgpu_pm_load_smu_firmware(adev, NULL); 3572 if (r) 3573 return r; 3574 } 3575 3576 gfx_v12_0_constants_init(adev); 3577 3578 if (adev->nbio.funcs->gc_doorbell_init) 3579 adev->nbio.funcs->gc_doorbell_init(adev); 3580 3581 r = gfx_v12_0_rlc_resume(adev); 3582 if (r) 3583 return r; 3584 3585 /* 3586 * init golden registers and rlc resume may override some registers, 3587 * reconfig them here 3588 */ 3589 gfx_v12_0_tcp_harvest(adev); 3590 3591 r = gfx_v12_0_cp_resume(adev); 3592 if (r) 3593 return r; 3594 3595 return r; 3596 } 3597 3598 static int gfx_v12_0_kiq_disable_kgq(struct amdgpu_device *adev) 3599 { 3600 struct amdgpu_kiq *kiq = &adev->gfx.kiq[0]; 3601 struct amdgpu_ring *kiq_ring = &kiq->ring; 3602 int i, r = 0; 3603 3604 if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues) 3605 return -EINVAL; 3606 3607 if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size * 3608 adev->gfx.num_gfx_rings)) 3609 return -ENOMEM; 3610 3611 for (i = 0; i < adev->gfx.num_gfx_rings; i++) 3612 kiq->pmf->kiq_unmap_queues(kiq_ring, &adev->gfx.gfx_ring[i], 3613 PREEMPT_QUEUES, 0, 0); 3614 3615 if (adev->gfx.kiq[0].ring.sched.ready) 3616 r = amdgpu_ring_test_helper(kiq_ring); 3617 3618 return r; 3619 } 3620 3621 static int gfx_v12_0_hw_fini(void *handle) 3622 { 3623 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 3624 int r; 3625 uint32_t tmp; 3626 3627 amdgpu_irq_put(adev, &adev->gfx.priv_reg_irq, 0); 3628 amdgpu_irq_put(adev, &adev->gfx.priv_inst_irq, 0); 3629 amdgpu_irq_put(adev, &adev->gfx.bad_op_irq, 0); 3630 3631 if (!adev->no_hw_access) { 3632 if (amdgpu_async_gfx_ring) { 3633 r = gfx_v12_0_kiq_disable_kgq(adev); 3634 if (r) 3635 DRM_ERROR("KGQ disable failed\n"); 3636 } 3637 3638 if (amdgpu_gfx_disable_kcq(adev, 0)) 3639 DRM_ERROR("KCQ disable failed\n"); 3640 3641 amdgpu_mes_kiq_hw_fini(adev); 3642 } 3643 3644 if (amdgpu_sriov_vf(adev)) { 3645 gfx_v12_0_cp_gfx_enable(adev, false); 3646 /* Program KIQ position of RLC_CP_SCHEDULERS during destroy */ 3647 tmp = RREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS); 3648 tmp &= 0xffffff00; 3649 WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp); 3650 3651 return 0; 3652 } 3653 gfx_v12_0_cp_enable(adev, false); 3654 gfx_v12_0_enable_gui_idle_interrupt(adev, false); 3655 3656 adev->gfxhub.funcs->gart_disable(adev); 3657 3658 adev->gfx.is_poweron = false; 3659 3660 return 0; 3661 } 3662 3663 static int gfx_v12_0_suspend(void *handle) 3664 { 3665 return gfx_v12_0_hw_fini(handle); 3666 } 3667 3668 static int gfx_v12_0_resume(void *handle) 3669 { 3670 return gfx_v12_0_hw_init(handle); 3671 } 3672 3673 static bool gfx_v12_0_is_idle(void *handle) 3674 { 3675 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 3676 3677 if (REG_GET_FIELD(RREG32_SOC15(GC, 0, regGRBM_STATUS), 3678 GRBM_STATUS, GUI_ACTIVE)) 3679 return false; 3680 else 3681 return true; 3682 } 3683 3684 static int gfx_v12_0_wait_for_idle(void *handle) 3685 { 3686 unsigned i; 3687 u32 tmp; 3688 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 3689 3690 for (i = 0; i < adev->usec_timeout; i++) { 3691 /* read MC_STATUS */ 3692 tmp = RREG32_SOC15(GC, 0, regGRBM_STATUS) & 3693 GRBM_STATUS__GUI_ACTIVE_MASK; 3694 3695 if (!REG_GET_FIELD(tmp, GRBM_STATUS, GUI_ACTIVE)) 3696 return 0; 3697 udelay(1); 3698 } 3699 return -ETIMEDOUT; 3700 } 3701 3702 static uint64_t gfx_v12_0_get_gpu_clock_counter(struct amdgpu_device *adev) 3703 { 3704 uint64_t clock = 0; 3705 3706 if (adev->smuio.funcs && 3707 adev->smuio.funcs->get_gpu_clock_counter) 3708 clock = adev->smuio.funcs->get_gpu_clock_counter(adev); 3709 else 3710 dev_warn(adev->dev, "query gpu clock counter is not supported\n"); 3711 3712 return clock; 3713 } 3714 3715 static int gfx_v12_0_early_init(void *handle) 3716 { 3717 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 3718 3719 adev->gfx.funcs = &gfx_v12_0_gfx_funcs; 3720 3721 adev->gfx.num_gfx_rings = GFX12_NUM_GFX_RINGS; 3722 adev->gfx.num_compute_rings = min(amdgpu_gfx_get_num_kcq(adev), 3723 AMDGPU_MAX_COMPUTE_RINGS); 3724 3725 gfx_v12_0_set_kiq_pm4_funcs(adev); 3726 gfx_v12_0_set_ring_funcs(adev); 3727 gfx_v12_0_set_irq_funcs(adev); 3728 gfx_v12_0_set_rlc_funcs(adev); 3729 gfx_v12_0_set_mqd_funcs(adev); 3730 gfx_v12_0_set_imu_funcs(adev); 3731 3732 gfx_v12_0_init_rlcg_reg_access_ctrl(adev); 3733 3734 return gfx_v12_0_init_microcode(adev); 3735 } 3736 3737 static int gfx_v12_0_late_init(void *handle) 3738 { 3739 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 3740 int r; 3741 3742 r = amdgpu_irq_get(adev, &adev->gfx.priv_reg_irq, 0); 3743 if (r) 3744 return r; 3745 3746 r = amdgpu_irq_get(adev, &adev->gfx.priv_inst_irq, 0); 3747 if (r) 3748 return r; 3749 3750 r = amdgpu_irq_get(adev, &adev->gfx.bad_op_irq, 0); 3751 if (r) 3752 return r; 3753 3754 return 0; 3755 } 3756 3757 static bool gfx_v12_0_is_rlc_enabled(struct amdgpu_device *adev) 3758 { 3759 uint32_t rlc_cntl; 3760 3761 /* if RLC is not enabled, do nothing */ 3762 rlc_cntl = RREG32_SOC15(GC, 0, regRLC_CNTL); 3763 return (REG_GET_FIELD(rlc_cntl, RLC_CNTL, RLC_ENABLE_F32)) ? true : false; 3764 } 3765 3766 static void gfx_v12_0_set_safe_mode(struct amdgpu_device *adev, 3767 int xcc_id) 3768 { 3769 uint32_t data; 3770 unsigned i; 3771 3772 data = RLC_SAFE_MODE__CMD_MASK; 3773 data |= (1 << RLC_SAFE_MODE__MESSAGE__SHIFT); 3774 3775 WREG32_SOC15(GC, 0, regRLC_SAFE_MODE, data); 3776 3777 /* wait for RLC_SAFE_MODE */ 3778 for (i = 0; i < adev->usec_timeout; i++) { 3779 if (!REG_GET_FIELD(RREG32_SOC15(GC, 0, regRLC_SAFE_MODE), 3780 RLC_SAFE_MODE, CMD)) 3781 break; 3782 udelay(1); 3783 } 3784 } 3785 3786 static void gfx_v12_0_unset_safe_mode(struct amdgpu_device *adev, 3787 int xcc_id) 3788 { 3789 WREG32_SOC15(GC, 0, regRLC_SAFE_MODE, RLC_SAFE_MODE__CMD_MASK); 3790 } 3791 3792 static void gfx_v12_0_update_perf_clk(struct amdgpu_device *adev, 3793 bool enable) 3794 { 3795 uint32_t def, data; 3796 3797 if (!(adev->cg_flags & AMD_CG_SUPPORT_GFX_PERF_CLK)) 3798 return; 3799 3800 def = data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE); 3801 3802 if (enable) 3803 data &= ~RLC_CGTT_MGCG_OVERRIDE__PERFMON_CLOCK_STATE_MASK; 3804 else 3805 data |= RLC_CGTT_MGCG_OVERRIDE__PERFMON_CLOCK_STATE_MASK; 3806 3807 if (def != data) 3808 WREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE, data); 3809 } 3810 3811 static void gfx_v12_0_update_spm_vmid(struct amdgpu_device *adev, 3812 struct amdgpu_ring *ring, 3813 unsigned vmid) 3814 { 3815 u32 reg, data; 3816 3817 reg = SOC15_REG_OFFSET(GC, 0, regRLC_SPM_MC_CNTL); 3818 if (amdgpu_sriov_is_pp_one_vf(adev)) 3819 data = RREG32_NO_KIQ(reg); 3820 else 3821 data = RREG32(reg); 3822 3823 data &= ~RLC_SPM_MC_CNTL__RLC_SPM_VMID_MASK; 3824 data |= (vmid & RLC_SPM_MC_CNTL__RLC_SPM_VMID_MASK) << RLC_SPM_MC_CNTL__RLC_SPM_VMID__SHIFT; 3825 3826 if (amdgpu_sriov_is_pp_one_vf(adev)) 3827 WREG32_SOC15_NO_KIQ(GC, 0, regRLC_SPM_MC_CNTL, data); 3828 else 3829 WREG32_SOC15(GC, 0, regRLC_SPM_MC_CNTL, data); 3830 3831 if (ring 3832 && amdgpu_sriov_is_pp_one_vf(adev) 3833 && ((ring->funcs->type == AMDGPU_RING_TYPE_GFX) 3834 || (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE))) { 3835 uint32_t reg = SOC15_REG_OFFSET(GC, 0, regRLC_SPM_MC_CNTL); 3836 amdgpu_ring_emit_wreg(ring, reg, data); 3837 } 3838 } 3839 3840 static const struct amdgpu_rlc_funcs gfx_v12_0_rlc_funcs = { 3841 .is_rlc_enabled = gfx_v12_0_is_rlc_enabled, 3842 .set_safe_mode = gfx_v12_0_set_safe_mode, 3843 .unset_safe_mode = gfx_v12_0_unset_safe_mode, 3844 .init = gfx_v12_0_rlc_init, 3845 .get_csb_size = gfx_v12_0_get_csb_size, 3846 .get_csb_buffer = gfx_v12_0_get_csb_buffer, 3847 .resume = gfx_v12_0_rlc_resume, 3848 .stop = gfx_v12_0_rlc_stop, 3849 .reset = gfx_v12_0_rlc_reset, 3850 .start = gfx_v12_0_rlc_start, 3851 .update_spm_vmid = gfx_v12_0_update_spm_vmid, 3852 }; 3853 3854 #if 0 3855 static void gfx_v12_cntl_power_gating(struct amdgpu_device *adev, bool enable) 3856 { 3857 /* TODO */ 3858 } 3859 3860 static void gfx_v12_cntl_pg(struct amdgpu_device *adev, bool enable) 3861 { 3862 /* TODO */ 3863 } 3864 #endif 3865 3866 static int gfx_v12_0_set_powergating_state(void *handle, 3867 enum amd_powergating_state state) 3868 { 3869 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 3870 bool enable = (state == AMD_PG_STATE_GATE); 3871 3872 if (amdgpu_sriov_vf(adev)) 3873 return 0; 3874 3875 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 3876 case IP_VERSION(12, 0, 0): 3877 case IP_VERSION(12, 0, 1): 3878 amdgpu_gfx_off_ctrl(adev, enable); 3879 break; 3880 default: 3881 break; 3882 } 3883 3884 return 0; 3885 } 3886 3887 static void gfx_v12_0_update_coarse_grain_clock_gating(struct amdgpu_device *adev, 3888 bool enable) 3889 { 3890 uint32_t def, data; 3891 3892 if (!(adev->cg_flags & 3893 (AMD_CG_SUPPORT_GFX_CGCG | 3894 AMD_CG_SUPPORT_GFX_CGLS | 3895 AMD_CG_SUPPORT_GFX_3D_CGCG | 3896 AMD_CG_SUPPORT_GFX_3D_CGLS))) 3897 return; 3898 3899 if (enable) { 3900 def = data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE); 3901 3902 /* unset CGCG override */ 3903 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGCG) 3904 data &= ~RLC_CGTT_MGCG_OVERRIDE__GFXIP_CGCG_OVERRIDE_MASK; 3905 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGLS) 3906 data &= ~RLC_CGTT_MGCG_OVERRIDE__GFXIP_CGLS_OVERRIDE_MASK; 3907 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGCG || 3908 adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGLS) 3909 data &= ~RLC_CGTT_MGCG_OVERRIDE__GFXIP_GFX3D_CG_OVERRIDE_MASK; 3910 3911 /* update CGCG override bits */ 3912 if (def != data) 3913 WREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE, data); 3914 3915 /* enable cgcg FSM(0x0000363F) */ 3916 def = data = RREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL); 3917 3918 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGCG) { 3919 data &= ~RLC_CGCG_CGLS_CTRL__CGCG_GFX_IDLE_THRESHOLD_MASK; 3920 data |= (0x36 << RLC_CGCG_CGLS_CTRL__CGCG_GFX_IDLE_THRESHOLD__SHIFT) | 3921 RLC_CGCG_CGLS_CTRL__CGCG_EN_MASK; 3922 } 3923 3924 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGLS) { 3925 data &= ~RLC_CGCG_CGLS_CTRL__CGLS_REP_COMPANSAT_DELAY_MASK; 3926 data |= (0x000F << RLC_CGCG_CGLS_CTRL__CGLS_REP_COMPANSAT_DELAY__SHIFT) | 3927 RLC_CGCG_CGLS_CTRL__CGLS_EN_MASK; 3928 } 3929 3930 if (def != data) 3931 WREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL, data); 3932 3933 /* Program RLC_CGCG_CGLS_CTRL_3D */ 3934 def = data = RREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL_3D); 3935 3936 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGCG) { 3937 data &= ~RLC_CGCG_CGLS_CTRL_3D__CGCG_GFX_IDLE_THRESHOLD_MASK; 3938 data |= (0x36 << RLC_CGCG_CGLS_CTRL_3D__CGCG_GFX_IDLE_THRESHOLD__SHIFT) | 3939 RLC_CGCG_CGLS_CTRL_3D__CGCG_EN_MASK; 3940 } 3941 3942 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGLS) { 3943 data &= ~RLC_CGCG_CGLS_CTRL_3D__CGLS_REP_COMPANSAT_DELAY_MASK; 3944 data |= (0xf << RLC_CGCG_CGLS_CTRL_3D__CGLS_REP_COMPANSAT_DELAY__SHIFT) | 3945 RLC_CGCG_CGLS_CTRL_3D__CGLS_EN_MASK; 3946 } 3947 3948 if (def != data) 3949 WREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL_3D, data); 3950 3951 /* set IDLE_POLL_COUNT(0x00900100) */ 3952 def = data = RREG32_SOC15(GC, 0, regCP_RB_WPTR_POLL_CNTL); 3953 3954 data &= ~(CP_RB_WPTR_POLL_CNTL__POLL_FREQUENCY_MASK | CP_RB_WPTR_POLL_CNTL__IDLE_POLL_COUNT_MASK); 3955 data |= (0x0100 << CP_RB_WPTR_POLL_CNTL__POLL_FREQUENCY__SHIFT) | 3956 (0x0090 << CP_RB_WPTR_POLL_CNTL__IDLE_POLL_COUNT__SHIFT); 3957 3958 if (def != data) 3959 WREG32_SOC15(GC, 0, regCP_RB_WPTR_POLL_CNTL, data); 3960 3961 data = RREG32_SOC15(GC, 0, regCP_INT_CNTL); 3962 data = REG_SET_FIELD(data, CP_INT_CNTL, CNTX_BUSY_INT_ENABLE, 1); 3963 data = REG_SET_FIELD(data, CP_INT_CNTL, CNTX_EMPTY_INT_ENABLE, 1); 3964 data = REG_SET_FIELD(data, CP_INT_CNTL, CMP_BUSY_INT_ENABLE, 1); 3965 data = REG_SET_FIELD(data, CP_INT_CNTL, GFX_IDLE_INT_ENABLE, 1); 3966 WREG32_SOC15(GC, 0, regCP_INT_CNTL, data); 3967 3968 data = RREG32_SOC15(GC, 0, regSDMA0_RLC_CGCG_CTRL); 3969 data = REG_SET_FIELD(data, SDMA0_RLC_CGCG_CTRL, CGCG_INT_ENABLE, 1); 3970 WREG32_SOC15(GC, 0, regSDMA0_RLC_CGCG_CTRL, data); 3971 3972 /* Some ASICs only have one SDMA instance, not need to configure SDMA1 */ 3973 if (adev->sdma.num_instances > 1) { 3974 data = RREG32_SOC15(GC, 0, regSDMA1_RLC_CGCG_CTRL); 3975 data = REG_SET_FIELD(data, SDMA1_RLC_CGCG_CTRL, CGCG_INT_ENABLE, 1); 3976 WREG32_SOC15(GC, 0, regSDMA1_RLC_CGCG_CTRL, data); 3977 } 3978 } else { 3979 /* Program RLC_CGCG_CGLS_CTRL */ 3980 def = data = RREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL); 3981 3982 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGCG) 3983 data &= ~RLC_CGCG_CGLS_CTRL__CGCG_EN_MASK; 3984 3985 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGLS) 3986 data &= ~RLC_CGCG_CGLS_CTRL__CGLS_EN_MASK; 3987 3988 if (def != data) 3989 WREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL, data); 3990 3991 /* Program RLC_CGCG_CGLS_CTRL_3D */ 3992 def = data = RREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL_3D); 3993 3994 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGCG) 3995 data &= ~RLC_CGCG_CGLS_CTRL_3D__CGCG_EN_MASK; 3996 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGLS) 3997 data &= ~RLC_CGCG_CGLS_CTRL_3D__CGLS_EN_MASK; 3998 3999 if (def != data) 4000 WREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL_3D, data); 4001 4002 data = RREG32_SOC15(GC, 0, regSDMA0_RLC_CGCG_CTRL); 4003 data &= ~SDMA0_RLC_CGCG_CTRL__CGCG_INT_ENABLE_MASK; 4004 WREG32_SOC15(GC, 0, regSDMA0_RLC_CGCG_CTRL, data); 4005 4006 /* Some ASICs only have one SDMA instance, not need to configure SDMA1 */ 4007 if (adev->sdma.num_instances > 1) { 4008 data = RREG32_SOC15(GC, 0, regSDMA1_RLC_CGCG_CTRL); 4009 data &= ~SDMA1_RLC_CGCG_CTRL__CGCG_INT_ENABLE_MASK; 4010 WREG32_SOC15(GC, 0, regSDMA1_RLC_CGCG_CTRL, data); 4011 } 4012 } 4013 } 4014 4015 static void gfx_v12_0_update_medium_grain_clock_gating(struct amdgpu_device *adev, 4016 bool enable) 4017 { 4018 uint32_t data, def; 4019 if (!(adev->cg_flags & (AMD_CG_SUPPORT_GFX_MGCG | AMD_CG_SUPPORT_GFX_MGLS))) 4020 return; 4021 4022 /* It is disabled by HW by default */ 4023 if (enable) { 4024 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_MGCG) { 4025 /* 1 - RLC_CGTT_MGCG_OVERRIDE */ 4026 def = data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE); 4027 4028 data &= ~(RLC_CGTT_MGCG_OVERRIDE__GRBM_CGTT_SCLK_OVERRIDE_MASK | 4029 RLC_CGTT_MGCG_OVERRIDE__RLC_CGTT_SCLK_OVERRIDE_MASK | 4030 RLC_CGTT_MGCG_OVERRIDE__GFXIP_MGCG_OVERRIDE_MASK); 4031 4032 if (def != data) 4033 WREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE, data); 4034 } 4035 } else { 4036 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_MGCG) { 4037 def = data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE); 4038 4039 data |= (RLC_CGTT_MGCG_OVERRIDE__RLC_CGTT_SCLK_OVERRIDE_MASK | 4040 RLC_CGTT_MGCG_OVERRIDE__GRBM_CGTT_SCLK_OVERRIDE_MASK | 4041 RLC_CGTT_MGCG_OVERRIDE__GFXIP_MGCG_OVERRIDE_MASK); 4042 4043 if (def != data) 4044 WREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE, data); 4045 } 4046 } 4047 } 4048 4049 static void gfx_v12_0_update_repeater_fgcg(struct amdgpu_device *adev, 4050 bool enable) 4051 { 4052 uint32_t def, data; 4053 4054 if (!(adev->cg_flags & AMD_CG_SUPPORT_REPEATER_FGCG)) 4055 return; 4056 4057 def = data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE); 4058 4059 if (enable) 4060 data &= ~(RLC_CGTT_MGCG_OVERRIDE__GFXIP_REPEATER_FGCG_OVERRIDE_MASK | 4061 RLC_CGTT_MGCG_OVERRIDE__RLC_REPEATER_FGCG_OVERRIDE_MASK); 4062 else 4063 data |= RLC_CGTT_MGCG_OVERRIDE__GFXIP_REPEATER_FGCG_OVERRIDE_MASK | 4064 RLC_CGTT_MGCG_OVERRIDE__RLC_REPEATER_FGCG_OVERRIDE_MASK; 4065 4066 if (def != data) 4067 WREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE, data); 4068 } 4069 4070 static void gfx_v12_0_update_sram_fgcg(struct amdgpu_device *adev, 4071 bool enable) 4072 { 4073 uint32_t def, data; 4074 4075 if (!(adev->cg_flags & AMD_CG_SUPPORT_GFX_FGCG)) 4076 return; 4077 4078 def = data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE); 4079 4080 if (enable) 4081 data &= ~RLC_CGTT_MGCG_OVERRIDE__GFXIP_FGCG_OVERRIDE_MASK; 4082 else 4083 data |= RLC_CGTT_MGCG_OVERRIDE__GFXIP_FGCG_OVERRIDE_MASK; 4084 4085 if (def != data) 4086 WREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE, data); 4087 } 4088 4089 static int gfx_v12_0_update_gfx_clock_gating(struct amdgpu_device *adev, 4090 bool enable) 4091 { 4092 amdgpu_gfx_rlc_enter_safe_mode(adev, 0); 4093 4094 gfx_v12_0_update_coarse_grain_clock_gating(adev, enable); 4095 4096 gfx_v12_0_update_medium_grain_clock_gating(adev, enable); 4097 4098 gfx_v12_0_update_repeater_fgcg(adev, enable); 4099 4100 gfx_v12_0_update_sram_fgcg(adev, enable); 4101 4102 gfx_v12_0_update_perf_clk(adev, enable); 4103 4104 if (adev->cg_flags & 4105 (AMD_CG_SUPPORT_GFX_MGCG | 4106 AMD_CG_SUPPORT_GFX_CGLS | 4107 AMD_CG_SUPPORT_GFX_CGCG | 4108 AMD_CG_SUPPORT_GFX_3D_CGCG | 4109 AMD_CG_SUPPORT_GFX_3D_CGLS)) 4110 gfx_v12_0_enable_gui_idle_interrupt(adev, enable); 4111 4112 amdgpu_gfx_rlc_exit_safe_mode(adev, 0); 4113 4114 return 0; 4115 } 4116 4117 static int gfx_v12_0_set_clockgating_state(void *handle, 4118 enum amd_clockgating_state state) 4119 { 4120 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 4121 4122 if (amdgpu_sriov_vf(adev)) 4123 return 0; 4124 4125 switch (adev->ip_versions[GC_HWIP][0]) { 4126 case IP_VERSION(12, 0, 0): 4127 case IP_VERSION(12, 0, 1): 4128 gfx_v12_0_update_gfx_clock_gating(adev, 4129 state == AMD_CG_STATE_GATE); 4130 break; 4131 default: 4132 break; 4133 } 4134 4135 return 0; 4136 } 4137 4138 static void gfx_v12_0_get_clockgating_state(void *handle, u64 *flags) 4139 { 4140 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 4141 int data; 4142 4143 /* AMD_CG_SUPPORT_GFX_MGCG */ 4144 data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE); 4145 if (!(data & RLC_CGTT_MGCG_OVERRIDE__GFXIP_MGCG_OVERRIDE_MASK)) 4146 *flags |= AMD_CG_SUPPORT_GFX_MGCG; 4147 4148 /* AMD_CG_SUPPORT_REPEATER_FGCG */ 4149 if (!(data & RLC_CGTT_MGCG_OVERRIDE__GFXIP_REPEATER_FGCG_OVERRIDE_MASK)) 4150 *flags |= AMD_CG_SUPPORT_REPEATER_FGCG; 4151 4152 /* AMD_CG_SUPPORT_GFX_FGCG */ 4153 if (!(data & RLC_CGTT_MGCG_OVERRIDE__GFXIP_FGCG_OVERRIDE_MASK)) 4154 *flags |= AMD_CG_SUPPORT_GFX_FGCG; 4155 4156 /* AMD_CG_SUPPORT_GFX_PERF_CLK */ 4157 if (!(data & RLC_CGTT_MGCG_OVERRIDE__PERFMON_CLOCK_STATE_MASK)) 4158 *flags |= AMD_CG_SUPPORT_GFX_PERF_CLK; 4159 4160 /* AMD_CG_SUPPORT_GFX_CGCG */ 4161 data = RREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL); 4162 if (data & RLC_CGCG_CGLS_CTRL__CGCG_EN_MASK) 4163 *flags |= AMD_CG_SUPPORT_GFX_CGCG; 4164 4165 /* AMD_CG_SUPPORT_GFX_CGLS */ 4166 if (data & RLC_CGCG_CGLS_CTRL__CGLS_EN_MASK) 4167 *flags |= AMD_CG_SUPPORT_GFX_CGLS; 4168 4169 /* AMD_CG_SUPPORT_GFX_3D_CGCG */ 4170 data = RREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL_3D); 4171 if (data & RLC_CGCG_CGLS_CTRL_3D__CGCG_EN_MASK) 4172 *flags |= AMD_CG_SUPPORT_GFX_3D_CGCG; 4173 4174 /* AMD_CG_SUPPORT_GFX_3D_CGLS */ 4175 if (data & RLC_CGCG_CGLS_CTRL_3D__CGLS_EN_MASK) 4176 *flags |= AMD_CG_SUPPORT_GFX_3D_CGLS; 4177 } 4178 4179 static u64 gfx_v12_0_ring_get_rptr_gfx(struct amdgpu_ring *ring) 4180 { 4181 /* gfx12 is 32bit rptr*/ 4182 return *(uint32_t *)ring->rptr_cpu_addr; 4183 } 4184 4185 static u64 gfx_v12_0_ring_get_wptr_gfx(struct amdgpu_ring *ring) 4186 { 4187 struct amdgpu_device *adev = ring->adev; 4188 u64 wptr; 4189 4190 /* XXX check if swapping is necessary on BE */ 4191 if (ring->use_doorbell) { 4192 wptr = atomic64_read((atomic64_t *)ring->wptr_cpu_addr); 4193 } else { 4194 wptr = RREG32_SOC15(GC, 0, regCP_RB0_WPTR); 4195 wptr += (u64)RREG32_SOC15(GC, 0, regCP_RB0_WPTR_HI) << 32; 4196 } 4197 4198 return wptr; 4199 } 4200 4201 static void gfx_v12_0_ring_set_wptr_gfx(struct amdgpu_ring *ring) 4202 { 4203 struct amdgpu_device *adev = ring->adev; 4204 uint32_t *wptr_saved; 4205 uint32_t *is_queue_unmap; 4206 uint64_t aggregated_db_index; 4207 uint32_t mqd_size = adev->mqds[AMDGPU_HW_IP_GFX].mqd_size; 4208 uint64_t wptr_tmp; 4209 4210 if (ring->is_mes_queue) { 4211 wptr_saved = (uint32_t *)(ring->mqd_ptr + mqd_size); 4212 is_queue_unmap = (uint32_t *)(ring->mqd_ptr + mqd_size + 4213 sizeof(uint32_t)); 4214 aggregated_db_index = 4215 amdgpu_mes_get_aggregated_doorbell_index(adev, 4216 ring->hw_prio); 4217 4218 wptr_tmp = ring->wptr & ring->buf_mask; 4219 atomic64_set((atomic64_t *)ring->wptr_cpu_addr, wptr_tmp); 4220 *wptr_saved = wptr_tmp; 4221 /* assume doorbell always being used by mes mapped queue */ 4222 if (*is_queue_unmap) { 4223 WDOORBELL64(aggregated_db_index, wptr_tmp); 4224 WDOORBELL64(ring->doorbell_index, wptr_tmp); 4225 } else { 4226 WDOORBELL64(ring->doorbell_index, wptr_tmp); 4227 4228 if (*is_queue_unmap) 4229 WDOORBELL64(aggregated_db_index, wptr_tmp); 4230 } 4231 } else { 4232 if (ring->use_doorbell) { 4233 /* XXX check if swapping is necessary on BE */ 4234 atomic64_set((atomic64_t *)ring->wptr_cpu_addr, 4235 ring->wptr); 4236 WDOORBELL64(ring->doorbell_index, ring->wptr); 4237 } else { 4238 WREG32_SOC15(GC, 0, regCP_RB0_WPTR, 4239 lower_32_bits(ring->wptr)); 4240 WREG32_SOC15(GC, 0, regCP_RB0_WPTR_HI, 4241 upper_32_bits(ring->wptr)); 4242 } 4243 } 4244 } 4245 4246 static u64 gfx_v12_0_ring_get_rptr_compute(struct amdgpu_ring *ring) 4247 { 4248 /* gfx12 hardware is 32bit rptr */ 4249 return *(uint32_t *)ring->rptr_cpu_addr; 4250 } 4251 4252 static u64 gfx_v12_0_ring_get_wptr_compute(struct amdgpu_ring *ring) 4253 { 4254 u64 wptr; 4255 4256 /* XXX check if swapping is necessary on BE */ 4257 if (ring->use_doorbell) 4258 wptr = atomic64_read((atomic64_t *)ring->wptr_cpu_addr); 4259 else 4260 BUG(); 4261 return wptr; 4262 } 4263 4264 static void gfx_v12_0_ring_set_wptr_compute(struct amdgpu_ring *ring) 4265 { 4266 struct amdgpu_device *adev = ring->adev; 4267 uint32_t *wptr_saved; 4268 uint32_t *is_queue_unmap; 4269 uint64_t aggregated_db_index; 4270 uint32_t mqd_size = adev->mqds[AMDGPU_HW_IP_COMPUTE].mqd_size; 4271 uint64_t wptr_tmp; 4272 4273 if (ring->is_mes_queue) { 4274 wptr_saved = (uint32_t *)(ring->mqd_ptr + mqd_size); 4275 is_queue_unmap = (uint32_t *)(ring->mqd_ptr + mqd_size + 4276 sizeof(uint32_t)); 4277 aggregated_db_index = 4278 amdgpu_mes_get_aggregated_doorbell_index(adev, 4279 ring->hw_prio); 4280 4281 wptr_tmp = ring->wptr & ring->buf_mask; 4282 atomic64_set((atomic64_t *)ring->wptr_cpu_addr, wptr_tmp); 4283 *wptr_saved = wptr_tmp; 4284 /* assume doorbell always used by mes mapped queue */ 4285 if (*is_queue_unmap) { 4286 WDOORBELL64(aggregated_db_index, wptr_tmp); 4287 WDOORBELL64(ring->doorbell_index, wptr_tmp); 4288 } else { 4289 WDOORBELL64(ring->doorbell_index, wptr_tmp); 4290 4291 if (*is_queue_unmap) 4292 WDOORBELL64(aggregated_db_index, wptr_tmp); 4293 } 4294 } else { 4295 /* XXX check if swapping is necessary on BE */ 4296 if (ring->use_doorbell) { 4297 atomic64_set((atomic64_t *)ring->wptr_cpu_addr, 4298 ring->wptr); 4299 WDOORBELL64(ring->doorbell_index, ring->wptr); 4300 } else { 4301 BUG(); /* only DOORBELL method supported on gfx12 now */ 4302 } 4303 } 4304 } 4305 4306 static void gfx_v12_0_ring_emit_hdp_flush(struct amdgpu_ring *ring) 4307 { 4308 struct amdgpu_device *adev = ring->adev; 4309 u32 ref_and_mask, reg_mem_engine; 4310 const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg; 4311 4312 if (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE) { 4313 switch (ring->me) { 4314 case 1: 4315 ref_and_mask = nbio_hf_reg->ref_and_mask_cp2 << ring->pipe; 4316 break; 4317 case 2: 4318 ref_and_mask = nbio_hf_reg->ref_and_mask_cp6 << ring->pipe; 4319 break; 4320 default: 4321 return; 4322 } 4323 reg_mem_engine = 0; 4324 } else { 4325 ref_and_mask = nbio_hf_reg->ref_and_mask_cp0; 4326 reg_mem_engine = 1; /* pfp */ 4327 } 4328 4329 gfx_v12_0_wait_reg_mem(ring, reg_mem_engine, 0, 1, 4330 adev->nbio.funcs->get_hdp_flush_req_offset(adev), 4331 adev->nbio.funcs->get_hdp_flush_done_offset(adev), 4332 ref_and_mask, ref_and_mask, 0x20); 4333 } 4334 4335 static void gfx_v12_0_ring_emit_ib_gfx(struct amdgpu_ring *ring, 4336 struct amdgpu_job *job, 4337 struct amdgpu_ib *ib, 4338 uint32_t flags) 4339 { 4340 unsigned vmid = AMDGPU_JOB_GET_VMID(job); 4341 u32 header, control = 0; 4342 4343 BUG_ON(ib->flags & AMDGPU_IB_FLAG_CE); 4344 4345 header = PACKET3(PACKET3_INDIRECT_BUFFER, 2); 4346 4347 control |= ib->length_dw | (vmid << 24); 4348 4349 if (ring->is_mes_queue) 4350 /* inherit vmid from mqd */ 4351 control |= 0x400000; 4352 4353 amdgpu_ring_write(ring, header); 4354 BUG_ON(ib->gpu_addr & 0x3); /* Dword align */ 4355 amdgpu_ring_write(ring, 4356 #ifdef __BIG_ENDIAN 4357 (2 << 0) | 4358 #endif 4359 lower_32_bits(ib->gpu_addr)); 4360 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr)); 4361 amdgpu_ring_write(ring, control); 4362 } 4363 4364 static void gfx_v12_0_ring_emit_ib_compute(struct amdgpu_ring *ring, 4365 struct amdgpu_job *job, 4366 struct amdgpu_ib *ib, 4367 uint32_t flags) 4368 { 4369 unsigned vmid = AMDGPU_JOB_GET_VMID(job); 4370 u32 control = INDIRECT_BUFFER_VALID | ib->length_dw | (vmid << 24); 4371 4372 if (ring->is_mes_queue) 4373 /* inherit vmid from mqd */ 4374 control |= 0x40000000; 4375 4376 amdgpu_ring_write(ring, PACKET3(PACKET3_INDIRECT_BUFFER, 2)); 4377 BUG_ON(ib->gpu_addr & 0x3); /* Dword align */ 4378 amdgpu_ring_write(ring, 4379 #ifdef __BIG_ENDIAN 4380 (2 << 0) | 4381 #endif 4382 lower_32_bits(ib->gpu_addr)); 4383 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr)); 4384 amdgpu_ring_write(ring, control); 4385 } 4386 4387 static void gfx_v12_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, 4388 u64 seq, unsigned flags) 4389 { 4390 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT; 4391 bool int_sel = flags & AMDGPU_FENCE_FLAG_INT; 4392 4393 /* RELEASE_MEM - flush caches, send int */ 4394 amdgpu_ring_write(ring, PACKET3(PACKET3_RELEASE_MEM, 6)); 4395 amdgpu_ring_write(ring, (PACKET3_RELEASE_MEM_GCR_SEQ | 4396 PACKET3_RELEASE_MEM_GCR_GL2_WB | 4397 PACKET3_RELEASE_MEM_CACHE_POLICY(3) | 4398 PACKET3_RELEASE_MEM_EVENT_TYPE(CACHE_FLUSH_AND_INV_TS_EVENT) | 4399 PACKET3_RELEASE_MEM_EVENT_INDEX(5))); 4400 amdgpu_ring_write(ring, (PACKET3_RELEASE_MEM_DATA_SEL(write64bit ? 2 : 1) | 4401 PACKET3_RELEASE_MEM_INT_SEL(int_sel ? 2 : 0))); 4402 4403 /* 4404 * the address should be Qword aligned if 64bit write, Dword 4405 * aligned if only send 32bit data low (discard data high) 4406 */ 4407 if (write64bit) 4408 BUG_ON(addr & 0x7); 4409 else 4410 BUG_ON(addr & 0x3); 4411 amdgpu_ring_write(ring, lower_32_bits(addr)); 4412 amdgpu_ring_write(ring, upper_32_bits(addr)); 4413 amdgpu_ring_write(ring, lower_32_bits(seq)); 4414 amdgpu_ring_write(ring, upper_32_bits(seq)); 4415 amdgpu_ring_write(ring, ring->is_mes_queue ? 4416 (ring->hw_queue_id | AMDGPU_FENCE_MES_QUEUE_FLAG) : 0); 4417 } 4418 4419 static void gfx_v12_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring) 4420 { 4421 int usepfp = (ring->funcs->type == AMDGPU_RING_TYPE_GFX); 4422 uint32_t seq = ring->fence_drv.sync_seq; 4423 uint64_t addr = ring->fence_drv.gpu_addr; 4424 4425 gfx_v12_0_wait_reg_mem(ring, usepfp, 1, 0, lower_32_bits(addr), 4426 upper_32_bits(addr), seq, 0xffffffff, 4); 4427 } 4428 4429 static void gfx_v12_0_ring_invalidate_tlbs(struct amdgpu_ring *ring, 4430 uint16_t pasid, uint32_t flush_type, 4431 bool all_hub, uint8_t dst_sel) 4432 { 4433 amdgpu_ring_write(ring, PACKET3(PACKET3_INVALIDATE_TLBS, 0)); 4434 amdgpu_ring_write(ring, 4435 PACKET3_INVALIDATE_TLBS_DST_SEL(dst_sel) | 4436 PACKET3_INVALIDATE_TLBS_ALL_HUB(all_hub) | 4437 PACKET3_INVALIDATE_TLBS_PASID(pasid) | 4438 PACKET3_INVALIDATE_TLBS_FLUSH_TYPE(flush_type)); 4439 } 4440 4441 static void gfx_v12_0_ring_emit_vm_flush(struct amdgpu_ring *ring, 4442 unsigned vmid, uint64_t pd_addr) 4443 { 4444 if (ring->is_mes_queue) 4445 gfx_v12_0_ring_invalidate_tlbs(ring, 0, 0, false, 0); 4446 else 4447 amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr); 4448 4449 /* compute doesn't have PFP */ 4450 if (ring->funcs->type == AMDGPU_RING_TYPE_GFX) { 4451 /* sync PFP to ME, otherwise we might get invalid PFP reads */ 4452 amdgpu_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0)); 4453 amdgpu_ring_write(ring, 0x0); 4454 } 4455 } 4456 4457 static void gfx_v12_0_ring_emit_fence_kiq(struct amdgpu_ring *ring, u64 addr, 4458 u64 seq, unsigned int flags) 4459 { 4460 struct amdgpu_device *adev = ring->adev; 4461 4462 /* we only allocate 32bit for each seq wb address */ 4463 BUG_ON(flags & AMDGPU_FENCE_FLAG_64BIT); 4464 4465 /* write fence seq to the "addr" */ 4466 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3)); 4467 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) | 4468 WRITE_DATA_DST_SEL(5) | WR_CONFIRM)); 4469 amdgpu_ring_write(ring, lower_32_bits(addr)); 4470 amdgpu_ring_write(ring, upper_32_bits(addr)); 4471 amdgpu_ring_write(ring, lower_32_bits(seq)); 4472 4473 if (flags & AMDGPU_FENCE_FLAG_INT) { 4474 /* set register to trigger INT */ 4475 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3)); 4476 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) | 4477 WRITE_DATA_DST_SEL(0) | WR_CONFIRM)); 4478 amdgpu_ring_write(ring, SOC15_REG_OFFSET(GC, 0, regCPC_INT_STATUS)); 4479 amdgpu_ring_write(ring, 0); 4480 amdgpu_ring_write(ring, 0x20000000); /* src_id is 178 */ 4481 } 4482 } 4483 4484 static void gfx_v12_0_ring_emit_cntxcntl(struct amdgpu_ring *ring, 4485 uint32_t flags) 4486 { 4487 uint32_t dw2 = 0; 4488 4489 dw2 |= 0x80000000; /* set load_enable otherwise this package is just NOPs */ 4490 if (flags & AMDGPU_HAVE_CTX_SWITCH) { 4491 /* set load_global_config & load_global_uconfig */ 4492 dw2 |= 0x8001; 4493 /* set load_cs_sh_regs */ 4494 dw2 |= 0x01000000; 4495 /* set load_per_context_state & load_gfx_sh_regs for GFX */ 4496 dw2 |= 0x10002; 4497 } 4498 4499 amdgpu_ring_write(ring, PACKET3(PACKET3_CONTEXT_CONTROL, 1)); 4500 amdgpu_ring_write(ring, dw2); 4501 amdgpu_ring_write(ring, 0); 4502 } 4503 4504 static unsigned gfx_v12_0_ring_emit_init_cond_exec(struct amdgpu_ring *ring, 4505 uint64_t addr) 4506 { 4507 unsigned ret; 4508 4509 amdgpu_ring_write(ring, PACKET3(PACKET3_COND_EXEC, 3)); 4510 amdgpu_ring_write(ring, lower_32_bits(addr)); 4511 amdgpu_ring_write(ring, upper_32_bits(addr)); 4512 /* discard following DWs if *cond_exec_gpu_addr==0 */ 4513 amdgpu_ring_write(ring, 0); 4514 ret = ring->wptr & ring->buf_mask; 4515 /* patch dummy value later */ 4516 amdgpu_ring_write(ring, 0); 4517 4518 return ret; 4519 } 4520 4521 static int gfx_v12_0_ring_preempt_ib(struct amdgpu_ring *ring) 4522 { 4523 int i, r = 0; 4524 struct amdgpu_device *adev = ring->adev; 4525 struct amdgpu_kiq *kiq = &adev->gfx.kiq[0]; 4526 struct amdgpu_ring *kiq_ring = &kiq->ring; 4527 unsigned long flags; 4528 4529 if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues) 4530 return -EINVAL; 4531 4532 spin_lock_irqsave(&kiq->ring_lock, flags); 4533 4534 if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size)) { 4535 spin_unlock_irqrestore(&kiq->ring_lock, flags); 4536 return -ENOMEM; 4537 } 4538 4539 /* assert preemption condition */ 4540 amdgpu_ring_set_preempt_cond_exec(ring, false); 4541 4542 /* assert IB preemption, emit the trailing fence */ 4543 kiq->pmf->kiq_unmap_queues(kiq_ring, ring, PREEMPT_QUEUES_NO_UNMAP, 4544 ring->trail_fence_gpu_addr, 4545 ++ring->trail_seq); 4546 amdgpu_ring_commit(kiq_ring); 4547 4548 spin_unlock_irqrestore(&kiq->ring_lock, flags); 4549 4550 /* poll the trailing fence */ 4551 for (i = 0; i < adev->usec_timeout; i++) { 4552 if (ring->trail_seq == 4553 le32_to_cpu(*(ring->trail_fence_cpu_addr))) 4554 break; 4555 udelay(1); 4556 } 4557 4558 if (i >= adev->usec_timeout) { 4559 r = -EINVAL; 4560 DRM_ERROR("ring %d failed to preempt ib\n", ring->idx); 4561 } 4562 4563 /* deassert preemption condition */ 4564 amdgpu_ring_set_preempt_cond_exec(ring, true); 4565 return r; 4566 } 4567 4568 static void gfx_v12_0_ring_emit_frame_cntl(struct amdgpu_ring *ring, 4569 bool start, 4570 bool secure) 4571 { 4572 uint32_t v = secure ? FRAME_TMZ : 0; 4573 4574 amdgpu_ring_write(ring, PACKET3(PACKET3_FRAME_CONTROL, 0)); 4575 amdgpu_ring_write(ring, v | FRAME_CMD(start ? 0 : 1)); 4576 } 4577 4578 static void gfx_v12_0_ring_emit_rreg(struct amdgpu_ring *ring, uint32_t reg, 4579 uint32_t reg_val_offs) 4580 { 4581 struct amdgpu_device *adev = ring->adev; 4582 4583 amdgpu_ring_write(ring, PACKET3(PACKET3_COPY_DATA, 4)); 4584 amdgpu_ring_write(ring, 0 | /* src: register*/ 4585 (5 << 8) | /* dst: memory */ 4586 (1 << 20)); /* write confirm */ 4587 amdgpu_ring_write(ring, reg); 4588 amdgpu_ring_write(ring, 0); 4589 amdgpu_ring_write(ring, lower_32_bits(adev->wb.gpu_addr + 4590 reg_val_offs * 4)); 4591 amdgpu_ring_write(ring, upper_32_bits(adev->wb.gpu_addr + 4592 reg_val_offs * 4)); 4593 } 4594 4595 static void gfx_v12_0_ring_emit_wreg(struct amdgpu_ring *ring, 4596 uint32_t reg, 4597 uint32_t val) 4598 { 4599 uint32_t cmd = 0; 4600 4601 switch (ring->funcs->type) { 4602 case AMDGPU_RING_TYPE_GFX: 4603 cmd = WRITE_DATA_ENGINE_SEL(1) | WR_CONFIRM; 4604 break; 4605 case AMDGPU_RING_TYPE_KIQ: 4606 cmd = (1 << 16); /* no inc addr */ 4607 break; 4608 default: 4609 cmd = WR_CONFIRM; 4610 break; 4611 } 4612 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3)); 4613 amdgpu_ring_write(ring, cmd); 4614 amdgpu_ring_write(ring, reg); 4615 amdgpu_ring_write(ring, 0); 4616 amdgpu_ring_write(ring, val); 4617 } 4618 4619 static void gfx_v12_0_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg, 4620 uint32_t val, uint32_t mask) 4621 { 4622 gfx_v12_0_wait_reg_mem(ring, 0, 0, 0, reg, 0, val, mask, 0x20); 4623 } 4624 4625 static void gfx_v12_0_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring, 4626 uint32_t reg0, uint32_t reg1, 4627 uint32_t ref, uint32_t mask) 4628 { 4629 int usepfp = (ring->funcs->type == AMDGPU_RING_TYPE_GFX); 4630 4631 gfx_v12_0_wait_reg_mem(ring, usepfp, 0, 1, reg0, reg1, 4632 ref, mask, 0x20); 4633 } 4634 4635 static void gfx_v12_0_ring_soft_recovery(struct amdgpu_ring *ring, 4636 unsigned vmid) 4637 { 4638 struct amdgpu_device *adev = ring->adev; 4639 uint32_t value = 0; 4640 4641 value = REG_SET_FIELD(value, SQ_CMD, CMD, 0x03); 4642 value = REG_SET_FIELD(value, SQ_CMD, MODE, 0x01); 4643 value = REG_SET_FIELD(value, SQ_CMD, CHECK_VMID, 1); 4644 value = REG_SET_FIELD(value, SQ_CMD, VM_ID, vmid); 4645 WREG32_SOC15(GC, 0, regSQ_CMD, value); 4646 } 4647 4648 static void 4649 gfx_v12_0_set_gfx_eop_interrupt_state(struct amdgpu_device *adev, 4650 uint32_t me, uint32_t pipe, 4651 enum amdgpu_interrupt_state state) 4652 { 4653 uint32_t cp_int_cntl, cp_int_cntl_reg; 4654 4655 if (!me) { 4656 switch (pipe) { 4657 case 0: 4658 cp_int_cntl_reg = SOC15_REG_OFFSET(GC, 0, regCP_INT_CNTL_RING0); 4659 break; 4660 default: 4661 DRM_DEBUG("invalid pipe %d\n", pipe); 4662 return; 4663 } 4664 } else { 4665 DRM_DEBUG("invalid me %d\n", me); 4666 return; 4667 } 4668 4669 switch (state) { 4670 case AMDGPU_IRQ_STATE_DISABLE: 4671 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg); 4672 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0, 4673 TIME_STAMP_INT_ENABLE, 0); 4674 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0, 4675 GENERIC0_INT_ENABLE, 0); 4676 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl); 4677 break; 4678 case AMDGPU_IRQ_STATE_ENABLE: 4679 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg); 4680 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0, 4681 TIME_STAMP_INT_ENABLE, 1); 4682 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0, 4683 GENERIC0_INT_ENABLE, 1); 4684 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl); 4685 break; 4686 default: 4687 break; 4688 } 4689 } 4690 4691 static void gfx_v12_0_set_compute_eop_interrupt_state(struct amdgpu_device *adev, 4692 int me, int pipe, 4693 enum amdgpu_interrupt_state state) 4694 { 4695 u32 mec_int_cntl, mec_int_cntl_reg; 4696 4697 /* 4698 * amdgpu controls only the first MEC. That's why this function only 4699 * handles the setting of interrupts for this specific MEC. All other 4700 * pipes' interrupts are set by amdkfd. 4701 */ 4702 4703 if (me == 1) { 4704 switch (pipe) { 4705 case 0: 4706 mec_int_cntl_reg = SOC15_REG_OFFSET(GC, 0, regCP_ME1_PIPE0_INT_CNTL); 4707 break; 4708 case 1: 4709 mec_int_cntl_reg = SOC15_REG_OFFSET(GC, 0, regCP_ME1_PIPE1_INT_CNTL); 4710 break; 4711 default: 4712 DRM_DEBUG("invalid pipe %d\n", pipe); 4713 return; 4714 } 4715 } else { 4716 DRM_DEBUG("invalid me %d\n", me); 4717 return; 4718 } 4719 4720 switch (state) { 4721 case AMDGPU_IRQ_STATE_DISABLE: 4722 mec_int_cntl = RREG32_SOC15_IP(GC, mec_int_cntl_reg); 4723 mec_int_cntl = REG_SET_FIELD(mec_int_cntl, CP_ME1_PIPE0_INT_CNTL, 4724 TIME_STAMP_INT_ENABLE, 0); 4725 mec_int_cntl = REG_SET_FIELD(mec_int_cntl, CP_ME1_PIPE0_INT_CNTL, 4726 GENERIC0_INT_ENABLE, 0); 4727 WREG32_SOC15_IP(GC, mec_int_cntl_reg, mec_int_cntl); 4728 break; 4729 case AMDGPU_IRQ_STATE_ENABLE: 4730 mec_int_cntl = RREG32_SOC15_IP(GC, mec_int_cntl_reg); 4731 mec_int_cntl = REG_SET_FIELD(mec_int_cntl, CP_ME1_PIPE0_INT_CNTL, 4732 TIME_STAMP_INT_ENABLE, 1); 4733 mec_int_cntl = REG_SET_FIELD(mec_int_cntl, CP_ME1_PIPE0_INT_CNTL, 4734 GENERIC0_INT_ENABLE, 1); 4735 WREG32_SOC15_IP(GC, mec_int_cntl_reg, mec_int_cntl); 4736 break; 4737 default: 4738 break; 4739 } 4740 } 4741 4742 static int gfx_v12_0_set_eop_interrupt_state(struct amdgpu_device *adev, 4743 struct amdgpu_irq_src *src, 4744 unsigned type, 4745 enum amdgpu_interrupt_state state) 4746 { 4747 switch (type) { 4748 case AMDGPU_CP_IRQ_GFX_ME0_PIPE0_EOP: 4749 gfx_v12_0_set_gfx_eop_interrupt_state(adev, 0, 0, state); 4750 break; 4751 case AMDGPU_CP_IRQ_GFX_ME0_PIPE1_EOP: 4752 gfx_v12_0_set_gfx_eop_interrupt_state(adev, 0, 1, state); 4753 break; 4754 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE0_EOP: 4755 gfx_v12_0_set_compute_eop_interrupt_state(adev, 1, 0, state); 4756 break; 4757 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE1_EOP: 4758 gfx_v12_0_set_compute_eop_interrupt_state(adev, 1, 1, state); 4759 break; 4760 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE2_EOP: 4761 gfx_v12_0_set_compute_eop_interrupt_state(adev, 1, 2, state); 4762 break; 4763 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE3_EOP: 4764 gfx_v12_0_set_compute_eop_interrupt_state(adev, 1, 3, state); 4765 break; 4766 default: 4767 break; 4768 } 4769 return 0; 4770 } 4771 4772 static int gfx_v12_0_eop_irq(struct amdgpu_device *adev, 4773 struct amdgpu_irq_src *source, 4774 struct amdgpu_iv_entry *entry) 4775 { 4776 int i; 4777 u8 me_id, pipe_id, queue_id; 4778 struct amdgpu_ring *ring; 4779 uint32_t mes_queue_id = entry->src_data[0]; 4780 4781 DRM_DEBUG("IH: CP EOP\n"); 4782 4783 if (adev->enable_mes && (mes_queue_id & AMDGPU_FENCE_MES_QUEUE_FLAG)) { 4784 struct amdgpu_mes_queue *queue; 4785 4786 mes_queue_id &= AMDGPU_FENCE_MES_QUEUE_ID_MASK; 4787 4788 spin_lock(&adev->mes.queue_id_lock); 4789 queue = idr_find(&adev->mes.queue_id_idr, mes_queue_id); 4790 if (queue) { 4791 DRM_DEBUG("process mes queue id = %d\n", mes_queue_id); 4792 amdgpu_fence_process(queue->ring); 4793 } 4794 spin_unlock(&adev->mes.queue_id_lock); 4795 } else { 4796 me_id = (entry->ring_id & 0x0c) >> 2; 4797 pipe_id = (entry->ring_id & 0x03) >> 0; 4798 queue_id = (entry->ring_id & 0x70) >> 4; 4799 4800 switch (me_id) { 4801 case 0: 4802 if (pipe_id == 0) 4803 amdgpu_fence_process(&adev->gfx.gfx_ring[0]); 4804 else 4805 amdgpu_fence_process(&adev->gfx.gfx_ring[1]); 4806 break; 4807 case 1: 4808 case 2: 4809 for (i = 0; i < adev->gfx.num_compute_rings; i++) { 4810 ring = &adev->gfx.compute_ring[i]; 4811 /* Per-queue interrupt is supported for MEC starting from VI. 4812 * The interrupt can only be enabled/disabled per pipe instead 4813 * of per queue. 4814 */ 4815 if ((ring->me == me_id) && 4816 (ring->pipe == pipe_id) && 4817 (ring->queue == queue_id)) 4818 amdgpu_fence_process(ring); 4819 } 4820 break; 4821 } 4822 } 4823 4824 return 0; 4825 } 4826 4827 static int gfx_v12_0_set_priv_reg_fault_state(struct amdgpu_device *adev, 4828 struct amdgpu_irq_src *source, 4829 unsigned int type, 4830 enum amdgpu_interrupt_state state) 4831 { 4832 u32 cp_int_cntl_reg, cp_int_cntl; 4833 int i, j; 4834 4835 switch (state) { 4836 case AMDGPU_IRQ_STATE_DISABLE: 4837 case AMDGPU_IRQ_STATE_ENABLE: 4838 for (i = 0; i < adev->gfx.me.num_me; i++) { 4839 for (j = 0; j < adev->gfx.me.num_pipe_per_me; j++) { 4840 cp_int_cntl_reg = gfx_v12_0_get_cpg_int_cntl(adev, i, j); 4841 4842 if (cp_int_cntl_reg) { 4843 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg); 4844 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0, 4845 PRIV_REG_INT_ENABLE, 4846 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0); 4847 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl); 4848 } 4849 } 4850 } 4851 for (i = 0; i < adev->gfx.mec.num_mec; i++) { 4852 for (j = 0; j < adev->gfx.mec.num_pipe_per_mec; j++) { 4853 /* MECs start at 1 */ 4854 cp_int_cntl_reg = gfx_v12_0_get_cpc_int_cntl(adev, i + 1, j); 4855 4856 if (cp_int_cntl_reg) { 4857 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg); 4858 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_ME1_PIPE0_INT_CNTL, 4859 PRIV_REG_INT_ENABLE, 4860 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0); 4861 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl); 4862 } 4863 } 4864 } 4865 break; 4866 default: 4867 break; 4868 } 4869 4870 return 0; 4871 } 4872 4873 static int gfx_v12_0_set_bad_op_fault_state(struct amdgpu_device *adev, 4874 struct amdgpu_irq_src *source, 4875 unsigned type, 4876 enum amdgpu_interrupt_state state) 4877 { 4878 u32 cp_int_cntl_reg, cp_int_cntl; 4879 int i, j; 4880 4881 switch (state) { 4882 case AMDGPU_IRQ_STATE_DISABLE: 4883 case AMDGPU_IRQ_STATE_ENABLE: 4884 for (i = 0; i < adev->gfx.me.num_me; i++) { 4885 for (j = 0; j < adev->gfx.me.num_pipe_per_me; j++) { 4886 cp_int_cntl_reg = gfx_v12_0_get_cpg_int_cntl(adev, i, j); 4887 4888 if (cp_int_cntl_reg) { 4889 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg); 4890 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0, 4891 OPCODE_ERROR_INT_ENABLE, 4892 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0); 4893 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl); 4894 } 4895 } 4896 } 4897 for (i = 0; i < adev->gfx.mec.num_mec; i++) { 4898 for (j = 0; j < adev->gfx.mec.num_pipe_per_mec; j++) { 4899 /* MECs start at 1 */ 4900 cp_int_cntl_reg = gfx_v12_0_get_cpc_int_cntl(adev, i + 1, j); 4901 4902 if (cp_int_cntl_reg) { 4903 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg); 4904 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_ME1_PIPE0_INT_CNTL, 4905 OPCODE_ERROR_INT_ENABLE, 4906 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0); 4907 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl); 4908 } 4909 } 4910 } 4911 break; 4912 default: 4913 break; 4914 } 4915 return 0; 4916 } 4917 4918 static int gfx_v12_0_set_priv_inst_fault_state(struct amdgpu_device *adev, 4919 struct amdgpu_irq_src *source, 4920 unsigned int type, 4921 enum amdgpu_interrupt_state state) 4922 { 4923 u32 cp_int_cntl_reg, cp_int_cntl; 4924 int i, j; 4925 4926 switch (state) { 4927 case AMDGPU_IRQ_STATE_DISABLE: 4928 case AMDGPU_IRQ_STATE_ENABLE: 4929 for (i = 0; i < adev->gfx.me.num_me; i++) { 4930 for (j = 0; j < adev->gfx.me.num_pipe_per_me; j++) { 4931 cp_int_cntl_reg = gfx_v12_0_get_cpg_int_cntl(adev, i, j); 4932 4933 if (cp_int_cntl_reg) { 4934 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg); 4935 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0, 4936 PRIV_INSTR_INT_ENABLE, 4937 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0); 4938 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl); 4939 } 4940 } 4941 } 4942 break; 4943 default: 4944 break; 4945 } 4946 4947 return 0; 4948 } 4949 4950 static void gfx_v12_0_handle_priv_fault(struct amdgpu_device *adev, 4951 struct amdgpu_iv_entry *entry) 4952 { 4953 u8 me_id, pipe_id, queue_id; 4954 struct amdgpu_ring *ring; 4955 int i; 4956 4957 me_id = (entry->ring_id & 0x0c) >> 2; 4958 pipe_id = (entry->ring_id & 0x03) >> 0; 4959 queue_id = (entry->ring_id & 0x70) >> 4; 4960 4961 switch (me_id) { 4962 case 0: 4963 for (i = 0; i < adev->gfx.num_gfx_rings; i++) { 4964 ring = &adev->gfx.gfx_ring[i]; 4965 if (ring->me == me_id && ring->pipe == pipe_id && 4966 ring->queue == queue_id) 4967 drm_sched_fault(&ring->sched); 4968 } 4969 break; 4970 case 1: 4971 case 2: 4972 for (i = 0; i < adev->gfx.num_compute_rings; i++) { 4973 ring = &adev->gfx.compute_ring[i]; 4974 if (ring->me == me_id && ring->pipe == pipe_id && 4975 ring->queue == queue_id) 4976 drm_sched_fault(&ring->sched); 4977 } 4978 break; 4979 default: 4980 BUG(); 4981 break; 4982 } 4983 } 4984 4985 static int gfx_v12_0_priv_reg_irq(struct amdgpu_device *adev, 4986 struct amdgpu_irq_src *source, 4987 struct amdgpu_iv_entry *entry) 4988 { 4989 DRM_ERROR("Illegal register access in command stream\n"); 4990 gfx_v12_0_handle_priv_fault(adev, entry); 4991 return 0; 4992 } 4993 4994 static int gfx_v12_0_bad_op_irq(struct amdgpu_device *adev, 4995 struct amdgpu_irq_src *source, 4996 struct amdgpu_iv_entry *entry) 4997 { 4998 DRM_ERROR("Illegal opcode in command stream \n"); 4999 gfx_v12_0_handle_priv_fault(adev, entry); 5000 return 0; 5001 } 5002 5003 static int gfx_v12_0_priv_inst_irq(struct amdgpu_device *adev, 5004 struct amdgpu_irq_src *source, 5005 struct amdgpu_iv_entry *entry) 5006 { 5007 DRM_ERROR("Illegal instruction in command stream\n"); 5008 gfx_v12_0_handle_priv_fault(adev, entry); 5009 return 0; 5010 } 5011 5012 static void gfx_v12_0_emit_mem_sync(struct amdgpu_ring *ring) 5013 { 5014 const unsigned int gcr_cntl = 5015 PACKET3_ACQUIRE_MEM_GCR_CNTL_GL2_INV(1) | 5016 PACKET3_ACQUIRE_MEM_GCR_CNTL_GL2_WB(1) | 5017 PACKET3_ACQUIRE_MEM_GCR_CNTL_GLM_INV(1) | 5018 PACKET3_ACQUIRE_MEM_GCR_CNTL_GLM_WB(1) | 5019 PACKET3_ACQUIRE_MEM_GCR_CNTL_GL1_INV(1) | 5020 PACKET3_ACQUIRE_MEM_GCR_CNTL_GLV_INV(1) | 5021 PACKET3_ACQUIRE_MEM_GCR_CNTL_GLK_INV(1) | 5022 PACKET3_ACQUIRE_MEM_GCR_CNTL_GLI_INV(1); 5023 5024 /* ACQUIRE_MEM - make one or more surfaces valid for use by the subsequent operations */ 5025 amdgpu_ring_write(ring, PACKET3(PACKET3_ACQUIRE_MEM, 6)); 5026 amdgpu_ring_write(ring, 0); /* CP_COHER_CNTL */ 5027 amdgpu_ring_write(ring, 0xffffffff); /* CP_COHER_SIZE */ 5028 amdgpu_ring_write(ring, 0xffffff); /* CP_COHER_SIZE_HI */ 5029 amdgpu_ring_write(ring, 0); /* CP_COHER_BASE */ 5030 amdgpu_ring_write(ring, 0); /* CP_COHER_BASE_HI */ 5031 amdgpu_ring_write(ring, 0x0000000A); /* POLL_INTERVAL */ 5032 amdgpu_ring_write(ring, gcr_cntl); /* GCR_CNTL */ 5033 } 5034 5035 static void gfx_v12_ring_insert_nop(struct amdgpu_ring *ring, uint32_t num_nop) 5036 { 5037 int i; 5038 5039 /* Header itself is a NOP packet */ 5040 if (num_nop == 1) { 5041 amdgpu_ring_write(ring, ring->funcs->nop); 5042 return; 5043 } 5044 5045 /* Max HW optimization till 0x3ffe, followed by remaining one NOP at a time*/ 5046 amdgpu_ring_write(ring, PACKET3(PACKET3_NOP, min(num_nop - 2, 0x3ffe))); 5047 5048 /* Header is at index 0, followed by num_nops - 1 NOP packet's */ 5049 for (i = 1; i < num_nop; i++) 5050 amdgpu_ring_write(ring, ring->funcs->nop); 5051 } 5052 5053 static void gfx_v12_ip_print(void *handle, struct drm_printer *p) 5054 { 5055 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 5056 uint32_t i, j, k, reg, index = 0; 5057 uint32_t reg_count = ARRAY_SIZE(gc_reg_list_12_0); 5058 5059 if (!adev->gfx.ip_dump_core) 5060 return; 5061 5062 for (i = 0; i < reg_count; i++) 5063 drm_printf(p, "%-50s \t 0x%08x\n", 5064 gc_reg_list_12_0[i].reg_name, 5065 adev->gfx.ip_dump_core[i]); 5066 5067 /* print compute queue registers for all instances */ 5068 if (!adev->gfx.ip_dump_compute_queues) 5069 return; 5070 5071 reg_count = ARRAY_SIZE(gc_cp_reg_list_12); 5072 drm_printf(p, "\nnum_mec: %d num_pipe: %d num_queue: %d\n", 5073 adev->gfx.mec.num_mec, 5074 adev->gfx.mec.num_pipe_per_mec, 5075 adev->gfx.mec.num_queue_per_pipe); 5076 5077 for (i = 0; i < adev->gfx.mec.num_mec; i++) { 5078 for (j = 0; j < adev->gfx.mec.num_pipe_per_mec; j++) { 5079 for (k = 0; k < adev->gfx.mec.num_queue_per_pipe; k++) { 5080 drm_printf(p, "\nmec %d, pipe %d, queue %d\n", i, j, k); 5081 for (reg = 0; reg < reg_count; reg++) { 5082 drm_printf(p, "%-50s \t 0x%08x\n", 5083 gc_cp_reg_list_12[reg].reg_name, 5084 adev->gfx.ip_dump_compute_queues[index + reg]); 5085 } 5086 index += reg_count; 5087 } 5088 } 5089 } 5090 5091 /* print gfx queue registers for all instances */ 5092 if (!adev->gfx.ip_dump_gfx_queues) 5093 return; 5094 5095 index = 0; 5096 reg_count = ARRAY_SIZE(gc_gfx_queue_reg_list_12); 5097 drm_printf(p, "\nnum_me: %d num_pipe: %d num_queue: %d\n", 5098 adev->gfx.me.num_me, 5099 adev->gfx.me.num_pipe_per_me, 5100 adev->gfx.me.num_queue_per_pipe); 5101 5102 for (i = 0; i < adev->gfx.me.num_me; i++) { 5103 for (j = 0; j < adev->gfx.me.num_pipe_per_me; j++) { 5104 for (k = 0; k < adev->gfx.me.num_queue_per_pipe; k++) { 5105 drm_printf(p, "\nme %d, pipe %d, queue %d\n", i, j, k); 5106 for (reg = 0; reg < reg_count; reg++) { 5107 drm_printf(p, "%-50s \t 0x%08x\n", 5108 gc_gfx_queue_reg_list_12[reg].reg_name, 5109 adev->gfx.ip_dump_gfx_queues[index + reg]); 5110 } 5111 index += reg_count; 5112 } 5113 } 5114 } 5115 } 5116 5117 static void gfx_v12_ip_dump(void *handle) 5118 { 5119 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 5120 uint32_t i, j, k, reg, index = 0; 5121 uint32_t reg_count = ARRAY_SIZE(gc_reg_list_12_0); 5122 5123 if (!adev->gfx.ip_dump_core) 5124 return; 5125 5126 amdgpu_gfx_off_ctrl(adev, false); 5127 for (i = 0; i < reg_count; i++) 5128 adev->gfx.ip_dump_core[i] = RREG32(SOC15_REG_ENTRY_OFFSET(gc_reg_list_12_0[i])); 5129 amdgpu_gfx_off_ctrl(adev, true); 5130 5131 /* dump compute queue registers for all instances */ 5132 if (!adev->gfx.ip_dump_compute_queues) 5133 return; 5134 5135 reg_count = ARRAY_SIZE(gc_cp_reg_list_12); 5136 amdgpu_gfx_off_ctrl(adev, false); 5137 mutex_lock(&adev->srbm_mutex); 5138 for (i = 0; i < adev->gfx.mec.num_mec; i++) { 5139 for (j = 0; j < adev->gfx.mec.num_pipe_per_mec; j++) { 5140 for (k = 0; k < adev->gfx.mec.num_queue_per_pipe; k++) { 5141 /* ME0 is for GFX so start from 1 for CP */ 5142 soc24_grbm_select(adev, adev->gfx.me.num_me + i, j, k, 0); 5143 for (reg = 0; reg < reg_count; reg++) { 5144 adev->gfx.ip_dump_compute_queues[index + reg] = 5145 RREG32(SOC15_REG_ENTRY_OFFSET( 5146 gc_cp_reg_list_12[reg])); 5147 } 5148 index += reg_count; 5149 } 5150 } 5151 } 5152 soc24_grbm_select(adev, 0, 0, 0, 0); 5153 mutex_unlock(&adev->srbm_mutex); 5154 amdgpu_gfx_off_ctrl(adev, true); 5155 5156 /* dump gfx queue registers for all instances */ 5157 if (!adev->gfx.ip_dump_gfx_queues) 5158 return; 5159 5160 index = 0; 5161 reg_count = ARRAY_SIZE(gc_gfx_queue_reg_list_12); 5162 amdgpu_gfx_off_ctrl(adev, false); 5163 mutex_lock(&adev->srbm_mutex); 5164 for (i = 0; i < adev->gfx.me.num_me; i++) { 5165 for (j = 0; j < adev->gfx.me.num_pipe_per_me; j++) { 5166 for (k = 0; k < adev->gfx.me.num_queue_per_pipe; k++) { 5167 soc24_grbm_select(adev, i, j, k, 0); 5168 5169 for (reg = 0; reg < reg_count; reg++) { 5170 adev->gfx.ip_dump_gfx_queues[index + reg] = 5171 RREG32(SOC15_REG_ENTRY_OFFSET( 5172 gc_gfx_queue_reg_list_12[reg])); 5173 } 5174 index += reg_count; 5175 } 5176 } 5177 } 5178 soc24_grbm_select(adev, 0, 0, 0, 0); 5179 mutex_unlock(&adev->srbm_mutex); 5180 amdgpu_gfx_off_ctrl(adev, true); 5181 } 5182 5183 static const struct amd_ip_funcs gfx_v12_0_ip_funcs = { 5184 .name = "gfx_v12_0", 5185 .early_init = gfx_v12_0_early_init, 5186 .late_init = gfx_v12_0_late_init, 5187 .sw_init = gfx_v12_0_sw_init, 5188 .sw_fini = gfx_v12_0_sw_fini, 5189 .hw_init = gfx_v12_0_hw_init, 5190 .hw_fini = gfx_v12_0_hw_fini, 5191 .suspend = gfx_v12_0_suspend, 5192 .resume = gfx_v12_0_resume, 5193 .is_idle = gfx_v12_0_is_idle, 5194 .wait_for_idle = gfx_v12_0_wait_for_idle, 5195 .set_clockgating_state = gfx_v12_0_set_clockgating_state, 5196 .set_powergating_state = gfx_v12_0_set_powergating_state, 5197 .get_clockgating_state = gfx_v12_0_get_clockgating_state, 5198 .dump_ip_state = gfx_v12_ip_dump, 5199 .print_ip_state = gfx_v12_ip_print, 5200 }; 5201 5202 static const struct amdgpu_ring_funcs gfx_v12_0_ring_funcs_gfx = { 5203 .type = AMDGPU_RING_TYPE_GFX, 5204 .align_mask = 0xff, 5205 .nop = PACKET3(PACKET3_NOP, 0x3FFF), 5206 .support_64bit_ptrs = true, 5207 .secure_submission_supported = true, 5208 .get_rptr = gfx_v12_0_ring_get_rptr_gfx, 5209 .get_wptr = gfx_v12_0_ring_get_wptr_gfx, 5210 .set_wptr = gfx_v12_0_ring_set_wptr_gfx, 5211 .emit_frame_size = /* totally 242 maximum if 16 IBs */ 5212 5 + /* COND_EXEC */ 5213 7 + /* PIPELINE_SYNC */ 5214 SOC15_FLUSH_GPU_TLB_NUM_WREG * 5 + 5215 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 7 + 5216 2 + /* VM_FLUSH */ 5217 8 + /* FENCE for VM_FLUSH */ 5218 5 + /* COND_EXEC */ 5219 7 + /* HDP_flush */ 5220 4 + /* VGT_flush */ 5221 31 + /* DE_META */ 5222 3 + /* CNTX_CTRL */ 5223 5 + /* HDP_INVL */ 5224 8 + 8 + /* FENCE x2 */ 5225 8, /* gfx_v12_0_emit_mem_sync */ 5226 .emit_ib_size = 4, /* gfx_v12_0_ring_emit_ib_gfx */ 5227 .emit_ib = gfx_v12_0_ring_emit_ib_gfx, 5228 .emit_fence = gfx_v12_0_ring_emit_fence, 5229 .emit_pipeline_sync = gfx_v12_0_ring_emit_pipeline_sync, 5230 .emit_vm_flush = gfx_v12_0_ring_emit_vm_flush, 5231 .emit_hdp_flush = gfx_v12_0_ring_emit_hdp_flush, 5232 .test_ring = gfx_v12_0_ring_test_ring, 5233 .test_ib = gfx_v12_0_ring_test_ib, 5234 .insert_nop = gfx_v12_ring_insert_nop, 5235 .pad_ib = amdgpu_ring_generic_pad_ib, 5236 .emit_cntxcntl = gfx_v12_0_ring_emit_cntxcntl, 5237 .init_cond_exec = gfx_v12_0_ring_emit_init_cond_exec, 5238 .preempt_ib = gfx_v12_0_ring_preempt_ib, 5239 .emit_frame_cntl = gfx_v12_0_ring_emit_frame_cntl, 5240 .emit_wreg = gfx_v12_0_ring_emit_wreg, 5241 .emit_reg_wait = gfx_v12_0_ring_emit_reg_wait, 5242 .emit_reg_write_reg_wait = gfx_v12_0_ring_emit_reg_write_reg_wait, 5243 .soft_recovery = gfx_v12_0_ring_soft_recovery, 5244 .emit_mem_sync = gfx_v12_0_emit_mem_sync, 5245 }; 5246 5247 static const struct amdgpu_ring_funcs gfx_v12_0_ring_funcs_compute = { 5248 .type = AMDGPU_RING_TYPE_COMPUTE, 5249 .align_mask = 0xff, 5250 .nop = PACKET3(PACKET3_NOP, 0x3FFF), 5251 .support_64bit_ptrs = true, 5252 .get_rptr = gfx_v12_0_ring_get_rptr_compute, 5253 .get_wptr = gfx_v12_0_ring_get_wptr_compute, 5254 .set_wptr = gfx_v12_0_ring_set_wptr_compute, 5255 .emit_frame_size = 5256 7 + /* gfx_v12_0_ring_emit_hdp_flush */ 5257 5 + /* hdp invalidate */ 5258 7 + /* gfx_v12_0_ring_emit_pipeline_sync */ 5259 SOC15_FLUSH_GPU_TLB_NUM_WREG * 5 + 5260 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 7 + 5261 2 + /* gfx_v12_0_ring_emit_vm_flush */ 5262 8 + 8 + 8 + /* gfx_v12_0_ring_emit_fence x3 for user fence, vm fence */ 5263 8, /* gfx_v12_0_emit_mem_sync */ 5264 .emit_ib_size = 7, /* gfx_v12_0_ring_emit_ib_compute */ 5265 .emit_ib = gfx_v12_0_ring_emit_ib_compute, 5266 .emit_fence = gfx_v12_0_ring_emit_fence, 5267 .emit_pipeline_sync = gfx_v12_0_ring_emit_pipeline_sync, 5268 .emit_vm_flush = gfx_v12_0_ring_emit_vm_flush, 5269 .emit_hdp_flush = gfx_v12_0_ring_emit_hdp_flush, 5270 .test_ring = gfx_v12_0_ring_test_ring, 5271 .test_ib = gfx_v12_0_ring_test_ib, 5272 .insert_nop = gfx_v12_ring_insert_nop, 5273 .pad_ib = amdgpu_ring_generic_pad_ib, 5274 .emit_wreg = gfx_v12_0_ring_emit_wreg, 5275 .emit_reg_wait = gfx_v12_0_ring_emit_reg_wait, 5276 .emit_reg_write_reg_wait = gfx_v12_0_ring_emit_reg_write_reg_wait, 5277 .soft_recovery = gfx_v12_0_ring_soft_recovery, 5278 .emit_mem_sync = gfx_v12_0_emit_mem_sync, 5279 }; 5280 5281 static const struct amdgpu_ring_funcs gfx_v12_0_ring_funcs_kiq = { 5282 .type = AMDGPU_RING_TYPE_KIQ, 5283 .align_mask = 0xff, 5284 .nop = PACKET3(PACKET3_NOP, 0x3FFF), 5285 .support_64bit_ptrs = true, 5286 .get_rptr = gfx_v12_0_ring_get_rptr_compute, 5287 .get_wptr = gfx_v12_0_ring_get_wptr_compute, 5288 .set_wptr = gfx_v12_0_ring_set_wptr_compute, 5289 .emit_frame_size = 5290 7 + /* gfx_v12_0_ring_emit_hdp_flush */ 5291 5 + /*hdp invalidate */ 5292 7 + /* gfx_v12_0_ring_emit_pipeline_sync */ 5293 SOC15_FLUSH_GPU_TLB_NUM_WREG * 5 + 5294 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 7 + 5295 2 + /* gfx_v12_0_ring_emit_vm_flush */ 5296 8 + 8 + 8, /* gfx_v12_0_ring_emit_fence_kiq x3 for user fence, vm fence */ 5297 .emit_ib_size = 7, /* gfx_v12_0_ring_emit_ib_compute */ 5298 .emit_ib = gfx_v12_0_ring_emit_ib_compute, 5299 .emit_fence = gfx_v12_0_ring_emit_fence_kiq, 5300 .test_ring = gfx_v12_0_ring_test_ring, 5301 .test_ib = gfx_v12_0_ring_test_ib, 5302 .insert_nop = amdgpu_ring_insert_nop, 5303 .pad_ib = amdgpu_ring_generic_pad_ib, 5304 .emit_rreg = gfx_v12_0_ring_emit_rreg, 5305 .emit_wreg = gfx_v12_0_ring_emit_wreg, 5306 .emit_reg_wait = gfx_v12_0_ring_emit_reg_wait, 5307 .emit_reg_write_reg_wait = gfx_v12_0_ring_emit_reg_write_reg_wait, 5308 }; 5309 5310 static void gfx_v12_0_set_ring_funcs(struct amdgpu_device *adev) 5311 { 5312 int i; 5313 5314 adev->gfx.kiq[0].ring.funcs = &gfx_v12_0_ring_funcs_kiq; 5315 5316 for (i = 0; i < adev->gfx.num_gfx_rings; i++) 5317 adev->gfx.gfx_ring[i].funcs = &gfx_v12_0_ring_funcs_gfx; 5318 5319 for (i = 0; i < adev->gfx.num_compute_rings; i++) 5320 adev->gfx.compute_ring[i].funcs = &gfx_v12_0_ring_funcs_compute; 5321 } 5322 5323 static const struct amdgpu_irq_src_funcs gfx_v12_0_eop_irq_funcs = { 5324 .set = gfx_v12_0_set_eop_interrupt_state, 5325 .process = gfx_v12_0_eop_irq, 5326 }; 5327 5328 static const struct amdgpu_irq_src_funcs gfx_v12_0_priv_reg_irq_funcs = { 5329 .set = gfx_v12_0_set_priv_reg_fault_state, 5330 .process = gfx_v12_0_priv_reg_irq, 5331 }; 5332 5333 static const struct amdgpu_irq_src_funcs gfx_v12_0_bad_op_irq_funcs = { 5334 .set = gfx_v12_0_set_bad_op_fault_state, 5335 .process = gfx_v12_0_bad_op_irq, 5336 }; 5337 5338 static const struct amdgpu_irq_src_funcs gfx_v12_0_priv_inst_irq_funcs = { 5339 .set = gfx_v12_0_set_priv_inst_fault_state, 5340 .process = gfx_v12_0_priv_inst_irq, 5341 }; 5342 5343 static void gfx_v12_0_set_irq_funcs(struct amdgpu_device *adev) 5344 { 5345 adev->gfx.eop_irq.num_types = AMDGPU_CP_IRQ_LAST; 5346 adev->gfx.eop_irq.funcs = &gfx_v12_0_eop_irq_funcs; 5347 5348 adev->gfx.priv_reg_irq.num_types = 1; 5349 adev->gfx.priv_reg_irq.funcs = &gfx_v12_0_priv_reg_irq_funcs; 5350 5351 adev->gfx.bad_op_irq.num_types = 1; 5352 adev->gfx.bad_op_irq.funcs = &gfx_v12_0_bad_op_irq_funcs; 5353 5354 adev->gfx.priv_inst_irq.num_types = 1; 5355 adev->gfx.priv_inst_irq.funcs = &gfx_v12_0_priv_inst_irq_funcs; 5356 } 5357 5358 static void gfx_v12_0_set_imu_funcs(struct amdgpu_device *adev) 5359 { 5360 if (adev->flags & AMD_IS_APU) 5361 adev->gfx.imu.mode = MISSION_MODE; 5362 else 5363 adev->gfx.imu.mode = DEBUG_MODE; 5364 5365 adev->gfx.imu.funcs = &gfx_v12_0_imu_funcs; 5366 } 5367 5368 static void gfx_v12_0_set_rlc_funcs(struct amdgpu_device *adev) 5369 { 5370 adev->gfx.rlc.funcs = &gfx_v12_0_rlc_funcs; 5371 } 5372 5373 static void gfx_v12_0_set_mqd_funcs(struct amdgpu_device *adev) 5374 { 5375 /* set gfx eng mqd */ 5376 adev->mqds[AMDGPU_HW_IP_GFX].mqd_size = 5377 sizeof(struct v12_gfx_mqd); 5378 adev->mqds[AMDGPU_HW_IP_GFX].init_mqd = 5379 gfx_v12_0_gfx_mqd_init; 5380 /* set compute eng mqd */ 5381 adev->mqds[AMDGPU_HW_IP_COMPUTE].mqd_size = 5382 sizeof(struct v12_compute_mqd); 5383 adev->mqds[AMDGPU_HW_IP_COMPUTE].init_mqd = 5384 gfx_v12_0_compute_mqd_init; 5385 } 5386 5387 static void gfx_v12_0_set_user_wgp_inactive_bitmap_per_sh(struct amdgpu_device *adev, 5388 u32 bitmap) 5389 { 5390 u32 data; 5391 5392 if (!bitmap) 5393 return; 5394 5395 data = bitmap << GC_USER_SHADER_ARRAY_CONFIG__INACTIVE_WGPS__SHIFT; 5396 data &= GC_USER_SHADER_ARRAY_CONFIG__INACTIVE_WGPS_MASK; 5397 5398 WREG32_SOC15(GC, 0, regGC_USER_SHADER_ARRAY_CONFIG, data); 5399 } 5400 5401 static u32 gfx_v12_0_get_wgp_active_bitmap_per_sh(struct amdgpu_device *adev) 5402 { 5403 u32 data, wgp_bitmask; 5404 data = RREG32_SOC15(GC, 0, regCC_GC_SHADER_ARRAY_CONFIG); 5405 data |= RREG32_SOC15(GC, 0, regGC_USER_SHADER_ARRAY_CONFIG); 5406 5407 data &= CC_GC_SHADER_ARRAY_CONFIG__INACTIVE_WGPS_MASK; 5408 data >>= CC_GC_SHADER_ARRAY_CONFIG__INACTIVE_WGPS__SHIFT; 5409 5410 wgp_bitmask = 5411 amdgpu_gfx_create_bitmask(adev->gfx.config.max_cu_per_sh >> 1); 5412 5413 return (~data) & wgp_bitmask; 5414 } 5415 5416 static u32 gfx_v12_0_get_cu_active_bitmap_per_sh(struct amdgpu_device *adev) 5417 { 5418 u32 wgp_idx, wgp_active_bitmap; 5419 u32 cu_bitmap_per_wgp, cu_active_bitmap; 5420 5421 wgp_active_bitmap = gfx_v12_0_get_wgp_active_bitmap_per_sh(adev); 5422 cu_active_bitmap = 0; 5423 5424 for (wgp_idx = 0; wgp_idx < 16; wgp_idx++) { 5425 /* if there is one WGP enabled, it means 2 CUs will be enabled */ 5426 cu_bitmap_per_wgp = 3 << (2 * wgp_idx); 5427 if (wgp_active_bitmap & (1 << wgp_idx)) 5428 cu_active_bitmap |= cu_bitmap_per_wgp; 5429 } 5430 5431 return cu_active_bitmap; 5432 } 5433 5434 static int gfx_v12_0_get_cu_info(struct amdgpu_device *adev, 5435 struct amdgpu_cu_info *cu_info) 5436 { 5437 int i, j, k, counter, active_cu_number = 0; 5438 u32 mask, bitmap; 5439 unsigned disable_masks[8 * 2]; 5440 5441 if (!adev || !cu_info) 5442 return -EINVAL; 5443 5444 amdgpu_gfx_parse_disable_cu(disable_masks, 8, 2); 5445 5446 mutex_lock(&adev->grbm_idx_mutex); 5447 for (i = 0; i < adev->gfx.config.max_shader_engines; i++) { 5448 for (j = 0; j < adev->gfx.config.max_sh_per_se; j++) { 5449 bitmap = i * adev->gfx.config.max_sh_per_se + j; 5450 if (!((gfx_v12_0_get_sa_active_bitmap(adev) >> bitmap) & 1)) 5451 continue; 5452 mask = 1; 5453 counter = 0; 5454 gfx_v12_0_select_se_sh(adev, i, j, 0xffffffff, 0); 5455 if (i < 8 && j < 2) 5456 gfx_v12_0_set_user_wgp_inactive_bitmap_per_sh( 5457 adev, disable_masks[i * 2 + j]); 5458 bitmap = gfx_v12_0_get_cu_active_bitmap_per_sh(adev); 5459 5460 /** 5461 * GFX12 could support more than 4 SEs, while the bitmap 5462 * in cu_info struct is 4x4 and ioctl interface struct 5463 * drm_amdgpu_info_device should keep stable. 5464 * So we use last two columns of bitmap to store cu mask for 5465 * SEs 4 to 7, the layout of the bitmap is as below: 5466 * SE0: {SH0,SH1} --> {bitmap[0][0], bitmap[0][1]} 5467 * SE1: {SH0,SH1} --> {bitmap[1][0], bitmap[1][1]} 5468 * SE2: {SH0,SH1} --> {bitmap[2][0], bitmap[2][1]} 5469 * SE3: {SH0,SH1} --> {bitmap[3][0], bitmap[3][1]} 5470 * SE4: {SH0,SH1} --> {bitmap[0][2], bitmap[0][3]} 5471 * SE5: {SH0,SH1} --> {bitmap[1][2], bitmap[1][3]} 5472 * SE6: {SH0,SH1} --> {bitmap[2][2], bitmap[2][3]} 5473 * SE7: {SH0,SH1} --> {bitmap[3][2], bitmap[3][3]} 5474 */ 5475 cu_info->bitmap[0][i % 4][j + (i / 4) * 2] = bitmap; 5476 5477 for (k = 0; k < adev->gfx.config.max_cu_per_sh; k++) { 5478 if (bitmap & mask) 5479 counter++; 5480 5481 mask <<= 1; 5482 } 5483 active_cu_number += counter; 5484 } 5485 } 5486 gfx_v12_0_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff, 0); 5487 mutex_unlock(&adev->grbm_idx_mutex); 5488 5489 cu_info->number = active_cu_number; 5490 cu_info->simd_per_cu = NUM_SIMD_PER_CU; 5491 5492 return 0; 5493 } 5494 5495 const struct amdgpu_ip_block_version gfx_v12_0_ip_block = { 5496 .type = AMD_IP_BLOCK_TYPE_GFX, 5497 .major = 12, 5498 .minor = 0, 5499 .rev = 0, 5500 .funcs = &gfx_v12_0_ip_funcs, 5501 }; 5502