1 /* 2 * Copyright 2021 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_v11_0.h" 34 #include "soc21.h" 35 #include "nvd.h" 36 37 #include "gc/gc_11_0_0_offset.h" 38 #include "gc/gc_11_0_0_sh_mask.h" 39 #include "smuio/smuio_13_0_6_offset.h" 40 #include "smuio/smuio_13_0_6_sh_mask.h" 41 #include "navi10_enum.h" 42 #include "ivsrcid/gfx/irqsrcs_gfx_11_0_0.h" 43 44 #include "soc15.h" 45 #include "soc15d.h" 46 #include "clearstate_gfx11.h" 47 #include "v11_structs.h" 48 #include "gfx_v11_0.h" 49 #include "gfx_v11_0_cleaner_shader.h" 50 #include "gfx_v11_0_3.h" 51 #include "nbio_v4_3.h" 52 #include "mes_v11_0.h" 53 54 #define GFX11_NUM_GFX_RINGS 1 55 #define GFX11_MEC_HPD_SIZE 2048 56 57 #define RLCG_UCODE_LOADING_START_ADDRESS 0x00002000L 58 #define RLC_PG_DELAY_3_DEFAULT_GC_11_0_1 0x1388 59 60 #define regCGTT_WD_CLK_CTRL 0x5086 61 #define regCGTT_WD_CLK_CTRL_BASE_IDX 1 62 #define regRLC_RLCS_BOOTLOAD_STATUS_gc_11_0_1 0x4e7e 63 #define regRLC_RLCS_BOOTLOAD_STATUS_gc_11_0_1_BASE_IDX 1 64 #define regPC_CONFIG_CNTL_1 0x194d 65 #define regPC_CONFIG_CNTL_1_BASE_IDX 1 66 67 MODULE_FIRMWARE("amdgpu/gc_11_0_0_pfp.bin"); 68 MODULE_FIRMWARE("amdgpu/gc_11_0_0_me.bin"); 69 MODULE_FIRMWARE("amdgpu/gc_11_0_0_mec.bin"); 70 MODULE_FIRMWARE("amdgpu/gc_11_0_0_rlc.bin"); 71 MODULE_FIRMWARE("amdgpu/gc_11_0_0_rlc_1.bin"); 72 MODULE_FIRMWARE("amdgpu/gc_11_0_0_toc.bin"); 73 MODULE_FIRMWARE("amdgpu/gc_11_0_1_pfp.bin"); 74 MODULE_FIRMWARE("amdgpu/gc_11_0_1_me.bin"); 75 MODULE_FIRMWARE("amdgpu/gc_11_0_1_mec.bin"); 76 MODULE_FIRMWARE("amdgpu/gc_11_0_1_rlc.bin"); 77 MODULE_FIRMWARE("amdgpu/gc_11_0_2_pfp.bin"); 78 MODULE_FIRMWARE("amdgpu/gc_11_0_2_me.bin"); 79 MODULE_FIRMWARE("amdgpu/gc_11_0_2_mec.bin"); 80 MODULE_FIRMWARE("amdgpu/gc_11_0_2_rlc.bin"); 81 MODULE_FIRMWARE("amdgpu/gc_11_0_3_pfp.bin"); 82 MODULE_FIRMWARE("amdgpu/gc_11_0_3_me.bin"); 83 MODULE_FIRMWARE("amdgpu/gc_11_0_3_mec.bin"); 84 MODULE_FIRMWARE("amdgpu/gc_11_0_3_rlc.bin"); 85 MODULE_FIRMWARE("amdgpu/gc_11_0_4_pfp.bin"); 86 MODULE_FIRMWARE("amdgpu/gc_11_0_4_me.bin"); 87 MODULE_FIRMWARE("amdgpu/gc_11_0_4_mec.bin"); 88 MODULE_FIRMWARE("amdgpu/gc_11_0_4_rlc.bin"); 89 MODULE_FIRMWARE("amdgpu/gc_11_5_0_pfp.bin"); 90 MODULE_FIRMWARE("amdgpu/gc_11_5_0_me.bin"); 91 MODULE_FIRMWARE("amdgpu/gc_11_5_0_mec.bin"); 92 MODULE_FIRMWARE("amdgpu/gc_11_5_0_rlc.bin"); 93 MODULE_FIRMWARE("amdgpu/gc_11_5_1_pfp.bin"); 94 MODULE_FIRMWARE("amdgpu/gc_11_5_1_me.bin"); 95 MODULE_FIRMWARE("amdgpu/gc_11_5_1_mec.bin"); 96 MODULE_FIRMWARE("amdgpu/gc_11_5_1_rlc.bin"); 97 MODULE_FIRMWARE("amdgpu/gc_11_5_2_pfp.bin"); 98 MODULE_FIRMWARE("amdgpu/gc_11_5_2_me.bin"); 99 MODULE_FIRMWARE("amdgpu/gc_11_5_2_mec.bin"); 100 MODULE_FIRMWARE("amdgpu/gc_11_5_2_rlc.bin"); 101 102 static const struct amdgpu_hwip_reg_entry gc_reg_list_11_0[] = { 103 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS), 104 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS2), 105 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS3), 106 SOC15_REG_ENTRY_STR(GC, 0, regCP_STALLED_STAT1), 107 SOC15_REG_ENTRY_STR(GC, 0, regCP_STALLED_STAT2), 108 SOC15_REG_ENTRY_STR(GC, 0, regCP_STALLED_STAT3), 109 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPC_STALLED_STAT1), 110 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPF_STALLED_STAT1), 111 SOC15_REG_ENTRY_STR(GC, 0, regCP_BUSY_STAT), 112 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPC_BUSY_STAT), 113 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPF_BUSY_STAT), 114 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPC_BUSY_STAT2), 115 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPF_BUSY_STAT2), 116 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPF_STATUS), 117 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_ERROR), 118 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HPD_STATUS0), 119 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_BASE), 120 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_RPTR), 121 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_WPTR), 122 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB0_BASE), 123 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB0_RPTR), 124 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB0_WPTR), 125 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB1_BASE), 126 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB1_RPTR), 127 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB1_WPTR), 128 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_CMD_BUFSZ), 129 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB2_CMD_BUFSZ), 130 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BASE_LO), 131 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BASE_HI), 132 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BUFSZ), 133 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB2_BASE_LO), 134 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB2_BASE_HI), 135 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB2_BUFSZ), 136 SOC15_REG_ENTRY_STR(GC, 0, regCPF_UTCL1_STATUS), 137 SOC15_REG_ENTRY_STR(GC, 0, regCPC_UTCL1_STATUS), 138 SOC15_REG_ENTRY_STR(GC, 0, regCPG_UTCL1_STATUS), 139 SOC15_REG_ENTRY_STR(GC, 0, regGDS_PROTECTION_FAULT), 140 SOC15_REG_ENTRY_STR(GC, 0, regGDS_VM_PROTECTION_FAULT), 141 SOC15_REG_ENTRY_STR(GC, 0, regIA_UTCL1_STATUS), 142 SOC15_REG_ENTRY_STR(GC, 0, regIA_UTCL1_STATUS_2), 143 SOC15_REG_ENTRY_STR(GC, 0, regPA_CL_CNTL_STATUS), 144 SOC15_REG_ENTRY_STR(GC, 0, regRLC_UTCL1_STATUS), 145 SOC15_REG_ENTRY_STR(GC, 0, regRMI_UTCL1_STATUS), 146 SOC15_REG_ENTRY_STR(GC, 0, regSQC_CACHES), 147 SOC15_REG_ENTRY_STR(GC, 0, regSQG_STATUS), 148 SOC15_REG_ENTRY_STR(GC, 0, regWD_UTCL1_STATUS), 149 SOC15_REG_ENTRY_STR(GC, 0, regGCVM_L2_PROTECTION_FAULT_CNTL), 150 SOC15_REG_ENTRY_STR(GC, 0, regGCVM_L2_PROTECTION_FAULT_STATUS), 151 SOC15_REG_ENTRY_STR(GC, 0, regCP_DEBUG), 152 SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_CNTL), 153 SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_CNTL), 154 SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC1_INSTR_PNTR), 155 SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_DEBUG_INTERRUPT_INSTR_PNTR), 156 SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_INSTR_PNTR), 157 SOC15_REG_ENTRY_STR(GC, 0, regCP_ME_INSTR_PNTR), 158 SOC15_REG_ENTRY_STR(GC, 0, regCP_PFP_INSTR_PNTR), 159 SOC15_REG_ENTRY_STR(GC, 0, regCP_CPC_STATUS), 160 /* cp header registers */ 161 SOC15_REG_ENTRY_STR(GC, 0, regCP_MEC_ME1_HEADER_DUMP), 162 SOC15_REG_ENTRY_STR(GC, 0, regCP_PFP_HEADER_DUMP), 163 SOC15_REG_ENTRY_STR(GC, 0, regCP_ME_HEADER_DUMP), 164 SOC15_REG_ENTRY_STR(GC, 0, regCP_MES_HEADER_DUMP), 165 /* SE status registers */ 166 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS_SE0), 167 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS_SE1), 168 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS_SE2), 169 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS_SE3), 170 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS_SE4), 171 SOC15_REG_ENTRY_STR(GC, 0, regGRBM_STATUS_SE5) 172 }; 173 174 static const struct amdgpu_hwip_reg_entry gc_cp_reg_list_11[] = { 175 /* compute registers */ 176 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_VMID), 177 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PERSISTENT_STATE), 178 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PIPE_PRIORITY), 179 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_QUEUE_PRIORITY), 180 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_QUANTUM), 181 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_BASE), 182 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_BASE_HI), 183 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_RPTR), 184 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR), 185 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR_HI), 186 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL), 187 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_CONTROL), 188 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_IB_BASE_ADDR), 189 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_IB_BASE_ADDR_HI), 190 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_IB_RPTR), 191 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_IB_CONTROL), 192 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_DEQUEUE_REQUEST), 193 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_BASE_ADDR), 194 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_BASE_ADDR_HI), 195 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_CONTROL), 196 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_RPTR), 197 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_WPTR), 198 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_EVENTS), 199 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CTX_SAVE_BASE_ADDR_LO), 200 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CTX_SAVE_BASE_ADDR_HI), 201 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CTX_SAVE_CONTROL), 202 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CNTL_STACK_OFFSET), 203 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CNTL_STACK_SIZE), 204 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_WG_STATE_OFFSET), 205 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_CTX_SAVE_SIZE), 206 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_GDS_RESOURCE_STATE), 207 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_ERROR), 208 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_EOP_WPTR_MEM), 209 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_WPTR_LO), 210 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_PQ_WPTR_HI), 211 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_SUSPEND_CNTL_STACK_OFFSET), 212 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_SUSPEND_CNTL_STACK_DW_CNT), 213 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_SUSPEND_WG_STATE_OFFSET), 214 SOC15_REG_ENTRY_STR(GC, 0, regCP_HQD_DEQUEUE_STATUS) 215 }; 216 217 static const struct amdgpu_hwip_reg_entry gc_gfx_queue_reg_list_11[] = { 218 /* gfx queue registers */ 219 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_ACTIVE), 220 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_VMID), 221 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_QUEUE_PRIORITY), 222 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_QUANTUM), 223 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_BASE), 224 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_BASE_HI), 225 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_OFFSET), 226 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_CNTL), 227 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_CSMD_RPTR), 228 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_WPTR), 229 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_WPTR_HI), 230 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_DEQUEUE_REQUEST), 231 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_MAPPED), 232 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_QUE_MGR_CONTROL), 233 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_HQ_CONTROL0), 234 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_HQD_HQ_STATUS0), 235 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_MQD_BASE_ADDR), 236 SOC15_REG_ENTRY_STR(GC, 0, regCP_GFX_MQD_BASE_ADDR_HI), 237 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_WPTR_POLL_ADDR_LO), 238 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_WPTR_POLL_ADDR_HI), 239 SOC15_REG_ENTRY_STR(GC, 0, regCP_RB_RPTR), 240 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BASE_LO), 241 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BASE_HI), 242 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_CMD_BUFSZ), 243 SOC15_REG_ENTRY_STR(GC, 0, regCP_IB1_BUFSZ) 244 }; 245 246 static const struct soc15_reg_golden golden_settings_gc_11_0[] = { 247 SOC15_REG_GOLDEN_VALUE(GC, 0, regTCP_CNTL, 0x20000000, 0x20000000) 248 }; 249 250 static const struct soc15_reg_golden golden_settings_gc_11_0_1[] = 251 { 252 SOC15_REG_GOLDEN_VALUE(GC, 0, regCGTT_GS_NGG_CLK_CTRL, 0x9fff8fff, 0x00000010), 253 SOC15_REG_GOLDEN_VALUE(GC, 0, regCGTT_WD_CLK_CTRL, 0xffff8fff, 0x00000010), 254 SOC15_REG_GOLDEN_VALUE(GC, 0, regCPF_GCR_CNTL, 0x0007ffff, 0x0000c200), 255 SOC15_REG_GOLDEN_VALUE(GC, 0, regGL2C_CTRL3, 0xffff001b, 0x00f01988), 256 SOC15_REG_GOLDEN_VALUE(GC, 0, regPA_CL_ENHANCE, 0xf0ffffff, 0x00880007), 257 SOC15_REG_GOLDEN_VALUE(GC, 0, regPA_SC_ENHANCE_3, 0xfffffffd, 0x00000008), 258 SOC15_REG_GOLDEN_VALUE(GC, 0, regPA_SC_VRS_SURFACE_CNTL_1, 0xfff891ff, 0x55480100), 259 SOC15_REG_GOLDEN_VALUE(GC, 0, regTA_CNTL_AUX, 0xf7f7ffff, 0x01030000), 260 SOC15_REG_GOLDEN_VALUE(GC, 0, regTCP_CNTL2, 0xfcffffff, 0x0000000a) 261 }; 262 263 #define DEFAULT_SH_MEM_CONFIG \ 264 ((SH_MEM_ADDRESS_MODE_64 << SH_MEM_CONFIG__ADDRESS_MODE__SHIFT) | \ 265 (SH_MEM_ALIGNMENT_MODE_UNALIGNED << SH_MEM_CONFIG__ALIGNMENT_MODE__SHIFT) | \ 266 (3 << SH_MEM_CONFIG__INITIAL_INST_PREFETCH__SHIFT)) 267 268 static void gfx_v11_0_disable_gpa_mode(struct amdgpu_device *adev); 269 static void gfx_v11_0_set_ring_funcs(struct amdgpu_device *adev); 270 static void gfx_v11_0_set_irq_funcs(struct amdgpu_device *adev); 271 static void gfx_v11_0_set_gds_init(struct amdgpu_device *adev); 272 static void gfx_v11_0_set_rlc_funcs(struct amdgpu_device *adev); 273 static void gfx_v11_0_set_mqd_funcs(struct amdgpu_device *adev); 274 static void gfx_v11_0_set_imu_funcs(struct amdgpu_device *adev); 275 static int gfx_v11_0_get_cu_info(struct amdgpu_device *adev, 276 struct amdgpu_cu_info *cu_info); 277 static uint64_t gfx_v11_0_get_gpu_clock_counter(struct amdgpu_device *adev); 278 static void gfx_v11_0_select_se_sh(struct amdgpu_device *adev, u32 se_num, 279 u32 sh_num, u32 instance, int xcc_id); 280 static u32 gfx_v11_0_get_wgp_active_bitmap_per_sh(struct amdgpu_device *adev); 281 282 static void gfx_v11_0_ring_emit_de_meta(struct amdgpu_ring *ring, bool resume); 283 static void gfx_v11_0_ring_emit_frame_cntl(struct amdgpu_ring *ring, bool start, bool secure); 284 static void gfx_v11_0_ring_emit_wreg(struct amdgpu_ring *ring, uint32_t reg, 285 uint32_t val); 286 static int gfx_v11_0_wait_for_rlc_autoload_complete(struct amdgpu_device *adev); 287 static void gfx_v11_0_ring_invalidate_tlbs(struct amdgpu_ring *ring, 288 uint16_t pasid, uint32_t flush_type, 289 bool all_hub, uint8_t dst_sel); 290 static void gfx_v11_0_set_safe_mode(struct amdgpu_device *adev, int xcc_id); 291 static void gfx_v11_0_unset_safe_mode(struct amdgpu_device *adev, int xcc_id); 292 static void gfx_v11_0_update_perf_clk(struct amdgpu_device *adev, 293 bool enable); 294 295 static void gfx11_kiq_set_resources(struct amdgpu_ring *kiq_ring, uint64_t queue_mask) 296 { 297 struct amdgpu_device *adev = kiq_ring->adev; 298 u64 shader_mc_addr; 299 300 /* Cleaner shader MC address */ 301 shader_mc_addr = adev->gfx.cleaner_shader_gpu_addr >> 8; 302 303 amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_SET_RESOURCES, 6)); 304 amdgpu_ring_write(kiq_ring, PACKET3_SET_RESOURCES_VMID_MASK(0) | 305 PACKET3_SET_RESOURCES_UNMAP_LATENTY(0xa) | /* unmap_latency: 0xa (~ 1s) */ 306 PACKET3_SET_RESOURCES_QUEUE_TYPE(0)); /* vmid_mask:0 queue_type:0 (KIQ) */ 307 amdgpu_ring_write(kiq_ring, lower_32_bits(queue_mask)); /* queue mask lo */ 308 amdgpu_ring_write(kiq_ring, upper_32_bits(queue_mask)); /* queue mask hi */ 309 amdgpu_ring_write(kiq_ring, lower_32_bits(shader_mc_addr)); /* cleaner shader addr lo */ 310 amdgpu_ring_write(kiq_ring, upper_32_bits(shader_mc_addr)); /* cleaner shader addr hi */ 311 amdgpu_ring_write(kiq_ring, 0); /* oac mask */ 312 amdgpu_ring_write(kiq_ring, 0); /* gds heap base:0, gds heap size:0 */ 313 } 314 315 static void gfx11_kiq_map_queues(struct amdgpu_ring *kiq_ring, 316 struct amdgpu_ring *ring) 317 { 318 uint64_t mqd_addr = amdgpu_bo_gpu_offset(ring->mqd_obj); 319 uint64_t wptr_addr = ring->wptr_gpu_addr; 320 uint32_t me = 0, eng_sel = 0; 321 322 switch (ring->funcs->type) { 323 case AMDGPU_RING_TYPE_COMPUTE: 324 me = 1; 325 eng_sel = 0; 326 break; 327 case AMDGPU_RING_TYPE_GFX: 328 me = 0; 329 eng_sel = 4; 330 break; 331 case AMDGPU_RING_TYPE_MES: 332 me = 2; 333 eng_sel = 5; 334 break; 335 default: 336 WARN_ON(1); 337 } 338 339 amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_MAP_QUEUES, 5)); 340 /* Q_sel:0, vmid:0, vidmem: 1, engine:0, num_Q:1*/ 341 amdgpu_ring_write(kiq_ring, /* Q_sel: 0, vmid: 0, engine: 0, num_Q: 1 */ 342 PACKET3_MAP_QUEUES_QUEUE_SEL(0) | /* Queue_Sel */ 343 PACKET3_MAP_QUEUES_VMID(0) | /* VMID */ 344 PACKET3_MAP_QUEUES_QUEUE(ring->queue) | 345 PACKET3_MAP_QUEUES_PIPE(ring->pipe) | 346 PACKET3_MAP_QUEUES_ME((me)) | 347 PACKET3_MAP_QUEUES_QUEUE_TYPE(0) | /*queue_type: normal compute queue */ 348 PACKET3_MAP_QUEUES_ALLOC_FORMAT(0) | /* alloc format: all_on_one_pipe */ 349 PACKET3_MAP_QUEUES_ENGINE_SEL(eng_sel) | 350 PACKET3_MAP_QUEUES_NUM_QUEUES(1)); /* num_queues: must be 1 */ 351 amdgpu_ring_write(kiq_ring, PACKET3_MAP_QUEUES_DOORBELL_OFFSET(ring->doorbell_index)); 352 amdgpu_ring_write(kiq_ring, lower_32_bits(mqd_addr)); 353 amdgpu_ring_write(kiq_ring, upper_32_bits(mqd_addr)); 354 amdgpu_ring_write(kiq_ring, lower_32_bits(wptr_addr)); 355 amdgpu_ring_write(kiq_ring, upper_32_bits(wptr_addr)); 356 } 357 358 static void gfx11_kiq_unmap_queues(struct amdgpu_ring *kiq_ring, 359 struct amdgpu_ring *ring, 360 enum amdgpu_unmap_queues_action action, 361 u64 gpu_addr, u64 seq) 362 { 363 struct amdgpu_device *adev = kiq_ring->adev; 364 uint32_t eng_sel = ring->funcs->type == AMDGPU_RING_TYPE_GFX ? 4 : 0; 365 366 if (adev->enable_mes && !adev->gfx.kiq[0].ring.sched.ready) { 367 amdgpu_mes_unmap_legacy_queue(adev, ring, action, gpu_addr, seq); 368 return; 369 } 370 371 amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_UNMAP_QUEUES, 4)); 372 amdgpu_ring_write(kiq_ring, /* Q_sel: 0, vmid: 0, engine: 0, num_Q: 1 */ 373 PACKET3_UNMAP_QUEUES_ACTION(action) | 374 PACKET3_UNMAP_QUEUES_QUEUE_SEL(0) | 375 PACKET3_UNMAP_QUEUES_ENGINE_SEL(eng_sel) | 376 PACKET3_UNMAP_QUEUES_NUM_QUEUES(1)); 377 amdgpu_ring_write(kiq_ring, 378 PACKET3_UNMAP_QUEUES_DOORBELL_OFFSET0(ring->doorbell_index)); 379 380 if (action == PREEMPT_QUEUES_NO_UNMAP) { 381 amdgpu_ring_write(kiq_ring, lower_32_bits(gpu_addr)); 382 amdgpu_ring_write(kiq_ring, upper_32_bits(gpu_addr)); 383 amdgpu_ring_write(kiq_ring, seq); 384 } else { 385 amdgpu_ring_write(kiq_ring, 0); 386 amdgpu_ring_write(kiq_ring, 0); 387 amdgpu_ring_write(kiq_ring, 0); 388 } 389 } 390 391 static void gfx11_kiq_query_status(struct amdgpu_ring *kiq_ring, 392 struct amdgpu_ring *ring, 393 u64 addr, 394 u64 seq) 395 { 396 uint32_t eng_sel = ring->funcs->type == AMDGPU_RING_TYPE_GFX ? 4 : 0; 397 398 amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_QUERY_STATUS, 5)); 399 amdgpu_ring_write(kiq_ring, 400 PACKET3_QUERY_STATUS_CONTEXT_ID(0) | 401 PACKET3_QUERY_STATUS_INTERRUPT_SEL(0) | 402 PACKET3_QUERY_STATUS_COMMAND(2)); 403 amdgpu_ring_write(kiq_ring, /* Q_sel: 0, vmid: 0, engine: 0, num_Q: 1 */ 404 PACKET3_QUERY_STATUS_DOORBELL_OFFSET(ring->doorbell_index) | 405 PACKET3_QUERY_STATUS_ENG_SEL(eng_sel)); 406 amdgpu_ring_write(kiq_ring, lower_32_bits(addr)); 407 amdgpu_ring_write(kiq_ring, upper_32_bits(addr)); 408 amdgpu_ring_write(kiq_ring, lower_32_bits(seq)); 409 amdgpu_ring_write(kiq_ring, upper_32_bits(seq)); 410 } 411 412 static void gfx11_kiq_invalidate_tlbs(struct amdgpu_ring *kiq_ring, 413 uint16_t pasid, uint32_t flush_type, 414 bool all_hub) 415 { 416 gfx_v11_0_ring_invalidate_tlbs(kiq_ring, pasid, flush_type, all_hub, 1); 417 } 418 419 static const struct kiq_pm4_funcs gfx_v11_0_kiq_pm4_funcs = { 420 .kiq_set_resources = gfx11_kiq_set_resources, 421 .kiq_map_queues = gfx11_kiq_map_queues, 422 .kiq_unmap_queues = gfx11_kiq_unmap_queues, 423 .kiq_query_status = gfx11_kiq_query_status, 424 .kiq_invalidate_tlbs = gfx11_kiq_invalidate_tlbs, 425 .set_resources_size = 8, 426 .map_queues_size = 7, 427 .unmap_queues_size = 6, 428 .query_status_size = 7, 429 .invalidate_tlbs_size = 2, 430 }; 431 432 static void gfx_v11_0_set_kiq_pm4_funcs(struct amdgpu_device *adev) 433 { 434 adev->gfx.kiq[0].pmf = &gfx_v11_0_kiq_pm4_funcs; 435 } 436 437 static void gfx_v11_0_init_golden_registers(struct amdgpu_device *adev) 438 { 439 if (amdgpu_sriov_vf(adev)) 440 return; 441 442 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 443 case IP_VERSION(11, 0, 1): 444 case IP_VERSION(11, 0, 4): 445 soc15_program_register_sequence(adev, 446 golden_settings_gc_11_0_1, 447 (const u32)ARRAY_SIZE(golden_settings_gc_11_0_1)); 448 break; 449 default: 450 break; 451 } 452 soc15_program_register_sequence(adev, 453 golden_settings_gc_11_0, 454 (const u32)ARRAY_SIZE(golden_settings_gc_11_0)); 455 456 } 457 458 static void gfx_v11_0_write_data_to_reg(struct amdgpu_ring *ring, int eng_sel, 459 bool wc, uint32_t reg, uint32_t val) 460 { 461 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3)); 462 amdgpu_ring_write(ring, WRITE_DATA_ENGINE_SEL(eng_sel) | 463 WRITE_DATA_DST_SEL(0) | (wc ? WR_CONFIRM : 0)); 464 amdgpu_ring_write(ring, reg); 465 amdgpu_ring_write(ring, 0); 466 amdgpu_ring_write(ring, val); 467 } 468 469 static void gfx_v11_0_wait_reg_mem(struct amdgpu_ring *ring, int eng_sel, 470 int mem_space, int opt, uint32_t addr0, 471 uint32_t addr1, uint32_t ref, uint32_t mask, 472 uint32_t inv) 473 { 474 amdgpu_ring_write(ring, PACKET3(PACKET3_WAIT_REG_MEM, 5)); 475 amdgpu_ring_write(ring, 476 /* memory (1) or register (0) */ 477 (WAIT_REG_MEM_MEM_SPACE(mem_space) | 478 WAIT_REG_MEM_OPERATION(opt) | /* wait */ 479 WAIT_REG_MEM_FUNCTION(3) | /* equal */ 480 WAIT_REG_MEM_ENGINE(eng_sel))); 481 482 if (mem_space) 483 BUG_ON(addr0 & 0x3); /* Dword align */ 484 amdgpu_ring_write(ring, addr0); 485 amdgpu_ring_write(ring, addr1); 486 amdgpu_ring_write(ring, ref); 487 amdgpu_ring_write(ring, mask); 488 amdgpu_ring_write(ring, inv); /* poll interval */ 489 } 490 491 static void gfx_v11_ring_insert_nop(struct amdgpu_ring *ring, uint32_t num_nop) 492 { 493 /* Header itself is a NOP packet */ 494 if (num_nop == 1) { 495 amdgpu_ring_write(ring, ring->funcs->nop); 496 return; 497 } 498 499 /* Max HW optimization till 0x3ffe, followed by remaining one NOP at a time*/ 500 amdgpu_ring_write(ring, PACKET3(PACKET3_NOP, min(num_nop - 2, 0x3ffe))); 501 502 /* Header is at index 0, followed by num_nops - 1 NOP packet's */ 503 amdgpu_ring_insert_nop(ring, num_nop - 1); 504 } 505 506 static int gfx_v11_0_ring_test_ring(struct amdgpu_ring *ring) 507 { 508 struct amdgpu_device *adev = ring->adev; 509 uint32_t scratch = SOC15_REG_OFFSET(GC, 0, regSCRATCH_REG0); 510 uint32_t tmp = 0; 511 unsigned i; 512 int r; 513 514 WREG32(scratch, 0xCAFEDEAD); 515 r = amdgpu_ring_alloc(ring, 5); 516 if (r) { 517 DRM_ERROR("amdgpu: cp failed to lock ring %d (%d).\n", 518 ring->idx, r); 519 return r; 520 } 521 522 if (ring->funcs->type == AMDGPU_RING_TYPE_KIQ) { 523 gfx_v11_0_ring_emit_wreg(ring, scratch, 0xDEADBEEF); 524 } else { 525 amdgpu_ring_write(ring, PACKET3(PACKET3_SET_UCONFIG_REG, 1)); 526 amdgpu_ring_write(ring, scratch - 527 PACKET3_SET_UCONFIG_REG_START); 528 amdgpu_ring_write(ring, 0xDEADBEEF); 529 } 530 amdgpu_ring_commit(ring); 531 532 for (i = 0; i < adev->usec_timeout; i++) { 533 tmp = RREG32(scratch); 534 if (tmp == 0xDEADBEEF) 535 break; 536 if (amdgpu_emu_mode == 1) 537 msleep(1); 538 else 539 udelay(1); 540 } 541 542 if (i >= adev->usec_timeout) 543 r = -ETIMEDOUT; 544 return r; 545 } 546 547 static int gfx_v11_0_ring_test_ib(struct amdgpu_ring *ring, long timeout) 548 { 549 struct amdgpu_device *adev = ring->adev; 550 struct amdgpu_ib ib; 551 struct dma_fence *f = NULL; 552 unsigned index; 553 uint64_t gpu_addr; 554 volatile uint32_t *cpu_ptr; 555 long r; 556 557 /* MES KIQ fw hasn't indirect buffer support for now */ 558 if (adev->enable_mes_kiq && 559 ring->funcs->type == AMDGPU_RING_TYPE_KIQ) 560 return 0; 561 562 memset(&ib, 0, sizeof(ib)); 563 564 if (ring->is_mes_queue) { 565 uint32_t padding, offset; 566 567 offset = amdgpu_mes_ctx_get_offs(ring, AMDGPU_MES_CTX_IB_OFFS); 568 padding = amdgpu_mes_ctx_get_offs(ring, 569 AMDGPU_MES_CTX_PADDING_OFFS); 570 571 ib.gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset); 572 ib.ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset); 573 574 gpu_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, padding); 575 cpu_ptr = amdgpu_mes_ctx_get_offs_cpu_addr(ring, padding); 576 *cpu_ptr = cpu_to_le32(0xCAFEDEAD); 577 } else { 578 r = amdgpu_device_wb_get(adev, &index); 579 if (r) 580 return r; 581 582 gpu_addr = adev->wb.gpu_addr + (index * 4); 583 adev->wb.wb[index] = cpu_to_le32(0xCAFEDEAD); 584 cpu_ptr = &adev->wb.wb[index]; 585 586 r = amdgpu_ib_get(adev, NULL, 20, AMDGPU_IB_POOL_DIRECT, &ib); 587 if (r) { 588 DRM_ERROR("amdgpu: failed to get ib (%ld).\n", r); 589 goto err1; 590 } 591 } 592 593 ib.ptr[0] = PACKET3(PACKET3_WRITE_DATA, 3); 594 ib.ptr[1] = WRITE_DATA_DST_SEL(5) | WR_CONFIRM; 595 ib.ptr[2] = lower_32_bits(gpu_addr); 596 ib.ptr[3] = upper_32_bits(gpu_addr); 597 ib.ptr[4] = 0xDEADBEEF; 598 ib.length_dw = 5; 599 600 r = amdgpu_ib_schedule(ring, 1, &ib, NULL, &f); 601 if (r) 602 goto err2; 603 604 r = dma_fence_wait_timeout(f, false, timeout); 605 if (r == 0) { 606 r = -ETIMEDOUT; 607 goto err2; 608 } else if (r < 0) { 609 goto err2; 610 } 611 612 if (le32_to_cpu(*cpu_ptr) == 0xDEADBEEF) 613 r = 0; 614 else 615 r = -EINVAL; 616 err2: 617 if (!ring->is_mes_queue) 618 amdgpu_ib_free(adev, &ib, NULL); 619 dma_fence_put(f); 620 err1: 621 if (!ring->is_mes_queue) 622 amdgpu_device_wb_free(adev, index); 623 return r; 624 } 625 626 static void gfx_v11_0_free_microcode(struct amdgpu_device *adev) 627 { 628 amdgpu_ucode_release(&adev->gfx.pfp_fw); 629 amdgpu_ucode_release(&adev->gfx.me_fw); 630 amdgpu_ucode_release(&adev->gfx.rlc_fw); 631 amdgpu_ucode_release(&adev->gfx.mec_fw); 632 633 kfree(adev->gfx.rlc.register_list_format); 634 } 635 636 static int gfx_v11_0_init_toc_microcode(struct amdgpu_device *adev, const char *ucode_prefix) 637 { 638 const struct psp_firmware_header_v1_0 *toc_hdr; 639 int err = 0; 640 641 err = amdgpu_ucode_request(adev, &adev->psp.toc_fw, 642 "amdgpu/%s_toc.bin", ucode_prefix); 643 if (err) 644 goto out; 645 646 toc_hdr = (const struct psp_firmware_header_v1_0 *)adev->psp.toc_fw->data; 647 adev->psp.toc.fw_version = le32_to_cpu(toc_hdr->header.ucode_version); 648 adev->psp.toc.feature_version = le32_to_cpu(toc_hdr->sos.fw_version); 649 adev->psp.toc.size_bytes = le32_to_cpu(toc_hdr->header.ucode_size_bytes); 650 adev->psp.toc.start_addr = (uint8_t *)toc_hdr + 651 le32_to_cpu(toc_hdr->header.ucode_array_offset_bytes); 652 return 0; 653 out: 654 amdgpu_ucode_release(&adev->psp.toc_fw); 655 return err; 656 } 657 658 static void gfx_v11_0_check_fw_cp_gfx_shadow(struct amdgpu_device *adev) 659 { 660 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 661 case IP_VERSION(11, 0, 0): 662 case IP_VERSION(11, 0, 2): 663 case IP_VERSION(11, 0, 3): 664 if ((adev->gfx.me_fw_version >= 1505) && 665 (adev->gfx.pfp_fw_version >= 1600) && 666 (adev->gfx.mec_fw_version >= 512)) { 667 if (amdgpu_sriov_vf(adev)) 668 adev->gfx.cp_gfx_shadow = true; 669 else 670 adev->gfx.cp_gfx_shadow = false; 671 } 672 break; 673 default: 674 adev->gfx.cp_gfx_shadow = false; 675 break; 676 } 677 } 678 679 static int gfx_v11_0_init_microcode(struct amdgpu_device *adev) 680 { 681 char ucode_prefix[25]; 682 int err; 683 const struct rlc_firmware_header_v2_0 *rlc_hdr; 684 uint16_t version_major; 685 uint16_t version_minor; 686 687 DRM_DEBUG("\n"); 688 689 amdgpu_ucode_ip_version_decode(adev, GC_HWIP, ucode_prefix, sizeof(ucode_prefix)); 690 err = amdgpu_ucode_request(adev, &adev->gfx.pfp_fw, 691 "amdgpu/%s_pfp.bin", ucode_prefix); 692 if (err) 693 goto out; 694 /* check pfp fw hdr version to decide if enable rs64 for gfx11.*/ 695 adev->gfx.rs64_enable = amdgpu_ucode_hdr_version( 696 (union amdgpu_firmware_header *) 697 adev->gfx.pfp_fw->data, 2, 0); 698 if (adev->gfx.rs64_enable) { 699 dev_info(adev->dev, "CP RS64 enable\n"); 700 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_PFP); 701 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_PFP_P0_STACK); 702 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_PFP_P1_STACK); 703 } else { 704 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_PFP); 705 } 706 707 err = amdgpu_ucode_request(adev, &adev->gfx.me_fw, 708 "amdgpu/%s_me.bin", ucode_prefix); 709 if (err) 710 goto out; 711 if (adev->gfx.rs64_enable) { 712 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_ME); 713 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_ME_P0_STACK); 714 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_ME_P1_STACK); 715 } else { 716 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_ME); 717 } 718 719 if (!amdgpu_sriov_vf(adev)) { 720 if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(11, 0, 0) && 721 adev->pdev->revision == 0xCE) 722 err = amdgpu_ucode_request(adev, &adev->gfx.rlc_fw, 723 "amdgpu/gc_11_0_0_rlc_1.bin"); 724 else 725 err = amdgpu_ucode_request(adev, &adev->gfx.rlc_fw, 726 "amdgpu/%s_rlc.bin", ucode_prefix); 727 if (err) 728 goto out; 729 rlc_hdr = (const struct rlc_firmware_header_v2_0 *)adev->gfx.rlc_fw->data; 730 version_major = le16_to_cpu(rlc_hdr->header.header_version_major); 731 version_minor = le16_to_cpu(rlc_hdr->header.header_version_minor); 732 err = amdgpu_gfx_rlc_init_microcode(adev, version_major, version_minor); 733 if (err) 734 goto out; 735 } 736 737 err = amdgpu_ucode_request(adev, &adev->gfx.mec_fw, 738 "amdgpu/%s_mec.bin", ucode_prefix); 739 if (err) 740 goto out; 741 if (adev->gfx.rs64_enable) { 742 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_MEC); 743 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_MEC_P0_STACK); 744 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_MEC_P1_STACK); 745 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_MEC_P2_STACK); 746 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_RS64_MEC_P3_STACK); 747 } else { 748 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_MEC1); 749 amdgpu_gfx_cp_init_microcode(adev, AMDGPU_UCODE_ID_CP_MEC1_JT); 750 } 751 752 if (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) 753 err = gfx_v11_0_init_toc_microcode(adev, ucode_prefix); 754 755 /* only one MEC for gfx 11.0.0. */ 756 adev->gfx.mec2_fw = NULL; 757 758 gfx_v11_0_check_fw_cp_gfx_shadow(adev); 759 760 if (adev->gfx.imu.funcs && adev->gfx.imu.funcs->init_microcode) { 761 err = adev->gfx.imu.funcs->init_microcode(adev); 762 if (err) 763 DRM_ERROR("Failed to init imu firmware!\n"); 764 return err; 765 } 766 767 out: 768 if (err) { 769 amdgpu_ucode_release(&adev->gfx.pfp_fw); 770 amdgpu_ucode_release(&adev->gfx.me_fw); 771 amdgpu_ucode_release(&adev->gfx.rlc_fw); 772 amdgpu_ucode_release(&adev->gfx.mec_fw); 773 } 774 775 return err; 776 } 777 778 static u32 gfx_v11_0_get_csb_size(struct amdgpu_device *adev) 779 { 780 u32 count = 0; 781 const struct cs_section_def *sect = NULL; 782 const struct cs_extent_def *ext = NULL; 783 784 /* begin clear state */ 785 count += 2; 786 /* context control state */ 787 count += 3; 788 789 for (sect = gfx11_cs_data; sect->section != NULL; ++sect) { 790 for (ext = sect->section; ext->extent != NULL; ++ext) { 791 if (sect->id == SECT_CONTEXT) 792 count += 2 + ext->reg_count; 793 else 794 return 0; 795 } 796 } 797 798 /* set PA_SC_TILE_STEERING_OVERRIDE */ 799 count += 3; 800 /* end clear state */ 801 count += 2; 802 /* clear state */ 803 count += 2; 804 805 return count; 806 } 807 808 static void gfx_v11_0_get_csb_buffer(struct amdgpu_device *adev, 809 volatile u32 *buffer) 810 { 811 u32 count = 0, i; 812 const struct cs_section_def *sect = NULL; 813 const struct cs_extent_def *ext = NULL; 814 int ctx_reg_offset; 815 816 if (adev->gfx.rlc.cs_data == NULL) 817 return; 818 if (buffer == NULL) 819 return; 820 821 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_PREAMBLE_CNTL, 0)); 822 buffer[count++] = cpu_to_le32(PACKET3_PREAMBLE_BEGIN_CLEAR_STATE); 823 824 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_CONTEXT_CONTROL, 1)); 825 buffer[count++] = cpu_to_le32(0x80000000); 826 buffer[count++] = cpu_to_le32(0x80000000); 827 828 for (sect = adev->gfx.rlc.cs_data; sect->section != NULL; ++sect) { 829 for (ext = sect->section; ext->extent != NULL; ++ext) { 830 if (sect->id == SECT_CONTEXT) { 831 buffer[count++] = 832 cpu_to_le32(PACKET3(PACKET3_SET_CONTEXT_REG, ext->reg_count)); 833 buffer[count++] = cpu_to_le32(ext->reg_index - 834 PACKET3_SET_CONTEXT_REG_START); 835 for (i = 0; i < ext->reg_count; i++) 836 buffer[count++] = cpu_to_le32(ext->extent[i]); 837 } else { 838 return; 839 } 840 } 841 } 842 843 ctx_reg_offset = 844 SOC15_REG_OFFSET(GC, 0, regPA_SC_TILE_STEERING_OVERRIDE) - PACKET3_SET_CONTEXT_REG_START; 845 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_SET_CONTEXT_REG, 1)); 846 buffer[count++] = cpu_to_le32(ctx_reg_offset); 847 buffer[count++] = cpu_to_le32(adev->gfx.config.pa_sc_tile_steering_override); 848 849 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_PREAMBLE_CNTL, 0)); 850 buffer[count++] = cpu_to_le32(PACKET3_PREAMBLE_END_CLEAR_STATE); 851 852 buffer[count++] = cpu_to_le32(PACKET3(PACKET3_CLEAR_STATE, 0)); 853 buffer[count++] = cpu_to_le32(0); 854 } 855 856 static void gfx_v11_0_rlc_fini(struct amdgpu_device *adev) 857 { 858 /* clear state block */ 859 amdgpu_bo_free_kernel(&adev->gfx.rlc.clear_state_obj, 860 &adev->gfx.rlc.clear_state_gpu_addr, 861 (void **)&adev->gfx.rlc.cs_ptr); 862 863 /* jump table block */ 864 amdgpu_bo_free_kernel(&adev->gfx.rlc.cp_table_obj, 865 &adev->gfx.rlc.cp_table_gpu_addr, 866 (void **)&adev->gfx.rlc.cp_table_ptr); 867 } 868 869 static void gfx_v11_0_init_rlcg_reg_access_ctrl(struct amdgpu_device *adev) 870 { 871 struct amdgpu_rlcg_reg_access_ctrl *reg_access_ctrl; 872 873 reg_access_ctrl = &adev->gfx.rlc.reg_access_ctrl[0]; 874 reg_access_ctrl->scratch_reg0 = SOC15_REG_OFFSET(GC, 0, regSCRATCH_REG0); 875 reg_access_ctrl->scratch_reg1 = SOC15_REG_OFFSET(GC, 0, regSCRATCH_REG1); 876 reg_access_ctrl->scratch_reg2 = SOC15_REG_OFFSET(GC, 0, regSCRATCH_REG2); 877 reg_access_ctrl->scratch_reg3 = SOC15_REG_OFFSET(GC, 0, regSCRATCH_REG3); 878 reg_access_ctrl->grbm_cntl = SOC15_REG_OFFSET(GC, 0, regGRBM_GFX_CNTL); 879 reg_access_ctrl->grbm_idx = SOC15_REG_OFFSET(GC, 0, regGRBM_GFX_INDEX); 880 reg_access_ctrl->spare_int = SOC15_REG_OFFSET(GC, 0, regRLC_SPARE_INT_0); 881 adev->gfx.rlc.rlcg_reg_access_supported = true; 882 } 883 884 static int gfx_v11_0_rlc_init(struct amdgpu_device *adev) 885 { 886 const struct cs_section_def *cs_data; 887 int r; 888 889 adev->gfx.rlc.cs_data = gfx11_cs_data; 890 891 cs_data = adev->gfx.rlc.cs_data; 892 893 if (cs_data) { 894 /* init clear state block */ 895 r = amdgpu_gfx_rlc_init_csb(adev); 896 if (r) 897 return r; 898 } 899 900 /* init spm vmid with 0xf */ 901 if (adev->gfx.rlc.funcs->update_spm_vmid) 902 adev->gfx.rlc.funcs->update_spm_vmid(adev, NULL, 0xf); 903 904 return 0; 905 } 906 907 static void gfx_v11_0_mec_fini(struct amdgpu_device *adev) 908 { 909 amdgpu_bo_free_kernel(&adev->gfx.mec.hpd_eop_obj, NULL, NULL); 910 amdgpu_bo_free_kernel(&adev->gfx.mec.mec_fw_obj, NULL, NULL); 911 amdgpu_bo_free_kernel(&adev->gfx.mec.mec_fw_data_obj, NULL, NULL); 912 } 913 914 static void gfx_v11_0_me_init(struct amdgpu_device *adev) 915 { 916 bitmap_zero(adev->gfx.me.queue_bitmap, AMDGPU_MAX_GFX_QUEUES); 917 918 amdgpu_gfx_graphics_queue_acquire(adev); 919 } 920 921 static int gfx_v11_0_mec_init(struct amdgpu_device *adev) 922 { 923 int r; 924 u32 *hpd; 925 size_t mec_hpd_size; 926 927 bitmap_zero(adev->gfx.mec_bitmap[0].queue_bitmap, AMDGPU_MAX_COMPUTE_QUEUES); 928 929 /* take ownership of the relevant compute queues */ 930 amdgpu_gfx_compute_queue_acquire(adev); 931 mec_hpd_size = adev->gfx.num_compute_rings * GFX11_MEC_HPD_SIZE; 932 933 if (mec_hpd_size) { 934 r = amdgpu_bo_create_reserved(adev, mec_hpd_size, PAGE_SIZE, 935 AMDGPU_GEM_DOMAIN_GTT, 936 &adev->gfx.mec.hpd_eop_obj, 937 &adev->gfx.mec.hpd_eop_gpu_addr, 938 (void **)&hpd); 939 if (r) { 940 dev_warn(adev->dev, "(%d) create HDP EOP bo failed\n", r); 941 gfx_v11_0_mec_fini(adev); 942 return r; 943 } 944 945 memset(hpd, 0, mec_hpd_size); 946 947 amdgpu_bo_kunmap(adev->gfx.mec.hpd_eop_obj); 948 amdgpu_bo_unreserve(adev->gfx.mec.hpd_eop_obj); 949 } 950 951 return 0; 952 } 953 954 static uint32_t wave_read_ind(struct amdgpu_device *adev, uint32_t wave, uint32_t address) 955 { 956 WREG32_SOC15(GC, 0, regSQ_IND_INDEX, 957 (wave << SQ_IND_INDEX__WAVE_ID__SHIFT) | 958 (address << SQ_IND_INDEX__INDEX__SHIFT)); 959 return RREG32_SOC15(GC, 0, regSQ_IND_DATA); 960 } 961 962 static void wave_read_regs(struct amdgpu_device *adev, uint32_t wave, 963 uint32_t thread, uint32_t regno, 964 uint32_t num, uint32_t *out) 965 { 966 WREG32_SOC15(GC, 0, regSQ_IND_INDEX, 967 (wave << SQ_IND_INDEX__WAVE_ID__SHIFT) | 968 (regno << SQ_IND_INDEX__INDEX__SHIFT) | 969 (thread << SQ_IND_INDEX__WORKITEM_ID__SHIFT) | 970 (SQ_IND_INDEX__AUTO_INCR_MASK)); 971 while (num--) 972 *(out++) = RREG32_SOC15(GC, 0, regSQ_IND_DATA); 973 } 974 975 static void gfx_v11_0_read_wave_data(struct amdgpu_device *adev, uint32_t xcc_id, uint32_t simd, uint32_t wave, uint32_t *dst, int *no_fields) 976 { 977 /* in gfx11 the SIMD_ID is specified as part of the INSTANCE 978 * field when performing a select_se_sh so it should be 979 * zero here */ 980 WARN_ON(simd != 0); 981 982 /* type 3 wave data */ 983 dst[(*no_fields)++] = 3; 984 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_STATUS); 985 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_PC_LO); 986 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_PC_HI); 987 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_EXEC_LO); 988 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_EXEC_HI); 989 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_HW_ID1); 990 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_HW_ID2); 991 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_GPR_ALLOC); 992 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_LDS_ALLOC); 993 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_TRAPSTS); 994 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_IB_STS); 995 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_IB_STS2); 996 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_IB_DBG1); 997 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_M0); 998 dst[(*no_fields)++] = wave_read_ind(adev, wave, ixSQ_WAVE_MODE); 999 } 1000 1001 static void gfx_v11_0_read_wave_sgprs(struct amdgpu_device *adev, uint32_t xcc_id, uint32_t simd, 1002 uint32_t wave, uint32_t start, 1003 uint32_t size, uint32_t *dst) 1004 { 1005 WARN_ON(simd != 0); 1006 1007 wave_read_regs( 1008 adev, wave, 0, start + SQIND_WAVE_SGPRS_OFFSET, size, 1009 dst); 1010 } 1011 1012 static void gfx_v11_0_read_wave_vgprs(struct amdgpu_device *adev, uint32_t xcc_id, uint32_t simd, 1013 uint32_t wave, uint32_t thread, 1014 uint32_t start, uint32_t size, 1015 uint32_t *dst) 1016 { 1017 wave_read_regs( 1018 adev, wave, thread, 1019 start + SQIND_WAVE_VGPRS_OFFSET, size, dst); 1020 } 1021 1022 static void gfx_v11_0_select_me_pipe_q(struct amdgpu_device *adev, 1023 u32 me, u32 pipe, u32 q, u32 vm, u32 xcc_id) 1024 { 1025 soc21_grbm_select(adev, me, pipe, q, vm); 1026 } 1027 1028 /* all sizes are in bytes */ 1029 #define MQD_SHADOW_BASE_SIZE 73728 1030 #define MQD_SHADOW_BASE_ALIGNMENT 256 1031 #define MQD_FWWORKAREA_SIZE 484 1032 #define MQD_FWWORKAREA_ALIGNMENT 256 1033 1034 static int gfx_v11_0_get_gfx_shadow_info(struct amdgpu_device *adev, 1035 struct amdgpu_gfx_shadow_info *shadow_info) 1036 { 1037 if (adev->gfx.cp_gfx_shadow) { 1038 shadow_info->shadow_size = MQD_SHADOW_BASE_SIZE; 1039 shadow_info->shadow_alignment = MQD_SHADOW_BASE_ALIGNMENT; 1040 shadow_info->csa_size = MQD_FWWORKAREA_SIZE; 1041 shadow_info->csa_alignment = MQD_FWWORKAREA_ALIGNMENT; 1042 return 0; 1043 } else { 1044 memset(shadow_info, 0, sizeof(struct amdgpu_gfx_shadow_info)); 1045 return -ENOTSUPP; 1046 } 1047 } 1048 1049 static const struct amdgpu_gfx_funcs gfx_v11_0_gfx_funcs = { 1050 .get_gpu_clock_counter = &gfx_v11_0_get_gpu_clock_counter, 1051 .select_se_sh = &gfx_v11_0_select_se_sh, 1052 .read_wave_data = &gfx_v11_0_read_wave_data, 1053 .read_wave_sgprs = &gfx_v11_0_read_wave_sgprs, 1054 .read_wave_vgprs = &gfx_v11_0_read_wave_vgprs, 1055 .select_me_pipe_q = &gfx_v11_0_select_me_pipe_q, 1056 .update_perfmon_mgcg = &gfx_v11_0_update_perf_clk, 1057 .get_gfx_shadow_info = &gfx_v11_0_get_gfx_shadow_info, 1058 }; 1059 1060 static int gfx_v11_0_gpu_early_init(struct amdgpu_device *adev) 1061 { 1062 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 1063 case IP_VERSION(11, 0, 0): 1064 case IP_VERSION(11, 0, 2): 1065 adev->gfx.config.max_hw_contexts = 8; 1066 adev->gfx.config.sc_prim_fifo_size_frontend = 0x20; 1067 adev->gfx.config.sc_prim_fifo_size_backend = 0x100; 1068 adev->gfx.config.sc_hiz_tile_fifo_size = 0; 1069 adev->gfx.config.sc_earlyz_tile_fifo_size = 0x4C0; 1070 break; 1071 case IP_VERSION(11, 0, 3): 1072 adev->gfx.ras = &gfx_v11_0_3_ras; 1073 adev->gfx.config.max_hw_contexts = 8; 1074 adev->gfx.config.sc_prim_fifo_size_frontend = 0x20; 1075 adev->gfx.config.sc_prim_fifo_size_backend = 0x100; 1076 adev->gfx.config.sc_hiz_tile_fifo_size = 0; 1077 adev->gfx.config.sc_earlyz_tile_fifo_size = 0x4C0; 1078 break; 1079 case IP_VERSION(11, 0, 1): 1080 case IP_VERSION(11, 0, 4): 1081 case IP_VERSION(11, 5, 0): 1082 case IP_VERSION(11, 5, 1): 1083 case IP_VERSION(11, 5, 2): 1084 adev->gfx.config.max_hw_contexts = 8; 1085 adev->gfx.config.sc_prim_fifo_size_frontend = 0x20; 1086 adev->gfx.config.sc_prim_fifo_size_backend = 0x100; 1087 adev->gfx.config.sc_hiz_tile_fifo_size = 0x80; 1088 adev->gfx.config.sc_earlyz_tile_fifo_size = 0x300; 1089 break; 1090 default: 1091 BUG(); 1092 break; 1093 } 1094 1095 return 0; 1096 } 1097 1098 static int gfx_v11_0_gfx_ring_init(struct amdgpu_device *adev, int ring_id, 1099 int me, int pipe, int queue) 1100 { 1101 struct amdgpu_ring *ring; 1102 unsigned int irq_type; 1103 unsigned int hw_prio; 1104 1105 ring = &adev->gfx.gfx_ring[ring_id]; 1106 1107 ring->me = me; 1108 ring->pipe = pipe; 1109 ring->queue = queue; 1110 1111 ring->ring_obj = NULL; 1112 ring->use_doorbell = true; 1113 1114 if (!ring_id) 1115 ring->doorbell_index = adev->doorbell_index.gfx_ring0 << 1; 1116 else 1117 ring->doorbell_index = adev->doorbell_index.gfx_ring1 << 1; 1118 ring->vm_hub = AMDGPU_GFXHUB(0); 1119 sprintf(ring->name, "gfx_%d.%d.%d", ring->me, ring->pipe, ring->queue); 1120 1121 irq_type = AMDGPU_CP_IRQ_GFX_ME0_PIPE0_EOP + ring->pipe; 1122 hw_prio = amdgpu_gfx_is_high_priority_graphics_queue(adev, ring) ? 1123 AMDGPU_GFX_PIPE_PRIO_HIGH : AMDGPU_GFX_PIPE_PRIO_NORMAL; 1124 return amdgpu_ring_init(adev, ring, 1024, &adev->gfx.eop_irq, irq_type, 1125 hw_prio, NULL); 1126 } 1127 1128 static int gfx_v11_0_compute_ring_init(struct amdgpu_device *adev, int ring_id, 1129 int mec, int pipe, int queue) 1130 { 1131 int r; 1132 unsigned irq_type; 1133 struct amdgpu_ring *ring; 1134 unsigned int hw_prio; 1135 1136 ring = &adev->gfx.compute_ring[ring_id]; 1137 1138 /* mec0 is me1 */ 1139 ring->me = mec + 1; 1140 ring->pipe = pipe; 1141 ring->queue = queue; 1142 1143 ring->ring_obj = NULL; 1144 ring->use_doorbell = true; 1145 ring->doorbell_index = (adev->doorbell_index.mec_ring0 + ring_id) << 1; 1146 ring->eop_gpu_addr = adev->gfx.mec.hpd_eop_gpu_addr 1147 + (ring_id * GFX11_MEC_HPD_SIZE); 1148 ring->vm_hub = AMDGPU_GFXHUB(0); 1149 sprintf(ring->name, "comp_%d.%d.%d", ring->me, ring->pipe, ring->queue); 1150 1151 irq_type = AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE0_EOP 1152 + ((ring->me - 1) * adev->gfx.mec.num_pipe_per_mec) 1153 + ring->pipe; 1154 hw_prio = amdgpu_gfx_is_high_priority_compute_queue(adev, ring) ? 1155 AMDGPU_GFX_PIPE_PRIO_HIGH : AMDGPU_GFX_PIPE_PRIO_NORMAL; 1156 /* type-2 packets are deprecated on MEC, use type-3 instead */ 1157 r = amdgpu_ring_init(adev, ring, 1024, &adev->gfx.eop_irq, irq_type, 1158 hw_prio, NULL); 1159 if (r) 1160 return r; 1161 1162 return 0; 1163 } 1164 1165 static struct { 1166 SOC21_FIRMWARE_ID id; 1167 unsigned int offset; 1168 unsigned int size; 1169 } rlc_autoload_info[SOC21_FIRMWARE_ID_MAX]; 1170 1171 static void gfx_v11_0_parse_rlc_toc(struct amdgpu_device *adev, void *rlc_toc) 1172 { 1173 RLC_TABLE_OF_CONTENT *ucode = rlc_toc; 1174 1175 while (ucode && (ucode->id > SOC21_FIRMWARE_ID_INVALID) && 1176 (ucode->id < SOC21_FIRMWARE_ID_MAX)) { 1177 rlc_autoload_info[ucode->id].id = ucode->id; 1178 rlc_autoload_info[ucode->id].offset = ucode->offset * 4; 1179 rlc_autoload_info[ucode->id].size = ucode->size * 4; 1180 1181 ucode++; 1182 } 1183 } 1184 1185 static uint32_t gfx_v11_0_calc_toc_total_size(struct amdgpu_device *adev) 1186 { 1187 uint32_t total_size = 0; 1188 SOC21_FIRMWARE_ID id; 1189 1190 gfx_v11_0_parse_rlc_toc(adev, adev->psp.toc.start_addr); 1191 1192 for (id = SOC21_FIRMWARE_ID_RLC_G_UCODE; id < SOC21_FIRMWARE_ID_MAX; id++) 1193 total_size += rlc_autoload_info[id].size; 1194 1195 /* In case the offset in rlc toc ucode is aligned */ 1196 if (total_size < rlc_autoload_info[SOC21_FIRMWARE_ID_MAX-1].offset) 1197 total_size = rlc_autoload_info[SOC21_FIRMWARE_ID_MAX-1].offset + 1198 rlc_autoload_info[SOC21_FIRMWARE_ID_MAX-1].size; 1199 1200 return total_size; 1201 } 1202 1203 static int gfx_v11_0_rlc_autoload_buffer_init(struct amdgpu_device *adev) 1204 { 1205 int r; 1206 uint32_t total_size; 1207 1208 total_size = gfx_v11_0_calc_toc_total_size(adev); 1209 1210 r = amdgpu_bo_create_reserved(adev, total_size, 64 * 1024, 1211 AMDGPU_GEM_DOMAIN_VRAM | 1212 AMDGPU_GEM_DOMAIN_GTT, 1213 &adev->gfx.rlc.rlc_autoload_bo, 1214 &adev->gfx.rlc.rlc_autoload_gpu_addr, 1215 (void **)&adev->gfx.rlc.rlc_autoload_ptr); 1216 1217 if (r) { 1218 dev_err(adev->dev, "(%d) failed to create fw autoload bo\n", r); 1219 return r; 1220 } 1221 1222 return 0; 1223 } 1224 1225 static void gfx_v11_0_rlc_backdoor_autoload_copy_ucode(struct amdgpu_device *adev, 1226 SOC21_FIRMWARE_ID id, 1227 const void *fw_data, 1228 uint32_t fw_size, 1229 uint32_t *fw_autoload_mask) 1230 { 1231 uint32_t toc_offset; 1232 uint32_t toc_fw_size; 1233 char *ptr = adev->gfx.rlc.rlc_autoload_ptr; 1234 1235 if (id <= SOC21_FIRMWARE_ID_INVALID || id >= SOC21_FIRMWARE_ID_MAX) 1236 return; 1237 1238 toc_offset = rlc_autoload_info[id].offset; 1239 toc_fw_size = rlc_autoload_info[id].size; 1240 1241 if (fw_size == 0) 1242 fw_size = toc_fw_size; 1243 1244 if (fw_size > toc_fw_size) 1245 fw_size = toc_fw_size; 1246 1247 memcpy(ptr + toc_offset, fw_data, fw_size); 1248 1249 if (fw_size < toc_fw_size) 1250 memset(ptr + toc_offset + fw_size, 0, toc_fw_size - fw_size); 1251 1252 if ((id != SOC21_FIRMWARE_ID_RS64_PFP) && (id != SOC21_FIRMWARE_ID_RS64_ME)) 1253 *(uint64_t *)fw_autoload_mask |= 1ULL << id; 1254 } 1255 1256 static void gfx_v11_0_rlc_backdoor_autoload_copy_toc_ucode(struct amdgpu_device *adev, 1257 uint32_t *fw_autoload_mask) 1258 { 1259 void *data; 1260 uint32_t size; 1261 uint64_t *toc_ptr; 1262 1263 *(uint64_t *)fw_autoload_mask |= 0x1; 1264 1265 DRM_DEBUG("rlc autoload enabled fw: 0x%llx\n", *(uint64_t *)fw_autoload_mask); 1266 1267 data = adev->psp.toc.start_addr; 1268 size = rlc_autoload_info[SOC21_FIRMWARE_ID_RLC_TOC].size; 1269 1270 toc_ptr = (uint64_t *)data + size / 8 - 1; 1271 *toc_ptr = *(uint64_t *)fw_autoload_mask; 1272 1273 gfx_v11_0_rlc_backdoor_autoload_copy_ucode(adev, SOC21_FIRMWARE_ID_RLC_TOC, 1274 data, size, fw_autoload_mask); 1275 } 1276 1277 static void gfx_v11_0_rlc_backdoor_autoload_copy_gfx_ucode(struct amdgpu_device *adev, 1278 uint32_t *fw_autoload_mask) 1279 { 1280 const __le32 *fw_data; 1281 uint32_t fw_size; 1282 const struct gfx_firmware_header_v1_0 *cp_hdr; 1283 const struct gfx_firmware_header_v2_0 *cpv2_hdr; 1284 const struct rlc_firmware_header_v2_0 *rlc_hdr; 1285 const struct rlc_firmware_header_v2_2 *rlcv22_hdr; 1286 uint16_t version_major, version_minor; 1287 1288 if (adev->gfx.rs64_enable) { 1289 /* pfp ucode */ 1290 cpv2_hdr = (const struct gfx_firmware_header_v2_0 *) 1291 adev->gfx.pfp_fw->data; 1292 /* instruction */ 1293 fw_data = (const __le32 *)(adev->gfx.pfp_fw->data + 1294 le32_to_cpu(cpv2_hdr->ucode_offset_bytes)); 1295 fw_size = le32_to_cpu(cpv2_hdr->ucode_size_bytes); 1296 gfx_v11_0_rlc_backdoor_autoload_copy_ucode(adev, SOC21_FIRMWARE_ID_RS64_PFP, 1297 fw_data, fw_size, fw_autoload_mask); 1298 /* data */ 1299 fw_data = (const __le32 *)(adev->gfx.pfp_fw->data + 1300 le32_to_cpu(cpv2_hdr->data_offset_bytes)); 1301 fw_size = le32_to_cpu(cpv2_hdr->data_size_bytes); 1302 gfx_v11_0_rlc_backdoor_autoload_copy_ucode(adev, SOC21_FIRMWARE_ID_RS64_PFP_P0_STACK, 1303 fw_data, fw_size, fw_autoload_mask); 1304 gfx_v11_0_rlc_backdoor_autoload_copy_ucode(adev, SOC21_FIRMWARE_ID_RS64_PFP_P1_STACK, 1305 fw_data, fw_size, fw_autoload_mask); 1306 /* me ucode */ 1307 cpv2_hdr = (const struct gfx_firmware_header_v2_0 *) 1308 adev->gfx.me_fw->data; 1309 /* instruction */ 1310 fw_data = (const __le32 *)(adev->gfx.me_fw->data + 1311 le32_to_cpu(cpv2_hdr->ucode_offset_bytes)); 1312 fw_size = le32_to_cpu(cpv2_hdr->ucode_size_bytes); 1313 gfx_v11_0_rlc_backdoor_autoload_copy_ucode(adev, SOC21_FIRMWARE_ID_RS64_ME, 1314 fw_data, fw_size, fw_autoload_mask); 1315 /* data */ 1316 fw_data = (const __le32 *)(adev->gfx.me_fw->data + 1317 le32_to_cpu(cpv2_hdr->data_offset_bytes)); 1318 fw_size = le32_to_cpu(cpv2_hdr->data_size_bytes); 1319 gfx_v11_0_rlc_backdoor_autoload_copy_ucode(adev, SOC21_FIRMWARE_ID_RS64_ME_P0_STACK, 1320 fw_data, fw_size, fw_autoload_mask); 1321 gfx_v11_0_rlc_backdoor_autoload_copy_ucode(adev, SOC21_FIRMWARE_ID_RS64_ME_P1_STACK, 1322 fw_data, fw_size, fw_autoload_mask); 1323 /* mec ucode */ 1324 cpv2_hdr = (const struct gfx_firmware_header_v2_0 *) 1325 adev->gfx.mec_fw->data; 1326 /* instruction */ 1327 fw_data = (const __le32 *) (adev->gfx.mec_fw->data + 1328 le32_to_cpu(cpv2_hdr->ucode_offset_bytes)); 1329 fw_size = le32_to_cpu(cpv2_hdr->ucode_size_bytes); 1330 gfx_v11_0_rlc_backdoor_autoload_copy_ucode(adev, SOC21_FIRMWARE_ID_RS64_MEC, 1331 fw_data, fw_size, fw_autoload_mask); 1332 /* data */ 1333 fw_data = (const __le32 *) (adev->gfx.mec_fw->data + 1334 le32_to_cpu(cpv2_hdr->data_offset_bytes)); 1335 fw_size = le32_to_cpu(cpv2_hdr->data_size_bytes); 1336 gfx_v11_0_rlc_backdoor_autoload_copy_ucode(adev, SOC21_FIRMWARE_ID_RS64_MEC_P0_STACK, 1337 fw_data, fw_size, fw_autoload_mask); 1338 gfx_v11_0_rlc_backdoor_autoload_copy_ucode(adev, SOC21_FIRMWARE_ID_RS64_MEC_P1_STACK, 1339 fw_data, fw_size, fw_autoload_mask); 1340 gfx_v11_0_rlc_backdoor_autoload_copy_ucode(adev, SOC21_FIRMWARE_ID_RS64_MEC_P2_STACK, 1341 fw_data, fw_size, fw_autoload_mask); 1342 gfx_v11_0_rlc_backdoor_autoload_copy_ucode(adev, SOC21_FIRMWARE_ID_RS64_MEC_P3_STACK, 1343 fw_data, fw_size, fw_autoload_mask); 1344 } else { 1345 /* pfp ucode */ 1346 cp_hdr = (const struct gfx_firmware_header_v1_0 *) 1347 adev->gfx.pfp_fw->data; 1348 fw_data = (const __le32 *)(adev->gfx.pfp_fw->data + 1349 le32_to_cpu(cp_hdr->header.ucode_array_offset_bytes)); 1350 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes); 1351 gfx_v11_0_rlc_backdoor_autoload_copy_ucode(adev, SOC21_FIRMWARE_ID_CP_PFP, 1352 fw_data, fw_size, fw_autoload_mask); 1353 1354 /* me ucode */ 1355 cp_hdr = (const struct gfx_firmware_header_v1_0 *) 1356 adev->gfx.me_fw->data; 1357 fw_data = (const __le32 *)(adev->gfx.me_fw->data + 1358 le32_to_cpu(cp_hdr->header.ucode_array_offset_bytes)); 1359 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes); 1360 gfx_v11_0_rlc_backdoor_autoload_copy_ucode(adev, SOC21_FIRMWARE_ID_CP_ME, 1361 fw_data, fw_size, fw_autoload_mask); 1362 1363 /* mec ucode */ 1364 cp_hdr = (const struct gfx_firmware_header_v1_0 *) 1365 adev->gfx.mec_fw->data; 1366 fw_data = (const __le32 *) (adev->gfx.mec_fw->data + 1367 le32_to_cpu(cp_hdr->header.ucode_array_offset_bytes)); 1368 fw_size = le32_to_cpu(cp_hdr->header.ucode_size_bytes) - 1369 cp_hdr->jt_size * 4; 1370 gfx_v11_0_rlc_backdoor_autoload_copy_ucode(adev, SOC21_FIRMWARE_ID_CP_MEC, 1371 fw_data, fw_size, fw_autoload_mask); 1372 } 1373 1374 /* rlc ucode */ 1375 rlc_hdr = (const struct rlc_firmware_header_v2_0 *) 1376 adev->gfx.rlc_fw->data; 1377 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data + 1378 le32_to_cpu(rlc_hdr->header.ucode_array_offset_bytes)); 1379 fw_size = le32_to_cpu(rlc_hdr->header.ucode_size_bytes); 1380 gfx_v11_0_rlc_backdoor_autoload_copy_ucode(adev, SOC21_FIRMWARE_ID_RLC_G_UCODE, 1381 fw_data, fw_size, fw_autoload_mask); 1382 1383 version_major = le16_to_cpu(rlc_hdr->header.header_version_major); 1384 version_minor = le16_to_cpu(rlc_hdr->header.header_version_minor); 1385 if (version_major == 2) { 1386 if (version_minor >= 2) { 1387 rlcv22_hdr = (const struct rlc_firmware_header_v2_2 *)adev->gfx.rlc_fw->data; 1388 1389 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data + 1390 le32_to_cpu(rlcv22_hdr->rlc_iram_ucode_offset_bytes)); 1391 fw_size = le32_to_cpu(rlcv22_hdr->rlc_iram_ucode_size_bytes); 1392 gfx_v11_0_rlc_backdoor_autoload_copy_ucode(adev, SOC21_FIRMWARE_ID_RLX6_UCODE, 1393 fw_data, fw_size, fw_autoload_mask); 1394 1395 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data + 1396 le32_to_cpu(rlcv22_hdr->rlc_dram_ucode_offset_bytes)); 1397 fw_size = le32_to_cpu(rlcv22_hdr->rlc_dram_ucode_size_bytes); 1398 gfx_v11_0_rlc_backdoor_autoload_copy_ucode(adev, SOC21_FIRMWARE_ID_RLX6_DRAM_BOOT, 1399 fw_data, fw_size, fw_autoload_mask); 1400 } 1401 } 1402 } 1403 1404 static void gfx_v11_0_rlc_backdoor_autoload_copy_sdma_ucode(struct amdgpu_device *adev, 1405 uint32_t *fw_autoload_mask) 1406 { 1407 const __le32 *fw_data; 1408 uint32_t fw_size; 1409 const struct sdma_firmware_header_v2_0 *sdma_hdr; 1410 1411 sdma_hdr = (const struct sdma_firmware_header_v2_0 *) 1412 adev->sdma.instance[0].fw->data; 1413 fw_data = (const __le32 *) (adev->sdma.instance[0].fw->data + 1414 le32_to_cpu(sdma_hdr->header.ucode_array_offset_bytes)); 1415 fw_size = le32_to_cpu(sdma_hdr->ctx_ucode_size_bytes); 1416 1417 gfx_v11_0_rlc_backdoor_autoload_copy_ucode(adev, 1418 SOC21_FIRMWARE_ID_SDMA_UCODE_TH0, fw_data, fw_size, fw_autoload_mask); 1419 1420 fw_data = (const __le32 *) (adev->sdma.instance[0].fw->data + 1421 le32_to_cpu(sdma_hdr->ctl_ucode_offset)); 1422 fw_size = le32_to_cpu(sdma_hdr->ctl_ucode_size_bytes); 1423 1424 gfx_v11_0_rlc_backdoor_autoload_copy_ucode(adev, 1425 SOC21_FIRMWARE_ID_SDMA_UCODE_TH1, fw_data, fw_size, fw_autoload_mask); 1426 } 1427 1428 static void gfx_v11_0_rlc_backdoor_autoload_copy_mes_ucode(struct amdgpu_device *adev, 1429 uint32_t *fw_autoload_mask) 1430 { 1431 const __le32 *fw_data; 1432 unsigned fw_size; 1433 const struct mes_firmware_header_v1_0 *mes_hdr; 1434 int pipe, ucode_id, data_id; 1435 1436 for (pipe = 0; pipe < 2; pipe++) { 1437 if (pipe==0) { 1438 ucode_id = SOC21_FIRMWARE_ID_RS64_MES_P0; 1439 data_id = SOC21_FIRMWARE_ID_RS64_MES_P0_STACK; 1440 } else { 1441 ucode_id = SOC21_FIRMWARE_ID_RS64_MES_P1; 1442 data_id = SOC21_FIRMWARE_ID_RS64_MES_P1_STACK; 1443 } 1444 1445 mes_hdr = (const struct mes_firmware_header_v1_0 *) 1446 adev->mes.fw[pipe]->data; 1447 1448 fw_data = (const __le32 *)(adev->mes.fw[pipe]->data + 1449 le32_to_cpu(mes_hdr->mes_ucode_offset_bytes)); 1450 fw_size = le32_to_cpu(mes_hdr->mes_ucode_size_bytes); 1451 1452 gfx_v11_0_rlc_backdoor_autoload_copy_ucode(adev, 1453 ucode_id, fw_data, fw_size, fw_autoload_mask); 1454 1455 fw_data = (const __le32 *)(adev->mes.fw[pipe]->data + 1456 le32_to_cpu(mes_hdr->mes_ucode_data_offset_bytes)); 1457 fw_size = le32_to_cpu(mes_hdr->mes_ucode_data_size_bytes); 1458 1459 gfx_v11_0_rlc_backdoor_autoload_copy_ucode(adev, 1460 data_id, fw_data, fw_size, fw_autoload_mask); 1461 } 1462 } 1463 1464 static int gfx_v11_0_rlc_backdoor_autoload_enable(struct amdgpu_device *adev) 1465 { 1466 uint32_t rlc_g_offset, rlc_g_size; 1467 uint64_t gpu_addr; 1468 uint32_t autoload_fw_id[2]; 1469 1470 memset(autoload_fw_id, 0, sizeof(uint32_t) * 2); 1471 1472 /* RLC autoload sequence 2: copy ucode */ 1473 gfx_v11_0_rlc_backdoor_autoload_copy_sdma_ucode(adev, autoload_fw_id); 1474 gfx_v11_0_rlc_backdoor_autoload_copy_gfx_ucode(adev, autoload_fw_id); 1475 gfx_v11_0_rlc_backdoor_autoload_copy_mes_ucode(adev, autoload_fw_id); 1476 gfx_v11_0_rlc_backdoor_autoload_copy_toc_ucode(adev, autoload_fw_id); 1477 1478 rlc_g_offset = rlc_autoload_info[SOC21_FIRMWARE_ID_RLC_G_UCODE].offset; 1479 rlc_g_size = rlc_autoload_info[SOC21_FIRMWARE_ID_RLC_G_UCODE].size; 1480 gpu_addr = adev->gfx.rlc.rlc_autoload_gpu_addr + rlc_g_offset; 1481 1482 WREG32_SOC15(GC, 0, regGFX_IMU_RLC_BOOTLOADER_ADDR_HI, upper_32_bits(gpu_addr)); 1483 WREG32_SOC15(GC, 0, regGFX_IMU_RLC_BOOTLOADER_ADDR_LO, lower_32_bits(gpu_addr)); 1484 1485 WREG32_SOC15(GC, 0, regGFX_IMU_RLC_BOOTLOADER_SIZE, rlc_g_size); 1486 1487 /* RLC autoload sequence 3: load IMU fw */ 1488 if (adev->gfx.imu.funcs->load_microcode) 1489 adev->gfx.imu.funcs->load_microcode(adev); 1490 /* RLC autoload sequence 4 init IMU fw */ 1491 if (adev->gfx.imu.funcs->setup_imu) 1492 adev->gfx.imu.funcs->setup_imu(adev); 1493 if (adev->gfx.imu.funcs->start_imu) 1494 adev->gfx.imu.funcs->start_imu(adev); 1495 1496 /* RLC autoload sequence 5 disable gpa mode */ 1497 gfx_v11_0_disable_gpa_mode(adev); 1498 1499 return 0; 1500 } 1501 1502 static void gfx_v11_0_alloc_ip_dump(struct amdgpu_device *adev) 1503 { 1504 uint32_t reg_count = ARRAY_SIZE(gc_reg_list_11_0); 1505 uint32_t *ptr; 1506 uint32_t inst; 1507 1508 ptr = kcalloc(reg_count, sizeof(uint32_t), GFP_KERNEL); 1509 if (!ptr) { 1510 DRM_ERROR("Failed to allocate memory for GFX IP Dump\n"); 1511 adev->gfx.ip_dump_core = NULL; 1512 } else { 1513 adev->gfx.ip_dump_core = ptr; 1514 } 1515 1516 /* Allocate memory for compute queue registers for all the instances */ 1517 reg_count = ARRAY_SIZE(gc_cp_reg_list_11); 1518 inst = adev->gfx.mec.num_mec * adev->gfx.mec.num_pipe_per_mec * 1519 adev->gfx.mec.num_queue_per_pipe; 1520 1521 ptr = kcalloc(reg_count * inst, sizeof(uint32_t), GFP_KERNEL); 1522 if (!ptr) { 1523 DRM_ERROR("Failed to allocate memory for Compute Queues IP Dump\n"); 1524 adev->gfx.ip_dump_compute_queues = NULL; 1525 } else { 1526 adev->gfx.ip_dump_compute_queues = ptr; 1527 } 1528 1529 /* Allocate memory for gfx queue registers for all the instances */ 1530 reg_count = ARRAY_SIZE(gc_gfx_queue_reg_list_11); 1531 inst = adev->gfx.me.num_me * adev->gfx.me.num_pipe_per_me * 1532 adev->gfx.me.num_queue_per_pipe; 1533 1534 ptr = kcalloc(reg_count * inst, sizeof(uint32_t), GFP_KERNEL); 1535 if (!ptr) { 1536 DRM_ERROR("Failed to allocate memory for GFX Queues IP Dump\n"); 1537 adev->gfx.ip_dump_gfx_queues = NULL; 1538 } else { 1539 adev->gfx.ip_dump_gfx_queues = ptr; 1540 } 1541 } 1542 1543 static int gfx_v11_0_sw_init(struct amdgpu_ip_block *ip_block) 1544 { 1545 int i, j, k, r, ring_id = 0; 1546 int xcc_id = 0; 1547 struct amdgpu_device *adev = ip_block->adev; 1548 1549 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 1550 case IP_VERSION(11, 0, 0): 1551 case IP_VERSION(11, 0, 2): 1552 case IP_VERSION(11, 0, 3): 1553 adev->gfx.me.num_me = 1; 1554 adev->gfx.me.num_pipe_per_me = 1; 1555 adev->gfx.me.num_queue_per_pipe = 1; 1556 adev->gfx.mec.num_mec = 2; 1557 adev->gfx.mec.num_pipe_per_mec = 4; 1558 adev->gfx.mec.num_queue_per_pipe = 4; 1559 break; 1560 case IP_VERSION(11, 0, 1): 1561 case IP_VERSION(11, 0, 4): 1562 case IP_VERSION(11, 5, 0): 1563 case IP_VERSION(11, 5, 1): 1564 case IP_VERSION(11, 5, 2): 1565 adev->gfx.me.num_me = 1; 1566 adev->gfx.me.num_pipe_per_me = 1; 1567 adev->gfx.me.num_queue_per_pipe = 1; 1568 adev->gfx.mec.num_mec = 1; 1569 adev->gfx.mec.num_pipe_per_mec = 4; 1570 adev->gfx.mec.num_queue_per_pipe = 4; 1571 break; 1572 default: 1573 adev->gfx.me.num_me = 1; 1574 adev->gfx.me.num_pipe_per_me = 1; 1575 adev->gfx.me.num_queue_per_pipe = 1; 1576 adev->gfx.mec.num_mec = 1; 1577 adev->gfx.mec.num_pipe_per_mec = 4; 1578 adev->gfx.mec.num_queue_per_pipe = 8; 1579 break; 1580 } 1581 1582 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 1583 case IP_VERSION(11, 0, 0): 1584 case IP_VERSION(11, 0, 2): 1585 case IP_VERSION(11, 0, 3): 1586 adev->gfx.cleaner_shader_ptr = gfx_11_0_3_cleaner_shader_hex; 1587 adev->gfx.cleaner_shader_size = sizeof(gfx_11_0_3_cleaner_shader_hex); 1588 if (adev->gfx.me_fw_version >= 2280 && 1589 adev->gfx.pfp_fw_version >= 2370 && 1590 adev->gfx.mec_fw_version >= 2450 && 1591 adev->mes.fw_version[0] >= 99) { 1592 adev->gfx.enable_cleaner_shader = true; 1593 r = amdgpu_gfx_cleaner_shader_sw_init(adev, adev->gfx.cleaner_shader_size); 1594 if (r) { 1595 adev->gfx.enable_cleaner_shader = false; 1596 dev_err(adev->dev, "Failed to initialize cleaner shader\n"); 1597 } 1598 } 1599 break; 1600 default: 1601 adev->gfx.enable_cleaner_shader = false; 1602 break; 1603 } 1604 1605 /* Enable CG flag in one VF mode for enabling RLC safe mode enter/exit */ 1606 if (amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(11, 0, 3) && 1607 amdgpu_sriov_is_pp_one_vf(adev)) 1608 adev->cg_flags = AMD_CG_SUPPORT_GFX_CGCG; 1609 1610 /* EOP Event */ 1611 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GRBM_CP, 1612 GFX_11_0_0__SRCID__CP_EOP_INTERRUPT, 1613 &adev->gfx.eop_irq); 1614 if (r) 1615 return r; 1616 1617 /* Bad opcode Event */ 1618 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GRBM_CP, 1619 GFX_11_0_0__SRCID__CP_BAD_OPCODE_ERROR, 1620 &adev->gfx.bad_op_irq); 1621 if (r) 1622 return r; 1623 1624 /* Privileged reg */ 1625 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GRBM_CP, 1626 GFX_11_0_0__SRCID__CP_PRIV_REG_FAULT, 1627 &adev->gfx.priv_reg_irq); 1628 if (r) 1629 return r; 1630 1631 /* Privileged inst */ 1632 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GRBM_CP, 1633 GFX_11_0_0__SRCID__CP_PRIV_INSTR_FAULT, 1634 &adev->gfx.priv_inst_irq); 1635 if (r) 1636 return r; 1637 1638 /* FED error */ 1639 r = amdgpu_irq_add_id(adev, SOC21_IH_CLIENTID_GFX, 1640 GFX_11_0_0__SRCID__RLC_GC_FED_INTERRUPT, 1641 &adev->gfx.rlc_gc_fed_irq); 1642 if (r) 1643 return r; 1644 1645 adev->gfx.gfx_current_status = AMDGPU_GFX_NORMAL_MODE; 1646 1647 gfx_v11_0_me_init(adev); 1648 1649 r = gfx_v11_0_rlc_init(adev); 1650 if (r) { 1651 DRM_ERROR("Failed to init rlc BOs!\n"); 1652 return r; 1653 } 1654 1655 r = gfx_v11_0_mec_init(adev); 1656 if (r) { 1657 DRM_ERROR("Failed to init MEC BOs!\n"); 1658 return r; 1659 } 1660 1661 /* set up the gfx ring */ 1662 for (i = 0; i < adev->gfx.me.num_me; i++) { 1663 for (j = 0; j < adev->gfx.me.num_queue_per_pipe; j++) { 1664 for (k = 0; k < adev->gfx.me.num_pipe_per_me; k++) { 1665 if (!amdgpu_gfx_is_me_queue_enabled(adev, i, k, j)) 1666 continue; 1667 1668 r = gfx_v11_0_gfx_ring_init(adev, ring_id, 1669 i, k, j); 1670 if (r) 1671 return r; 1672 ring_id++; 1673 } 1674 } 1675 } 1676 1677 ring_id = 0; 1678 /* set up the compute queues - allocate horizontally across pipes */ 1679 for (i = 0; i < adev->gfx.mec.num_mec; ++i) { 1680 for (j = 0; j < adev->gfx.mec.num_queue_per_pipe; j++) { 1681 for (k = 0; k < adev->gfx.mec.num_pipe_per_mec; k++) { 1682 if (!amdgpu_gfx_is_mec_queue_enabled(adev, 0, i, 1683 k, j)) 1684 continue; 1685 1686 r = gfx_v11_0_compute_ring_init(adev, ring_id, 1687 i, k, j); 1688 if (r) 1689 return r; 1690 1691 ring_id++; 1692 } 1693 } 1694 } 1695 1696 adev->gfx.gfx_supported_reset = 1697 amdgpu_get_soft_full_reset_mask(&adev->gfx.gfx_ring[0]); 1698 adev->gfx.compute_supported_reset = 1699 amdgpu_get_soft_full_reset_mask(&adev->gfx.compute_ring[0]); 1700 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 1701 case IP_VERSION(11, 0, 0): 1702 case IP_VERSION(11, 0, 2): 1703 case IP_VERSION(11, 0, 3): 1704 if ((adev->gfx.me_fw_version >= 2280) && 1705 (adev->gfx.mec_fw_version >= 2410)) { 1706 adev->gfx.compute_supported_reset |= AMDGPU_RESET_TYPE_PER_QUEUE; 1707 adev->gfx.gfx_supported_reset |= AMDGPU_RESET_TYPE_PER_QUEUE; 1708 } 1709 break; 1710 default: 1711 break; 1712 } 1713 1714 if (!adev->enable_mes_kiq) { 1715 r = amdgpu_gfx_kiq_init(adev, GFX11_MEC_HPD_SIZE, 0); 1716 if (r) { 1717 DRM_ERROR("Failed to init KIQ BOs!\n"); 1718 return r; 1719 } 1720 1721 r = amdgpu_gfx_kiq_init_ring(adev, xcc_id); 1722 if (r) 1723 return r; 1724 } 1725 1726 r = amdgpu_gfx_mqd_sw_init(adev, sizeof(struct v11_compute_mqd), 0); 1727 if (r) 1728 return r; 1729 1730 /* allocate visible FB for rlc auto-loading fw */ 1731 if (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) { 1732 r = gfx_v11_0_rlc_autoload_buffer_init(adev); 1733 if (r) 1734 return r; 1735 } 1736 1737 r = gfx_v11_0_gpu_early_init(adev); 1738 if (r) 1739 return r; 1740 1741 if (amdgpu_gfx_ras_sw_init(adev)) { 1742 dev_err(adev->dev, "Failed to initialize gfx ras block!\n"); 1743 return -EINVAL; 1744 } 1745 1746 gfx_v11_0_alloc_ip_dump(adev); 1747 1748 r = amdgpu_gfx_sysfs_init(adev); 1749 if (r) 1750 return r; 1751 1752 return 0; 1753 } 1754 1755 static void gfx_v11_0_pfp_fini(struct amdgpu_device *adev) 1756 { 1757 amdgpu_bo_free_kernel(&adev->gfx.pfp.pfp_fw_obj, 1758 &adev->gfx.pfp.pfp_fw_gpu_addr, 1759 (void **)&adev->gfx.pfp.pfp_fw_ptr); 1760 1761 amdgpu_bo_free_kernel(&adev->gfx.pfp.pfp_fw_data_obj, 1762 &adev->gfx.pfp.pfp_fw_data_gpu_addr, 1763 (void **)&adev->gfx.pfp.pfp_fw_data_ptr); 1764 } 1765 1766 static void gfx_v11_0_me_fini(struct amdgpu_device *adev) 1767 { 1768 amdgpu_bo_free_kernel(&adev->gfx.me.me_fw_obj, 1769 &adev->gfx.me.me_fw_gpu_addr, 1770 (void **)&adev->gfx.me.me_fw_ptr); 1771 1772 amdgpu_bo_free_kernel(&adev->gfx.me.me_fw_data_obj, 1773 &adev->gfx.me.me_fw_data_gpu_addr, 1774 (void **)&adev->gfx.me.me_fw_data_ptr); 1775 } 1776 1777 static void gfx_v11_0_rlc_autoload_buffer_fini(struct amdgpu_device *adev) 1778 { 1779 amdgpu_bo_free_kernel(&adev->gfx.rlc.rlc_autoload_bo, 1780 &adev->gfx.rlc.rlc_autoload_gpu_addr, 1781 (void **)&adev->gfx.rlc.rlc_autoload_ptr); 1782 } 1783 1784 static int gfx_v11_0_sw_fini(struct amdgpu_ip_block *ip_block) 1785 { 1786 int i; 1787 struct amdgpu_device *adev = ip_block->adev; 1788 1789 for (i = 0; i < adev->gfx.num_gfx_rings; i++) 1790 amdgpu_ring_fini(&adev->gfx.gfx_ring[i]); 1791 for (i = 0; i < adev->gfx.num_compute_rings; i++) 1792 amdgpu_ring_fini(&adev->gfx.compute_ring[i]); 1793 1794 amdgpu_gfx_mqd_sw_fini(adev, 0); 1795 1796 if (!adev->enable_mes_kiq) { 1797 amdgpu_gfx_kiq_free_ring(&adev->gfx.kiq[0].ring); 1798 amdgpu_gfx_kiq_fini(adev, 0); 1799 } 1800 1801 amdgpu_gfx_cleaner_shader_sw_fini(adev); 1802 1803 gfx_v11_0_pfp_fini(adev); 1804 gfx_v11_0_me_fini(adev); 1805 gfx_v11_0_rlc_fini(adev); 1806 gfx_v11_0_mec_fini(adev); 1807 1808 if (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) 1809 gfx_v11_0_rlc_autoload_buffer_fini(adev); 1810 1811 gfx_v11_0_free_microcode(adev); 1812 1813 amdgpu_gfx_sysfs_fini(adev); 1814 1815 kfree(adev->gfx.ip_dump_core); 1816 kfree(adev->gfx.ip_dump_compute_queues); 1817 kfree(adev->gfx.ip_dump_gfx_queues); 1818 1819 return 0; 1820 } 1821 1822 static void gfx_v11_0_select_se_sh(struct amdgpu_device *adev, u32 se_num, 1823 u32 sh_num, u32 instance, int xcc_id) 1824 { 1825 u32 data; 1826 1827 if (instance == 0xffffffff) 1828 data = REG_SET_FIELD(0, GRBM_GFX_INDEX, 1829 INSTANCE_BROADCAST_WRITES, 1); 1830 else 1831 data = REG_SET_FIELD(0, GRBM_GFX_INDEX, INSTANCE_INDEX, 1832 instance); 1833 1834 if (se_num == 0xffffffff) 1835 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SE_BROADCAST_WRITES, 1836 1); 1837 else 1838 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SE_INDEX, se_num); 1839 1840 if (sh_num == 0xffffffff) 1841 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SA_BROADCAST_WRITES, 1842 1); 1843 else 1844 data = REG_SET_FIELD(data, GRBM_GFX_INDEX, SA_INDEX, sh_num); 1845 1846 WREG32_SOC15(GC, 0, regGRBM_GFX_INDEX, data); 1847 } 1848 1849 static u32 gfx_v11_0_get_sa_active_bitmap(struct amdgpu_device *adev) 1850 { 1851 u32 gc_disabled_sa_mask, gc_user_disabled_sa_mask, sa_mask; 1852 1853 gc_disabled_sa_mask = RREG32_SOC15(GC, 0, regCC_GC_SA_UNIT_DISABLE); 1854 gc_disabled_sa_mask = REG_GET_FIELD(gc_disabled_sa_mask, 1855 CC_GC_SA_UNIT_DISABLE, 1856 SA_DISABLE); 1857 gc_user_disabled_sa_mask = RREG32_SOC15(GC, 0, regGC_USER_SA_UNIT_DISABLE); 1858 gc_user_disabled_sa_mask = REG_GET_FIELD(gc_user_disabled_sa_mask, 1859 GC_USER_SA_UNIT_DISABLE, 1860 SA_DISABLE); 1861 sa_mask = amdgpu_gfx_create_bitmask(adev->gfx.config.max_sh_per_se * 1862 adev->gfx.config.max_shader_engines); 1863 1864 return sa_mask & (~(gc_disabled_sa_mask | gc_user_disabled_sa_mask)); 1865 } 1866 1867 static u32 gfx_v11_0_get_rb_active_bitmap(struct amdgpu_device *adev) 1868 { 1869 u32 gc_disabled_rb_mask, gc_user_disabled_rb_mask; 1870 u32 rb_mask; 1871 1872 gc_disabled_rb_mask = RREG32_SOC15(GC, 0, regCC_RB_BACKEND_DISABLE); 1873 gc_disabled_rb_mask = REG_GET_FIELD(gc_disabled_rb_mask, 1874 CC_RB_BACKEND_DISABLE, 1875 BACKEND_DISABLE); 1876 gc_user_disabled_rb_mask = RREG32_SOC15(GC, 0, regGC_USER_RB_BACKEND_DISABLE); 1877 gc_user_disabled_rb_mask = REG_GET_FIELD(gc_user_disabled_rb_mask, 1878 GC_USER_RB_BACKEND_DISABLE, 1879 BACKEND_DISABLE); 1880 rb_mask = amdgpu_gfx_create_bitmask(adev->gfx.config.max_backends_per_se * 1881 adev->gfx.config.max_shader_engines); 1882 1883 return rb_mask & (~(gc_disabled_rb_mask | gc_user_disabled_rb_mask)); 1884 } 1885 1886 static void gfx_v11_0_setup_rb(struct amdgpu_device *adev) 1887 { 1888 u32 rb_bitmap_width_per_sa; 1889 u32 max_sa; 1890 u32 active_sa_bitmap; 1891 u32 global_active_rb_bitmap; 1892 u32 active_rb_bitmap = 0; 1893 u32 i; 1894 1895 /* query sa bitmap from SA_UNIT_DISABLE registers */ 1896 active_sa_bitmap = gfx_v11_0_get_sa_active_bitmap(adev); 1897 /* query rb bitmap from RB_BACKEND_DISABLE registers */ 1898 global_active_rb_bitmap = gfx_v11_0_get_rb_active_bitmap(adev); 1899 1900 /* generate active rb bitmap according to active sa bitmap */ 1901 max_sa = adev->gfx.config.max_shader_engines * 1902 adev->gfx.config.max_sh_per_se; 1903 rb_bitmap_width_per_sa = adev->gfx.config.max_backends_per_se / 1904 adev->gfx.config.max_sh_per_se; 1905 for (i = 0; i < max_sa; i++) { 1906 if (active_sa_bitmap & (1 << i)) 1907 active_rb_bitmap |= (0x3 << (i * rb_bitmap_width_per_sa)); 1908 } 1909 1910 active_rb_bitmap &= global_active_rb_bitmap; 1911 adev->gfx.config.backend_enable_mask = active_rb_bitmap; 1912 adev->gfx.config.num_rbs = hweight32(active_rb_bitmap); 1913 } 1914 1915 #define DEFAULT_SH_MEM_BASES (0x6000) 1916 #define LDS_APP_BASE 0x1 1917 #define SCRATCH_APP_BASE 0x2 1918 1919 static void gfx_v11_0_init_compute_vmid(struct amdgpu_device *adev) 1920 { 1921 int i; 1922 uint32_t sh_mem_bases; 1923 uint32_t data; 1924 1925 /* 1926 * Configure apertures: 1927 * LDS: 0x60000000'00000000 - 0x60000001'00000000 (4GB) 1928 * Scratch: 0x60000001'00000000 - 0x60000002'00000000 (4GB) 1929 * GPUVM: 0x60010000'00000000 - 0x60020000'00000000 (1TB) 1930 */ 1931 sh_mem_bases = (LDS_APP_BASE << SH_MEM_BASES__SHARED_BASE__SHIFT) | 1932 SCRATCH_APP_BASE; 1933 1934 mutex_lock(&adev->srbm_mutex); 1935 for (i = adev->vm_manager.first_kfd_vmid; i < AMDGPU_NUM_VMID; i++) { 1936 soc21_grbm_select(adev, 0, 0, 0, i); 1937 /* CP and shaders */ 1938 WREG32_SOC15(GC, 0, regSH_MEM_CONFIG, DEFAULT_SH_MEM_CONFIG); 1939 WREG32_SOC15(GC, 0, regSH_MEM_BASES, sh_mem_bases); 1940 1941 /* Enable trap for each kfd vmid. */ 1942 data = RREG32_SOC15(GC, 0, regSPI_GDBG_PER_VMID_CNTL); 1943 data = REG_SET_FIELD(data, SPI_GDBG_PER_VMID_CNTL, TRAP_EN, 1); 1944 WREG32_SOC15(GC, 0, regSPI_GDBG_PER_VMID_CNTL, data); 1945 } 1946 soc21_grbm_select(adev, 0, 0, 0, 0); 1947 mutex_unlock(&adev->srbm_mutex); 1948 1949 /* 1950 * Initialize all compute VMIDs to have no GDS, GWS, or OA 1951 * access. These should be enabled by FW for target VMIDs. 1952 */ 1953 for (i = adev->vm_manager.first_kfd_vmid; i < AMDGPU_NUM_VMID; i++) { 1954 WREG32_SOC15_OFFSET(GC, 0, regGDS_VMID0_BASE, 2 * i, 0); 1955 WREG32_SOC15_OFFSET(GC, 0, regGDS_VMID0_SIZE, 2 * i, 0); 1956 WREG32_SOC15_OFFSET(GC, 0, regGDS_GWS_VMID0, i, 0); 1957 WREG32_SOC15_OFFSET(GC, 0, regGDS_OA_VMID0, i, 0); 1958 } 1959 } 1960 1961 static void gfx_v11_0_init_gds_vmid(struct amdgpu_device *adev) 1962 { 1963 int vmid; 1964 1965 /* 1966 * Initialize all compute and user-gfx VMIDs to have no GDS, GWS, or OA 1967 * access. Compute VMIDs should be enabled by FW for target VMIDs, 1968 * the driver can enable them for graphics. VMID0 should maintain 1969 * access so that HWS firmware can save/restore entries. 1970 */ 1971 for (vmid = 1; vmid < 16; vmid++) { 1972 WREG32_SOC15_OFFSET(GC, 0, regGDS_VMID0_BASE, 2 * vmid, 0); 1973 WREG32_SOC15_OFFSET(GC, 0, regGDS_VMID0_SIZE, 2 * vmid, 0); 1974 WREG32_SOC15_OFFSET(GC, 0, regGDS_GWS_VMID0, vmid, 0); 1975 WREG32_SOC15_OFFSET(GC, 0, regGDS_OA_VMID0, vmid, 0); 1976 } 1977 } 1978 1979 static void gfx_v11_0_tcp_harvest(struct amdgpu_device *adev) 1980 { 1981 /* TODO: harvest feature to be added later. */ 1982 } 1983 1984 static void gfx_v11_0_get_tcc_info(struct amdgpu_device *adev) 1985 { 1986 /* TCCs are global (not instanced). */ 1987 uint32_t tcc_disable = RREG32_SOC15(GC, 0, regCGTS_TCC_DISABLE) | 1988 RREG32_SOC15(GC, 0, regCGTS_USER_TCC_DISABLE); 1989 1990 adev->gfx.config.tcc_disabled_mask = 1991 REG_GET_FIELD(tcc_disable, CGTS_TCC_DISABLE, TCC_DISABLE) | 1992 (REG_GET_FIELD(tcc_disable, CGTS_TCC_DISABLE, HI_TCC_DISABLE) << 16); 1993 } 1994 1995 static void gfx_v11_0_constants_init(struct amdgpu_device *adev) 1996 { 1997 u32 tmp; 1998 int i; 1999 2000 if (!amdgpu_sriov_vf(adev)) 2001 WREG32_FIELD15_PREREG(GC, 0, GRBM_CNTL, READ_TIMEOUT, 0xff); 2002 2003 gfx_v11_0_setup_rb(adev); 2004 gfx_v11_0_get_cu_info(adev, &adev->gfx.cu_info); 2005 gfx_v11_0_get_tcc_info(adev); 2006 adev->gfx.config.pa_sc_tile_steering_override = 0; 2007 2008 /* Set whether texture coordinate truncation is conformant. */ 2009 tmp = RREG32_SOC15(GC, 0, regTA_CNTL2); 2010 adev->gfx.config.ta_cntl2_truncate_coord_mode = 2011 REG_GET_FIELD(tmp, TA_CNTL2, TRUNCATE_COORD_MODE); 2012 2013 /* XXX SH_MEM regs */ 2014 /* where to put LDS, scratch, GPUVM in FSA64 space */ 2015 mutex_lock(&adev->srbm_mutex); 2016 for (i = 0; i < adev->vm_manager.id_mgr[AMDGPU_GFXHUB(0)].num_ids; i++) { 2017 soc21_grbm_select(adev, 0, 0, 0, i); 2018 /* CP and shaders */ 2019 WREG32_SOC15(GC, 0, regSH_MEM_CONFIG, DEFAULT_SH_MEM_CONFIG); 2020 if (i != 0) { 2021 tmp = REG_SET_FIELD(0, SH_MEM_BASES, PRIVATE_BASE, 2022 (adev->gmc.private_aperture_start >> 48)); 2023 tmp = REG_SET_FIELD(tmp, SH_MEM_BASES, SHARED_BASE, 2024 (adev->gmc.shared_aperture_start >> 48)); 2025 WREG32_SOC15(GC, 0, regSH_MEM_BASES, tmp); 2026 } 2027 } 2028 soc21_grbm_select(adev, 0, 0, 0, 0); 2029 2030 mutex_unlock(&adev->srbm_mutex); 2031 2032 gfx_v11_0_init_compute_vmid(adev); 2033 gfx_v11_0_init_gds_vmid(adev); 2034 } 2035 2036 static u32 gfx_v11_0_get_cpg_int_cntl(struct amdgpu_device *adev, 2037 int me, int pipe) 2038 { 2039 if (me != 0) 2040 return 0; 2041 2042 switch (pipe) { 2043 case 0: 2044 return SOC15_REG_OFFSET(GC, 0, regCP_INT_CNTL_RING0); 2045 case 1: 2046 return SOC15_REG_OFFSET(GC, 0, regCP_INT_CNTL_RING1); 2047 default: 2048 return 0; 2049 } 2050 } 2051 2052 static u32 gfx_v11_0_get_cpc_int_cntl(struct amdgpu_device *adev, 2053 int me, int pipe) 2054 { 2055 /* 2056 * amdgpu controls only the first MEC. That's why this function only 2057 * handles the setting of interrupts for this specific MEC. All other 2058 * pipes' interrupts are set by amdkfd. 2059 */ 2060 if (me != 1) 2061 return 0; 2062 2063 switch (pipe) { 2064 case 0: 2065 return SOC15_REG_OFFSET(GC, 0, regCP_ME1_PIPE0_INT_CNTL); 2066 case 1: 2067 return SOC15_REG_OFFSET(GC, 0, regCP_ME1_PIPE1_INT_CNTL); 2068 case 2: 2069 return SOC15_REG_OFFSET(GC, 0, regCP_ME1_PIPE2_INT_CNTL); 2070 case 3: 2071 return SOC15_REG_OFFSET(GC, 0, regCP_ME1_PIPE3_INT_CNTL); 2072 default: 2073 return 0; 2074 } 2075 } 2076 2077 static void gfx_v11_0_enable_gui_idle_interrupt(struct amdgpu_device *adev, 2078 bool enable) 2079 { 2080 u32 tmp, cp_int_cntl_reg; 2081 int i, j; 2082 2083 if (amdgpu_sriov_vf(adev)) 2084 return; 2085 2086 for (i = 0; i < adev->gfx.me.num_me; i++) { 2087 for (j = 0; j < adev->gfx.me.num_pipe_per_me; j++) { 2088 cp_int_cntl_reg = gfx_v11_0_get_cpg_int_cntl(adev, i, j); 2089 2090 if (cp_int_cntl_reg) { 2091 tmp = RREG32_SOC15_IP(GC, cp_int_cntl_reg); 2092 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL_RING0, CNTX_BUSY_INT_ENABLE, 2093 enable ? 1 : 0); 2094 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL_RING0, CNTX_EMPTY_INT_ENABLE, 2095 enable ? 1 : 0); 2096 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL_RING0, CMP_BUSY_INT_ENABLE, 2097 enable ? 1 : 0); 2098 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL_RING0, GFX_IDLE_INT_ENABLE, 2099 enable ? 1 : 0); 2100 WREG32_SOC15_IP(GC, cp_int_cntl_reg, tmp); 2101 } 2102 } 2103 } 2104 } 2105 2106 static int gfx_v11_0_init_csb(struct amdgpu_device *adev) 2107 { 2108 adev->gfx.rlc.funcs->get_csb_buffer(adev, adev->gfx.rlc.cs_ptr); 2109 2110 WREG32_SOC15(GC, 0, regRLC_CSIB_ADDR_HI, 2111 adev->gfx.rlc.clear_state_gpu_addr >> 32); 2112 WREG32_SOC15(GC, 0, regRLC_CSIB_ADDR_LO, 2113 adev->gfx.rlc.clear_state_gpu_addr & 0xfffffffc); 2114 WREG32_SOC15(GC, 0, regRLC_CSIB_LENGTH, adev->gfx.rlc.clear_state_size); 2115 2116 return 0; 2117 } 2118 2119 static void gfx_v11_0_rlc_stop(struct amdgpu_device *adev) 2120 { 2121 u32 tmp = RREG32_SOC15(GC, 0, regRLC_CNTL); 2122 2123 tmp = REG_SET_FIELD(tmp, RLC_CNTL, RLC_ENABLE_F32, 0); 2124 WREG32_SOC15(GC, 0, regRLC_CNTL, tmp); 2125 } 2126 2127 static void gfx_v11_0_rlc_reset(struct amdgpu_device *adev) 2128 { 2129 WREG32_FIELD15_PREREG(GC, 0, GRBM_SOFT_RESET, SOFT_RESET_RLC, 1); 2130 udelay(50); 2131 WREG32_FIELD15_PREREG(GC, 0, GRBM_SOFT_RESET, SOFT_RESET_RLC, 0); 2132 udelay(50); 2133 } 2134 2135 static void gfx_v11_0_rlc_smu_handshake_cntl(struct amdgpu_device *adev, 2136 bool enable) 2137 { 2138 uint32_t rlc_pg_cntl; 2139 2140 rlc_pg_cntl = RREG32_SOC15(GC, 0, regRLC_PG_CNTL); 2141 2142 if (!enable) { 2143 /* RLC_PG_CNTL[23] = 0 (default) 2144 * RLC will wait for handshake acks with SMU 2145 * GFXOFF will be enabled 2146 * RLC_PG_CNTL[23] = 1 2147 * RLC will not issue any message to SMU 2148 * hence no handshake between SMU & RLC 2149 * GFXOFF will be disabled 2150 */ 2151 rlc_pg_cntl |= RLC_PG_CNTL__SMU_HANDSHAKE_DISABLE_MASK; 2152 } else 2153 rlc_pg_cntl &= ~RLC_PG_CNTL__SMU_HANDSHAKE_DISABLE_MASK; 2154 WREG32_SOC15(GC, 0, regRLC_PG_CNTL, rlc_pg_cntl); 2155 } 2156 2157 static void gfx_v11_0_rlc_start(struct amdgpu_device *adev) 2158 { 2159 /* TODO: enable rlc & smu handshake until smu 2160 * and gfxoff feature works as expected */ 2161 if (!(amdgpu_pp_feature_mask & PP_GFXOFF_MASK)) 2162 gfx_v11_0_rlc_smu_handshake_cntl(adev, false); 2163 2164 WREG32_FIELD15_PREREG(GC, 0, RLC_CNTL, RLC_ENABLE_F32, 1); 2165 udelay(50); 2166 } 2167 2168 static void gfx_v11_0_rlc_enable_srm(struct amdgpu_device *adev) 2169 { 2170 uint32_t tmp; 2171 2172 /* enable Save Restore Machine */ 2173 tmp = RREG32(SOC15_REG_OFFSET(GC, 0, regRLC_SRM_CNTL)); 2174 tmp |= RLC_SRM_CNTL__AUTO_INCR_ADDR_MASK; 2175 tmp |= RLC_SRM_CNTL__SRM_ENABLE_MASK; 2176 WREG32(SOC15_REG_OFFSET(GC, 0, regRLC_SRM_CNTL), tmp); 2177 } 2178 2179 static void gfx_v11_0_load_rlcg_microcode(struct amdgpu_device *adev) 2180 { 2181 const struct rlc_firmware_header_v2_0 *hdr; 2182 const __le32 *fw_data; 2183 unsigned i, fw_size; 2184 2185 hdr = (const struct rlc_firmware_header_v2_0 *)adev->gfx.rlc_fw->data; 2186 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data + 2187 le32_to_cpu(hdr->header.ucode_array_offset_bytes)); 2188 fw_size = le32_to_cpu(hdr->header.ucode_size_bytes) / 4; 2189 2190 WREG32_SOC15(GC, 0, regRLC_GPM_UCODE_ADDR, 2191 RLCG_UCODE_LOADING_START_ADDRESS); 2192 2193 for (i = 0; i < fw_size; i++) 2194 WREG32_SOC15(GC, 0, regRLC_GPM_UCODE_DATA, 2195 le32_to_cpup(fw_data++)); 2196 2197 WREG32_SOC15(GC, 0, regRLC_GPM_UCODE_ADDR, adev->gfx.rlc_fw_version); 2198 } 2199 2200 static void gfx_v11_0_load_rlc_iram_dram_microcode(struct amdgpu_device *adev) 2201 { 2202 const struct rlc_firmware_header_v2_2 *hdr; 2203 const __le32 *fw_data; 2204 unsigned i, fw_size; 2205 u32 tmp; 2206 2207 hdr = (const struct rlc_firmware_header_v2_2 *)adev->gfx.rlc_fw->data; 2208 2209 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data + 2210 le32_to_cpu(hdr->rlc_iram_ucode_offset_bytes)); 2211 fw_size = le32_to_cpu(hdr->rlc_iram_ucode_size_bytes) / 4; 2212 2213 WREG32_SOC15(GC, 0, regRLC_LX6_IRAM_ADDR, 0); 2214 2215 for (i = 0; i < fw_size; i++) { 2216 if ((amdgpu_emu_mode == 1) && (i % 100 == 99)) 2217 msleep(1); 2218 WREG32_SOC15(GC, 0, regRLC_LX6_IRAM_DATA, 2219 le32_to_cpup(fw_data++)); 2220 } 2221 2222 WREG32_SOC15(GC, 0, regRLC_LX6_IRAM_ADDR, adev->gfx.rlc_fw_version); 2223 2224 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data + 2225 le32_to_cpu(hdr->rlc_dram_ucode_offset_bytes)); 2226 fw_size = le32_to_cpu(hdr->rlc_dram_ucode_size_bytes) / 4; 2227 2228 WREG32_SOC15(GC, 0, regRLC_LX6_DRAM_ADDR, 0); 2229 for (i = 0; i < fw_size; i++) { 2230 if ((amdgpu_emu_mode == 1) && (i % 100 == 99)) 2231 msleep(1); 2232 WREG32_SOC15(GC, 0, regRLC_LX6_DRAM_DATA, 2233 le32_to_cpup(fw_data++)); 2234 } 2235 2236 WREG32_SOC15(GC, 0, regRLC_LX6_IRAM_ADDR, adev->gfx.rlc_fw_version); 2237 2238 tmp = RREG32_SOC15(GC, 0, regRLC_LX6_CNTL); 2239 tmp = REG_SET_FIELD(tmp, RLC_LX6_CNTL, PDEBUG_ENABLE, 1); 2240 tmp = REG_SET_FIELD(tmp, RLC_LX6_CNTL, BRESET, 0); 2241 WREG32_SOC15(GC, 0, regRLC_LX6_CNTL, tmp); 2242 } 2243 2244 static void gfx_v11_0_load_rlcp_rlcv_microcode(struct amdgpu_device *adev) 2245 { 2246 const struct rlc_firmware_header_v2_3 *hdr; 2247 const __le32 *fw_data; 2248 unsigned i, fw_size; 2249 u32 tmp; 2250 2251 hdr = (const struct rlc_firmware_header_v2_3 *)adev->gfx.rlc_fw->data; 2252 2253 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data + 2254 le32_to_cpu(hdr->rlcp_ucode_offset_bytes)); 2255 fw_size = le32_to_cpu(hdr->rlcp_ucode_size_bytes) / 4; 2256 2257 WREG32_SOC15(GC, 0, regRLC_PACE_UCODE_ADDR, 0); 2258 2259 for (i = 0; i < fw_size; i++) { 2260 if ((amdgpu_emu_mode == 1) && (i % 100 == 99)) 2261 msleep(1); 2262 WREG32_SOC15(GC, 0, regRLC_PACE_UCODE_DATA, 2263 le32_to_cpup(fw_data++)); 2264 } 2265 2266 WREG32_SOC15(GC, 0, regRLC_PACE_UCODE_ADDR, adev->gfx.rlc_fw_version); 2267 2268 tmp = RREG32_SOC15(GC, 0, regRLC_GPM_THREAD_ENABLE); 2269 tmp = REG_SET_FIELD(tmp, RLC_GPM_THREAD_ENABLE, THREAD1_ENABLE, 1); 2270 WREG32_SOC15(GC, 0, regRLC_GPM_THREAD_ENABLE, tmp); 2271 2272 fw_data = (const __le32 *)(adev->gfx.rlc_fw->data + 2273 le32_to_cpu(hdr->rlcv_ucode_offset_bytes)); 2274 fw_size = le32_to_cpu(hdr->rlcv_ucode_size_bytes) / 4; 2275 2276 WREG32_SOC15(GC, 0, regRLC_GPU_IOV_UCODE_ADDR, 0); 2277 2278 for (i = 0; i < fw_size; i++) { 2279 if ((amdgpu_emu_mode == 1) && (i % 100 == 99)) 2280 msleep(1); 2281 WREG32_SOC15(GC, 0, regRLC_GPU_IOV_UCODE_DATA, 2282 le32_to_cpup(fw_data++)); 2283 } 2284 2285 WREG32_SOC15(GC, 0, regRLC_GPU_IOV_UCODE_ADDR, adev->gfx.rlc_fw_version); 2286 2287 tmp = RREG32_SOC15(GC, 0, regRLC_GPU_IOV_F32_CNTL); 2288 tmp = REG_SET_FIELD(tmp, RLC_GPU_IOV_F32_CNTL, ENABLE, 1); 2289 WREG32_SOC15(GC, 0, regRLC_GPU_IOV_F32_CNTL, tmp); 2290 } 2291 2292 static int gfx_v11_0_rlc_load_microcode(struct amdgpu_device *adev) 2293 { 2294 const struct rlc_firmware_header_v2_0 *hdr; 2295 uint16_t version_major; 2296 uint16_t version_minor; 2297 2298 if (!adev->gfx.rlc_fw) 2299 return -EINVAL; 2300 2301 hdr = (const struct rlc_firmware_header_v2_0 *)adev->gfx.rlc_fw->data; 2302 amdgpu_ucode_print_rlc_hdr(&hdr->header); 2303 2304 version_major = le16_to_cpu(hdr->header.header_version_major); 2305 version_minor = le16_to_cpu(hdr->header.header_version_minor); 2306 2307 if (version_major == 2) { 2308 gfx_v11_0_load_rlcg_microcode(adev); 2309 if (amdgpu_dpm == 1) { 2310 if (version_minor >= 2) 2311 gfx_v11_0_load_rlc_iram_dram_microcode(adev); 2312 if (version_minor == 3) 2313 gfx_v11_0_load_rlcp_rlcv_microcode(adev); 2314 } 2315 2316 return 0; 2317 } 2318 2319 return -EINVAL; 2320 } 2321 2322 static int gfx_v11_0_rlc_resume(struct amdgpu_device *adev) 2323 { 2324 int r; 2325 2326 if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP) { 2327 gfx_v11_0_init_csb(adev); 2328 2329 if (!amdgpu_sriov_vf(adev)) /* enable RLC SRM */ 2330 gfx_v11_0_rlc_enable_srm(adev); 2331 } else { 2332 if (amdgpu_sriov_vf(adev)) { 2333 gfx_v11_0_init_csb(adev); 2334 return 0; 2335 } 2336 2337 adev->gfx.rlc.funcs->stop(adev); 2338 2339 /* disable CG */ 2340 WREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL, 0); 2341 2342 /* disable PG */ 2343 WREG32_SOC15(GC, 0, regRLC_PG_CNTL, 0); 2344 2345 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) { 2346 /* legacy rlc firmware loading */ 2347 r = gfx_v11_0_rlc_load_microcode(adev); 2348 if (r) 2349 return r; 2350 } 2351 2352 gfx_v11_0_init_csb(adev); 2353 2354 adev->gfx.rlc.funcs->start(adev); 2355 } 2356 return 0; 2357 } 2358 2359 static int gfx_v11_0_config_me_cache(struct amdgpu_device *adev, uint64_t addr) 2360 { 2361 uint32_t usec_timeout = 50000; /* wait for 50ms */ 2362 uint32_t tmp; 2363 int i; 2364 2365 /* Trigger an invalidation of the L1 instruction caches */ 2366 tmp = RREG32_SOC15(GC, 0, regCP_ME_IC_OP_CNTL); 2367 tmp = REG_SET_FIELD(tmp, CP_ME_IC_OP_CNTL, INVALIDATE_CACHE, 1); 2368 WREG32_SOC15(GC, 0, regCP_ME_IC_OP_CNTL, tmp); 2369 2370 /* Wait for invalidation complete */ 2371 for (i = 0; i < usec_timeout; i++) { 2372 tmp = RREG32_SOC15(GC, 0, regCP_ME_IC_OP_CNTL); 2373 if (1 == REG_GET_FIELD(tmp, CP_ME_IC_OP_CNTL, 2374 INVALIDATE_CACHE_COMPLETE)) 2375 break; 2376 udelay(1); 2377 } 2378 2379 if (i >= usec_timeout) { 2380 dev_err(adev->dev, "failed to invalidate instruction cache\n"); 2381 return -EINVAL; 2382 } 2383 2384 if (amdgpu_emu_mode == 1) 2385 adev->hdp.funcs->flush_hdp(adev, NULL); 2386 2387 tmp = RREG32_SOC15(GC, 0, regCP_ME_IC_BASE_CNTL); 2388 tmp = REG_SET_FIELD(tmp, CP_ME_IC_BASE_CNTL, VMID, 0); 2389 tmp = REG_SET_FIELD(tmp, CP_ME_IC_BASE_CNTL, CACHE_POLICY, 0); 2390 tmp = REG_SET_FIELD(tmp, CP_ME_IC_BASE_CNTL, EXE_DISABLE, 0); 2391 tmp = REG_SET_FIELD(tmp, CP_ME_IC_BASE_CNTL, ADDRESS_CLAMP, 1); 2392 WREG32_SOC15(GC, 0, regCP_ME_IC_BASE_CNTL, tmp); 2393 2394 /* Program me ucode address into intruction cache address register */ 2395 WREG32_SOC15(GC, 0, regCP_ME_IC_BASE_LO, 2396 lower_32_bits(addr) & 0xFFFFF000); 2397 WREG32_SOC15(GC, 0, regCP_ME_IC_BASE_HI, 2398 upper_32_bits(addr)); 2399 2400 return 0; 2401 } 2402 2403 static int gfx_v11_0_config_pfp_cache(struct amdgpu_device *adev, uint64_t addr) 2404 { 2405 uint32_t usec_timeout = 50000; /* wait for 50ms */ 2406 uint32_t tmp; 2407 int i; 2408 2409 /* Trigger an invalidation of the L1 instruction caches */ 2410 tmp = RREG32_SOC15(GC, 0, regCP_PFP_IC_OP_CNTL); 2411 tmp = REG_SET_FIELD(tmp, CP_PFP_IC_OP_CNTL, INVALIDATE_CACHE, 1); 2412 WREG32_SOC15(GC, 0, regCP_PFP_IC_OP_CNTL, tmp); 2413 2414 /* Wait for invalidation complete */ 2415 for (i = 0; i < usec_timeout; i++) { 2416 tmp = RREG32_SOC15(GC, 0, regCP_PFP_IC_OP_CNTL); 2417 if (1 == REG_GET_FIELD(tmp, CP_PFP_IC_OP_CNTL, 2418 INVALIDATE_CACHE_COMPLETE)) 2419 break; 2420 udelay(1); 2421 } 2422 2423 if (i >= usec_timeout) { 2424 dev_err(adev->dev, "failed to invalidate instruction cache\n"); 2425 return -EINVAL; 2426 } 2427 2428 if (amdgpu_emu_mode == 1) 2429 adev->hdp.funcs->flush_hdp(adev, NULL); 2430 2431 tmp = RREG32_SOC15(GC, 0, regCP_PFP_IC_BASE_CNTL); 2432 tmp = REG_SET_FIELD(tmp, CP_PFP_IC_BASE_CNTL, VMID, 0); 2433 tmp = REG_SET_FIELD(tmp, CP_PFP_IC_BASE_CNTL, CACHE_POLICY, 0); 2434 tmp = REG_SET_FIELD(tmp, CP_PFP_IC_BASE_CNTL, EXE_DISABLE, 0); 2435 tmp = REG_SET_FIELD(tmp, CP_PFP_IC_BASE_CNTL, ADDRESS_CLAMP, 1); 2436 WREG32_SOC15(GC, 0, regCP_PFP_IC_BASE_CNTL, tmp); 2437 2438 /* Program pfp ucode address into intruction cache address register */ 2439 WREG32_SOC15(GC, 0, regCP_PFP_IC_BASE_LO, 2440 lower_32_bits(addr) & 0xFFFFF000); 2441 WREG32_SOC15(GC, 0, regCP_PFP_IC_BASE_HI, 2442 upper_32_bits(addr)); 2443 2444 return 0; 2445 } 2446 2447 static int gfx_v11_0_config_mec_cache(struct amdgpu_device *adev, uint64_t addr) 2448 { 2449 uint32_t usec_timeout = 50000; /* wait for 50ms */ 2450 uint32_t tmp; 2451 int i; 2452 2453 /* Trigger an invalidation of the L1 instruction caches */ 2454 tmp = RREG32_SOC15(GC, 0, regCP_CPC_IC_OP_CNTL); 2455 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_OP_CNTL, INVALIDATE_CACHE, 1); 2456 2457 WREG32_SOC15(GC, 0, regCP_CPC_IC_OP_CNTL, tmp); 2458 2459 /* Wait for invalidation complete */ 2460 for (i = 0; i < usec_timeout; i++) { 2461 tmp = RREG32_SOC15(GC, 0, regCP_CPC_IC_OP_CNTL); 2462 if (1 == REG_GET_FIELD(tmp, CP_CPC_IC_OP_CNTL, 2463 INVALIDATE_CACHE_COMPLETE)) 2464 break; 2465 udelay(1); 2466 } 2467 2468 if (i >= usec_timeout) { 2469 dev_err(adev->dev, "failed to invalidate instruction cache\n"); 2470 return -EINVAL; 2471 } 2472 2473 if (amdgpu_emu_mode == 1) 2474 adev->hdp.funcs->flush_hdp(adev, NULL); 2475 2476 tmp = RREG32_SOC15(GC, 0, regCP_CPC_IC_BASE_CNTL); 2477 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_BASE_CNTL, CACHE_POLICY, 0); 2478 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_BASE_CNTL, EXE_DISABLE, 0); 2479 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_BASE_CNTL, ADDRESS_CLAMP, 1); 2480 WREG32_SOC15(GC, 0, regCP_CPC_IC_BASE_CNTL, tmp); 2481 2482 /* Program mec1 ucode address into intruction cache address register */ 2483 WREG32_SOC15(GC, 0, regCP_CPC_IC_BASE_LO, 2484 lower_32_bits(addr) & 0xFFFFF000); 2485 WREG32_SOC15(GC, 0, regCP_CPC_IC_BASE_HI, 2486 upper_32_bits(addr)); 2487 2488 return 0; 2489 } 2490 2491 static int gfx_v11_0_config_pfp_cache_rs64(struct amdgpu_device *adev, uint64_t addr, uint64_t addr2) 2492 { 2493 uint32_t usec_timeout = 50000; /* wait for 50ms */ 2494 uint32_t tmp; 2495 unsigned i, pipe_id; 2496 const struct gfx_firmware_header_v2_0 *pfp_hdr; 2497 2498 pfp_hdr = (const struct gfx_firmware_header_v2_0 *) 2499 adev->gfx.pfp_fw->data; 2500 2501 WREG32_SOC15(GC, 0, regCP_PFP_IC_BASE_LO, 2502 lower_32_bits(addr)); 2503 WREG32_SOC15(GC, 0, regCP_PFP_IC_BASE_HI, 2504 upper_32_bits(addr)); 2505 2506 tmp = RREG32_SOC15(GC, 0, regCP_PFP_IC_BASE_CNTL); 2507 tmp = REG_SET_FIELD(tmp, CP_PFP_IC_BASE_CNTL, VMID, 0); 2508 tmp = REG_SET_FIELD(tmp, CP_PFP_IC_BASE_CNTL, CACHE_POLICY, 0); 2509 tmp = REG_SET_FIELD(tmp, CP_PFP_IC_BASE_CNTL, EXE_DISABLE, 0); 2510 WREG32_SOC15(GC, 0, regCP_PFP_IC_BASE_CNTL, tmp); 2511 2512 /* 2513 * Programming any of the CP_PFP_IC_BASE registers 2514 * forces invalidation of the ME L1 I$. Wait for the 2515 * invalidation complete 2516 */ 2517 for (i = 0; i < usec_timeout; i++) { 2518 tmp = RREG32_SOC15(GC, 0, regCP_PFP_IC_OP_CNTL); 2519 if (1 == REG_GET_FIELD(tmp, CP_PFP_IC_OP_CNTL, 2520 INVALIDATE_CACHE_COMPLETE)) 2521 break; 2522 udelay(1); 2523 } 2524 2525 if (i >= usec_timeout) { 2526 dev_err(adev->dev, "failed to invalidate instruction cache\n"); 2527 return -EINVAL; 2528 } 2529 2530 /* Prime the L1 instruction caches */ 2531 tmp = RREG32_SOC15(GC, 0, regCP_PFP_IC_OP_CNTL); 2532 tmp = REG_SET_FIELD(tmp, CP_PFP_IC_OP_CNTL, PRIME_ICACHE, 1); 2533 WREG32_SOC15(GC, 0, regCP_PFP_IC_OP_CNTL, tmp); 2534 /* Waiting for cache primed*/ 2535 for (i = 0; i < usec_timeout; i++) { 2536 tmp = RREG32_SOC15(GC, 0, regCP_PFP_IC_OP_CNTL); 2537 if (1 == REG_GET_FIELD(tmp, CP_PFP_IC_OP_CNTL, 2538 ICACHE_PRIMED)) 2539 break; 2540 udelay(1); 2541 } 2542 2543 if (i >= usec_timeout) { 2544 dev_err(adev->dev, "failed to prime instruction cache\n"); 2545 return -EINVAL; 2546 } 2547 2548 mutex_lock(&adev->srbm_mutex); 2549 for (pipe_id = 0; pipe_id < adev->gfx.me.num_pipe_per_me; pipe_id++) { 2550 soc21_grbm_select(adev, 0, pipe_id, 0, 0); 2551 WREG32_SOC15(GC, 0, regCP_PFP_PRGRM_CNTR_START, 2552 (pfp_hdr->ucode_start_addr_hi << 30) | 2553 (pfp_hdr->ucode_start_addr_lo >> 2)); 2554 WREG32_SOC15(GC, 0, regCP_PFP_PRGRM_CNTR_START_HI, 2555 pfp_hdr->ucode_start_addr_hi >> 2); 2556 2557 /* 2558 * Program CP_ME_CNTL to reset given PIPE to take 2559 * effect of CP_PFP_PRGRM_CNTR_START. 2560 */ 2561 tmp = RREG32_SOC15(GC, 0, regCP_ME_CNTL); 2562 if (pipe_id == 0) 2563 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 2564 PFP_PIPE0_RESET, 1); 2565 else 2566 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 2567 PFP_PIPE1_RESET, 1); 2568 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 2569 2570 /* Clear pfp pipe0 reset bit. */ 2571 if (pipe_id == 0) 2572 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 2573 PFP_PIPE0_RESET, 0); 2574 else 2575 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 2576 PFP_PIPE1_RESET, 0); 2577 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 2578 2579 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE0_LO, 2580 lower_32_bits(addr2)); 2581 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE0_HI, 2582 upper_32_bits(addr2)); 2583 } 2584 soc21_grbm_select(adev, 0, 0, 0, 0); 2585 mutex_unlock(&adev->srbm_mutex); 2586 2587 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE_CNTL); 2588 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_BASE_CNTL, VMID, 0); 2589 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_BASE_CNTL, CACHE_POLICY, 0); 2590 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE_CNTL, tmp); 2591 2592 /* Invalidate the data caches */ 2593 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL); 2594 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_OP_CNTL, INVALIDATE_DCACHE, 1); 2595 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL, tmp); 2596 2597 for (i = 0; i < usec_timeout; i++) { 2598 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL); 2599 if (1 == REG_GET_FIELD(tmp, CP_GFX_RS64_DC_OP_CNTL, 2600 INVALIDATE_DCACHE_COMPLETE)) 2601 break; 2602 udelay(1); 2603 } 2604 2605 if (i >= usec_timeout) { 2606 dev_err(adev->dev, "failed to invalidate RS64 data cache\n"); 2607 return -EINVAL; 2608 } 2609 2610 return 0; 2611 } 2612 2613 static int gfx_v11_0_config_me_cache_rs64(struct amdgpu_device *adev, uint64_t addr, uint64_t addr2) 2614 { 2615 uint32_t usec_timeout = 50000; /* wait for 50ms */ 2616 uint32_t tmp; 2617 unsigned i, pipe_id; 2618 const struct gfx_firmware_header_v2_0 *me_hdr; 2619 2620 me_hdr = (const struct gfx_firmware_header_v2_0 *) 2621 adev->gfx.me_fw->data; 2622 2623 WREG32_SOC15(GC, 0, regCP_ME_IC_BASE_LO, 2624 lower_32_bits(addr)); 2625 WREG32_SOC15(GC, 0, regCP_ME_IC_BASE_HI, 2626 upper_32_bits(addr)); 2627 2628 tmp = RREG32_SOC15(GC, 0, regCP_ME_IC_BASE_CNTL); 2629 tmp = REG_SET_FIELD(tmp, CP_ME_IC_BASE_CNTL, VMID, 0); 2630 tmp = REG_SET_FIELD(tmp, CP_ME_IC_BASE_CNTL, CACHE_POLICY, 0); 2631 tmp = REG_SET_FIELD(tmp, CP_ME_IC_BASE_CNTL, EXE_DISABLE, 0); 2632 WREG32_SOC15(GC, 0, regCP_ME_IC_BASE_CNTL, tmp); 2633 2634 /* 2635 * Programming any of the CP_ME_IC_BASE registers 2636 * forces invalidation of the ME L1 I$. Wait for the 2637 * invalidation complete 2638 */ 2639 for (i = 0; i < usec_timeout; i++) { 2640 tmp = RREG32_SOC15(GC, 0, regCP_ME_IC_OP_CNTL); 2641 if (1 == REG_GET_FIELD(tmp, CP_ME_IC_OP_CNTL, 2642 INVALIDATE_CACHE_COMPLETE)) 2643 break; 2644 udelay(1); 2645 } 2646 2647 if (i >= usec_timeout) { 2648 dev_err(adev->dev, "failed to invalidate instruction cache\n"); 2649 return -EINVAL; 2650 } 2651 2652 /* Prime the instruction caches */ 2653 tmp = RREG32_SOC15(GC, 0, regCP_ME_IC_OP_CNTL); 2654 tmp = REG_SET_FIELD(tmp, CP_ME_IC_OP_CNTL, PRIME_ICACHE, 1); 2655 WREG32_SOC15(GC, 0, regCP_ME_IC_OP_CNTL, tmp); 2656 2657 /* Waiting for instruction cache primed*/ 2658 for (i = 0; i < usec_timeout; i++) { 2659 tmp = RREG32_SOC15(GC, 0, regCP_ME_IC_OP_CNTL); 2660 if (1 == REG_GET_FIELD(tmp, CP_ME_IC_OP_CNTL, 2661 ICACHE_PRIMED)) 2662 break; 2663 udelay(1); 2664 } 2665 2666 if (i >= usec_timeout) { 2667 dev_err(adev->dev, "failed to prime instruction cache\n"); 2668 return -EINVAL; 2669 } 2670 2671 mutex_lock(&adev->srbm_mutex); 2672 for (pipe_id = 0; pipe_id < adev->gfx.me.num_pipe_per_me; pipe_id++) { 2673 soc21_grbm_select(adev, 0, pipe_id, 0, 0); 2674 WREG32_SOC15(GC, 0, regCP_ME_PRGRM_CNTR_START, 2675 (me_hdr->ucode_start_addr_hi << 30) | 2676 (me_hdr->ucode_start_addr_lo >> 2) ); 2677 WREG32_SOC15(GC, 0, regCP_ME_PRGRM_CNTR_START_HI, 2678 me_hdr->ucode_start_addr_hi>>2); 2679 2680 /* 2681 * Program CP_ME_CNTL to reset given PIPE to take 2682 * effect of CP_PFP_PRGRM_CNTR_START. 2683 */ 2684 tmp = RREG32_SOC15(GC, 0, regCP_ME_CNTL); 2685 if (pipe_id == 0) 2686 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 2687 ME_PIPE0_RESET, 1); 2688 else 2689 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 2690 ME_PIPE1_RESET, 1); 2691 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 2692 2693 /* Clear pfp pipe0 reset bit. */ 2694 if (pipe_id == 0) 2695 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 2696 ME_PIPE0_RESET, 0); 2697 else 2698 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 2699 ME_PIPE1_RESET, 0); 2700 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 2701 2702 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE1_LO, 2703 lower_32_bits(addr2)); 2704 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE1_HI, 2705 upper_32_bits(addr2)); 2706 } 2707 soc21_grbm_select(adev, 0, 0, 0, 0); 2708 mutex_unlock(&adev->srbm_mutex); 2709 2710 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE_CNTL); 2711 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_BASE_CNTL, VMID, 0); 2712 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_BASE_CNTL, CACHE_POLICY, 0); 2713 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE_CNTL, tmp); 2714 2715 /* Invalidate the data caches */ 2716 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL); 2717 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_OP_CNTL, INVALIDATE_DCACHE, 1); 2718 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL, tmp); 2719 2720 for (i = 0; i < usec_timeout; i++) { 2721 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL); 2722 if (1 == REG_GET_FIELD(tmp, CP_GFX_RS64_DC_OP_CNTL, 2723 INVALIDATE_DCACHE_COMPLETE)) 2724 break; 2725 udelay(1); 2726 } 2727 2728 if (i >= usec_timeout) { 2729 dev_err(adev->dev, "failed to invalidate RS64 data cache\n"); 2730 return -EINVAL; 2731 } 2732 2733 return 0; 2734 } 2735 2736 static int gfx_v11_0_config_mec_cache_rs64(struct amdgpu_device *adev, uint64_t addr, uint64_t addr2) 2737 { 2738 uint32_t usec_timeout = 50000; /* wait for 50ms */ 2739 uint32_t tmp; 2740 unsigned i; 2741 const struct gfx_firmware_header_v2_0 *mec_hdr; 2742 2743 mec_hdr = (const struct gfx_firmware_header_v2_0 *) 2744 adev->gfx.mec_fw->data; 2745 2746 tmp = RREG32_SOC15(GC, 0, regCP_CPC_IC_BASE_CNTL); 2747 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_BASE_CNTL, VMID, 0); 2748 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_BASE_CNTL, EXE_DISABLE, 0); 2749 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_BASE_CNTL, CACHE_POLICY, 0); 2750 WREG32_SOC15(GC, 0, regCP_CPC_IC_BASE_CNTL, tmp); 2751 2752 tmp = RREG32_SOC15(GC, 0, regCP_MEC_DC_BASE_CNTL); 2753 tmp = REG_SET_FIELD(tmp, CP_MEC_DC_BASE_CNTL, VMID, 0); 2754 tmp = REG_SET_FIELD(tmp, CP_MEC_DC_BASE_CNTL, CACHE_POLICY, 0); 2755 WREG32_SOC15(GC, 0, regCP_MEC_DC_BASE_CNTL, tmp); 2756 2757 mutex_lock(&adev->srbm_mutex); 2758 for (i = 0; i < adev->gfx.mec.num_pipe_per_mec; i++) { 2759 soc21_grbm_select(adev, 1, i, 0, 0); 2760 2761 WREG32_SOC15(GC, 0, regCP_MEC_MDBASE_LO, addr2); 2762 WREG32_SOC15(GC, 0, regCP_MEC_MDBASE_HI, 2763 upper_32_bits(addr2)); 2764 2765 WREG32_SOC15(GC, 0, regCP_MEC_RS64_PRGRM_CNTR_START, 2766 mec_hdr->ucode_start_addr_lo >> 2 | 2767 mec_hdr->ucode_start_addr_hi << 30); 2768 WREG32_SOC15(GC, 0, regCP_MEC_RS64_PRGRM_CNTR_START_HI, 2769 mec_hdr->ucode_start_addr_hi >> 2); 2770 2771 WREG32_SOC15(GC, 0, regCP_CPC_IC_BASE_LO, addr); 2772 WREG32_SOC15(GC, 0, regCP_CPC_IC_BASE_HI, 2773 upper_32_bits(addr)); 2774 } 2775 mutex_unlock(&adev->srbm_mutex); 2776 soc21_grbm_select(adev, 0, 0, 0, 0); 2777 2778 /* Trigger an invalidation of the L1 instruction caches */ 2779 tmp = RREG32_SOC15(GC, 0, regCP_MEC_DC_OP_CNTL); 2780 tmp = REG_SET_FIELD(tmp, CP_MEC_DC_OP_CNTL, INVALIDATE_DCACHE, 1); 2781 WREG32_SOC15(GC, 0, regCP_MEC_DC_OP_CNTL, tmp); 2782 2783 /* Wait for invalidation complete */ 2784 for (i = 0; i < usec_timeout; i++) { 2785 tmp = RREG32_SOC15(GC, 0, regCP_MEC_DC_OP_CNTL); 2786 if (1 == REG_GET_FIELD(tmp, CP_MEC_DC_OP_CNTL, 2787 INVALIDATE_DCACHE_COMPLETE)) 2788 break; 2789 udelay(1); 2790 } 2791 2792 if (i >= usec_timeout) { 2793 dev_err(adev->dev, "failed to invalidate instruction cache\n"); 2794 return -EINVAL; 2795 } 2796 2797 /* Trigger an invalidation of the L1 instruction caches */ 2798 tmp = RREG32_SOC15(GC, 0, regCP_CPC_IC_OP_CNTL); 2799 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_OP_CNTL, INVALIDATE_CACHE, 1); 2800 WREG32_SOC15(GC, 0, regCP_CPC_IC_OP_CNTL, tmp); 2801 2802 /* Wait for invalidation complete */ 2803 for (i = 0; i < usec_timeout; i++) { 2804 tmp = RREG32_SOC15(GC, 0, regCP_CPC_IC_OP_CNTL); 2805 if (1 == REG_GET_FIELD(tmp, CP_CPC_IC_OP_CNTL, 2806 INVALIDATE_CACHE_COMPLETE)) 2807 break; 2808 udelay(1); 2809 } 2810 2811 if (i >= usec_timeout) { 2812 dev_err(adev->dev, "failed to invalidate instruction cache\n"); 2813 return -EINVAL; 2814 } 2815 2816 return 0; 2817 } 2818 2819 static void gfx_v11_0_config_gfx_rs64(struct amdgpu_device *adev) 2820 { 2821 const struct gfx_firmware_header_v2_0 *pfp_hdr; 2822 const struct gfx_firmware_header_v2_0 *me_hdr; 2823 const struct gfx_firmware_header_v2_0 *mec_hdr; 2824 uint32_t pipe_id, tmp; 2825 2826 mec_hdr = (const struct gfx_firmware_header_v2_0 *) 2827 adev->gfx.mec_fw->data; 2828 me_hdr = (const struct gfx_firmware_header_v2_0 *) 2829 adev->gfx.me_fw->data; 2830 pfp_hdr = (const struct gfx_firmware_header_v2_0 *) 2831 adev->gfx.pfp_fw->data; 2832 2833 /* config pfp program start addr */ 2834 for (pipe_id = 0; pipe_id < 2; pipe_id++) { 2835 soc21_grbm_select(adev, 0, pipe_id, 0, 0); 2836 WREG32_SOC15(GC, 0, regCP_PFP_PRGRM_CNTR_START, 2837 (pfp_hdr->ucode_start_addr_hi << 30) | 2838 (pfp_hdr->ucode_start_addr_lo >> 2)); 2839 WREG32_SOC15(GC, 0, regCP_PFP_PRGRM_CNTR_START_HI, 2840 pfp_hdr->ucode_start_addr_hi >> 2); 2841 } 2842 soc21_grbm_select(adev, 0, 0, 0, 0); 2843 2844 /* reset pfp pipe */ 2845 tmp = RREG32_SOC15(GC, 0, regCP_ME_CNTL); 2846 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, PFP_PIPE0_RESET, 1); 2847 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, PFP_PIPE1_RESET, 1); 2848 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 2849 2850 /* clear pfp pipe reset */ 2851 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, PFP_PIPE0_RESET, 0); 2852 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, PFP_PIPE1_RESET, 0); 2853 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 2854 2855 /* config me program start addr */ 2856 for (pipe_id = 0; pipe_id < 2; pipe_id++) { 2857 soc21_grbm_select(adev, 0, pipe_id, 0, 0); 2858 WREG32_SOC15(GC, 0, regCP_ME_PRGRM_CNTR_START, 2859 (me_hdr->ucode_start_addr_hi << 30) | 2860 (me_hdr->ucode_start_addr_lo >> 2) ); 2861 WREG32_SOC15(GC, 0, regCP_ME_PRGRM_CNTR_START_HI, 2862 me_hdr->ucode_start_addr_hi>>2); 2863 } 2864 soc21_grbm_select(adev, 0, 0, 0, 0); 2865 2866 /* reset me pipe */ 2867 tmp = RREG32_SOC15(GC, 0, regCP_ME_CNTL); 2868 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, ME_PIPE0_RESET, 1); 2869 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, ME_PIPE1_RESET, 1); 2870 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 2871 2872 /* clear me pipe reset */ 2873 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, ME_PIPE0_RESET, 0); 2874 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, ME_PIPE1_RESET, 0); 2875 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 2876 2877 /* config mec program start addr */ 2878 for (pipe_id = 0; pipe_id < 4; pipe_id++) { 2879 soc21_grbm_select(adev, 1, pipe_id, 0, 0); 2880 WREG32_SOC15(GC, 0, regCP_MEC_RS64_PRGRM_CNTR_START, 2881 mec_hdr->ucode_start_addr_lo >> 2 | 2882 mec_hdr->ucode_start_addr_hi << 30); 2883 WREG32_SOC15(GC, 0, regCP_MEC_RS64_PRGRM_CNTR_START_HI, 2884 mec_hdr->ucode_start_addr_hi >> 2); 2885 } 2886 soc21_grbm_select(adev, 0, 0, 0, 0); 2887 2888 /* reset mec pipe */ 2889 tmp = RREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL); 2890 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE0_RESET, 1); 2891 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE1_RESET, 1); 2892 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE2_RESET, 1); 2893 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE3_RESET, 1); 2894 WREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL, tmp); 2895 2896 /* clear mec pipe reset */ 2897 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE0_RESET, 0); 2898 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE1_RESET, 0); 2899 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE2_RESET, 0); 2900 tmp = REG_SET_FIELD(tmp, CP_MEC_RS64_CNTL, MEC_PIPE3_RESET, 0); 2901 WREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL, tmp); 2902 } 2903 2904 static int gfx_v11_0_wait_for_rlc_autoload_complete(struct amdgpu_device *adev) 2905 { 2906 uint32_t cp_status; 2907 uint32_t bootload_status; 2908 int i, r; 2909 uint64_t addr, addr2; 2910 2911 for (i = 0; i < adev->usec_timeout; i++) { 2912 cp_status = RREG32_SOC15(GC, 0, regCP_STAT); 2913 2914 if (amdgpu_ip_version(adev, GC_HWIP, 0) == 2915 IP_VERSION(11, 0, 1) || 2916 amdgpu_ip_version(adev, GC_HWIP, 0) == 2917 IP_VERSION(11, 0, 4) || 2918 amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(11, 5, 0) || 2919 amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(11, 5, 1) || 2920 amdgpu_ip_version(adev, GC_HWIP, 0) == IP_VERSION(11, 5, 2)) 2921 bootload_status = RREG32_SOC15(GC, 0, 2922 regRLC_RLCS_BOOTLOAD_STATUS_gc_11_0_1); 2923 else 2924 bootload_status = RREG32_SOC15(GC, 0, regRLC_RLCS_BOOTLOAD_STATUS); 2925 2926 if ((cp_status == 0) && 2927 (REG_GET_FIELD(bootload_status, 2928 RLC_RLCS_BOOTLOAD_STATUS, BOOTLOAD_COMPLETE) == 1)) { 2929 break; 2930 } 2931 udelay(1); 2932 } 2933 2934 if (i >= adev->usec_timeout) { 2935 dev_err(adev->dev, "rlc autoload: gc ucode autoload timeout\n"); 2936 return -ETIMEDOUT; 2937 } 2938 2939 if (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) { 2940 if (adev->gfx.rs64_enable) { 2941 addr = adev->gfx.rlc.rlc_autoload_gpu_addr + 2942 rlc_autoload_info[SOC21_FIRMWARE_ID_RS64_ME].offset; 2943 addr2 = adev->gfx.rlc.rlc_autoload_gpu_addr + 2944 rlc_autoload_info[SOC21_FIRMWARE_ID_RS64_ME_P0_STACK].offset; 2945 r = gfx_v11_0_config_me_cache_rs64(adev, addr, addr2); 2946 if (r) 2947 return r; 2948 addr = adev->gfx.rlc.rlc_autoload_gpu_addr + 2949 rlc_autoload_info[SOC21_FIRMWARE_ID_RS64_PFP].offset; 2950 addr2 = adev->gfx.rlc.rlc_autoload_gpu_addr + 2951 rlc_autoload_info[SOC21_FIRMWARE_ID_RS64_PFP_P0_STACK].offset; 2952 r = gfx_v11_0_config_pfp_cache_rs64(adev, addr, addr2); 2953 if (r) 2954 return r; 2955 addr = adev->gfx.rlc.rlc_autoload_gpu_addr + 2956 rlc_autoload_info[SOC21_FIRMWARE_ID_RS64_MEC].offset; 2957 addr2 = adev->gfx.rlc.rlc_autoload_gpu_addr + 2958 rlc_autoload_info[SOC21_FIRMWARE_ID_RS64_MEC_P0_STACK].offset; 2959 r = gfx_v11_0_config_mec_cache_rs64(adev, addr, addr2); 2960 if (r) 2961 return r; 2962 } else { 2963 addr = adev->gfx.rlc.rlc_autoload_gpu_addr + 2964 rlc_autoload_info[SOC21_FIRMWARE_ID_CP_ME].offset; 2965 r = gfx_v11_0_config_me_cache(adev, addr); 2966 if (r) 2967 return r; 2968 addr = adev->gfx.rlc.rlc_autoload_gpu_addr + 2969 rlc_autoload_info[SOC21_FIRMWARE_ID_CP_PFP].offset; 2970 r = gfx_v11_0_config_pfp_cache(adev, addr); 2971 if (r) 2972 return r; 2973 addr = adev->gfx.rlc.rlc_autoload_gpu_addr + 2974 rlc_autoload_info[SOC21_FIRMWARE_ID_CP_MEC].offset; 2975 r = gfx_v11_0_config_mec_cache(adev, addr); 2976 if (r) 2977 return r; 2978 } 2979 } 2980 2981 return 0; 2982 } 2983 2984 static int gfx_v11_0_cp_gfx_enable(struct amdgpu_device *adev, bool enable) 2985 { 2986 int i; 2987 u32 tmp = RREG32_SOC15(GC, 0, regCP_ME_CNTL); 2988 2989 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, ME_HALT, enable ? 0 : 1); 2990 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, PFP_HALT, enable ? 0 : 1); 2991 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 2992 2993 for (i = 0; i < adev->usec_timeout; i++) { 2994 if (RREG32_SOC15(GC, 0, regCP_STAT) == 0) 2995 break; 2996 udelay(1); 2997 } 2998 2999 if (i >= adev->usec_timeout) 3000 DRM_ERROR("failed to %s cp gfx\n", enable ? "unhalt" : "halt"); 3001 3002 return 0; 3003 } 3004 3005 static int gfx_v11_0_cp_gfx_load_pfp_microcode(struct amdgpu_device *adev) 3006 { 3007 int r; 3008 const struct gfx_firmware_header_v1_0 *pfp_hdr; 3009 const __le32 *fw_data; 3010 unsigned i, fw_size; 3011 3012 pfp_hdr = (const struct gfx_firmware_header_v1_0 *) 3013 adev->gfx.pfp_fw->data; 3014 3015 amdgpu_ucode_print_gfx_hdr(&pfp_hdr->header); 3016 3017 fw_data = (const __le32 *)(adev->gfx.pfp_fw->data + 3018 le32_to_cpu(pfp_hdr->header.ucode_array_offset_bytes)); 3019 fw_size = le32_to_cpu(pfp_hdr->header.ucode_size_bytes); 3020 3021 r = amdgpu_bo_create_reserved(adev, pfp_hdr->header.ucode_size_bytes, 3022 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT, 3023 &adev->gfx.pfp.pfp_fw_obj, 3024 &adev->gfx.pfp.pfp_fw_gpu_addr, 3025 (void **)&adev->gfx.pfp.pfp_fw_ptr); 3026 if (r) { 3027 dev_err(adev->dev, "(%d) failed to create pfp fw bo\n", r); 3028 gfx_v11_0_pfp_fini(adev); 3029 return r; 3030 } 3031 3032 memcpy(adev->gfx.pfp.pfp_fw_ptr, fw_data, fw_size); 3033 3034 amdgpu_bo_kunmap(adev->gfx.pfp.pfp_fw_obj); 3035 amdgpu_bo_unreserve(adev->gfx.pfp.pfp_fw_obj); 3036 3037 gfx_v11_0_config_pfp_cache(adev, adev->gfx.pfp.pfp_fw_gpu_addr); 3038 3039 WREG32_SOC15(GC, 0, regCP_HYP_PFP_UCODE_ADDR, 0); 3040 3041 for (i = 0; i < pfp_hdr->jt_size; i++) 3042 WREG32_SOC15(GC, 0, regCP_HYP_PFP_UCODE_DATA, 3043 le32_to_cpup(fw_data + pfp_hdr->jt_offset + i)); 3044 3045 WREG32_SOC15(GC, 0, regCP_HYP_PFP_UCODE_ADDR, adev->gfx.pfp_fw_version); 3046 3047 return 0; 3048 } 3049 3050 static int gfx_v11_0_cp_gfx_load_pfp_microcode_rs64(struct amdgpu_device *adev) 3051 { 3052 int r; 3053 const struct gfx_firmware_header_v2_0 *pfp_hdr; 3054 const __le32 *fw_ucode, *fw_data; 3055 unsigned i, pipe_id, fw_ucode_size, fw_data_size; 3056 uint32_t tmp; 3057 uint32_t usec_timeout = 50000; /* wait for 50ms */ 3058 3059 pfp_hdr = (const struct gfx_firmware_header_v2_0 *) 3060 adev->gfx.pfp_fw->data; 3061 3062 amdgpu_ucode_print_gfx_hdr(&pfp_hdr->header); 3063 3064 /* instruction */ 3065 fw_ucode = (const __le32 *)(adev->gfx.pfp_fw->data + 3066 le32_to_cpu(pfp_hdr->ucode_offset_bytes)); 3067 fw_ucode_size = le32_to_cpu(pfp_hdr->ucode_size_bytes); 3068 /* data */ 3069 fw_data = (const __le32 *)(adev->gfx.pfp_fw->data + 3070 le32_to_cpu(pfp_hdr->data_offset_bytes)); 3071 fw_data_size = le32_to_cpu(pfp_hdr->data_size_bytes); 3072 3073 /* 64kb align */ 3074 r = amdgpu_bo_create_reserved(adev, fw_ucode_size, 3075 64 * 1024, 3076 AMDGPU_GEM_DOMAIN_VRAM | 3077 AMDGPU_GEM_DOMAIN_GTT, 3078 &adev->gfx.pfp.pfp_fw_obj, 3079 &adev->gfx.pfp.pfp_fw_gpu_addr, 3080 (void **)&adev->gfx.pfp.pfp_fw_ptr); 3081 if (r) { 3082 dev_err(adev->dev, "(%d) failed to create pfp ucode fw bo\n", r); 3083 gfx_v11_0_pfp_fini(adev); 3084 return r; 3085 } 3086 3087 r = amdgpu_bo_create_reserved(adev, fw_data_size, 3088 64 * 1024, 3089 AMDGPU_GEM_DOMAIN_VRAM | 3090 AMDGPU_GEM_DOMAIN_GTT, 3091 &adev->gfx.pfp.pfp_fw_data_obj, 3092 &adev->gfx.pfp.pfp_fw_data_gpu_addr, 3093 (void **)&adev->gfx.pfp.pfp_fw_data_ptr); 3094 if (r) { 3095 dev_err(adev->dev, "(%d) failed to create pfp data fw bo\n", r); 3096 gfx_v11_0_pfp_fini(adev); 3097 return r; 3098 } 3099 3100 memcpy(adev->gfx.pfp.pfp_fw_ptr, fw_ucode, fw_ucode_size); 3101 memcpy(adev->gfx.pfp.pfp_fw_data_ptr, fw_data, fw_data_size); 3102 3103 amdgpu_bo_kunmap(adev->gfx.pfp.pfp_fw_obj); 3104 amdgpu_bo_kunmap(adev->gfx.pfp.pfp_fw_data_obj); 3105 amdgpu_bo_unreserve(adev->gfx.pfp.pfp_fw_obj); 3106 amdgpu_bo_unreserve(adev->gfx.pfp.pfp_fw_data_obj); 3107 3108 if (amdgpu_emu_mode == 1) 3109 adev->hdp.funcs->flush_hdp(adev, NULL); 3110 3111 WREG32_SOC15(GC, 0, regCP_PFP_IC_BASE_LO, 3112 lower_32_bits(adev->gfx.pfp.pfp_fw_gpu_addr)); 3113 WREG32_SOC15(GC, 0, regCP_PFP_IC_BASE_HI, 3114 upper_32_bits(adev->gfx.pfp.pfp_fw_gpu_addr)); 3115 3116 tmp = RREG32_SOC15(GC, 0, regCP_PFP_IC_BASE_CNTL); 3117 tmp = REG_SET_FIELD(tmp, CP_PFP_IC_BASE_CNTL, VMID, 0); 3118 tmp = REG_SET_FIELD(tmp, CP_PFP_IC_BASE_CNTL, CACHE_POLICY, 0); 3119 tmp = REG_SET_FIELD(tmp, CP_PFP_IC_BASE_CNTL, EXE_DISABLE, 0); 3120 WREG32_SOC15(GC, 0, regCP_PFP_IC_BASE_CNTL, tmp); 3121 3122 /* 3123 * Programming any of the CP_PFP_IC_BASE registers 3124 * forces invalidation of the ME L1 I$. Wait for the 3125 * invalidation complete 3126 */ 3127 for (i = 0; i < usec_timeout; i++) { 3128 tmp = RREG32_SOC15(GC, 0, regCP_PFP_IC_OP_CNTL); 3129 if (1 == REG_GET_FIELD(tmp, CP_PFP_IC_OP_CNTL, 3130 INVALIDATE_CACHE_COMPLETE)) 3131 break; 3132 udelay(1); 3133 } 3134 3135 if (i >= usec_timeout) { 3136 dev_err(adev->dev, "failed to invalidate instruction cache\n"); 3137 return -EINVAL; 3138 } 3139 3140 /* Prime the L1 instruction caches */ 3141 tmp = RREG32_SOC15(GC, 0, regCP_PFP_IC_OP_CNTL); 3142 tmp = REG_SET_FIELD(tmp, CP_PFP_IC_OP_CNTL, PRIME_ICACHE, 1); 3143 WREG32_SOC15(GC, 0, regCP_PFP_IC_OP_CNTL, tmp); 3144 /* Waiting for cache primed*/ 3145 for (i = 0; i < usec_timeout; i++) { 3146 tmp = RREG32_SOC15(GC, 0, regCP_PFP_IC_OP_CNTL); 3147 if (1 == REG_GET_FIELD(tmp, CP_PFP_IC_OP_CNTL, 3148 ICACHE_PRIMED)) 3149 break; 3150 udelay(1); 3151 } 3152 3153 if (i >= usec_timeout) { 3154 dev_err(adev->dev, "failed to prime instruction cache\n"); 3155 return -EINVAL; 3156 } 3157 3158 mutex_lock(&adev->srbm_mutex); 3159 for (pipe_id = 0; pipe_id < adev->gfx.me.num_pipe_per_me; pipe_id++) { 3160 soc21_grbm_select(adev, 0, pipe_id, 0, 0); 3161 WREG32_SOC15(GC, 0, regCP_PFP_PRGRM_CNTR_START, 3162 (pfp_hdr->ucode_start_addr_hi << 30) | 3163 (pfp_hdr->ucode_start_addr_lo >> 2) ); 3164 WREG32_SOC15(GC, 0, regCP_PFP_PRGRM_CNTR_START_HI, 3165 pfp_hdr->ucode_start_addr_hi>>2); 3166 3167 /* 3168 * Program CP_ME_CNTL to reset given PIPE to take 3169 * effect of CP_PFP_PRGRM_CNTR_START. 3170 */ 3171 tmp = RREG32_SOC15(GC, 0, regCP_ME_CNTL); 3172 if (pipe_id == 0) 3173 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 3174 PFP_PIPE0_RESET, 1); 3175 else 3176 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 3177 PFP_PIPE1_RESET, 1); 3178 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 3179 3180 /* Clear pfp pipe0 reset bit. */ 3181 if (pipe_id == 0) 3182 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 3183 PFP_PIPE0_RESET, 0); 3184 else 3185 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 3186 PFP_PIPE1_RESET, 0); 3187 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 3188 3189 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE0_LO, 3190 lower_32_bits(adev->gfx.pfp.pfp_fw_data_gpu_addr)); 3191 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE0_HI, 3192 upper_32_bits(adev->gfx.pfp.pfp_fw_data_gpu_addr)); 3193 } 3194 soc21_grbm_select(adev, 0, 0, 0, 0); 3195 mutex_unlock(&adev->srbm_mutex); 3196 3197 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE_CNTL); 3198 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_BASE_CNTL, VMID, 0); 3199 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_BASE_CNTL, CACHE_POLICY, 0); 3200 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE_CNTL, tmp); 3201 3202 /* Invalidate the data caches */ 3203 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL); 3204 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_OP_CNTL, INVALIDATE_DCACHE, 1); 3205 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL, tmp); 3206 3207 for (i = 0; i < usec_timeout; i++) { 3208 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL); 3209 if (1 == REG_GET_FIELD(tmp, CP_GFX_RS64_DC_OP_CNTL, 3210 INVALIDATE_DCACHE_COMPLETE)) 3211 break; 3212 udelay(1); 3213 } 3214 3215 if (i >= usec_timeout) { 3216 dev_err(adev->dev, "failed to invalidate RS64 data cache\n"); 3217 return -EINVAL; 3218 } 3219 3220 return 0; 3221 } 3222 3223 static int gfx_v11_0_cp_gfx_load_me_microcode(struct amdgpu_device *adev) 3224 { 3225 int r; 3226 const struct gfx_firmware_header_v1_0 *me_hdr; 3227 const __le32 *fw_data; 3228 unsigned i, fw_size; 3229 3230 me_hdr = (const struct gfx_firmware_header_v1_0 *) 3231 adev->gfx.me_fw->data; 3232 3233 amdgpu_ucode_print_gfx_hdr(&me_hdr->header); 3234 3235 fw_data = (const __le32 *)(adev->gfx.me_fw->data + 3236 le32_to_cpu(me_hdr->header.ucode_array_offset_bytes)); 3237 fw_size = le32_to_cpu(me_hdr->header.ucode_size_bytes); 3238 3239 r = amdgpu_bo_create_reserved(adev, me_hdr->header.ucode_size_bytes, 3240 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT, 3241 &adev->gfx.me.me_fw_obj, 3242 &adev->gfx.me.me_fw_gpu_addr, 3243 (void **)&adev->gfx.me.me_fw_ptr); 3244 if (r) { 3245 dev_err(adev->dev, "(%d) failed to create me fw bo\n", r); 3246 gfx_v11_0_me_fini(adev); 3247 return r; 3248 } 3249 3250 memcpy(adev->gfx.me.me_fw_ptr, fw_data, fw_size); 3251 3252 amdgpu_bo_kunmap(adev->gfx.me.me_fw_obj); 3253 amdgpu_bo_unreserve(adev->gfx.me.me_fw_obj); 3254 3255 gfx_v11_0_config_me_cache(adev, adev->gfx.me.me_fw_gpu_addr); 3256 3257 WREG32_SOC15(GC, 0, regCP_HYP_ME_UCODE_ADDR, 0); 3258 3259 for (i = 0; i < me_hdr->jt_size; i++) 3260 WREG32_SOC15(GC, 0, regCP_HYP_ME_UCODE_DATA, 3261 le32_to_cpup(fw_data + me_hdr->jt_offset + i)); 3262 3263 WREG32_SOC15(GC, 0, regCP_HYP_ME_UCODE_ADDR, adev->gfx.me_fw_version); 3264 3265 return 0; 3266 } 3267 3268 static int gfx_v11_0_cp_gfx_load_me_microcode_rs64(struct amdgpu_device *adev) 3269 { 3270 int r; 3271 const struct gfx_firmware_header_v2_0 *me_hdr; 3272 const __le32 *fw_ucode, *fw_data; 3273 unsigned i, pipe_id, fw_ucode_size, fw_data_size; 3274 uint32_t tmp; 3275 uint32_t usec_timeout = 50000; /* wait for 50ms */ 3276 3277 me_hdr = (const struct gfx_firmware_header_v2_0 *) 3278 adev->gfx.me_fw->data; 3279 3280 amdgpu_ucode_print_gfx_hdr(&me_hdr->header); 3281 3282 /* instruction */ 3283 fw_ucode = (const __le32 *)(adev->gfx.me_fw->data + 3284 le32_to_cpu(me_hdr->ucode_offset_bytes)); 3285 fw_ucode_size = le32_to_cpu(me_hdr->ucode_size_bytes); 3286 /* data */ 3287 fw_data = (const __le32 *)(adev->gfx.me_fw->data + 3288 le32_to_cpu(me_hdr->data_offset_bytes)); 3289 fw_data_size = le32_to_cpu(me_hdr->data_size_bytes); 3290 3291 /* 64kb align*/ 3292 r = amdgpu_bo_create_reserved(adev, fw_ucode_size, 3293 64 * 1024, 3294 AMDGPU_GEM_DOMAIN_VRAM | 3295 AMDGPU_GEM_DOMAIN_GTT, 3296 &adev->gfx.me.me_fw_obj, 3297 &adev->gfx.me.me_fw_gpu_addr, 3298 (void **)&adev->gfx.me.me_fw_ptr); 3299 if (r) { 3300 dev_err(adev->dev, "(%d) failed to create me ucode bo\n", r); 3301 gfx_v11_0_me_fini(adev); 3302 return r; 3303 } 3304 3305 r = amdgpu_bo_create_reserved(adev, fw_data_size, 3306 64 * 1024, 3307 AMDGPU_GEM_DOMAIN_VRAM | 3308 AMDGPU_GEM_DOMAIN_GTT, 3309 &adev->gfx.me.me_fw_data_obj, 3310 &adev->gfx.me.me_fw_data_gpu_addr, 3311 (void **)&adev->gfx.me.me_fw_data_ptr); 3312 if (r) { 3313 dev_err(adev->dev, "(%d) failed to create me data bo\n", r); 3314 gfx_v11_0_pfp_fini(adev); 3315 return r; 3316 } 3317 3318 memcpy(adev->gfx.me.me_fw_ptr, fw_ucode, fw_ucode_size); 3319 memcpy(adev->gfx.me.me_fw_data_ptr, fw_data, fw_data_size); 3320 3321 amdgpu_bo_kunmap(adev->gfx.me.me_fw_obj); 3322 amdgpu_bo_kunmap(adev->gfx.me.me_fw_data_obj); 3323 amdgpu_bo_unreserve(adev->gfx.me.me_fw_obj); 3324 amdgpu_bo_unreserve(adev->gfx.me.me_fw_data_obj); 3325 3326 if (amdgpu_emu_mode == 1) 3327 adev->hdp.funcs->flush_hdp(adev, NULL); 3328 3329 WREG32_SOC15(GC, 0, regCP_ME_IC_BASE_LO, 3330 lower_32_bits(adev->gfx.me.me_fw_gpu_addr)); 3331 WREG32_SOC15(GC, 0, regCP_ME_IC_BASE_HI, 3332 upper_32_bits(adev->gfx.me.me_fw_gpu_addr)); 3333 3334 tmp = RREG32_SOC15(GC, 0, regCP_ME_IC_BASE_CNTL); 3335 tmp = REG_SET_FIELD(tmp, CP_ME_IC_BASE_CNTL, VMID, 0); 3336 tmp = REG_SET_FIELD(tmp, CP_ME_IC_BASE_CNTL, CACHE_POLICY, 0); 3337 tmp = REG_SET_FIELD(tmp, CP_ME_IC_BASE_CNTL, EXE_DISABLE, 0); 3338 WREG32_SOC15(GC, 0, regCP_ME_IC_BASE_CNTL, tmp); 3339 3340 /* 3341 * Programming any of the CP_ME_IC_BASE registers 3342 * forces invalidation of the ME L1 I$. Wait for the 3343 * invalidation complete 3344 */ 3345 for (i = 0; i < usec_timeout; i++) { 3346 tmp = RREG32_SOC15(GC, 0, regCP_ME_IC_OP_CNTL); 3347 if (1 == REG_GET_FIELD(tmp, CP_ME_IC_OP_CNTL, 3348 INVALIDATE_CACHE_COMPLETE)) 3349 break; 3350 udelay(1); 3351 } 3352 3353 if (i >= usec_timeout) { 3354 dev_err(adev->dev, "failed to invalidate instruction cache\n"); 3355 return -EINVAL; 3356 } 3357 3358 /* Prime the instruction caches */ 3359 tmp = RREG32_SOC15(GC, 0, regCP_ME_IC_OP_CNTL); 3360 tmp = REG_SET_FIELD(tmp, CP_ME_IC_OP_CNTL, PRIME_ICACHE, 1); 3361 WREG32_SOC15(GC, 0, regCP_ME_IC_OP_CNTL, tmp); 3362 3363 /* Waiting for instruction cache primed*/ 3364 for (i = 0; i < usec_timeout; i++) { 3365 tmp = RREG32_SOC15(GC, 0, regCP_ME_IC_OP_CNTL); 3366 if (1 == REG_GET_FIELD(tmp, CP_ME_IC_OP_CNTL, 3367 ICACHE_PRIMED)) 3368 break; 3369 udelay(1); 3370 } 3371 3372 if (i >= usec_timeout) { 3373 dev_err(adev->dev, "failed to prime instruction cache\n"); 3374 return -EINVAL; 3375 } 3376 3377 mutex_lock(&adev->srbm_mutex); 3378 for (pipe_id = 0; pipe_id < adev->gfx.me.num_pipe_per_me; pipe_id++) { 3379 soc21_grbm_select(adev, 0, pipe_id, 0, 0); 3380 WREG32_SOC15(GC, 0, regCP_ME_PRGRM_CNTR_START, 3381 (me_hdr->ucode_start_addr_hi << 30) | 3382 (me_hdr->ucode_start_addr_lo >> 2) ); 3383 WREG32_SOC15(GC, 0, regCP_ME_PRGRM_CNTR_START_HI, 3384 me_hdr->ucode_start_addr_hi>>2); 3385 3386 /* 3387 * Program CP_ME_CNTL to reset given PIPE to take 3388 * effect of CP_PFP_PRGRM_CNTR_START. 3389 */ 3390 tmp = RREG32_SOC15(GC, 0, regCP_ME_CNTL); 3391 if (pipe_id == 0) 3392 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 3393 ME_PIPE0_RESET, 1); 3394 else 3395 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 3396 ME_PIPE1_RESET, 1); 3397 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 3398 3399 /* Clear pfp pipe0 reset bit. */ 3400 if (pipe_id == 0) 3401 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 3402 ME_PIPE0_RESET, 0); 3403 else 3404 tmp = REG_SET_FIELD(tmp, CP_ME_CNTL, 3405 ME_PIPE1_RESET, 0); 3406 WREG32_SOC15(GC, 0, regCP_ME_CNTL, tmp); 3407 3408 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE1_LO, 3409 lower_32_bits(adev->gfx.me.me_fw_data_gpu_addr)); 3410 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE1_HI, 3411 upper_32_bits(adev->gfx.me.me_fw_data_gpu_addr)); 3412 } 3413 soc21_grbm_select(adev, 0, 0, 0, 0); 3414 mutex_unlock(&adev->srbm_mutex); 3415 3416 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE_CNTL); 3417 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_BASE_CNTL, VMID, 0); 3418 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_BASE_CNTL, CACHE_POLICY, 0); 3419 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_BASE_CNTL, tmp); 3420 3421 /* Invalidate the data caches */ 3422 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL); 3423 tmp = REG_SET_FIELD(tmp, CP_GFX_RS64_DC_OP_CNTL, INVALIDATE_DCACHE, 1); 3424 WREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL, tmp); 3425 3426 for (i = 0; i < usec_timeout; i++) { 3427 tmp = RREG32_SOC15(GC, 0, regCP_GFX_RS64_DC_OP_CNTL); 3428 if (1 == REG_GET_FIELD(tmp, CP_GFX_RS64_DC_OP_CNTL, 3429 INVALIDATE_DCACHE_COMPLETE)) 3430 break; 3431 udelay(1); 3432 } 3433 3434 if (i >= usec_timeout) { 3435 dev_err(adev->dev, "failed to invalidate RS64 data cache\n"); 3436 return -EINVAL; 3437 } 3438 3439 return 0; 3440 } 3441 3442 static int gfx_v11_0_cp_gfx_load_microcode(struct amdgpu_device *adev) 3443 { 3444 int r; 3445 3446 if (!adev->gfx.me_fw || !adev->gfx.pfp_fw) 3447 return -EINVAL; 3448 3449 gfx_v11_0_cp_gfx_enable(adev, false); 3450 3451 if (adev->gfx.rs64_enable) 3452 r = gfx_v11_0_cp_gfx_load_pfp_microcode_rs64(adev); 3453 else 3454 r = gfx_v11_0_cp_gfx_load_pfp_microcode(adev); 3455 if (r) { 3456 dev_err(adev->dev, "(%d) failed to load pfp fw\n", r); 3457 return r; 3458 } 3459 3460 if (adev->gfx.rs64_enable) 3461 r = gfx_v11_0_cp_gfx_load_me_microcode_rs64(adev); 3462 else 3463 r = gfx_v11_0_cp_gfx_load_me_microcode(adev); 3464 if (r) { 3465 dev_err(adev->dev, "(%d) failed to load me fw\n", r); 3466 return r; 3467 } 3468 3469 return 0; 3470 } 3471 3472 static int gfx_v11_0_cp_gfx_start(struct amdgpu_device *adev) 3473 { 3474 struct amdgpu_ring *ring; 3475 const struct cs_section_def *sect = NULL; 3476 const struct cs_extent_def *ext = NULL; 3477 int r, i; 3478 int ctx_reg_offset; 3479 3480 /* init the CP */ 3481 WREG32_SOC15(GC, 0, regCP_MAX_CONTEXT, 3482 adev->gfx.config.max_hw_contexts - 1); 3483 WREG32_SOC15(GC, 0, regCP_DEVICE_ID, 1); 3484 3485 if (!amdgpu_async_gfx_ring) 3486 gfx_v11_0_cp_gfx_enable(adev, true); 3487 3488 ring = &adev->gfx.gfx_ring[0]; 3489 r = amdgpu_ring_alloc(ring, gfx_v11_0_get_csb_size(adev)); 3490 if (r) { 3491 DRM_ERROR("amdgpu: cp failed to lock ring (%d).\n", r); 3492 return r; 3493 } 3494 3495 amdgpu_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0)); 3496 amdgpu_ring_write(ring, PACKET3_PREAMBLE_BEGIN_CLEAR_STATE); 3497 3498 amdgpu_ring_write(ring, PACKET3(PACKET3_CONTEXT_CONTROL, 1)); 3499 amdgpu_ring_write(ring, 0x80000000); 3500 amdgpu_ring_write(ring, 0x80000000); 3501 3502 for (sect = gfx11_cs_data; sect->section != NULL; ++sect) { 3503 for (ext = sect->section; ext->extent != NULL; ++ext) { 3504 if (sect->id == SECT_CONTEXT) { 3505 amdgpu_ring_write(ring, 3506 PACKET3(PACKET3_SET_CONTEXT_REG, 3507 ext->reg_count)); 3508 amdgpu_ring_write(ring, ext->reg_index - 3509 PACKET3_SET_CONTEXT_REG_START); 3510 for (i = 0; i < ext->reg_count; i++) 3511 amdgpu_ring_write(ring, ext->extent[i]); 3512 } 3513 } 3514 } 3515 3516 ctx_reg_offset = 3517 SOC15_REG_OFFSET(GC, 0, regPA_SC_TILE_STEERING_OVERRIDE) - PACKET3_SET_CONTEXT_REG_START; 3518 amdgpu_ring_write(ring, PACKET3(PACKET3_SET_CONTEXT_REG, 1)); 3519 amdgpu_ring_write(ring, ctx_reg_offset); 3520 amdgpu_ring_write(ring, adev->gfx.config.pa_sc_tile_steering_override); 3521 3522 amdgpu_ring_write(ring, PACKET3(PACKET3_PREAMBLE_CNTL, 0)); 3523 amdgpu_ring_write(ring, PACKET3_PREAMBLE_END_CLEAR_STATE); 3524 3525 amdgpu_ring_write(ring, PACKET3(PACKET3_CLEAR_STATE, 0)); 3526 amdgpu_ring_write(ring, 0); 3527 3528 amdgpu_ring_commit(ring); 3529 3530 /* submit cs packet to copy state 0 to next available state */ 3531 if (adev->gfx.num_gfx_rings > 1) { 3532 /* maximum supported gfx ring is 2 */ 3533 ring = &adev->gfx.gfx_ring[1]; 3534 r = amdgpu_ring_alloc(ring, 2); 3535 if (r) { 3536 DRM_ERROR("amdgpu: cp failed to lock ring (%d).\n", r); 3537 return r; 3538 } 3539 3540 amdgpu_ring_write(ring, PACKET3(PACKET3_CLEAR_STATE, 0)); 3541 amdgpu_ring_write(ring, 0); 3542 3543 amdgpu_ring_commit(ring); 3544 } 3545 return 0; 3546 } 3547 3548 static void gfx_v11_0_cp_gfx_switch_pipe(struct amdgpu_device *adev, 3549 CP_PIPE_ID pipe) 3550 { 3551 u32 tmp; 3552 3553 tmp = RREG32_SOC15(GC, 0, regGRBM_GFX_CNTL); 3554 tmp = REG_SET_FIELD(tmp, GRBM_GFX_CNTL, PIPEID, pipe); 3555 3556 WREG32_SOC15(GC, 0, regGRBM_GFX_CNTL, tmp); 3557 } 3558 3559 static void gfx_v11_0_cp_gfx_set_doorbell(struct amdgpu_device *adev, 3560 struct amdgpu_ring *ring) 3561 { 3562 u32 tmp; 3563 3564 tmp = RREG32_SOC15(GC, 0, regCP_RB_DOORBELL_CONTROL); 3565 if (ring->use_doorbell) { 3566 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL, 3567 DOORBELL_OFFSET, ring->doorbell_index); 3568 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL, 3569 DOORBELL_EN, 1); 3570 } else { 3571 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL, 3572 DOORBELL_EN, 0); 3573 } 3574 WREG32_SOC15(GC, 0, regCP_RB_DOORBELL_CONTROL, tmp); 3575 3576 tmp = REG_SET_FIELD(0, CP_RB_DOORBELL_RANGE_LOWER, 3577 DOORBELL_RANGE_LOWER, ring->doorbell_index); 3578 WREG32_SOC15(GC, 0, regCP_RB_DOORBELL_RANGE_LOWER, tmp); 3579 3580 WREG32_SOC15(GC, 0, regCP_RB_DOORBELL_RANGE_UPPER, 3581 CP_RB_DOORBELL_RANGE_UPPER__DOORBELL_RANGE_UPPER_MASK); 3582 } 3583 3584 static int gfx_v11_0_cp_gfx_resume(struct amdgpu_device *adev) 3585 { 3586 struct amdgpu_ring *ring; 3587 u32 tmp; 3588 u32 rb_bufsz; 3589 u64 rb_addr, rptr_addr, wptr_gpu_addr; 3590 3591 /* Set the write pointer delay */ 3592 WREG32_SOC15(GC, 0, regCP_RB_WPTR_DELAY, 0); 3593 3594 /* set the RB to use vmid 0 */ 3595 WREG32_SOC15(GC, 0, regCP_RB_VMID, 0); 3596 3597 /* Init gfx ring 0 for pipe 0 */ 3598 mutex_lock(&adev->srbm_mutex); 3599 gfx_v11_0_cp_gfx_switch_pipe(adev, PIPE_ID0); 3600 3601 /* Set ring buffer size */ 3602 ring = &adev->gfx.gfx_ring[0]; 3603 rb_bufsz = order_base_2(ring->ring_size / 8); 3604 tmp = REG_SET_FIELD(0, CP_RB0_CNTL, RB_BUFSZ, rb_bufsz); 3605 tmp = REG_SET_FIELD(tmp, CP_RB0_CNTL, RB_BLKSZ, rb_bufsz - 2); 3606 WREG32_SOC15(GC, 0, regCP_RB0_CNTL, tmp); 3607 3608 /* Initialize the ring buffer's write pointers */ 3609 ring->wptr = 0; 3610 WREG32_SOC15(GC, 0, regCP_RB0_WPTR, lower_32_bits(ring->wptr)); 3611 WREG32_SOC15(GC, 0, regCP_RB0_WPTR_HI, upper_32_bits(ring->wptr)); 3612 3613 /* set the wb address whether it's enabled or not */ 3614 rptr_addr = ring->rptr_gpu_addr; 3615 WREG32_SOC15(GC, 0, regCP_RB0_RPTR_ADDR, lower_32_bits(rptr_addr)); 3616 WREG32_SOC15(GC, 0, regCP_RB0_RPTR_ADDR_HI, upper_32_bits(rptr_addr) & 3617 CP_RB_RPTR_ADDR_HI__RB_RPTR_ADDR_HI_MASK); 3618 3619 wptr_gpu_addr = ring->wptr_gpu_addr; 3620 WREG32_SOC15(GC, 0, regCP_RB_WPTR_POLL_ADDR_LO, 3621 lower_32_bits(wptr_gpu_addr)); 3622 WREG32_SOC15(GC, 0, regCP_RB_WPTR_POLL_ADDR_HI, 3623 upper_32_bits(wptr_gpu_addr)); 3624 3625 mdelay(1); 3626 WREG32_SOC15(GC, 0, regCP_RB0_CNTL, tmp); 3627 3628 rb_addr = ring->gpu_addr >> 8; 3629 WREG32_SOC15(GC, 0, regCP_RB0_BASE, rb_addr); 3630 WREG32_SOC15(GC, 0, regCP_RB0_BASE_HI, upper_32_bits(rb_addr)); 3631 3632 WREG32_SOC15(GC, 0, regCP_RB_ACTIVE, 1); 3633 3634 gfx_v11_0_cp_gfx_set_doorbell(adev, ring); 3635 mutex_unlock(&adev->srbm_mutex); 3636 3637 /* Init gfx ring 1 for pipe 1 */ 3638 if (adev->gfx.num_gfx_rings > 1) { 3639 mutex_lock(&adev->srbm_mutex); 3640 gfx_v11_0_cp_gfx_switch_pipe(adev, PIPE_ID1); 3641 /* maximum supported gfx ring is 2 */ 3642 ring = &adev->gfx.gfx_ring[1]; 3643 rb_bufsz = order_base_2(ring->ring_size / 8); 3644 tmp = REG_SET_FIELD(0, CP_RB1_CNTL, RB_BUFSZ, rb_bufsz); 3645 tmp = REG_SET_FIELD(tmp, CP_RB1_CNTL, RB_BLKSZ, rb_bufsz - 2); 3646 WREG32_SOC15(GC, 0, regCP_RB1_CNTL, tmp); 3647 /* Initialize the ring buffer's write pointers */ 3648 ring->wptr = 0; 3649 WREG32_SOC15(GC, 0, regCP_RB1_WPTR, lower_32_bits(ring->wptr)); 3650 WREG32_SOC15(GC, 0, regCP_RB1_WPTR_HI, upper_32_bits(ring->wptr)); 3651 /* Set the wb address whether it's enabled or not */ 3652 rptr_addr = ring->rptr_gpu_addr; 3653 WREG32_SOC15(GC, 0, regCP_RB1_RPTR_ADDR, lower_32_bits(rptr_addr)); 3654 WREG32_SOC15(GC, 0, regCP_RB1_RPTR_ADDR_HI, upper_32_bits(rptr_addr) & 3655 CP_RB1_RPTR_ADDR_HI__RB_RPTR_ADDR_HI_MASK); 3656 wptr_gpu_addr = ring->wptr_gpu_addr; 3657 WREG32_SOC15(GC, 0, regCP_RB_WPTR_POLL_ADDR_LO, 3658 lower_32_bits(wptr_gpu_addr)); 3659 WREG32_SOC15(GC, 0, regCP_RB_WPTR_POLL_ADDR_HI, 3660 upper_32_bits(wptr_gpu_addr)); 3661 3662 mdelay(1); 3663 WREG32_SOC15(GC, 0, regCP_RB1_CNTL, tmp); 3664 3665 rb_addr = ring->gpu_addr >> 8; 3666 WREG32_SOC15(GC, 0, regCP_RB1_BASE, rb_addr); 3667 WREG32_SOC15(GC, 0, regCP_RB1_BASE_HI, upper_32_bits(rb_addr)); 3668 WREG32_SOC15(GC, 0, regCP_RB1_ACTIVE, 1); 3669 3670 gfx_v11_0_cp_gfx_set_doorbell(adev, ring); 3671 mutex_unlock(&adev->srbm_mutex); 3672 } 3673 /* Switch to pipe 0 */ 3674 mutex_lock(&adev->srbm_mutex); 3675 gfx_v11_0_cp_gfx_switch_pipe(adev, PIPE_ID0); 3676 mutex_unlock(&adev->srbm_mutex); 3677 3678 /* start the ring */ 3679 gfx_v11_0_cp_gfx_start(adev); 3680 3681 return 0; 3682 } 3683 3684 static void gfx_v11_0_cp_compute_enable(struct amdgpu_device *adev, bool enable) 3685 { 3686 u32 data; 3687 3688 if (adev->gfx.rs64_enable) { 3689 data = RREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL); 3690 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_INVALIDATE_ICACHE, 3691 enable ? 0 : 1); 3692 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE0_RESET, 3693 enable ? 0 : 1); 3694 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE1_RESET, 3695 enable ? 0 : 1); 3696 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE2_RESET, 3697 enable ? 0 : 1); 3698 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE3_RESET, 3699 enable ? 0 : 1); 3700 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE0_ACTIVE, 3701 enable ? 1 : 0); 3702 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE1_ACTIVE, 3703 enable ? 1 : 0); 3704 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE2_ACTIVE, 3705 enable ? 1 : 0); 3706 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_PIPE3_ACTIVE, 3707 enable ? 1 : 0); 3708 data = REG_SET_FIELD(data, CP_MEC_RS64_CNTL, MEC_HALT, 3709 enable ? 0 : 1); 3710 WREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL, data); 3711 } else { 3712 data = RREG32_SOC15(GC, 0, regCP_MEC_CNTL); 3713 3714 if (enable) { 3715 data = REG_SET_FIELD(data, CP_MEC_CNTL, MEC_ME1_HALT, 0); 3716 if (!adev->enable_mes_kiq) 3717 data = REG_SET_FIELD(data, CP_MEC_CNTL, 3718 MEC_ME2_HALT, 0); 3719 } else { 3720 data = REG_SET_FIELD(data, CP_MEC_CNTL, MEC_ME1_HALT, 1); 3721 data = REG_SET_FIELD(data, CP_MEC_CNTL, MEC_ME2_HALT, 1); 3722 } 3723 WREG32_SOC15(GC, 0, regCP_MEC_CNTL, data); 3724 } 3725 3726 udelay(50); 3727 } 3728 3729 static int gfx_v11_0_cp_compute_load_microcode(struct amdgpu_device *adev) 3730 { 3731 const struct gfx_firmware_header_v1_0 *mec_hdr; 3732 const __le32 *fw_data; 3733 unsigned i, fw_size; 3734 u32 *fw = NULL; 3735 int r; 3736 3737 if (!adev->gfx.mec_fw) 3738 return -EINVAL; 3739 3740 gfx_v11_0_cp_compute_enable(adev, false); 3741 3742 mec_hdr = (const struct gfx_firmware_header_v1_0 *)adev->gfx.mec_fw->data; 3743 amdgpu_ucode_print_gfx_hdr(&mec_hdr->header); 3744 3745 fw_data = (const __le32 *) 3746 (adev->gfx.mec_fw->data + 3747 le32_to_cpu(mec_hdr->header.ucode_array_offset_bytes)); 3748 fw_size = le32_to_cpu(mec_hdr->header.ucode_size_bytes); 3749 3750 r = amdgpu_bo_create_reserved(adev, mec_hdr->header.ucode_size_bytes, 3751 PAGE_SIZE, AMDGPU_GEM_DOMAIN_GTT, 3752 &adev->gfx.mec.mec_fw_obj, 3753 &adev->gfx.mec.mec_fw_gpu_addr, 3754 (void **)&fw); 3755 if (r) { 3756 dev_err(adev->dev, "(%d) failed to create mec fw bo\n", r); 3757 gfx_v11_0_mec_fini(adev); 3758 return r; 3759 } 3760 3761 memcpy(fw, fw_data, fw_size); 3762 3763 amdgpu_bo_kunmap(adev->gfx.mec.mec_fw_obj); 3764 amdgpu_bo_unreserve(adev->gfx.mec.mec_fw_obj); 3765 3766 gfx_v11_0_config_mec_cache(adev, adev->gfx.mec.mec_fw_gpu_addr); 3767 3768 /* MEC1 */ 3769 WREG32_SOC15(GC, 0, regCP_MEC_ME1_UCODE_ADDR, 0); 3770 3771 for (i = 0; i < mec_hdr->jt_size; i++) 3772 WREG32_SOC15(GC, 0, regCP_MEC_ME1_UCODE_DATA, 3773 le32_to_cpup(fw_data + mec_hdr->jt_offset + i)); 3774 3775 WREG32_SOC15(GC, 0, regCP_MEC_ME1_UCODE_ADDR, adev->gfx.mec_fw_version); 3776 3777 return 0; 3778 } 3779 3780 static int gfx_v11_0_cp_compute_load_microcode_rs64(struct amdgpu_device *adev) 3781 { 3782 const struct gfx_firmware_header_v2_0 *mec_hdr; 3783 const __le32 *fw_ucode, *fw_data; 3784 u32 tmp, fw_ucode_size, fw_data_size; 3785 u32 i, usec_timeout = 50000; /* Wait for 50 ms */ 3786 u32 *fw_ucode_ptr, *fw_data_ptr; 3787 int r; 3788 3789 if (!adev->gfx.mec_fw) 3790 return -EINVAL; 3791 3792 gfx_v11_0_cp_compute_enable(adev, false); 3793 3794 mec_hdr = (const struct gfx_firmware_header_v2_0 *)adev->gfx.mec_fw->data; 3795 amdgpu_ucode_print_gfx_hdr(&mec_hdr->header); 3796 3797 fw_ucode = (const __le32 *) (adev->gfx.mec_fw->data + 3798 le32_to_cpu(mec_hdr->ucode_offset_bytes)); 3799 fw_ucode_size = le32_to_cpu(mec_hdr->ucode_size_bytes); 3800 3801 fw_data = (const __le32 *) (adev->gfx.mec_fw->data + 3802 le32_to_cpu(mec_hdr->data_offset_bytes)); 3803 fw_data_size = le32_to_cpu(mec_hdr->data_size_bytes); 3804 3805 r = amdgpu_bo_create_reserved(adev, fw_ucode_size, 3806 64 * 1024, 3807 AMDGPU_GEM_DOMAIN_VRAM | 3808 AMDGPU_GEM_DOMAIN_GTT, 3809 &adev->gfx.mec.mec_fw_obj, 3810 &adev->gfx.mec.mec_fw_gpu_addr, 3811 (void **)&fw_ucode_ptr); 3812 if (r) { 3813 dev_err(adev->dev, "(%d) failed to create mec fw ucode bo\n", r); 3814 gfx_v11_0_mec_fini(adev); 3815 return r; 3816 } 3817 3818 r = amdgpu_bo_create_reserved(adev, fw_data_size, 3819 64 * 1024, 3820 AMDGPU_GEM_DOMAIN_VRAM | 3821 AMDGPU_GEM_DOMAIN_GTT, 3822 &adev->gfx.mec.mec_fw_data_obj, 3823 &adev->gfx.mec.mec_fw_data_gpu_addr, 3824 (void **)&fw_data_ptr); 3825 if (r) { 3826 dev_err(adev->dev, "(%d) failed to create mec fw ucode bo\n", r); 3827 gfx_v11_0_mec_fini(adev); 3828 return r; 3829 } 3830 3831 memcpy(fw_ucode_ptr, fw_ucode, fw_ucode_size); 3832 memcpy(fw_data_ptr, fw_data, fw_data_size); 3833 3834 amdgpu_bo_kunmap(adev->gfx.mec.mec_fw_obj); 3835 amdgpu_bo_kunmap(adev->gfx.mec.mec_fw_data_obj); 3836 amdgpu_bo_unreserve(adev->gfx.mec.mec_fw_obj); 3837 amdgpu_bo_unreserve(adev->gfx.mec.mec_fw_data_obj); 3838 3839 tmp = RREG32_SOC15(GC, 0, regCP_CPC_IC_BASE_CNTL); 3840 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_BASE_CNTL, VMID, 0); 3841 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_BASE_CNTL, EXE_DISABLE, 0); 3842 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_BASE_CNTL, CACHE_POLICY, 0); 3843 WREG32_SOC15(GC, 0, regCP_CPC_IC_BASE_CNTL, tmp); 3844 3845 tmp = RREG32_SOC15(GC, 0, regCP_MEC_DC_BASE_CNTL); 3846 tmp = REG_SET_FIELD(tmp, CP_MEC_DC_BASE_CNTL, VMID, 0); 3847 tmp = REG_SET_FIELD(tmp, CP_MEC_DC_BASE_CNTL, CACHE_POLICY, 0); 3848 WREG32_SOC15(GC, 0, regCP_MEC_DC_BASE_CNTL, tmp); 3849 3850 mutex_lock(&adev->srbm_mutex); 3851 for (i = 0; i < adev->gfx.mec.num_pipe_per_mec; i++) { 3852 soc21_grbm_select(adev, 1, i, 0, 0); 3853 3854 WREG32_SOC15(GC, 0, regCP_MEC_MDBASE_LO, adev->gfx.mec.mec_fw_data_gpu_addr); 3855 WREG32_SOC15(GC, 0, regCP_MEC_MDBASE_HI, 3856 upper_32_bits(adev->gfx.mec.mec_fw_data_gpu_addr)); 3857 3858 WREG32_SOC15(GC, 0, regCP_MEC_RS64_PRGRM_CNTR_START, 3859 mec_hdr->ucode_start_addr_lo >> 2 | 3860 mec_hdr->ucode_start_addr_hi << 30); 3861 WREG32_SOC15(GC, 0, regCP_MEC_RS64_PRGRM_CNTR_START_HI, 3862 mec_hdr->ucode_start_addr_hi >> 2); 3863 3864 WREG32_SOC15(GC, 0, regCP_CPC_IC_BASE_LO, adev->gfx.mec.mec_fw_gpu_addr); 3865 WREG32_SOC15(GC, 0, regCP_CPC_IC_BASE_HI, 3866 upper_32_bits(adev->gfx.mec.mec_fw_gpu_addr)); 3867 } 3868 mutex_unlock(&adev->srbm_mutex); 3869 soc21_grbm_select(adev, 0, 0, 0, 0); 3870 3871 /* Trigger an invalidation of the L1 instruction caches */ 3872 tmp = RREG32_SOC15(GC, 0, regCP_MEC_DC_OP_CNTL); 3873 tmp = REG_SET_FIELD(tmp, CP_MEC_DC_OP_CNTL, INVALIDATE_DCACHE, 1); 3874 WREG32_SOC15(GC, 0, regCP_MEC_DC_OP_CNTL, tmp); 3875 3876 /* Wait for invalidation complete */ 3877 for (i = 0; i < usec_timeout; i++) { 3878 tmp = RREG32_SOC15(GC, 0, regCP_MEC_DC_OP_CNTL); 3879 if (1 == REG_GET_FIELD(tmp, CP_MEC_DC_OP_CNTL, 3880 INVALIDATE_DCACHE_COMPLETE)) 3881 break; 3882 udelay(1); 3883 } 3884 3885 if (i >= usec_timeout) { 3886 dev_err(adev->dev, "failed to invalidate instruction cache\n"); 3887 return -EINVAL; 3888 } 3889 3890 /* Trigger an invalidation of the L1 instruction caches */ 3891 tmp = RREG32_SOC15(GC, 0, regCP_CPC_IC_OP_CNTL); 3892 tmp = REG_SET_FIELD(tmp, CP_CPC_IC_OP_CNTL, INVALIDATE_CACHE, 1); 3893 WREG32_SOC15(GC, 0, regCP_CPC_IC_OP_CNTL, tmp); 3894 3895 /* Wait for invalidation complete */ 3896 for (i = 0; i < usec_timeout; i++) { 3897 tmp = RREG32_SOC15(GC, 0, regCP_CPC_IC_OP_CNTL); 3898 if (1 == REG_GET_FIELD(tmp, CP_CPC_IC_OP_CNTL, 3899 INVALIDATE_CACHE_COMPLETE)) 3900 break; 3901 udelay(1); 3902 } 3903 3904 if (i >= usec_timeout) { 3905 dev_err(adev->dev, "failed to invalidate instruction cache\n"); 3906 return -EINVAL; 3907 } 3908 3909 return 0; 3910 } 3911 3912 static void gfx_v11_0_kiq_setting(struct amdgpu_ring *ring) 3913 { 3914 uint32_t tmp; 3915 struct amdgpu_device *adev = ring->adev; 3916 3917 /* tell RLC which is KIQ queue */ 3918 tmp = RREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS); 3919 tmp &= 0xffffff00; 3920 tmp |= (ring->me << 5) | (ring->pipe << 3) | (ring->queue); 3921 WREG32_SOC15(GC, 0, regRLC_CP_SCHEDULERS, tmp | 0x80); 3922 } 3923 3924 static void gfx_v11_0_cp_set_doorbell_range(struct amdgpu_device *adev) 3925 { 3926 /* set graphics engine doorbell range */ 3927 WREG32_SOC15(GC, 0, regCP_RB_DOORBELL_RANGE_LOWER, 3928 (adev->doorbell_index.gfx_ring0 * 2) << 2); 3929 WREG32_SOC15(GC, 0, regCP_RB_DOORBELL_RANGE_UPPER, 3930 (adev->doorbell_index.gfx_userqueue_end * 2) << 2); 3931 3932 /* set compute engine doorbell range */ 3933 WREG32_SOC15(GC, 0, regCP_MEC_DOORBELL_RANGE_LOWER, 3934 (adev->doorbell_index.kiq * 2) << 2); 3935 WREG32_SOC15(GC, 0, regCP_MEC_DOORBELL_RANGE_UPPER, 3936 (adev->doorbell_index.userqueue_end * 2) << 2); 3937 } 3938 3939 static void gfx_v11_0_gfx_mqd_set_priority(struct amdgpu_device *adev, 3940 struct v11_gfx_mqd *mqd, 3941 struct amdgpu_mqd_prop *prop) 3942 { 3943 bool priority = 0; 3944 u32 tmp; 3945 3946 /* set up default queue priority level 3947 * 0x0 = low priority, 0x1 = high priority 3948 */ 3949 if (prop->hqd_pipe_priority == AMDGPU_GFX_PIPE_PRIO_HIGH) 3950 priority = 1; 3951 3952 tmp = RREG32_SOC15(GC, 0, regCP_GFX_HQD_QUEUE_PRIORITY); 3953 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_QUEUE_PRIORITY, PRIORITY_LEVEL, priority); 3954 mqd->cp_gfx_hqd_queue_priority = tmp; 3955 } 3956 3957 static int gfx_v11_0_gfx_mqd_init(struct amdgpu_device *adev, void *m, 3958 struct amdgpu_mqd_prop *prop) 3959 { 3960 struct v11_gfx_mqd *mqd = m; 3961 uint64_t hqd_gpu_addr, wb_gpu_addr; 3962 uint32_t tmp; 3963 uint32_t rb_bufsz; 3964 3965 /* set up gfx hqd wptr */ 3966 mqd->cp_gfx_hqd_wptr = 0; 3967 mqd->cp_gfx_hqd_wptr_hi = 0; 3968 3969 /* set the pointer to the MQD */ 3970 mqd->cp_mqd_base_addr = prop->mqd_gpu_addr & 0xfffffffc; 3971 mqd->cp_mqd_base_addr_hi = upper_32_bits(prop->mqd_gpu_addr); 3972 3973 /* set up mqd control */ 3974 tmp = RREG32_SOC15(GC, 0, regCP_GFX_MQD_CONTROL); 3975 tmp = REG_SET_FIELD(tmp, CP_GFX_MQD_CONTROL, VMID, 0); 3976 tmp = REG_SET_FIELD(tmp, CP_GFX_MQD_CONTROL, PRIV_STATE, 1); 3977 tmp = REG_SET_FIELD(tmp, CP_GFX_MQD_CONTROL, CACHE_POLICY, 0); 3978 mqd->cp_gfx_mqd_control = tmp; 3979 3980 /* set up gfx_hqd_vimd with 0x0 to indicate the ring buffer's vmid */ 3981 tmp = RREG32_SOC15(GC, 0, regCP_GFX_HQD_VMID); 3982 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_VMID, VMID, 0); 3983 mqd->cp_gfx_hqd_vmid = 0; 3984 3985 /* set up gfx queue priority */ 3986 gfx_v11_0_gfx_mqd_set_priority(adev, mqd, prop); 3987 3988 /* set up time quantum */ 3989 tmp = RREG32_SOC15(GC, 0, regCP_GFX_HQD_QUANTUM); 3990 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_QUANTUM, QUANTUM_EN, 1); 3991 mqd->cp_gfx_hqd_quantum = tmp; 3992 3993 /* set up gfx hqd base. this is similar as CP_RB_BASE */ 3994 hqd_gpu_addr = prop->hqd_base_gpu_addr >> 8; 3995 mqd->cp_gfx_hqd_base = hqd_gpu_addr; 3996 mqd->cp_gfx_hqd_base_hi = upper_32_bits(hqd_gpu_addr); 3997 3998 /* set up hqd_rptr_addr/_hi, similar as CP_RB_RPTR */ 3999 wb_gpu_addr = prop->rptr_gpu_addr; 4000 mqd->cp_gfx_hqd_rptr_addr = wb_gpu_addr & 0xfffffffc; 4001 mqd->cp_gfx_hqd_rptr_addr_hi = 4002 upper_32_bits(wb_gpu_addr) & 0xffff; 4003 4004 /* set up rb_wptr_poll addr */ 4005 wb_gpu_addr = prop->wptr_gpu_addr; 4006 mqd->cp_rb_wptr_poll_addr_lo = wb_gpu_addr & 0xfffffffc; 4007 mqd->cp_rb_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr) & 0xffff; 4008 4009 /* set up the gfx_hqd_control, similar as CP_RB0_CNTL */ 4010 rb_bufsz = order_base_2(prop->queue_size / 4) - 1; 4011 tmp = RREG32_SOC15(GC, 0, regCP_GFX_HQD_CNTL); 4012 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_CNTL, RB_BUFSZ, rb_bufsz); 4013 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_CNTL, RB_BLKSZ, rb_bufsz - 2); 4014 #ifdef __BIG_ENDIAN 4015 tmp = REG_SET_FIELD(tmp, CP_GFX_HQD_CNTL, BUF_SWAP, 1); 4016 #endif 4017 mqd->cp_gfx_hqd_cntl = tmp; 4018 4019 /* set up cp_doorbell_control */ 4020 tmp = RREG32_SOC15(GC, 0, regCP_RB_DOORBELL_CONTROL); 4021 if (prop->use_doorbell) { 4022 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL, 4023 DOORBELL_OFFSET, prop->doorbell_index); 4024 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL, 4025 DOORBELL_EN, 1); 4026 } else 4027 tmp = REG_SET_FIELD(tmp, CP_RB_DOORBELL_CONTROL, 4028 DOORBELL_EN, 0); 4029 mqd->cp_rb_doorbell_control = tmp; 4030 4031 /* reset read and write pointers, similar to CP_RB0_WPTR/_RPTR */ 4032 mqd->cp_gfx_hqd_rptr = RREG32_SOC15(GC, 0, regCP_GFX_HQD_RPTR); 4033 4034 /* active the queue */ 4035 mqd->cp_gfx_hqd_active = 1; 4036 4037 return 0; 4038 } 4039 4040 static int gfx_v11_0_kgq_init_queue(struct amdgpu_ring *ring, bool reset) 4041 { 4042 struct amdgpu_device *adev = ring->adev; 4043 struct v11_gfx_mqd *mqd = ring->mqd_ptr; 4044 int mqd_idx = ring - &adev->gfx.gfx_ring[0]; 4045 4046 if (!reset && !amdgpu_in_reset(adev) && !adev->in_suspend) { 4047 memset((void *)mqd, 0, sizeof(*mqd)); 4048 mutex_lock(&adev->srbm_mutex); 4049 soc21_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0); 4050 amdgpu_ring_init_mqd(ring); 4051 soc21_grbm_select(adev, 0, 0, 0, 0); 4052 mutex_unlock(&adev->srbm_mutex); 4053 if (adev->gfx.me.mqd_backup[mqd_idx]) 4054 memcpy_fromio(adev->gfx.me.mqd_backup[mqd_idx], mqd, sizeof(*mqd)); 4055 } else { 4056 /* restore mqd with the backup copy */ 4057 if (adev->gfx.me.mqd_backup[mqd_idx]) 4058 memcpy_toio(mqd, adev->gfx.me.mqd_backup[mqd_idx], sizeof(*mqd)); 4059 /* reset the ring */ 4060 ring->wptr = 0; 4061 *ring->wptr_cpu_addr = 0; 4062 amdgpu_ring_clear_ring(ring); 4063 } 4064 4065 return 0; 4066 } 4067 4068 static int gfx_v11_0_cp_async_gfx_ring_resume(struct amdgpu_device *adev) 4069 { 4070 int r, i; 4071 struct amdgpu_ring *ring; 4072 4073 for (i = 0; i < adev->gfx.num_gfx_rings; i++) { 4074 ring = &adev->gfx.gfx_ring[i]; 4075 4076 r = amdgpu_bo_reserve(ring->mqd_obj, false); 4077 if (unlikely(r != 0)) 4078 return r; 4079 4080 r = amdgpu_bo_kmap(ring->mqd_obj, (void **)&ring->mqd_ptr); 4081 if (!r) { 4082 r = gfx_v11_0_kgq_init_queue(ring, false); 4083 amdgpu_bo_kunmap(ring->mqd_obj); 4084 ring->mqd_ptr = NULL; 4085 } 4086 amdgpu_bo_unreserve(ring->mqd_obj); 4087 if (r) 4088 return r; 4089 } 4090 4091 r = amdgpu_gfx_enable_kgq(adev, 0); 4092 if (r) 4093 return r; 4094 4095 return gfx_v11_0_cp_gfx_start(adev); 4096 } 4097 4098 static int gfx_v11_0_compute_mqd_init(struct amdgpu_device *adev, void *m, 4099 struct amdgpu_mqd_prop *prop) 4100 { 4101 struct v11_compute_mqd *mqd = m; 4102 uint64_t hqd_gpu_addr, wb_gpu_addr, eop_base_addr; 4103 uint32_t tmp; 4104 4105 mqd->header = 0xC0310800; 4106 mqd->compute_pipelinestat_enable = 0x00000001; 4107 mqd->compute_static_thread_mgmt_se0 = 0xffffffff; 4108 mqd->compute_static_thread_mgmt_se1 = 0xffffffff; 4109 mqd->compute_static_thread_mgmt_se2 = 0xffffffff; 4110 mqd->compute_static_thread_mgmt_se3 = 0xffffffff; 4111 mqd->compute_misc_reserved = 0x00000007; 4112 4113 eop_base_addr = prop->eop_gpu_addr >> 8; 4114 mqd->cp_hqd_eop_base_addr_lo = eop_base_addr; 4115 mqd->cp_hqd_eop_base_addr_hi = upper_32_bits(eop_base_addr); 4116 4117 /* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */ 4118 tmp = RREG32_SOC15(GC, 0, regCP_HQD_EOP_CONTROL); 4119 tmp = REG_SET_FIELD(tmp, CP_HQD_EOP_CONTROL, EOP_SIZE, 4120 (order_base_2(GFX11_MEC_HPD_SIZE / 4) - 1)); 4121 4122 mqd->cp_hqd_eop_control = tmp; 4123 4124 /* enable doorbell? */ 4125 tmp = RREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL); 4126 4127 if (prop->use_doorbell) { 4128 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 4129 DOORBELL_OFFSET, prop->doorbell_index); 4130 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 4131 DOORBELL_EN, 1); 4132 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 4133 DOORBELL_SOURCE, 0); 4134 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 4135 DOORBELL_HIT, 0); 4136 } else { 4137 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 4138 DOORBELL_EN, 0); 4139 } 4140 4141 mqd->cp_hqd_pq_doorbell_control = tmp; 4142 4143 /* disable the queue if it's active */ 4144 mqd->cp_hqd_dequeue_request = 0; 4145 mqd->cp_hqd_pq_rptr = 0; 4146 mqd->cp_hqd_pq_wptr_lo = 0; 4147 mqd->cp_hqd_pq_wptr_hi = 0; 4148 4149 /* set the pointer to the MQD */ 4150 mqd->cp_mqd_base_addr_lo = prop->mqd_gpu_addr & 0xfffffffc; 4151 mqd->cp_mqd_base_addr_hi = upper_32_bits(prop->mqd_gpu_addr); 4152 4153 /* set MQD vmid to 0 */ 4154 tmp = RREG32_SOC15(GC, 0, regCP_MQD_CONTROL); 4155 tmp = REG_SET_FIELD(tmp, CP_MQD_CONTROL, VMID, 0); 4156 mqd->cp_mqd_control = tmp; 4157 4158 /* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */ 4159 hqd_gpu_addr = prop->hqd_base_gpu_addr >> 8; 4160 mqd->cp_hqd_pq_base_lo = hqd_gpu_addr; 4161 mqd->cp_hqd_pq_base_hi = upper_32_bits(hqd_gpu_addr); 4162 4163 /* set up the HQD, this is similar to CP_RB0_CNTL */ 4164 tmp = RREG32_SOC15(GC, 0, regCP_HQD_PQ_CONTROL); 4165 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, QUEUE_SIZE, 4166 (order_base_2(prop->queue_size / 4) - 1)); 4167 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, RPTR_BLOCK_SIZE, 4168 (order_base_2(AMDGPU_GPU_PAGE_SIZE / 4) - 1)); 4169 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, UNORD_DISPATCH, 1); 4170 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, TUNNEL_DISPATCH, 4171 prop->allow_tunneling); 4172 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, PRIV_STATE, 1); 4173 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_CONTROL, KMD_QUEUE, 1); 4174 mqd->cp_hqd_pq_control = tmp; 4175 4176 /* set the wb address whether it's enabled or not */ 4177 wb_gpu_addr = prop->rptr_gpu_addr; 4178 mqd->cp_hqd_pq_rptr_report_addr_lo = wb_gpu_addr & 0xfffffffc; 4179 mqd->cp_hqd_pq_rptr_report_addr_hi = 4180 upper_32_bits(wb_gpu_addr) & 0xffff; 4181 4182 /* only used if CP_PQ_WPTR_POLL_CNTL.CP_PQ_WPTR_POLL_CNTL__EN_MASK=1 */ 4183 wb_gpu_addr = prop->wptr_gpu_addr; 4184 mqd->cp_hqd_pq_wptr_poll_addr_lo = wb_gpu_addr & 0xfffffffc; 4185 mqd->cp_hqd_pq_wptr_poll_addr_hi = upper_32_bits(wb_gpu_addr) & 0xffff; 4186 4187 tmp = 0; 4188 /* enable the doorbell if requested */ 4189 if (prop->use_doorbell) { 4190 tmp = RREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL); 4191 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 4192 DOORBELL_OFFSET, prop->doorbell_index); 4193 4194 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 4195 DOORBELL_EN, 1); 4196 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 4197 DOORBELL_SOURCE, 0); 4198 tmp = REG_SET_FIELD(tmp, CP_HQD_PQ_DOORBELL_CONTROL, 4199 DOORBELL_HIT, 0); 4200 } 4201 4202 mqd->cp_hqd_pq_doorbell_control = tmp; 4203 4204 /* reset read and write pointers, similar to CP_RB0_WPTR/_RPTR */ 4205 mqd->cp_hqd_pq_rptr = RREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR); 4206 4207 /* set the vmid for the queue */ 4208 mqd->cp_hqd_vmid = 0; 4209 4210 tmp = RREG32_SOC15(GC, 0, regCP_HQD_PERSISTENT_STATE); 4211 tmp = REG_SET_FIELD(tmp, CP_HQD_PERSISTENT_STATE, PRELOAD_SIZE, 0x55); 4212 mqd->cp_hqd_persistent_state = tmp; 4213 4214 /* set MIN_IB_AVAIL_SIZE */ 4215 tmp = RREG32_SOC15(GC, 0, regCP_HQD_IB_CONTROL); 4216 tmp = REG_SET_FIELD(tmp, CP_HQD_IB_CONTROL, MIN_IB_AVAIL_SIZE, 3); 4217 mqd->cp_hqd_ib_control = tmp; 4218 4219 /* set static priority for a compute queue/ring */ 4220 mqd->cp_hqd_pipe_priority = prop->hqd_pipe_priority; 4221 mqd->cp_hqd_queue_priority = prop->hqd_queue_priority; 4222 4223 mqd->cp_hqd_active = prop->hqd_active; 4224 4225 return 0; 4226 } 4227 4228 static int gfx_v11_0_kiq_init_register(struct amdgpu_ring *ring) 4229 { 4230 struct amdgpu_device *adev = ring->adev; 4231 struct v11_compute_mqd *mqd = ring->mqd_ptr; 4232 int j; 4233 4234 /* inactivate the queue */ 4235 if (amdgpu_sriov_vf(adev)) 4236 WREG32_SOC15(GC, 0, regCP_HQD_ACTIVE, 0); 4237 4238 /* disable wptr polling */ 4239 WREG32_FIELD15_PREREG(GC, 0, CP_PQ_WPTR_POLL_CNTL, EN, 0); 4240 4241 /* write the EOP addr */ 4242 WREG32_SOC15(GC, 0, regCP_HQD_EOP_BASE_ADDR, 4243 mqd->cp_hqd_eop_base_addr_lo); 4244 WREG32_SOC15(GC, 0, regCP_HQD_EOP_BASE_ADDR_HI, 4245 mqd->cp_hqd_eop_base_addr_hi); 4246 4247 /* set the EOP size, register value is 2^(EOP_SIZE+1) dwords */ 4248 WREG32_SOC15(GC, 0, regCP_HQD_EOP_CONTROL, 4249 mqd->cp_hqd_eop_control); 4250 4251 /* enable doorbell? */ 4252 WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, 4253 mqd->cp_hqd_pq_doorbell_control); 4254 4255 /* disable the queue if it's active */ 4256 if (RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1) { 4257 WREG32_SOC15(GC, 0, regCP_HQD_DEQUEUE_REQUEST, 1); 4258 for (j = 0; j < adev->usec_timeout; j++) { 4259 if (!(RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1)) 4260 break; 4261 udelay(1); 4262 } 4263 WREG32_SOC15(GC, 0, regCP_HQD_DEQUEUE_REQUEST, 4264 mqd->cp_hqd_dequeue_request); 4265 WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR, 4266 mqd->cp_hqd_pq_rptr); 4267 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_LO, 4268 mqd->cp_hqd_pq_wptr_lo); 4269 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_HI, 4270 mqd->cp_hqd_pq_wptr_hi); 4271 } 4272 4273 /* set the pointer to the MQD */ 4274 WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR, 4275 mqd->cp_mqd_base_addr_lo); 4276 WREG32_SOC15(GC, 0, regCP_MQD_BASE_ADDR_HI, 4277 mqd->cp_mqd_base_addr_hi); 4278 4279 /* set MQD vmid to 0 */ 4280 WREG32_SOC15(GC, 0, regCP_MQD_CONTROL, 4281 mqd->cp_mqd_control); 4282 4283 /* set the pointer to the HQD, this is similar CP_RB0_BASE/_HI */ 4284 WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE, 4285 mqd->cp_hqd_pq_base_lo); 4286 WREG32_SOC15(GC, 0, regCP_HQD_PQ_BASE_HI, 4287 mqd->cp_hqd_pq_base_hi); 4288 4289 /* set up the HQD, this is similar to CP_RB0_CNTL */ 4290 WREG32_SOC15(GC, 0, regCP_HQD_PQ_CONTROL, 4291 mqd->cp_hqd_pq_control); 4292 4293 /* set the wb address whether it's enabled or not */ 4294 WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR, 4295 mqd->cp_hqd_pq_rptr_report_addr_lo); 4296 WREG32_SOC15(GC, 0, regCP_HQD_PQ_RPTR_REPORT_ADDR_HI, 4297 mqd->cp_hqd_pq_rptr_report_addr_hi); 4298 4299 /* only used if CP_PQ_WPTR_POLL_CNTL.CP_PQ_WPTR_POLL_CNTL__EN_MASK=1 */ 4300 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR, 4301 mqd->cp_hqd_pq_wptr_poll_addr_lo); 4302 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_POLL_ADDR_HI, 4303 mqd->cp_hqd_pq_wptr_poll_addr_hi); 4304 4305 /* enable the doorbell if requested */ 4306 if (ring->use_doorbell) { 4307 WREG32_SOC15(GC, 0, regCP_MEC_DOORBELL_RANGE_LOWER, 4308 (adev->doorbell_index.kiq * 2) << 2); 4309 WREG32_SOC15(GC, 0, regCP_MEC_DOORBELL_RANGE_UPPER, 4310 (adev->doorbell_index.userqueue_end * 2) << 2); 4311 } 4312 4313 WREG32_SOC15(GC, 0, regCP_HQD_PQ_DOORBELL_CONTROL, 4314 mqd->cp_hqd_pq_doorbell_control); 4315 4316 /* reset read and write pointers, similar to CP_RB0_WPTR/_RPTR */ 4317 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_LO, 4318 mqd->cp_hqd_pq_wptr_lo); 4319 WREG32_SOC15(GC, 0, regCP_HQD_PQ_WPTR_HI, 4320 mqd->cp_hqd_pq_wptr_hi); 4321 4322 /* set the vmid for the queue */ 4323 WREG32_SOC15(GC, 0, regCP_HQD_VMID, mqd->cp_hqd_vmid); 4324 4325 WREG32_SOC15(GC, 0, regCP_HQD_PERSISTENT_STATE, 4326 mqd->cp_hqd_persistent_state); 4327 4328 /* activate the queue */ 4329 WREG32_SOC15(GC, 0, regCP_HQD_ACTIVE, 4330 mqd->cp_hqd_active); 4331 4332 if (ring->use_doorbell) 4333 WREG32_FIELD15_PREREG(GC, 0, CP_PQ_STATUS, DOORBELL_ENABLE, 1); 4334 4335 return 0; 4336 } 4337 4338 static int gfx_v11_0_kiq_init_queue(struct amdgpu_ring *ring) 4339 { 4340 struct amdgpu_device *adev = ring->adev; 4341 struct v11_compute_mqd *mqd = ring->mqd_ptr; 4342 4343 gfx_v11_0_kiq_setting(ring); 4344 4345 if (amdgpu_in_reset(adev)) { /* for GPU_RESET case */ 4346 /* reset MQD to a clean status */ 4347 if (adev->gfx.kiq[0].mqd_backup) 4348 memcpy_toio(mqd, adev->gfx.kiq[0].mqd_backup, sizeof(*mqd)); 4349 4350 /* reset ring buffer */ 4351 ring->wptr = 0; 4352 amdgpu_ring_clear_ring(ring); 4353 4354 mutex_lock(&adev->srbm_mutex); 4355 soc21_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0); 4356 gfx_v11_0_kiq_init_register(ring); 4357 soc21_grbm_select(adev, 0, 0, 0, 0); 4358 mutex_unlock(&adev->srbm_mutex); 4359 } else { 4360 memset((void *)mqd, 0, sizeof(*mqd)); 4361 if (amdgpu_sriov_vf(adev) && adev->in_suspend) 4362 amdgpu_ring_clear_ring(ring); 4363 mutex_lock(&adev->srbm_mutex); 4364 soc21_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0); 4365 amdgpu_ring_init_mqd(ring); 4366 gfx_v11_0_kiq_init_register(ring); 4367 soc21_grbm_select(adev, 0, 0, 0, 0); 4368 mutex_unlock(&adev->srbm_mutex); 4369 4370 if (adev->gfx.kiq[0].mqd_backup) 4371 memcpy_fromio(adev->gfx.kiq[0].mqd_backup, mqd, sizeof(*mqd)); 4372 } 4373 4374 return 0; 4375 } 4376 4377 static int gfx_v11_0_kcq_init_queue(struct amdgpu_ring *ring, bool reset) 4378 { 4379 struct amdgpu_device *adev = ring->adev; 4380 struct v11_compute_mqd *mqd = ring->mqd_ptr; 4381 int mqd_idx = ring - &adev->gfx.compute_ring[0]; 4382 4383 if (!reset && !amdgpu_in_reset(adev) && !adev->in_suspend) { 4384 memset((void *)mqd, 0, sizeof(*mqd)); 4385 mutex_lock(&adev->srbm_mutex); 4386 soc21_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0); 4387 amdgpu_ring_init_mqd(ring); 4388 soc21_grbm_select(adev, 0, 0, 0, 0); 4389 mutex_unlock(&adev->srbm_mutex); 4390 4391 if (adev->gfx.mec.mqd_backup[mqd_idx]) 4392 memcpy_fromio(adev->gfx.mec.mqd_backup[mqd_idx], mqd, sizeof(*mqd)); 4393 } else { 4394 /* restore MQD to a clean status */ 4395 if (adev->gfx.mec.mqd_backup[mqd_idx]) 4396 memcpy_toio(mqd, adev->gfx.mec.mqd_backup[mqd_idx], sizeof(*mqd)); 4397 /* reset ring buffer */ 4398 ring->wptr = 0; 4399 atomic64_set((atomic64_t *)ring->wptr_cpu_addr, 0); 4400 amdgpu_ring_clear_ring(ring); 4401 } 4402 4403 return 0; 4404 } 4405 4406 static int gfx_v11_0_kiq_resume(struct amdgpu_device *adev) 4407 { 4408 struct amdgpu_ring *ring; 4409 int r; 4410 4411 ring = &adev->gfx.kiq[0].ring; 4412 4413 r = amdgpu_bo_reserve(ring->mqd_obj, false); 4414 if (unlikely(r != 0)) 4415 return r; 4416 4417 r = amdgpu_bo_kmap(ring->mqd_obj, (void **)&ring->mqd_ptr); 4418 if (unlikely(r != 0)) { 4419 amdgpu_bo_unreserve(ring->mqd_obj); 4420 return r; 4421 } 4422 4423 gfx_v11_0_kiq_init_queue(ring); 4424 amdgpu_bo_kunmap(ring->mqd_obj); 4425 ring->mqd_ptr = NULL; 4426 amdgpu_bo_unreserve(ring->mqd_obj); 4427 ring->sched.ready = true; 4428 return 0; 4429 } 4430 4431 static int gfx_v11_0_kcq_resume(struct amdgpu_device *adev) 4432 { 4433 struct amdgpu_ring *ring = NULL; 4434 int r = 0, i; 4435 4436 if (!amdgpu_async_gfx_ring) 4437 gfx_v11_0_cp_compute_enable(adev, true); 4438 4439 for (i = 0; i < adev->gfx.num_compute_rings; i++) { 4440 ring = &adev->gfx.compute_ring[i]; 4441 4442 r = amdgpu_bo_reserve(ring->mqd_obj, false); 4443 if (unlikely(r != 0)) 4444 goto done; 4445 r = amdgpu_bo_kmap(ring->mqd_obj, (void **)&ring->mqd_ptr); 4446 if (!r) { 4447 r = gfx_v11_0_kcq_init_queue(ring, false); 4448 amdgpu_bo_kunmap(ring->mqd_obj); 4449 ring->mqd_ptr = NULL; 4450 } 4451 amdgpu_bo_unreserve(ring->mqd_obj); 4452 if (r) 4453 goto done; 4454 } 4455 4456 r = amdgpu_gfx_enable_kcq(adev, 0); 4457 done: 4458 return r; 4459 } 4460 4461 static int gfx_v11_0_cp_resume(struct amdgpu_device *adev) 4462 { 4463 int r, i; 4464 struct amdgpu_ring *ring; 4465 4466 if (!(adev->flags & AMD_IS_APU)) 4467 gfx_v11_0_enable_gui_idle_interrupt(adev, false); 4468 4469 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) { 4470 /* legacy firmware loading */ 4471 r = gfx_v11_0_cp_gfx_load_microcode(adev); 4472 if (r) 4473 return r; 4474 4475 if (adev->gfx.rs64_enable) 4476 r = gfx_v11_0_cp_compute_load_microcode_rs64(adev); 4477 else 4478 r = gfx_v11_0_cp_compute_load_microcode(adev); 4479 if (r) 4480 return r; 4481 } 4482 4483 gfx_v11_0_cp_set_doorbell_range(adev); 4484 4485 if (amdgpu_async_gfx_ring) { 4486 gfx_v11_0_cp_compute_enable(adev, true); 4487 gfx_v11_0_cp_gfx_enable(adev, true); 4488 } 4489 4490 if (adev->enable_mes_kiq && adev->mes.kiq_hw_init) 4491 r = amdgpu_mes_kiq_hw_init(adev); 4492 else 4493 r = gfx_v11_0_kiq_resume(adev); 4494 if (r) 4495 return r; 4496 4497 r = gfx_v11_0_kcq_resume(adev); 4498 if (r) 4499 return r; 4500 4501 if (!amdgpu_async_gfx_ring) { 4502 r = gfx_v11_0_cp_gfx_resume(adev); 4503 if (r) 4504 return r; 4505 } else { 4506 r = gfx_v11_0_cp_async_gfx_ring_resume(adev); 4507 if (r) 4508 return r; 4509 } 4510 4511 for (i = 0; i < adev->gfx.num_gfx_rings; i++) { 4512 ring = &adev->gfx.gfx_ring[i]; 4513 r = amdgpu_ring_test_helper(ring); 4514 if (r) 4515 return r; 4516 } 4517 4518 for (i = 0; i < adev->gfx.num_compute_rings; i++) { 4519 ring = &adev->gfx.compute_ring[i]; 4520 r = amdgpu_ring_test_helper(ring); 4521 if (r) 4522 return r; 4523 } 4524 4525 return 0; 4526 } 4527 4528 static void gfx_v11_0_cp_enable(struct amdgpu_device *adev, bool enable) 4529 { 4530 gfx_v11_0_cp_gfx_enable(adev, enable); 4531 gfx_v11_0_cp_compute_enable(adev, enable); 4532 } 4533 4534 static int gfx_v11_0_gfxhub_enable(struct amdgpu_device *adev) 4535 { 4536 int r; 4537 bool value; 4538 4539 r = adev->gfxhub.funcs->gart_enable(adev); 4540 if (r) 4541 return r; 4542 4543 adev->hdp.funcs->flush_hdp(adev, NULL); 4544 4545 value = (amdgpu_vm_fault_stop == AMDGPU_VM_FAULT_STOP_ALWAYS) ? 4546 false : true; 4547 4548 adev->gfxhub.funcs->set_fault_enable_default(adev, value); 4549 /* TODO investigate why this and the hdp flush above is needed, 4550 * are we missing a flush somewhere else? */ 4551 adev->gmc.gmc_funcs->flush_gpu_tlb(adev, 0, AMDGPU_GFXHUB(0), 0); 4552 4553 return 0; 4554 } 4555 4556 static void gfx_v11_0_select_cp_fw_arch(struct amdgpu_device *adev) 4557 { 4558 u32 tmp; 4559 4560 /* select RS64 */ 4561 if (adev->gfx.rs64_enable) { 4562 tmp = RREG32_SOC15(GC, 0, regCP_GFX_CNTL); 4563 tmp = REG_SET_FIELD(tmp, CP_GFX_CNTL, ENGINE_SEL, 1); 4564 WREG32_SOC15(GC, 0, regCP_GFX_CNTL, tmp); 4565 4566 tmp = RREG32_SOC15(GC, 0, regCP_MEC_ISA_CNTL); 4567 tmp = REG_SET_FIELD(tmp, CP_MEC_ISA_CNTL, ISA_MODE, 1); 4568 WREG32_SOC15(GC, 0, regCP_MEC_ISA_CNTL, tmp); 4569 } 4570 4571 if (amdgpu_emu_mode == 1) 4572 msleep(100); 4573 } 4574 4575 static int get_gb_addr_config(struct amdgpu_device * adev) 4576 { 4577 u32 gb_addr_config; 4578 4579 gb_addr_config = RREG32_SOC15(GC, 0, regGB_ADDR_CONFIG); 4580 if (gb_addr_config == 0) 4581 return -EINVAL; 4582 4583 adev->gfx.config.gb_addr_config_fields.num_pkrs = 4584 1 << REG_GET_FIELD(gb_addr_config, GB_ADDR_CONFIG, NUM_PKRS); 4585 4586 adev->gfx.config.gb_addr_config = gb_addr_config; 4587 4588 adev->gfx.config.gb_addr_config_fields.num_pipes = 1 << 4589 REG_GET_FIELD(adev->gfx.config.gb_addr_config, 4590 GB_ADDR_CONFIG, NUM_PIPES); 4591 4592 adev->gfx.config.max_tile_pipes = 4593 adev->gfx.config.gb_addr_config_fields.num_pipes; 4594 4595 adev->gfx.config.gb_addr_config_fields.max_compress_frags = 1 << 4596 REG_GET_FIELD(adev->gfx.config.gb_addr_config, 4597 GB_ADDR_CONFIG, MAX_COMPRESSED_FRAGS); 4598 adev->gfx.config.gb_addr_config_fields.num_rb_per_se = 1 << 4599 REG_GET_FIELD(adev->gfx.config.gb_addr_config, 4600 GB_ADDR_CONFIG, NUM_RB_PER_SE); 4601 adev->gfx.config.gb_addr_config_fields.num_se = 1 << 4602 REG_GET_FIELD(adev->gfx.config.gb_addr_config, 4603 GB_ADDR_CONFIG, NUM_SHADER_ENGINES); 4604 adev->gfx.config.gb_addr_config_fields.pipe_interleave_size = 1 << (8 + 4605 REG_GET_FIELD(adev->gfx.config.gb_addr_config, 4606 GB_ADDR_CONFIG, PIPE_INTERLEAVE_SIZE)); 4607 4608 return 0; 4609 } 4610 4611 static void gfx_v11_0_disable_gpa_mode(struct amdgpu_device *adev) 4612 { 4613 uint32_t data; 4614 4615 data = RREG32_SOC15(GC, 0, regCPC_PSP_DEBUG); 4616 data |= CPC_PSP_DEBUG__GPA_OVERRIDE_MASK; 4617 WREG32_SOC15(GC, 0, regCPC_PSP_DEBUG, data); 4618 4619 data = RREG32_SOC15(GC, 0, regCPG_PSP_DEBUG); 4620 data |= CPG_PSP_DEBUG__GPA_OVERRIDE_MASK; 4621 WREG32_SOC15(GC, 0, regCPG_PSP_DEBUG, data); 4622 } 4623 4624 static int gfx_v11_0_hw_init(struct amdgpu_ip_block *ip_block) 4625 { 4626 int r; 4627 struct amdgpu_device *adev = ip_block->adev; 4628 4629 amdgpu_gfx_cleaner_shader_init(adev, adev->gfx.cleaner_shader_size, 4630 adev->gfx.cleaner_shader_ptr); 4631 4632 if (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) { 4633 if (adev->gfx.imu.funcs) { 4634 /* RLC autoload sequence 1: Program rlc ram */ 4635 if (adev->gfx.imu.funcs->program_rlc_ram) 4636 adev->gfx.imu.funcs->program_rlc_ram(adev); 4637 /* rlc autoload firmware */ 4638 r = gfx_v11_0_rlc_backdoor_autoload_enable(adev); 4639 if (r) 4640 return r; 4641 } 4642 } else { 4643 if (adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) { 4644 if (adev->gfx.imu.funcs && (amdgpu_dpm > 0)) { 4645 if (adev->gfx.imu.funcs->load_microcode) 4646 adev->gfx.imu.funcs->load_microcode(adev); 4647 if (adev->gfx.imu.funcs->setup_imu) 4648 adev->gfx.imu.funcs->setup_imu(adev); 4649 if (adev->gfx.imu.funcs->start_imu) 4650 adev->gfx.imu.funcs->start_imu(adev); 4651 } 4652 4653 /* disable gpa mode in backdoor loading */ 4654 gfx_v11_0_disable_gpa_mode(adev); 4655 } 4656 } 4657 4658 if ((adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO) || 4659 (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP)) { 4660 r = gfx_v11_0_wait_for_rlc_autoload_complete(adev); 4661 if (r) { 4662 dev_err(adev->dev, "(%d) failed to wait rlc autoload complete\n", r); 4663 return r; 4664 } 4665 } 4666 4667 adev->gfx.is_poweron = true; 4668 4669 if(get_gb_addr_config(adev)) 4670 DRM_WARN("Invalid gb_addr_config !\n"); 4671 4672 if (adev->firmware.load_type == AMDGPU_FW_LOAD_PSP && 4673 adev->gfx.rs64_enable) 4674 gfx_v11_0_config_gfx_rs64(adev); 4675 4676 r = gfx_v11_0_gfxhub_enable(adev); 4677 if (r) 4678 return r; 4679 4680 if (!amdgpu_emu_mode) 4681 gfx_v11_0_init_golden_registers(adev); 4682 4683 if ((adev->firmware.load_type == AMDGPU_FW_LOAD_DIRECT) || 4684 (adev->firmware.load_type == AMDGPU_FW_LOAD_RLC_BACKDOOR_AUTO && amdgpu_dpm == 1)) { 4685 /** 4686 * For gfx 11, rlc firmware loading relies on smu firmware is 4687 * loaded firstly, so in direct type, it has to load smc ucode 4688 * here before rlc. 4689 */ 4690 r = amdgpu_pm_load_smu_firmware(adev, NULL); 4691 if (r) 4692 return r; 4693 } 4694 4695 gfx_v11_0_constants_init(adev); 4696 4697 if (adev->firmware.load_type != AMDGPU_FW_LOAD_PSP) 4698 gfx_v11_0_select_cp_fw_arch(adev); 4699 4700 if (adev->nbio.funcs->gc_doorbell_init) 4701 adev->nbio.funcs->gc_doorbell_init(adev); 4702 4703 r = gfx_v11_0_rlc_resume(adev); 4704 if (r) 4705 return r; 4706 4707 /* 4708 * init golden registers and rlc resume may override some registers, 4709 * reconfig them here 4710 */ 4711 gfx_v11_0_tcp_harvest(adev); 4712 4713 r = gfx_v11_0_cp_resume(adev); 4714 if (r) 4715 return r; 4716 4717 /* get IMU version from HW if it's not set */ 4718 if (!adev->gfx.imu_fw_version) 4719 adev->gfx.imu_fw_version = RREG32_SOC15(GC, 0, regGFX_IMU_SCRATCH_0); 4720 4721 return r; 4722 } 4723 4724 static int gfx_v11_0_hw_fini(struct amdgpu_ip_block *ip_block) 4725 { 4726 struct amdgpu_device *adev = ip_block->adev; 4727 4728 amdgpu_irq_put(adev, &adev->gfx.priv_reg_irq, 0); 4729 amdgpu_irq_put(adev, &adev->gfx.priv_inst_irq, 0); 4730 amdgpu_irq_put(adev, &adev->gfx.bad_op_irq, 0); 4731 4732 if (!adev->no_hw_access) { 4733 if (amdgpu_async_gfx_ring) { 4734 if (amdgpu_gfx_disable_kgq(adev, 0)) 4735 DRM_ERROR("KGQ disable failed\n"); 4736 } 4737 4738 if (amdgpu_gfx_disable_kcq(adev, 0)) 4739 DRM_ERROR("KCQ disable failed\n"); 4740 4741 amdgpu_mes_kiq_hw_fini(adev); 4742 } 4743 4744 if (amdgpu_sriov_vf(adev)) 4745 /* Remove the steps disabling CPG and clearing KIQ position, 4746 * so that CP could perform IDLE-SAVE during switch. Those 4747 * steps are necessary to avoid a DMAR error in gfx9 but it is 4748 * not reproduced on gfx11. 4749 */ 4750 return 0; 4751 4752 gfx_v11_0_cp_enable(adev, false); 4753 gfx_v11_0_enable_gui_idle_interrupt(adev, false); 4754 4755 adev->gfxhub.funcs->gart_disable(adev); 4756 4757 adev->gfx.is_poweron = false; 4758 4759 return 0; 4760 } 4761 4762 static int gfx_v11_0_suspend(struct amdgpu_ip_block *ip_block) 4763 { 4764 return gfx_v11_0_hw_fini(ip_block); 4765 } 4766 4767 static int gfx_v11_0_resume(struct amdgpu_ip_block *ip_block) 4768 { 4769 return gfx_v11_0_hw_init(ip_block); 4770 } 4771 4772 static bool gfx_v11_0_is_idle(void *handle) 4773 { 4774 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 4775 4776 if (REG_GET_FIELD(RREG32_SOC15(GC, 0, regGRBM_STATUS), 4777 GRBM_STATUS, GUI_ACTIVE)) 4778 return false; 4779 else 4780 return true; 4781 } 4782 4783 static int gfx_v11_0_wait_for_idle(struct amdgpu_ip_block *ip_block) 4784 { 4785 unsigned i; 4786 u32 tmp; 4787 struct amdgpu_device *adev = ip_block->adev; 4788 4789 for (i = 0; i < adev->usec_timeout; i++) { 4790 /* read MC_STATUS */ 4791 tmp = RREG32_SOC15(GC, 0, regGRBM_STATUS) & 4792 GRBM_STATUS__GUI_ACTIVE_MASK; 4793 4794 if (!REG_GET_FIELD(tmp, GRBM_STATUS, GUI_ACTIVE)) 4795 return 0; 4796 udelay(1); 4797 } 4798 return -ETIMEDOUT; 4799 } 4800 4801 int gfx_v11_0_request_gfx_index_mutex(struct amdgpu_device *adev, 4802 bool req) 4803 { 4804 u32 i, tmp, val; 4805 4806 for (i = 0; i < adev->usec_timeout; i++) { 4807 /* Request with MeId=2, PipeId=0 */ 4808 tmp = REG_SET_FIELD(0, CP_GFX_INDEX_MUTEX, REQUEST, req); 4809 tmp = REG_SET_FIELD(tmp, CP_GFX_INDEX_MUTEX, CLIENTID, 4); 4810 WREG32_SOC15(GC, 0, regCP_GFX_INDEX_MUTEX, tmp); 4811 4812 val = RREG32_SOC15(GC, 0, regCP_GFX_INDEX_MUTEX); 4813 if (req) { 4814 if (val == tmp) 4815 break; 4816 } else { 4817 tmp = REG_SET_FIELD(tmp, CP_GFX_INDEX_MUTEX, 4818 REQUEST, 1); 4819 4820 /* unlocked or locked by firmware */ 4821 if (val != tmp) 4822 break; 4823 } 4824 udelay(1); 4825 } 4826 4827 if (i >= adev->usec_timeout) 4828 return -EINVAL; 4829 4830 return 0; 4831 } 4832 4833 static int gfx_v11_0_soft_reset(struct amdgpu_ip_block *ip_block) 4834 { 4835 u32 grbm_soft_reset = 0; 4836 u32 tmp; 4837 int r, i, j, k; 4838 struct amdgpu_device *adev = ip_block->adev; 4839 4840 amdgpu_gfx_rlc_enter_safe_mode(adev, 0); 4841 4842 tmp = RREG32_SOC15(GC, 0, regCP_INT_CNTL); 4843 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL, CMP_BUSY_INT_ENABLE, 0); 4844 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL, CNTX_BUSY_INT_ENABLE, 0); 4845 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL, CNTX_EMPTY_INT_ENABLE, 0); 4846 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL, GFX_IDLE_INT_ENABLE, 0); 4847 WREG32_SOC15(GC, 0, regCP_INT_CNTL, tmp); 4848 4849 mutex_lock(&adev->srbm_mutex); 4850 for (i = 0; i < adev->gfx.mec.num_mec; ++i) { 4851 for (j = 0; j < adev->gfx.mec.num_queue_per_pipe; j++) { 4852 for (k = 0; k < adev->gfx.mec.num_pipe_per_mec; k++) { 4853 soc21_grbm_select(adev, i, k, j, 0); 4854 4855 WREG32_SOC15(GC, 0, regCP_HQD_DEQUEUE_REQUEST, 0x2); 4856 WREG32_SOC15(GC, 0, regSPI_COMPUTE_QUEUE_RESET, 0x1); 4857 } 4858 } 4859 } 4860 for (i = 0; i < adev->gfx.me.num_me; ++i) { 4861 for (j = 0; j < adev->gfx.me.num_queue_per_pipe; j++) { 4862 for (k = 0; k < adev->gfx.me.num_pipe_per_me; k++) { 4863 soc21_grbm_select(adev, i, k, j, 0); 4864 4865 WREG32_SOC15(GC, 0, regCP_GFX_HQD_DEQUEUE_REQUEST, 0x1); 4866 } 4867 } 4868 } 4869 soc21_grbm_select(adev, 0, 0, 0, 0); 4870 mutex_unlock(&adev->srbm_mutex); 4871 4872 /* Try to acquire the gfx mutex before access to CP_VMID_RESET */ 4873 mutex_lock(&adev->gfx.reset_sem_mutex); 4874 r = gfx_v11_0_request_gfx_index_mutex(adev, true); 4875 if (r) { 4876 mutex_unlock(&adev->gfx.reset_sem_mutex); 4877 DRM_ERROR("Failed to acquire the gfx mutex during soft reset\n"); 4878 return r; 4879 } 4880 4881 WREG32_SOC15(GC, 0, regCP_VMID_RESET, 0xfffffffe); 4882 4883 // Read CP_VMID_RESET register three times. 4884 // to get sufficient time for GFX_HQD_ACTIVE reach 0 4885 RREG32_SOC15(GC, 0, regCP_VMID_RESET); 4886 RREG32_SOC15(GC, 0, regCP_VMID_RESET); 4887 RREG32_SOC15(GC, 0, regCP_VMID_RESET); 4888 4889 /* release the gfx mutex */ 4890 r = gfx_v11_0_request_gfx_index_mutex(adev, false); 4891 mutex_unlock(&adev->gfx.reset_sem_mutex); 4892 if (r) { 4893 DRM_ERROR("Failed to release the gfx mutex during soft reset\n"); 4894 return r; 4895 } 4896 4897 for (i = 0; i < adev->usec_timeout; i++) { 4898 if (!RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) && 4899 !RREG32_SOC15(GC, 0, regCP_GFX_HQD_ACTIVE)) 4900 break; 4901 udelay(1); 4902 } 4903 if (i >= adev->usec_timeout) { 4904 printk("Failed to wait all pipes clean\n"); 4905 return -EINVAL; 4906 } 4907 4908 /********** trigger soft reset ***********/ 4909 grbm_soft_reset = RREG32_SOC15(GC, 0, regGRBM_SOFT_RESET); 4910 grbm_soft_reset = REG_SET_FIELD(grbm_soft_reset, GRBM_SOFT_RESET, 4911 SOFT_RESET_CP, 1); 4912 grbm_soft_reset = REG_SET_FIELD(grbm_soft_reset, GRBM_SOFT_RESET, 4913 SOFT_RESET_GFX, 1); 4914 grbm_soft_reset = REG_SET_FIELD(grbm_soft_reset, GRBM_SOFT_RESET, 4915 SOFT_RESET_CPF, 1); 4916 grbm_soft_reset = REG_SET_FIELD(grbm_soft_reset, GRBM_SOFT_RESET, 4917 SOFT_RESET_CPC, 1); 4918 grbm_soft_reset = REG_SET_FIELD(grbm_soft_reset, GRBM_SOFT_RESET, 4919 SOFT_RESET_CPG, 1); 4920 WREG32_SOC15(GC, 0, regGRBM_SOFT_RESET, grbm_soft_reset); 4921 /********** exit soft reset ***********/ 4922 grbm_soft_reset = RREG32_SOC15(GC, 0, regGRBM_SOFT_RESET); 4923 grbm_soft_reset = REG_SET_FIELD(grbm_soft_reset, GRBM_SOFT_RESET, 4924 SOFT_RESET_CP, 0); 4925 grbm_soft_reset = REG_SET_FIELD(grbm_soft_reset, GRBM_SOFT_RESET, 4926 SOFT_RESET_GFX, 0); 4927 grbm_soft_reset = REG_SET_FIELD(grbm_soft_reset, GRBM_SOFT_RESET, 4928 SOFT_RESET_CPF, 0); 4929 grbm_soft_reset = REG_SET_FIELD(grbm_soft_reset, GRBM_SOFT_RESET, 4930 SOFT_RESET_CPC, 0); 4931 grbm_soft_reset = REG_SET_FIELD(grbm_soft_reset, GRBM_SOFT_RESET, 4932 SOFT_RESET_CPG, 0); 4933 WREG32_SOC15(GC, 0, regGRBM_SOFT_RESET, grbm_soft_reset); 4934 4935 tmp = RREG32_SOC15(GC, 0, regCP_SOFT_RESET_CNTL); 4936 tmp = REG_SET_FIELD(tmp, CP_SOFT_RESET_CNTL, CMP_HQD_REG_RESET, 0x1); 4937 WREG32_SOC15(GC, 0, regCP_SOFT_RESET_CNTL, tmp); 4938 4939 WREG32_SOC15(GC, 0, regCP_ME_CNTL, 0x0); 4940 WREG32_SOC15(GC, 0, regCP_MEC_RS64_CNTL, 0x0); 4941 4942 for (i = 0; i < adev->usec_timeout; i++) { 4943 if (!RREG32_SOC15(GC, 0, regCP_VMID_RESET)) 4944 break; 4945 udelay(1); 4946 } 4947 if (i >= adev->usec_timeout) { 4948 printk("Failed to wait CP_VMID_RESET to 0\n"); 4949 return -EINVAL; 4950 } 4951 4952 tmp = RREG32_SOC15(GC, 0, regCP_INT_CNTL); 4953 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL, CMP_BUSY_INT_ENABLE, 1); 4954 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL, CNTX_BUSY_INT_ENABLE, 1); 4955 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL, CNTX_EMPTY_INT_ENABLE, 1); 4956 tmp = REG_SET_FIELD(tmp, CP_INT_CNTL, GFX_IDLE_INT_ENABLE, 1); 4957 WREG32_SOC15(GC, 0, regCP_INT_CNTL, tmp); 4958 4959 amdgpu_gfx_rlc_exit_safe_mode(adev, 0); 4960 4961 return gfx_v11_0_cp_resume(adev); 4962 } 4963 4964 static bool gfx_v11_0_check_soft_reset(struct amdgpu_ip_block *ip_block) 4965 { 4966 int i, r; 4967 struct amdgpu_device *adev = ip_block->adev; 4968 struct amdgpu_ring *ring; 4969 long tmo = msecs_to_jiffies(1000); 4970 4971 for (i = 0; i < adev->gfx.num_gfx_rings; i++) { 4972 ring = &adev->gfx.gfx_ring[i]; 4973 r = amdgpu_ring_test_ib(ring, tmo); 4974 if (r) 4975 return true; 4976 } 4977 4978 for (i = 0; i < adev->gfx.num_compute_rings; i++) { 4979 ring = &adev->gfx.compute_ring[i]; 4980 r = amdgpu_ring_test_ib(ring, tmo); 4981 if (r) 4982 return true; 4983 } 4984 4985 return false; 4986 } 4987 4988 static int gfx_v11_0_post_soft_reset(struct amdgpu_ip_block *ip_block) 4989 { 4990 struct amdgpu_device *adev = ip_block->adev; 4991 /** 4992 * GFX soft reset will impact MES, need resume MES when do GFX soft reset 4993 */ 4994 return amdgpu_mes_resume(adev); 4995 } 4996 4997 static uint64_t gfx_v11_0_get_gpu_clock_counter(struct amdgpu_device *adev) 4998 { 4999 uint64_t clock; 5000 uint64_t clock_counter_lo, clock_counter_hi_pre, clock_counter_hi_after; 5001 5002 if (amdgpu_sriov_vf(adev)) { 5003 amdgpu_gfx_off_ctrl(adev, false); 5004 mutex_lock(&adev->gfx.gpu_clock_mutex); 5005 clock_counter_hi_pre = (uint64_t)RREG32_SOC15(GC, 0, regCP_MES_MTIME_HI); 5006 clock_counter_lo = (uint64_t)RREG32_SOC15(GC, 0, regCP_MES_MTIME_LO); 5007 clock_counter_hi_after = (uint64_t)RREG32_SOC15(GC, 0, regCP_MES_MTIME_HI); 5008 if (clock_counter_hi_pre != clock_counter_hi_after) 5009 clock_counter_lo = (uint64_t)RREG32_SOC15(GC, 0, regCP_MES_MTIME_LO); 5010 mutex_unlock(&adev->gfx.gpu_clock_mutex); 5011 amdgpu_gfx_off_ctrl(adev, true); 5012 } else { 5013 preempt_disable(); 5014 clock_counter_hi_pre = (uint64_t)RREG32_SOC15(SMUIO, 0, regGOLDEN_TSC_COUNT_UPPER); 5015 clock_counter_lo = (uint64_t)RREG32_SOC15(SMUIO, 0, regGOLDEN_TSC_COUNT_LOWER); 5016 clock_counter_hi_after = (uint64_t)RREG32_SOC15(SMUIO, 0, regGOLDEN_TSC_COUNT_UPPER); 5017 if (clock_counter_hi_pre != clock_counter_hi_after) 5018 clock_counter_lo = (uint64_t)RREG32_SOC15(SMUIO, 0, regGOLDEN_TSC_COUNT_LOWER); 5019 preempt_enable(); 5020 } 5021 clock = clock_counter_lo | (clock_counter_hi_after << 32ULL); 5022 5023 return clock; 5024 } 5025 5026 static void gfx_v11_0_ring_emit_gds_switch(struct amdgpu_ring *ring, 5027 uint32_t vmid, 5028 uint32_t gds_base, uint32_t gds_size, 5029 uint32_t gws_base, uint32_t gws_size, 5030 uint32_t oa_base, uint32_t oa_size) 5031 { 5032 struct amdgpu_device *adev = ring->adev; 5033 5034 /* GDS Base */ 5035 gfx_v11_0_write_data_to_reg(ring, 0, false, 5036 SOC15_REG_OFFSET(GC, 0, regGDS_VMID0_BASE) + 2 * vmid, 5037 gds_base); 5038 5039 /* GDS Size */ 5040 gfx_v11_0_write_data_to_reg(ring, 0, false, 5041 SOC15_REG_OFFSET(GC, 0, regGDS_VMID0_SIZE) + 2 * vmid, 5042 gds_size); 5043 5044 /* GWS */ 5045 gfx_v11_0_write_data_to_reg(ring, 0, false, 5046 SOC15_REG_OFFSET(GC, 0, regGDS_GWS_VMID0) + vmid, 5047 gws_size << GDS_GWS_VMID0__SIZE__SHIFT | gws_base); 5048 5049 /* OA */ 5050 gfx_v11_0_write_data_to_reg(ring, 0, false, 5051 SOC15_REG_OFFSET(GC, 0, regGDS_OA_VMID0) + vmid, 5052 (1 << (oa_size + oa_base)) - (1 << oa_base)); 5053 } 5054 5055 static int gfx_v11_0_early_init(struct amdgpu_ip_block *ip_block) 5056 { 5057 struct amdgpu_device *adev = ip_block->adev; 5058 5059 adev->gfx.funcs = &gfx_v11_0_gfx_funcs; 5060 5061 adev->gfx.num_gfx_rings = GFX11_NUM_GFX_RINGS; 5062 adev->gfx.num_compute_rings = min(amdgpu_gfx_get_num_kcq(adev), 5063 AMDGPU_MAX_COMPUTE_RINGS); 5064 5065 gfx_v11_0_set_kiq_pm4_funcs(adev); 5066 gfx_v11_0_set_ring_funcs(adev); 5067 gfx_v11_0_set_irq_funcs(adev); 5068 gfx_v11_0_set_gds_init(adev); 5069 gfx_v11_0_set_rlc_funcs(adev); 5070 gfx_v11_0_set_mqd_funcs(adev); 5071 gfx_v11_0_set_imu_funcs(adev); 5072 5073 gfx_v11_0_init_rlcg_reg_access_ctrl(adev); 5074 5075 return gfx_v11_0_init_microcode(adev); 5076 } 5077 5078 static int gfx_v11_0_late_init(struct amdgpu_ip_block *ip_block) 5079 { 5080 struct amdgpu_device *adev = ip_block->adev; 5081 int r; 5082 5083 r = amdgpu_irq_get(adev, &adev->gfx.priv_reg_irq, 0); 5084 if (r) 5085 return r; 5086 5087 r = amdgpu_irq_get(adev, &adev->gfx.priv_inst_irq, 0); 5088 if (r) 5089 return r; 5090 5091 r = amdgpu_irq_get(adev, &adev->gfx.bad_op_irq, 0); 5092 if (r) 5093 return r; 5094 return 0; 5095 } 5096 5097 static bool gfx_v11_0_is_rlc_enabled(struct amdgpu_device *adev) 5098 { 5099 uint32_t rlc_cntl; 5100 5101 /* if RLC is not enabled, do nothing */ 5102 rlc_cntl = RREG32_SOC15(GC, 0, regRLC_CNTL); 5103 return (REG_GET_FIELD(rlc_cntl, RLC_CNTL, RLC_ENABLE_F32)) ? true : false; 5104 } 5105 5106 static void gfx_v11_0_set_safe_mode(struct amdgpu_device *adev, int xcc_id) 5107 { 5108 uint32_t data; 5109 unsigned i; 5110 5111 data = RLC_SAFE_MODE__CMD_MASK; 5112 data |= (1 << RLC_SAFE_MODE__MESSAGE__SHIFT); 5113 5114 WREG32_SOC15(GC, 0, regRLC_SAFE_MODE, data); 5115 5116 /* wait for RLC_SAFE_MODE */ 5117 for (i = 0; i < adev->usec_timeout; i++) { 5118 if (!REG_GET_FIELD(RREG32_SOC15(GC, 0, regRLC_SAFE_MODE), 5119 RLC_SAFE_MODE, CMD)) 5120 break; 5121 udelay(1); 5122 } 5123 } 5124 5125 static void gfx_v11_0_unset_safe_mode(struct amdgpu_device *adev, int xcc_id) 5126 { 5127 WREG32_SOC15(GC, 0, regRLC_SAFE_MODE, RLC_SAFE_MODE__CMD_MASK); 5128 } 5129 5130 static void gfx_v11_0_update_perf_clk(struct amdgpu_device *adev, 5131 bool enable) 5132 { 5133 uint32_t def, data; 5134 5135 if (!(adev->cg_flags & AMD_CG_SUPPORT_GFX_PERF_CLK)) 5136 return; 5137 5138 def = data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE); 5139 5140 if (enable) 5141 data &= ~RLC_CGTT_MGCG_OVERRIDE__PERFMON_CLOCK_STATE_MASK; 5142 else 5143 data |= RLC_CGTT_MGCG_OVERRIDE__PERFMON_CLOCK_STATE_MASK; 5144 5145 if (def != data) 5146 WREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE, data); 5147 } 5148 5149 static void gfx_v11_0_update_sram_fgcg(struct amdgpu_device *adev, 5150 bool enable) 5151 { 5152 uint32_t def, data; 5153 5154 if (!(adev->cg_flags & AMD_CG_SUPPORT_GFX_FGCG)) 5155 return; 5156 5157 def = data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE); 5158 5159 if (enable) 5160 data &= ~RLC_CGTT_MGCG_OVERRIDE__GFXIP_FGCG_OVERRIDE_MASK; 5161 else 5162 data |= RLC_CGTT_MGCG_OVERRIDE__GFXIP_FGCG_OVERRIDE_MASK; 5163 5164 if (def != data) 5165 WREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE, data); 5166 } 5167 5168 static void gfx_v11_0_update_repeater_fgcg(struct amdgpu_device *adev, 5169 bool enable) 5170 { 5171 uint32_t def, data; 5172 5173 if (!(adev->cg_flags & AMD_CG_SUPPORT_REPEATER_FGCG)) 5174 return; 5175 5176 def = data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE); 5177 5178 if (enable) 5179 data &= ~RLC_CGTT_MGCG_OVERRIDE__GFXIP_REPEATER_FGCG_OVERRIDE_MASK; 5180 else 5181 data |= RLC_CGTT_MGCG_OVERRIDE__GFXIP_REPEATER_FGCG_OVERRIDE_MASK; 5182 5183 if (def != data) 5184 WREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE, data); 5185 } 5186 5187 static void gfx_v11_0_update_medium_grain_clock_gating(struct amdgpu_device *adev, 5188 bool enable) 5189 { 5190 uint32_t data, def; 5191 5192 if (!(adev->cg_flags & (AMD_CG_SUPPORT_GFX_MGCG | AMD_CG_SUPPORT_GFX_MGLS))) 5193 return; 5194 5195 /* It is disabled by HW by default */ 5196 if (enable) { 5197 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_MGCG) { 5198 /* 1 - RLC_CGTT_MGCG_OVERRIDE */ 5199 def = data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE); 5200 5201 data &= ~(RLC_CGTT_MGCG_OVERRIDE__GRBM_CGTT_SCLK_OVERRIDE_MASK | 5202 RLC_CGTT_MGCG_OVERRIDE__RLC_CGTT_SCLK_OVERRIDE_MASK | 5203 RLC_CGTT_MGCG_OVERRIDE__GFXIP_MGCG_OVERRIDE_MASK); 5204 5205 if (def != data) 5206 WREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE, data); 5207 } 5208 } else { 5209 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_MGCG) { 5210 def = data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE); 5211 5212 data |= (RLC_CGTT_MGCG_OVERRIDE__RLC_CGTT_SCLK_OVERRIDE_MASK | 5213 RLC_CGTT_MGCG_OVERRIDE__GRBM_CGTT_SCLK_OVERRIDE_MASK | 5214 RLC_CGTT_MGCG_OVERRIDE__GFXIP_MGCG_OVERRIDE_MASK); 5215 5216 if (def != data) 5217 WREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE, data); 5218 } 5219 } 5220 } 5221 5222 static void gfx_v11_0_update_coarse_grain_clock_gating(struct amdgpu_device *adev, 5223 bool enable) 5224 { 5225 uint32_t def, data; 5226 5227 if (!(adev->cg_flags & 5228 (AMD_CG_SUPPORT_GFX_CGCG | 5229 AMD_CG_SUPPORT_GFX_CGLS | 5230 AMD_CG_SUPPORT_GFX_3D_CGCG | 5231 AMD_CG_SUPPORT_GFX_3D_CGLS))) 5232 return; 5233 5234 if (enable) { 5235 def = data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE); 5236 5237 /* unset CGCG override */ 5238 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGCG) 5239 data &= ~RLC_CGTT_MGCG_OVERRIDE__GFXIP_CGCG_OVERRIDE_MASK; 5240 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGLS) 5241 data &= ~RLC_CGTT_MGCG_OVERRIDE__GFXIP_CGLS_OVERRIDE_MASK; 5242 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGCG || 5243 adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGLS) 5244 data &= ~RLC_CGTT_MGCG_OVERRIDE__GFXIP_GFX3D_CG_OVERRIDE_MASK; 5245 5246 /* update CGCG override bits */ 5247 if (def != data) 5248 WREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE, data); 5249 5250 /* enable cgcg FSM(0x0000363F) */ 5251 def = data = RREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL); 5252 5253 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGCG) { 5254 data &= ~RLC_CGCG_CGLS_CTRL__CGCG_GFX_IDLE_THRESHOLD_MASK; 5255 data |= (0x36 << RLC_CGCG_CGLS_CTRL__CGCG_GFX_IDLE_THRESHOLD__SHIFT) | 5256 RLC_CGCG_CGLS_CTRL__CGCG_EN_MASK; 5257 } 5258 5259 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGLS) { 5260 data &= ~RLC_CGCG_CGLS_CTRL__CGLS_REP_COMPANSAT_DELAY_MASK; 5261 data |= (0x000F << RLC_CGCG_CGLS_CTRL__CGLS_REP_COMPANSAT_DELAY__SHIFT) | 5262 RLC_CGCG_CGLS_CTRL__CGLS_EN_MASK; 5263 } 5264 5265 if (def != data) 5266 WREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL, data); 5267 5268 /* Program RLC_CGCG_CGLS_CTRL_3D */ 5269 def = data = RREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL_3D); 5270 5271 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGCG) { 5272 data &= ~RLC_CGCG_CGLS_CTRL_3D__CGCG_GFX_IDLE_THRESHOLD_MASK; 5273 data |= (0x36 << RLC_CGCG_CGLS_CTRL_3D__CGCG_GFX_IDLE_THRESHOLD__SHIFT) | 5274 RLC_CGCG_CGLS_CTRL_3D__CGCG_EN_MASK; 5275 } 5276 5277 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGLS) { 5278 data &= ~RLC_CGCG_CGLS_CTRL_3D__CGLS_REP_COMPANSAT_DELAY_MASK; 5279 data |= (0xf << RLC_CGCG_CGLS_CTRL_3D__CGLS_REP_COMPANSAT_DELAY__SHIFT) | 5280 RLC_CGCG_CGLS_CTRL_3D__CGLS_EN_MASK; 5281 } 5282 5283 if (def != data) 5284 WREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL_3D, data); 5285 5286 /* set IDLE_POLL_COUNT(0x00900100) */ 5287 def = data = RREG32_SOC15(GC, 0, regCP_RB_WPTR_POLL_CNTL); 5288 5289 data &= ~(CP_RB_WPTR_POLL_CNTL__POLL_FREQUENCY_MASK | CP_RB_WPTR_POLL_CNTL__IDLE_POLL_COUNT_MASK); 5290 data |= (0x0100 << CP_RB_WPTR_POLL_CNTL__POLL_FREQUENCY__SHIFT) | 5291 (0x0090 << CP_RB_WPTR_POLL_CNTL__IDLE_POLL_COUNT__SHIFT); 5292 5293 if (def != data) 5294 WREG32_SOC15(GC, 0, regCP_RB_WPTR_POLL_CNTL, data); 5295 5296 data = RREG32_SOC15(GC, 0, regCP_INT_CNTL); 5297 data = REG_SET_FIELD(data, CP_INT_CNTL, CNTX_BUSY_INT_ENABLE, 1); 5298 data = REG_SET_FIELD(data, CP_INT_CNTL, CNTX_EMPTY_INT_ENABLE, 1); 5299 data = REG_SET_FIELD(data, CP_INT_CNTL, CMP_BUSY_INT_ENABLE, 1); 5300 data = REG_SET_FIELD(data, CP_INT_CNTL, GFX_IDLE_INT_ENABLE, 1); 5301 WREG32_SOC15(GC, 0, regCP_INT_CNTL, data); 5302 5303 data = RREG32_SOC15(GC, 0, regSDMA0_RLC_CGCG_CTRL); 5304 data = REG_SET_FIELD(data, SDMA0_RLC_CGCG_CTRL, CGCG_INT_ENABLE, 1); 5305 WREG32_SOC15(GC, 0, regSDMA0_RLC_CGCG_CTRL, data); 5306 5307 /* Some ASICs only have one SDMA instance, not need to configure SDMA1 */ 5308 if (adev->sdma.num_instances > 1) { 5309 data = RREG32_SOC15(GC, 0, regSDMA1_RLC_CGCG_CTRL); 5310 data = REG_SET_FIELD(data, SDMA1_RLC_CGCG_CTRL, CGCG_INT_ENABLE, 1); 5311 WREG32_SOC15(GC, 0, regSDMA1_RLC_CGCG_CTRL, data); 5312 } 5313 } else { 5314 /* Program RLC_CGCG_CGLS_CTRL */ 5315 def = data = RREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL); 5316 5317 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGCG) 5318 data &= ~RLC_CGCG_CGLS_CTRL__CGCG_EN_MASK; 5319 5320 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_CGLS) 5321 data &= ~RLC_CGCG_CGLS_CTRL__CGLS_EN_MASK; 5322 5323 if (def != data) 5324 WREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL, data); 5325 5326 /* Program RLC_CGCG_CGLS_CTRL_3D */ 5327 def = data = RREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL_3D); 5328 5329 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGCG) 5330 data &= ~RLC_CGCG_CGLS_CTRL_3D__CGCG_EN_MASK; 5331 if (adev->cg_flags & AMD_CG_SUPPORT_GFX_3D_CGLS) 5332 data &= ~RLC_CGCG_CGLS_CTRL_3D__CGLS_EN_MASK; 5333 5334 if (def != data) 5335 WREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL_3D, data); 5336 5337 data = RREG32_SOC15(GC, 0, regSDMA0_RLC_CGCG_CTRL); 5338 data &= ~SDMA0_RLC_CGCG_CTRL__CGCG_INT_ENABLE_MASK; 5339 WREG32_SOC15(GC, 0, regSDMA0_RLC_CGCG_CTRL, data); 5340 5341 /* Some ASICs only have one SDMA instance, not need to configure SDMA1 */ 5342 if (adev->sdma.num_instances > 1) { 5343 data = RREG32_SOC15(GC, 0, regSDMA1_RLC_CGCG_CTRL); 5344 data &= ~SDMA1_RLC_CGCG_CTRL__CGCG_INT_ENABLE_MASK; 5345 WREG32_SOC15(GC, 0, regSDMA1_RLC_CGCG_CTRL, data); 5346 } 5347 } 5348 } 5349 5350 static int gfx_v11_0_update_gfx_clock_gating(struct amdgpu_device *adev, 5351 bool enable) 5352 { 5353 amdgpu_gfx_rlc_enter_safe_mode(adev, 0); 5354 5355 gfx_v11_0_update_coarse_grain_clock_gating(adev, enable); 5356 5357 gfx_v11_0_update_medium_grain_clock_gating(adev, enable); 5358 5359 gfx_v11_0_update_repeater_fgcg(adev, enable); 5360 5361 gfx_v11_0_update_sram_fgcg(adev, enable); 5362 5363 gfx_v11_0_update_perf_clk(adev, enable); 5364 5365 if (adev->cg_flags & 5366 (AMD_CG_SUPPORT_GFX_MGCG | 5367 AMD_CG_SUPPORT_GFX_CGLS | 5368 AMD_CG_SUPPORT_GFX_CGCG | 5369 AMD_CG_SUPPORT_GFX_3D_CGCG | 5370 AMD_CG_SUPPORT_GFX_3D_CGLS)) 5371 gfx_v11_0_enable_gui_idle_interrupt(adev, enable); 5372 5373 amdgpu_gfx_rlc_exit_safe_mode(adev, 0); 5374 5375 return 0; 5376 } 5377 5378 static void gfx_v11_0_update_spm_vmid(struct amdgpu_device *adev, struct amdgpu_ring *ring, unsigned vmid) 5379 { 5380 u32 reg, pre_data, data; 5381 5382 amdgpu_gfx_off_ctrl(adev, false); 5383 reg = SOC15_REG_OFFSET(GC, 0, regRLC_SPM_MC_CNTL); 5384 if (amdgpu_sriov_is_pp_one_vf(adev) && !amdgpu_sriov_runtime(adev)) 5385 pre_data = RREG32_NO_KIQ(reg); 5386 else 5387 pre_data = RREG32(reg); 5388 5389 data = pre_data & (~RLC_SPM_MC_CNTL__RLC_SPM_VMID_MASK); 5390 data |= (vmid & RLC_SPM_MC_CNTL__RLC_SPM_VMID_MASK) << RLC_SPM_MC_CNTL__RLC_SPM_VMID__SHIFT; 5391 5392 if (pre_data != data) { 5393 if (amdgpu_sriov_is_pp_one_vf(adev) && !amdgpu_sriov_runtime(adev)) { 5394 WREG32_SOC15_NO_KIQ(GC, 0, regRLC_SPM_MC_CNTL, data); 5395 } else 5396 WREG32_SOC15(GC, 0, regRLC_SPM_MC_CNTL, data); 5397 } 5398 amdgpu_gfx_off_ctrl(adev, true); 5399 5400 if (ring 5401 && amdgpu_sriov_is_pp_one_vf(adev) 5402 && (pre_data != data) 5403 && ((ring->funcs->type == AMDGPU_RING_TYPE_GFX) 5404 || (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE))) { 5405 amdgpu_ring_emit_wreg(ring, reg, data); 5406 } 5407 } 5408 5409 static const struct amdgpu_rlc_funcs gfx_v11_0_rlc_funcs = { 5410 .is_rlc_enabled = gfx_v11_0_is_rlc_enabled, 5411 .set_safe_mode = gfx_v11_0_set_safe_mode, 5412 .unset_safe_mode = gfx_v11_0_unset_safe_mode, 5413 .init = gfx_v11_0_rlc_init, 5414 .get_csb_size = gfx_v11_0_get_csb_size, 5415 .get_csb_buffer = gfx_v11_0_get_csb_buffer, 5416 .resume = gfx_v11_0_rlc_resume, 5417 .stop = gfx_v11_0_rlc_stop, 5418 .reset = gfx_v11_0_rlc_reset, 5419 .start = gfx_v11_0_rlc_start, 5420 .update_spm_vmid = gfx_v11_0_update_spm_vmid, 5421 }; 5422 5423 static void gfx_v11_cntl_power_gating(struct amdgpu_device *adev, bool enable) 5424 { 5425 u32 data = RREG32_SOC15(GC, 0, regRLC_PG_CNTL); 5426 5427 if (enable && (adev->pg_flags & AMD_PG_SUPPORT_GFX_PG)) 5428 data |= RLC_PG_CNTL__GFX_POWER_GATING_ENABLE_MASK; 5429 else 5430 data &= ~RLC_PG_CNTL__GFX_POWER_GATING_ENABLE_MASK; 5431 5432 WREG32_SOC15(GC, 0, regRLC_PG_CNTL, data); 5433 5434 // Program RLC_PG_DELAY3 for CGPG hysteresis 5435 if (enable && (adev->pg_flags & AMD_PG_SUPPORT_GFX_PG)) { 5436 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 5437 case IP_VERSION(11, 0, 1): 5438 case IP_VERSION(11, 0, 4): 5439 case IP_VERSION(11, 5, 0): 5440 case IP_VERSION(11, 5, 1): 5441 case IP_VERSION(11, 5, 2): 5442 WREG32_SOC15(GC, 0, regRLC_PG_DELAY_3, RLC_PG_DELAY_3_DEFAULT_GC_11_0_1); 5443 break; 5444 default: 5445 break; 5446 } 5447 } 5448 } 5449 5450 static void gfx_v11_cntl_pg(struct amdgpu_device *adev, bool enable) 5451 { 5452 amdgpu_gfx_rlc_enter_safe_mode(adev, 0); 5453 5454 gfx_v11_cntl_power_gating(adev, enable); 5455 5456 amdgpu_gfx_rlc_exit_safe_mode(adev, 0); 5457 } 5458 5459 static int gfx_v11_0_set_powergating_state(struct amdgpu_ip_block *ip_block, 5460 enum amd_powergating_state state) 5461 { 5462 struct amdgpu_device *adev = ip_block->adev; 5463 bool enable = (state == AMD_PG_STATE_GATE); 5464 5465 if (amdgpu_sriov_vf(adev)) 5466 return 0; 5467 5468 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 5469 case IP_VERSION(11, 0, 0): 5470 case IP_VERSION(11, 0, 2): 5471 case IP_VERSION(11, 0, 3): 5472 amdgpu_gfx_off_ctrl(adev, enable); 5473 break; 5474 case IP_VERSION(11, 0, 1): 5475 case IP_VERSION(11, 0, 4): 5476 case IP_VERSION(11, 5, 0): 5477 case IP_VERSION(11, 5, 1): 5478 case IP_VERSION(11, 5, 2): 5479 if (!enable) 5480 amdgpu_gfx_off_ctrl(adev, false); 5481 5482 gfx_v11_cntl_pg(adev, enable); 5483 5484 if (enable) 5485 amdgpu_gfx_off_ctrl(adev, true); 5486 5487 break; 5488 default: 5489 break; 5490 } 5491 5492 return 0; 5493 } 5494 5495 static int gfx_v11_0_set_clockgating_state(struct amdgpu_ip_block *ip_block, 5496 enum amd_clockgating_state state) 5497 { 5498 struct amdgpu_device *adev = ip_block->adev; 5499 5500 if (amdgpu_sriov_vf(adev)) 5501 return 0; 5502 5503 switch (amdgpu_ip_version(adev, GC_HWIP, 0)) { 5504 case IP_VERSION(11, 0, 0): 5505 case IP_VERSION(11, 0, 1): 5506 case IP_VERSION(11, 0, 2): 5507 case IP_VERSION(11, 0, 3): 5508 case IP_VERSION(11, 0, 4): 5509 case IP_VERSION(11, 5, 0): 5510 case IP_VERSION(11, 5, 1): 5511 case IP_VERSION(11, 5, 2): 5512 gfx_v11_0_update_gfx_clock_gating(adev, 5513 state == AMD_CG_STATE_GATE); 5514 break; 5515 default: 5516 break; 5517 } 5518 5519 return 0; 5520 } 5521 5522 static void gfx_v11_0_get_clockgating_state(void *handle, u64 *flags) 5523 { 5524 struct amdgpu_device *adev = (struct amdgpu_device *)handle; 5525 int data; 5526 5527 /* AMD_CG_SUPPORT_GFX_MGCG */ 5528 data = RREG32_SOC15(GC, 0, regRLC_CGTT_MGCG_OVERRIDE); 5529 if (!(data & RLC_CGTT_MGCG_OVERRIDE__GFXIP_MGCG_OVERRIDE_MASK)) 5530 *flags |= AMD_CG_SUPPORT_GFX_MGCG; 5531 5532 /* AMD_CG_SUPPORT_REPEATER_FGCG */ 5533 if (!(data & RLC_CGTT_MGCG_OVERRIDE__GFXIP_REPEATER_FGCG_OVERRIDE_MASK)) 5534 *flags |= AMD_CG_SUPPORT_REPEATER_FGCG; 5535 5536 /* AMD_CG_SUPPORT_GFX_FGCG */ 5537 if (!(data & RLC_CGTT_MGCG_OVERRIDE__GFXIP_FGCG_OVERRIDE_MASK)) 5538 *flags |= AMD_CG_SUPPORT_GFX_FGCG; 5539 5540 /* AMD_CG_SUPPORT_GFX_PERF_CLK */ 5541 if (!(data & RLC_CGTT_MGCG_OVERRIDE__PERFMON_CLOCK_STATE_MASK)) 5542 *flags |= AMD_CG_SUPPORT_GFX_PERF_CLK; 5543 5544 /* AMD_CG_SUPPORT_GFX_CGCG */ 5545 data = RREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL); 5546 if (data & RLC_CGCG_CGLS_CTRL__CGCG_EN_MASK) 5547 *flags |= AMD_CG_SUPPORT_GFX_CGCG; 5548 5549 /* AMD_CG_SUPPORT_GFX_CGLS */ 5550 if (data & RLC_CGCG_CGLS_CTRL__CGLS_EN_MASK) 5551 *flags |= AMD_CG_SUPPORT_GFX_CGLS; 5552 5553 /* AMD_CG_SUPPORT_GFX_3D_CGCG */ 5554 data = RREG32_SOC15(GC, 0, regRLC_CGCG_CGLS_CTRL_3D); 5555 if (data & RLC_CGCG_CGLS_CTRL_3D__CGCG_EN_MASK) 5556 *flags |= AMD_CG_SUPPORT_GFX_3D_CGCG; 5557 5558 /* AMD_CG_SUPPORT_GFX_3D_CGLS */ 5559 if (data & RLC_CGCG_CGLS_CTRL_3D__CGLS_EN_MASK) 5560 *flags |= AMD_CG_SUPPORT_GFX_3D_CGLS; 5561 } 5562 5563 static u64 gfx_v11_0_ring_get_rptr_gfx(struct amdgpu_ring *ring) 5564 { 5565 /* gfx11 is 32bit rptr*/ 5566 return *(uint32_t *)ring->rptr_cpu_addr; 5567 } 5568 5569 static u64 gfx_v11_0_ring_get_wptr_gfx(struct amdgpu_ring *ring) 5570 { 5571 struct amdgpu_device *adev = ring->adev; 5572 u64 wptr; 5573 5574 /* XXX check if swapping is necessary on BE */ 5575 if (ring->use_doorbell) { 5576 wptr = atomic64_read((atomic64_t *)ring->wptr_cpu_addr); 5577 } else { 5578 wptr = RREG32_SOC15(GC, 0, regCP_RB0_WPTR); 5579 wptr += (u64)RREG32_SOC15(GC, 0, regCP_RB0_WPTR_HI) << 32; 5580 } 5581 5582 return wptr; 5583 } 5584 5585 static void gfx_v11_0_ring_set_wptr_gfx(struct amdgpu_ring *ring) 5586 { 5587 struct amdgpu_device *adev = ring->adev; 5588 5589 if (ring->use_doorbell) { 5590 /* XXX check if swapping is necessary on BE */ 5591 atomic64_set((atomic64_t *)ring->wptr_cpu_addr, 5592 ring->wptr); 5593 WDOORBELL64(ring->doorbell_index, ring->wptr); 5594 } else { 5595 WREG32_SOC15(GC, 0, regCP_RB0_WPTR, 5596 lower_32_bits(ring->wptr)); 5597 WREG32_SOC15(GC, 0, regCP_RB0_WPTR_HI, 5598 upper_32_bits(ring->wptr)); 5599 } 5600 } 5601 5602 static u64 gfx_v11_0_ring_get_rptr_compute(struct amdgpu_ring *ring) 5603 { 5604 /* gfx11 hardware is 32bit rptr */ 5605 return *(uint32_t *)ring->rptr_cpu_addr; 5606 } 5607 5608 static u64 gfx_v11_0_ring_get_wptr_compute(struct amdgpu_ring *ring) 5609 { 5610 u64 wptr; 5611 5612 /* XXX check if swapping is necessary on BE */ 5613 if (ring->use_doorbell) 5614 wptr = atomic64_read((atomic64_t *)ring->wptr_cpu_addr); 5615 else 5616 BUG(); 5617 return wptr; 5618 } 5619 5620 static void gfx_v11_0_ring_set_wptr_compute(struct amdgpu_ring *ring) 5621 { 5622 struct amdgpu_device *adev = ring->adev; 5623 5624 /* XXX check if swapping is necessary on BE */ 5625 if (ring->use_doorbell) { 5626 atomic64_set((atomic64_t *)ring->wptr_cpu_addr, 5627 ring->wptr); 5628 WDOORBELL64(ring->doorbell_index, ring->wptr); 5629 } else { 5630 BUG(); /* only DOORBELL method supported on gfx11 now */ 5631 } 5632 } 5633 5634 static void gfx_v11_0_ring_emit_hdp_flush(struct amdgpu_ring *ring) 5635 { 5636 struct amdgpu_device *adev = ring->adev; 5637 u32 ref_and_mask, reg_mem_engine; 5638 const struct nbio_hdp_flush_reg *nbio_hf_reg = adev->nbio.hdp_flush_reg; 5639 5640 if (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE) { 5641 switch (ring->me) { 5642 case 1: 5643 ref_and_mask = nbio_hf_reg->ref_and_mask_cp2 << ring->pipe; 5644 break; 5645 case 2: 5646 ref_and_mask = nbio_hf_reg->ref_and_mask_cp6 << ring->pipe; 5647 break; 5648 default: 5649 return; 5650 } 5651 reg_mem_engine = 0; 5652 } else { 5653 ref_and_mask = nbio_hf_reg->ref_and_mask_cp0 << ring->pipe; 5654 reg_mem_engine = 1; /* pfp */ 5655 } 5656 5657 gfx_v11_0_wait_reg_mem(ring, reg_mem_engine, 0, 1, 5658 adev->nbio.funcs->get_hdp_flush_req_offset(adev), 5659 adev->nbio.funcs->get_hdp_flush_done_offset(adev), 5660 ref_and_mask, ref_and_mask, 0x20); 5661 } 5662 5663 static void gfx_v11_0_ring_emit_ib_gfx(struct amdgpu_ring *ring, 5664 struct amdgpu_job *job, 5665 struct amdgpu_ib *ib, 5666 uint32_t flags) 5667 { 5668 unsigned vmid = AMDGPU_JOB_GET_VMID(job); 5669 u32 header, control = 0; 5670 5671 BUG_ON(ib->flags & AMDGPU_IB_FLAG_CE); 5672 5673 header = PACKET3(PACKET3_INDIRECT_BUFFER, 2); 5674 5675 control |= ib->length_dw | (vmid << 24); 5676 5677 if (ring->adev->gfx.mcbp && (ib->flags & AMDGPU_IB_FLAG_PREEMPT)) { 5678 control |= INDIRECT_BUFFER_PRE_ENB(1); 5679 5680 if (flags & AMDGPU_IB_PREEMPTED) 5681 control |= INDIRECT_BUFFER_PRE_RESUME(1); 5682 5683 if (vmid) 5684 gfx_v11_0_ring_emit_de_meta(ring, 5685 (!amdgpu_sriov_vf(ring->adev) && flags & AMDGPU_IB_PREEMPTED) ? true : false); 5686 } 5687 5688 if (ring->is_mes_queue) 5689 /* inherit vmid from mqd */ 5690 control |= 0x400000; 5691 5692 amdgpu_ring_write(ring, header); 5693 BUG_ON(ib->gpu_addr & 0x3); /* Dword align */ 5694 amdgpu_ring_write(ring, 5695 #ifdef __BIG_ENDIAN 5696 (2 << 0) | 5697 #endif 5698 lower_32_bits(ib->gpu_addr)); 5699 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr)); 5700 amdgpu_ring_write(ring, control); 5701 } 5702 5703 static void gfx_v11_0_ring_emit_ib_compute(struct amdgpu_ring *ring, 5704 struct amdgpu_job *job, 5705 struct amdgpu_ib *ib, 5706 uint32_t flags) 5707 { 5708 unsigned vmid = AMDGPU_JOB_GET_VMID(job); 5709 u32 control = INDIRECT_BUFFER_VALID | ib->length_dw | (vmid << 24); 5710 5711 if (ring->is_mes_queue) 5712 /* inherit vmid from mqd */ 5713 control |= 0x40000000; 5714 5715 /* Currently, there is a high possibility to get wave ID mismatch 5716 * between ME and GDS, leading to a hw deadlock, because ME generates 5717 * different wave IDs than the GDS expects. This situation happens 5718 * randomly when at least 5 compute pipes use GDS ordered append. 5719 * The wave IDs generated by ME are also wrong after suspend/resume. 5720 * Those are probably bugs somewhere else in the kernel driver. 5721 * 5722 * Writing GDS_COMPUTE_MAX_WAVE_ID resets wave ID counters in ME and 5723 * GDS to 0 for this ring (me/pipe). 5724 */ 5725 if (ib->flags & AMDGPU_IB_FLAG_RESET_GDS_MAX_WAVE_ID) { 5726 amdgpu_ring_write(ring, PACKET3(PACKET3_SET_CONFIG_REG, 1)); 5727 amdgpu_ring_write(ring, regGDS_COMPUTE_MAX_WAVE_ID); 5728 amdgpu_ring_write(ring, ring->adev->gds.gds_compute_max_wave_id); 5729 } 5730 5731 amdgpu_ring_write(ring, PACKET3(PACKET3_INDIRECT_BUFFER, 2)); 5732 BUG_ON(ib->gpu_addr & 0x3); /* Dword align */ 5733 amdgpu_ring_write(ring, 5734 #ifdef __BIG_ENDIAN 5735 (2 << 0) | 5736 #endif 5737 lower_32_bits(ib->gpu_addr)); 5738 amdgpu_ring_write(ring, upper_32_bits(ib->gpu_addr)); 5739 amdgpu_ring_write(ring, control); 5740 } 5741 5742 static void gfx_v11_0_ring_emit_fence(struct amdgpu_ring *ring, u64 addr, 5743 u64 seq, unsigned flags) 5744 { 5745 bool write64bit = flags & AMDGPU_FENCE_FLAG_64BIT; 5746 bool int_sel = flags & AMDGPU_FENCE_FLAG_INT; 5747 5748 /* RELEASE_MEM - flush caches, send int */ 5749 amdgpu_ring_write(ring, PACKET3(PACKET3_RELEASE_MEM, 6)); 5750 amdgpu_ring_write(ring, (PACKET3_RELEASE_MEM_GCR_SEQ | 5751 PACKET3_RELEASE_MEM_GCR_GL2_WB | 5752 PACKET3_RELEASE_MEM_GCR_GLM_INV | /* must be set with GLM_WB */ 5753 PACKET3_RELEASE_MEM_GCR_GLM_WB | 5754 PACKET3_RELEASE_MEM_CACHE_POLICY(3) | 5755 PACKET3_RELEASE_MEM_EVENT_TYPE(CACHE_FLUSH_AND_INV_TS_EVENT) | 5756 PACKET3_RELEASE_MEM_EVENT_INDEX(5))); 5757 amdgpu_ring_write(ring, (PACKET3_RELEASE_MEM_DATA_SEL(write64bit ? 2 : 1) | 5758 PACKET3_RELEASE_MEM_INT_SEL(int_sel ? 2 : 0))); 5759 5760 /* 5761 * the address should be Qword aligned if 64bit write, Dword 5762 * aligned if only send 32bit data low (discard data high) 5763 */ 5764 if (write64bit) 5765 BUG_ON(addr & 0x7); 5766 else 5767 BUG_ON(addr & 0x3); 5768 amdgpu_ring_write(ring, lower_32_bits(addr)); 5769 amdgpu_ring_write(ring, upper_32_bits(addr)); 5770 amdgpu_ring_write(ring, lower_32_bits(seq)); 5771 amdgpu_ring_write(ring, upper_32_bits(seq)); 5772 amdgpu_ring_write(ring, ring->is_mes_queue ? 5773 (ring->hw_queue_id | AMDGPU_FENCE_MES_QUEUE_FLAG) : 0); 5774 } 5775 5776 static void gfx_v11_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring) 5777 { 5778 int usepfp = (ring->funcs->type == AMDGPU_RING_TYPE_GFX); 5779 uint32_t seq = ring->fence_drv.sync_seq; 5780 uint64_t addr = ring->fence_drv.gpu_addr; 5781 5782 gfx_v11_0_wait_reg_mem(ring, usepfp, 1, 0, lower_32_bits(addr), 5783 upper_32_bits(addr), seq, 0xffffffff, 4); 5784 } 5785 5786 static void gfx_v11_0_ring_invalidate_tlbs(struct amdgpu_ring *ring, 5787 uint16_t pasid, uint32_t flush_type, 5788 bool all_hub, uint8_t dst_sel) 5789 { 5790 amdgpu_ring_write(ring, PACKET3(PACKET3_INVALIDATE_TLBS, 0)); 5791 amdgpu_ring_write(ring, 5792 PACKET3_INVALIDATE_TLBS_DST_SEL(dst_sel) | 5793 PACKET3_INVALIDATE_TLBS_ALL_HUB(all_hub) | 5794 PACKET3_INVALIDATE_TLBS_PASID(pasid) | 5795 PACKET3_INVALIDATE_TLBS_FLUSH_TYPE(flush_type)); 5796 } 5797 5798 static void gfx_v11_0_ring_emit_vm_flush(struct amdgpu_ring *ring, 5799 unsigned vmid, uint64_t pd_addr) 5800 { 5801 if (ring->is_mes_queue) 5802 gfx_v11_0_ring_invalidate_tlbs(ring, 0, 0, false, 0); 5803 else 5804 amdgpu_gmc_emit_flush_gpu_tlb(ring, vmid, pd_addr); 5805 5806 /* compute doesn't have PFP */ 5807 if (ring->funcs->type == AMDGPU_RING_TYPE_GFX) { 5808 /* sync PFP to ME, otherwise we might get invalid PFP reads */ 5809 amdgpu_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0)); 5810 amdgpu_ring_write(ring, 0x0); 5811 } 5812 5813 /* Make sure that we can't skip the SET_Q_MODE packets when the VM 5814 * changed in any way. 5815 */ 5816 ring->set_q_mode_offs = 0; 5817 ring->set_q_mode_ptr = NULL; 5818 } 5819 5820 static void gfx_v11_0_ring_emit_fence_kiq(struct amdgpu_ring *ring, u64 addr, 5821 u64 seq, unsigned int flags) 5822 { 5823 struct amdgpu_device *adev = ring->adev; 5824 5825 /* we only allocate 32bit for each seq wb address */ 5826 BUG_ON(flags & AMDGPU_FENCE_FLAG_64BIT); 5827 5828 /* write fence seq to the "addr" */ 5829 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3)); 5830 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) | 5831 WRITE_DATA_DST_SEL(5) | WR_CONFIRM)); 5832 amdgpu_ring_write(ring, lower_32_bits(addr)); 5833 amdgpu_ring_write(ring, upper_32_bits(addr)); 5834 amdgpu_ring_write(ring, lower_32_bits(seq)); 5835 5836 if (flags & AMDGPU_FENCE_FLAG_INT) { 5837 /* set register to trigger INT */ 5838 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3)); 5839 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(0) | 5840 WRITE_DATA_DST_SEL(0) | WR_CONFIRM)); 5841 amdgpu_ring_write(ring, SOC15_REG_OFFSET(GC, 0, regCPC_INT_STATUS)); 5842 amdgpu_ring_write(ring, 0); 5843 amdgpu_ring_write(ring, 0x20000000); /* src_id is 178 */ 5844 } 5845 } 5846 5847 static void gfx_v11_0_ring_emit_cntxcntl(struct amdgpu_ring *ring, 5848 uint32_t flags) 5849 { 5850 uint32_t dw2 = 0; 5851 5852 dw2 |= 0x80000000; /* set load_enable otherwise this package is just NOPs */ 5853 if (flags & AMDGPU_HAVE_CTX_SWITCH) { 5854 /* set load_global_config & load_global_uconfig */ 5855 dw2 |= 0x8001; 5856 /* set load_cs_sh_regs */ 5857 dw2 |= 0x01000000; 5858 /* set load_per_context_state & load_gfx_sh_regs for GFX */ 5859 dw2 |= 0x10002; 5860 } 5861 5862 amdgpu_ring_write(ring, PACKET3(PACKET3_CONTEXT_CONTROL, 1)); 5863 amdgpu_ring_write(ring, dw2); 5864 amdgpu_ring_write(ring, 0); 5865 } 5866 5867 static unsigned gfx_v11_0_ring_emit_init_cond_exec(struct amdgpu_ring *ring, 5868 uint64_t addr) 5869 { 5870 unsigned ret; 5871 5872 amdgpu_ring_write(ring, PACKET3(PACKET3_COND_EXEC, 3)); 5873 amdgpu_ring_write(ring, lower_32_bits(addr)); 5874 amdgpu_ring_write(ring, upper_32_bits(addr)); 5875 /* discard following DWs if *cond_exec_gpu_addr==0 */ 5876 amdgpu_ring_write(ring, 0); 5877 ret = ring->wptr & ring->buf_mask; 5878 /* patch dummy value later */ 5879 amdgpu_ring_write(ring, 0); 5880 5881 return ret; 5882 } 5883 5884 static void gfx_v11_0_ring_emit_gfx_shadow(struct amdgpu_ring *ring, 5885 u64 shadow_va, u64 csa_va, 5886 u64 gds_va, bool init_shadow, 5887 int vmid) 5888 { 5889 struct amdgpu_device *adev = ring->adev; 5890 unsigned int offs, end; 5891 5892 if (!adev->gfx.cp_gfx_shadow || !ring->ring_obj) 5893 return; 5894 5895 /* 5896 * The logic here isn't easy to understand because we need to keep state 5897 * accross multiple executions of the function as well as between the 5898 * CPU and GPU. The general idea is that the newly written GPU command 5899 * has a condition on the previous one and only executed if really 5900 * necessary. 5901 */ 5902 5903 /* 5904 * The dw in the NOP controls if the next SET_Q_MODE packet should be 5905 * executed or not. Reserve 64bits just to be on the save side. 5906 */ 5907 amdgpu_ring_write(ring, PACKET3(PACKET3_NOP, 1)); 5908 offs = ring->wptr & ring->buf_mask; 5909 5910 /* 5911 * We start with skipping the prefix SET_Q_MODE and always executing 5912 * the postfix SET_Q_MODE packet. This is changed below with a 5913 * WRITE_DATA command when the postfix executed. 5914 */ 5915 amdgpu_ring_write(ring, shadow_va ? 1 : 0); 5916 amdgpu_ring_write(ring, 0); 5917 5918 if (ring->set_q_mode_offs) { 5919 uint64_t addr; 5920 5921 addr = amdgpu_bo_gpu_offset(ring->ring_obj); 5922 addr += ring->set_q_mode_offs << 2; 5923 end = gfx_v11_0_ring_emit_init_cond_exec(ring, addr); 5924 } 5925 5926 /* 5927 * When the postfix SET_Q_MODE packet executes we need to make sure that the 5928 * next prefix SET_Q_MODE packet executes as well. 5929 */ 5930 if (!shadow_va) { 5931 uint64_t addr; 5932 5933 addr = amdgpu_bo_gpu_offset(ring->ring_obj); 5934 addr += offs << 2; 5935 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3)); 5936 amdgpu_ring_write(ring, WRITE_DATA_DST_SEL(5) | WR_CONFIRM); 5937 amdgpu_ring_write(ring, lower_32_bits(addr)); 5938 amdgpu_ring_write(ring, upper_32_bits(addr)); 5939 amdgpu_ring_write(ring, 0x1); 5940 } 5941 5942 amdgpu_ring_write(ring, PACKET3(PACKET3_SET_Q_PREEMPTION_MODE, 7)); 5943 amdgpu_ring_write(ring, lower_32_bits(shadow_va)); 5944 amdgpu_ring_write(ring, upper_32_bits(shadow_va)); 5945 amdgpu_ring_write(ring, lower_32_bits(gds_va)); 5946 amdgpu_ring_write(ring, upper_32_bits(gds_va)); 5947 amdgpu_ring_write(ring, lower_32_bits(csa_va)); 5948 amdgpu_ring_write(ring, upper_32_bits(csa_va)); 5949 amdgpu_ring_write(ring, shadow_va ? 5950 PACKET3_SET_Q_PREEMPTION_MODE_IB_VMID(vmid) : 0); 5951 amdgpu_ring_write(ring, init_shadow ? 5952 PACKET3_SET_Q_PREEMPTION_MODE_INIT_SHADOW_MEM : 0); 5953 5954 if (ring->set_q_mode_offs) 5955 amdgpu_ring_patch_cond_exec(ring, end); 5956 5957 if (shadow_va) { 5958 uint64_t token = shadow_va ^ csa_va ^ gds_va ^ vmid; 5959 5960 /* 5961 * If the tokens match try to skip the last postfix SET_Q_MODE 5962 * packet to avoid saving/restoring the state all the time. 5963 */ 5964 if (ring->set_q_mode_ptr && ring->set_q_mode_token == token) 5965 *ring->set_q_mode_ptr = 0; 5966 5967 ring->set_q_mode_token = token; 5968 } else { 5969 ring->set_q_mode_ptr = &ring->ring[ring->set_q_mode_offs]; 5970 } 5971 5972 ring->set_q_mode_offs = offs; 5973 } 5974 5975 static int gfx_v11_0_ring_preempt_ib(struct amdgpu_ring *ring) 5976 { 5977 int i, r = 0; 5978 struct amdgpu_device *adev = ring->adev; 5979 struct amdgpu_kiq *kiq = &adev->gfx.kiq[0]; 5980 struct amdgpu_ring *kiq_ring = &kiq->ring; 5981 unsigned long flags; 5982 5983 if (adev->enable_mes) 5984 return -EINVAL; 5985 5986 if (!kiq->pmf || !kiq->pmf->kiq_unmap_queues) 5987 return -EINVAL; 5988 5989 spin_lock_irqsave(&kiq->ring_lock, flags); 5990 5991 if (amdgpu_ring_alloc(kiq_ring, kiq->pmf->unmap_queues_size)) { 5992 spin_unlock_irqrestore(&kiq->ring_lock, flags); 5993 return -ENOMEM; 5994 } 5995 5996 /* assert preemption condition */ 5997 amdgpu_ring_set_preempt_cond_exec(ring, false); 5998 5999 /* assert IB preemption, emit the trailing fence */ 6000 kiq->pmf->kiq_unmap_queues(kiq_ring, ring, PREEMPT_QUEUES_NO_UNMAP, 6001 ring->trail_fence_gpu_addr, 6002 ++ring->trail_seq); 6003 amdgpu_ring_commit(kiq_ring); 6004 6005 spin_unlock_irqrestore(&kiq->ring_lock, flags); 6006 6007 /* poll the trailing fence */ 6008 for (i = 0; i < adev->usec_timeout; i++) { 6009 if (ring->trail_seq == 6010 le32_to_cpu(*(ring->trail_fence_cpu_addr))) 6011 break; 6012 udelay(1); 6013 } 6014 6015 if (i >= adev->usec_timeout) { 6016 r = -EINVAL; 6017 DRM_ERROR("ring %d failed to preempt ib\n", ring->idx); 6018 } 6019 6020 /* deassert preemption condition */ 6021 amdgpu_ring_set_preempt_cond_exec(ring, true); 6022 return r; 6023 } 6024 6025 static void gfx_v11_0_ring_emit_de_meta(struct amdgpu_ring *ring, bool resume) 6026 { 6027 struct amdgpu_device *adev = ring->adev; 6028 struct v10_de_ib_state de_payload = {0}; 6029 uint64_t offset, gds_addr, de_payload_gpu_addr; 6030 void *de_payload_cpu_addr; 6031 int cnt; 6032 6033 if (ring->is_mes_queue) { 6034 offset = offsetof(struct amdgpu_mes_ctx_meta_data, 6035 gfx[0].gfx_meta_data) + 6036 offsetof(struct v10_gfx_meta_data, de_payload); 6037 de_payload_gpu_addr = 6038 amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset); 6039 de_payload_cpu_addr = 6040 amdgpu_mes_ctx_get_offs_cpu_addr(ring, offset); 6041 6042 offset = offsetof(struct amdgpu_mes_ctx_meta_data, 6043 gfx[0].gds_backup) + 6044 offsetof(struct v10_gfx_meta_data, de_payload); 6045 gds_addr = amdgpu_mes_ctx_get_offs_gpu_addr(ring, offset); 6046 } else { 6047 offset = offsetof(struct v10_gfx_meta_data, de_payload); 6048 de_payload_gpu_addr = amdgpu_csa_vaddr(ring->adev) + offset; 6049 de_payload_cpu_addr = adev->virt.csa_cpu_addr + offset; 6050 6051 gds_addr = ALIGN(amdgpu_csa_vaddr(ring->adev) + 6052 AMDGPU_CSA_SIZE - adev->gds.gds_size, 6053 PAGE_SIZE); 6054 } 6055 6056 de_payload.gds_backup_addrlo = lower_32_bits(gds_addr); 6057 de_payload.gds_backup_addrhi = upper_32_bits(gds_addr); 6058 6059 cnt = (sizeof(de_payload) >> 2) + 4 - 2; 6060 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, cnt)); 6061 amdgpu_ring_write(ring, (WRITE_DATA_ENGINE_SEL(1) | 6062 WRITE_DATA_DST_SEL(8) | 6063 WR_CONFIRM) | 6064 WRITE_DATA_CACHE_POLICY(0)); 6065 amdgpu_ring_write(ring, lower_32_bits(de_payload_gpu_addr)); 6066 amdgpu_ring_write(ring, upper_32_bits(de_payload_gpu_addr)); 6067 6068 if (resume) 6069 amdgpu_ring_write_multiple(ring, de_payload_cpu_addr, 6070 sizeof(de_payload) >> 2); 6071 else 6072 amdgpu_ring_write_multiple(ring, (void *)&de_payload, 6073 sizeof(de_payload) >> 2); 6074 } 6075 6076 static void gfx_v11_0_ring_emit_frame_cntl(struct amdgpu_ring *ring, bool start, 6077 bool secure) 6078 { 6079 uint32_t v = secure ? FRAME_TMZ : 0; 6080 6081 amdgpu_ring_write(ring, PACKET3(PACKET3_FRAME_CONTROL, 0)); 6082 amdgpu_ring_write(ring, v | FRAME_CMD(start ? 0 : 1)); 6083 } 6084 6085 static void gfx_v11_0_ring_emit_rreg(struct amdgpu_ring *ring, uint32_t reg, 6086 uint32_t reg_val_offs) 6087 { 6088 struct amdgpu_device *adev = ring->adev; 6089 6090 amdgpu_ring_write(ring, PACKET3(PACKET3_COPY_DATA, 4)); 6091 amdgpu_ring_write(ring, 0 | /* src: register*/ 6092 (5 << 8) | /* dst: memory */ 6093 (1 << 20)); /* write confirm */ 6094 amdgpu_ring_write(ring, reg); 6095 amdgpu_ring_write(ring, 0); 6096 amdgpu_ring_write(ring, lower_32_bits(adev->wb.gpu_addr + 6097 reg_val_offs * 4)); 6098 amdgpu_ring_write(ring, upper_32_bits(adev->wb.gpu_addr + 6099 reg_val_offs * 4)); 6100 } 6101 6102 static void gfx_v11_0_ring_emit_wreg(struct amdgpu_ring *ring, uint32_t reg, 6103 uint32_t val) 6104 { 6105 uint32_t cmd = 0; 6106 6107 switch (ring->funcs->type) { 6108 case AMDGPU_RING_TYPE_GFX: 6109 cmd = WRITE_DATA_ENGINE_SEL(1) | WR_CONFIRM; 6110 break; 6111 case AMDGPU_RING_TYPE_KIQ: 6112 cmd = (1 << 16); /* no inc addr */ 6113 break; 6114 default: 6115 cmd = WR_CONFIRM; 6116 break; 6117 } 6118 amdgpu_ring_write(ring, PACKET3(PACKET3_WRITE_DATA, 3)); 6119 amdgpu_ring_write(ring, cmd); 6120 amdgpu_ring_write(ring, reg); 6121 amdgpu_ring_write(ring, 0); 6122 amdgpu_ring_write(ring, val); 6123 } 6124 6125 static void gfx_v11_0_ring_emit_reg_wait(struct amdgpu_ring *ring, uint32_t reg, 6126 uint32_t val, uint32_t mask) 6127 { 6128 gfx_v11_0_wait_reg_mem(ring, 0, 0, 0, reg, 0, val, mask, 0x20); 6129 } 6130 6131 static void gfx_v11_0_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring, 6132 uint32_t reg0, uint32_t reg1, 6133 uint32_t ref, uint32_t mask) 6134 { 6135 int usepfp = (ring->funcs->type == AMDGPU_RING_TYPE_GFX); 6136 6137 gfx_v11_0_wait_reg_mem(ring, usepfp, 0, 1, reg0, reg1, 6138 ref, mask, 0x20); 6139 } 6140 6141 static void gfx_v11_0_ring_soft_recovery(struct amdgpu_ring *ring, 6142 unsigned vmid) 6143 { 6144 struct amdgpu_device *adev = ring->adev; 6145 uint32_t value = 0; 6146 6147 value = REG_SET_FIELD(value, SQ_CMD, CMD, 0x03); 6148 value = REG_SET_FIELD(value, SQ_CMD, MODE, 0x01); 6149 value = REG_SET_FIELD(value, SQ_CMD, CHECK_VMID, 1); 6150 value = REG_SET_FIELD(value, SQ_CMD, VM_ID, vmid); 6151 amdgpu_gfx_rlc_enter_safe_mode(adev, 0); 6152 WREG32_SOC15(GC, 0, regSQ_CMD, value); 6153 amdgpu_gfx_rlc_exit_safe_mode(adev, 0); 6154 } 6155 6156 static void 6157 gfx_v11_0_set_gfx_eop_interrupt_state(struct amdgpu_device *adev, 6158 uint32_t me, uint32_t pipe, 6159 enum amdgpu_interrupt_state state) 6160 { 6161 uint32_t cp_int_cntl, cp_int_cntl_reg; 6162 6163 if (!me) { 6164 switch (pipe) { 6165 case 0: 6166 cp_int_cntl_reg = SOC15_REG_OFFSET(GC, 0, regCP_INT_CNTL_RING0); 6167 break; 6168 case 1: 6169 cp_int_cntl_reg = SOC15_REG_OFFSET(GC, 0, regCP_INT_CNTL_RING1); 6170 break; 6171 default: 6172 DRM_DEBUG("invalid pipe %d\n", pipe); 6173 return; 6174 } 6175 } else { 6176 DRM_DEBUG("invalid me %d\n", me); 6177 return; 6178 } 6179 6180 switch (state) { 6181 case AMDGPU_IRQ_STATE_DISABLE: 6182 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg); 6183 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0, 6184 TIME_STAMP_INT_ENABLE, 0); 6185 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0, 6186 GENERIC0_INT_ENABLE, 0); 6187 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl); 6188 break; 6189 case AMDGPU_IRQ_STATE_ENABLE: 6190 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg); 6191 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0, 6192 TIME_STAMP_INT_ENABLE, 1); 6193 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0, 6194 GENERIC0_INT_ENABLE, 1); 6195 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl); 6196 break; 6197 default: 6198 break; 6199 } 6200 } 6201 6202 static void gfx_v11_0_set_compute_eop_interrupt_state(struct amdgpu_device *adev, 6203 int me, int pipe, 6204 enum amdgpu_interrupt_state state) 6205 { 6206 u32 mec_int_cntl, mec_int_cntl_reg; 6207 6208 /* 6209 * amdgpu controls only the first MEC. That's why this function only 6210 * handles the setting of interrupts for this specific MEC. All other 6211 * pipes' interrupts are set by amdkfd. 6212 */ 6213 6214 if (me == 1) { 6215 switch (pipe) { 6216 case 0: 6217 mec_int_cntl_reg = SOC15_REG_OFFSET(GC, 0, regCP_ME1_PIPE0_INT_CNTL); 6218 break; 6219 case 1: 6220 mec_int_cntl_reg = SOC15_REG_OFFSET(GC, 0, regCP_ME1_PIPE1_INT_CNTL); 6221 break; 6222 case 2: 6223 mec_int_cntl_reg = SOC15_REG_OFFSET(GC, 0, regCP_ME1_PIPE2_INT_CNTL); 6224 break; 6225 case 3: 6226 mec_int_cntl_reg = SOC15_REG_OFFSET(GC, 0, regCP_ME1_PIPE3_INT_CNTL); 6227 break; 6228 default: 6229 DRM_DEBUG("invalid pipe %d\n", pipe); 6230 return; 6231 } 6232 } else { 6233 DRM_DEBUG("invalid me %d\n", me); 6234 return; 6235 } 6236 6237 switch (state) { 6238 case AMDGPU_IRQ_STATE_DISABLE: 6239 mec_int_cntl = RREG32_SOC15_IP(GC, mec_int_cntl_reg); 6240 mec_int_cntl = REG_SET_FIELD(mec_int_cntl, CP_ME1_PIPE0_INT_CNTL, 6241 TIME_STAMP_INT_ENABLE, 0); 6242 mec_int_cntl = REG_SET_FIELD(mec_int_cntl, CP_ME1_PIPE0_INT_CNTL, 6243 GENERIC0_INT_ENABLE, 0); 6244 WREG32_SOC15_IP(GC, mec_int_cntl_reg, mec_int_cntl); 6245 break; 6246 case AMDGPU_IRQ_STATE_ENABLE: 6247 mec_int_cntl = RREG32_SOC15_IP(GC, mec_int_cntl_reg); 6248 mec_int_cntl = REG_SET_FIELD(mec_int_cntl, CP_ME1_PIPE0_INT_CNTL, 6249 TIME_STAMP_INT_ENABLE, 1); 6250 mec_int_cntl = REG_SET_FIELD(mec_int_cntl, CP_ME1_PIPE0_INT_CNTL, 6251 GENERIC0_INT_ENABLE, 1); 6252 WREG32_SOC15_IP(GC, mec_int_cntl_reg, mec_int_cntl); 6253 break; 6254 default: 6255 break; 6256 } 6257 } 6258 6259 static int gfx_v11_0_set_eop_interrupt_state(struct amdgpu_device *adev, 6260 struct amdgpu_irq_src *src, 6261 unsigned type, 6262 enum amdgpu_interrupt_state state) 6263 { 6264 switch (type) { 6265 case AMDGPU_CP_IRQ_GFX_ME0_PIPE0_EOP: 6266 gfx_v11_0_set_gfx_eop_interrupt_state(adev, 0, 0, state); 6267 break; 6268 case AMDGPU_CP_IRQ_GFX_ME0_PIPE1_EOP: 6269 gfx_v11_0_set_gfx_eop_interrupt_state(adev, 0, 1, state); 6270 break; 6271 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE0_EOP: 6272 gfx_v11_0_set_compute_eop_interrupt_state(adev, 1, 0, state); 6273 break; 6274 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE1_EOP: 6275 gfx_v11_0_set_compute_eop_interrupt_state(adev, 1, 1, state); 6276 break; 6277 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE2_EOP: 6278 gfx_v11_0_set_compute_eop_interrupt_state(adev, 1, 2, state); 6279 break; 6280 case AMDGPU_CP_IRQ_COMPUTE_MEC1_PIPE3_EOP: 6281 gfx_v11_0_set_compute_eop_interrupt_state(adev, 1, 3, state); 6282 break; 6283 default: 6284 break; 6285 } 6286 return 0; 6287 } 6288 6289 static int gfx_v11_0_eop_irq(struct amdgpu_device *adev, 6290 struct amdgpu_irq_src *source, 6291 struct amdgpu_iv_entry *entry) 6292 { 6293 int i; 6294 u8 me_id, pipe_id, queue_id; 6295 struct amdgpu_ring *ring; 6296 uint32_t mes_queue_id = entry->src_data[0]; 6297 6298 DRM_DEBUG("IH: CP EOP\n"); 6299 6300 if (adev->enable_mes && (mes_queue_id & AMDGPU_FENCE_MES_QUEUE_FLAG)) { 6301 struct amdgpu_mes_queue *queue; 6302 6303 mes_queue_id &= AMDGPU_FENCE_MES_QUEUE_ID_MASK; 6304 6305 spin_lock(&adev->mes.queue_id_lock); 6306 queue = idr_find(&adev->mes.queue_id_idr, mes_queue_id); 6307 if (queue) { 6308 DRM_DEBUG("process mes queue id = %d\n", mes_queue_id); 6309 amdgpu_fence_process(queue->ring); 6310 } 6311 spin_unlock(&adev->mes.queue_id_lock); 6312 } else { 6313 me_id = (entry->ring_id & 0x0c) >> 2; 6314 pipe_id = (entry->ring_id & 0x03) >> 0; 6315 queue_id = (entry->ring_id & 0x70) >> 4; 6316 6317 switch (me_id) { 6318 case 0: 6319 if (pipe_id == 0) 6320 amdgpu_fence_process(&adev->gfx.gfx_ring[0]); 6321 else 6322 amdgpu_fence_process(&adev->gfx.gfx_ring[1]); 6323 break; 6324 case 1: 6325 case 2: 6326 for (i = 0; i < adev->gfx.num_compute_rings; i++) { 6327 ring = &adev->gfx.compute_ring[i]; 6328 /* Per-queue interrupt is supported for MEC starting from VI. 6329 * The interrupt can only be enabled/disabled per pipe instead 6330 * of per queue. 6331 */ 6332 if ((ring->me == me_id) && 6333 (ring->pipe == pipe_id) && 6334 (ring->queue == queue_id)) 6335 amdgpu_fence_process(ring); 6336 } 6337 break; 6338 } 6339 } 6340 6341 return 0; 6342 } 6343 6344 static int gfx_v11_0_set_priv_reg_fault_state(struct amdgpu_device *adev, 6345 struct amdgpu_irq_src *source, 6346 unsigned int type, 6347 enum amdgpu_interrupt_state state) 6348 { 6349 u32 cp_int_cntl_reg, cp_int_cntl; 6350 int i, j; 6351 6352 switch (state) { 6353 case AMDGPU_IRQ_STATE_DISABLE: 6354 case AMDGPU_IRQ_STATE_ENABLE: 6355 for (i = 0; i < adev->gfx.me.num_me; i++) { 6356 for (j = 0; j < adev->gfx.me.num_pipe_per_me; j++) { 6357 cp_int_cntl_reg = gfx_v11_0_get_cpg_int_cntl(adev, i, j); 6358 6359 if (cp_int_cntl_reg) { 6360 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg); 6361 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0, 6362 PRIV_REG_INT_ENABLE, 6363 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0); 6364 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl); 6365 } 6366 } 6367 } 6368 for (i = 0; i < adev->gfx.mec.num_mec; i++) { 6369 for (j = 0; j < adev->gfx.mec.num_pipe_per_mec; j++) { 6370 /* MECs start at 1 */ 6371 cp_int_cntl_reg = gfx_v11_0_get_cpc_int_cntl(adev, i + 1, j); 6372 6373 if (cp_int_cntl_reg) { 6374 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg); 6375 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_ME1_PIPE0_INT_CNTL, 6376 PRIV_REG_INT_ENABLE, 6377 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0); 6378 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl); 6379 } 6380 } 6381 } 6382 break; 6383 default: 6384 break; 6385 } 6386 6387 return 0; 6388 } 6389 6390 static int gfx_v11_0_set_bad_op_fault_state(struct amdgpu_device *adev, 6391 struct amdgpu_irq_src *source, 6392 unsigned type, 6393 enum amdgpu_interrupt_state state) 6394 { 6395 u32 cp_int_cntl_reg, cp_int_cntl; 6396 int i, j; 6397 6398 switch (state) { 6399 case AMDGPU_IRQ_STATE_DISABLE: 6400 case AMDGPU_IRQ_STATE_ENABLE: 6401 for (i = 0; i < adev->gfx.me.num_me; i++) { 6402 for (j = 0; j < adev->gfx.me.num_pipe_per_me; j++) { 6403 cp_int_cntl_reg = gfx_v11_0_get_cpg_int_cntl(adev, i, j); 6404 6405 if (cp_int_cntl_reg) { 6406 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg); 6407 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0, 6408 OPCODE_ERROR_INT_ENABLE, 6409 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0); 6410 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl); 6411 } 6412 } 6413 } 6414 for (i = 0; i < adev->gfx.mec.num_mec; i++) { 6415 for (j = 0; j < adev->gfx.mec.num_pipe_per_mec; j++) { 6416 /* MECs start at 1 */ 6417 cp_int_cntl_reg = gfx_v11_0_get_cpc_int_cntl(adev, i + 1, j); 6418 6419 if (cp_int_cntl_reg) { 6420 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg); 6421 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_ME1_PIPE0_INT_CNTL, 6422 OPCODE_ERROR_INT_ENABLE, 6423 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0); 6424 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl); 6425 } 6426 } 6427 } 6428 break; 6429 default: 6430 break; 6431 } 6432 return 0; 6433 } 6434 6435 static int gfx_v11_0_set_priv_inst_fault_state(struct amdgpu_device *adev, 6436 struct amdgpu_irq_src *source, 6437 unsigned int type, 6438 enum amdgpu_interrupt_state state) 6439 { 6440 u32 cp_int_cntl_reg, cp_int_cntl; 6441 int i, j; 6442 6443 switch (state) { 6444 case AMDGPU_IRQ_STATE_DISABLE: 6445 case AMDGPU_IRQ_STATE_ENABLE: 6446 for (i = 0; i < adev->gfx.me.num_me; i++) { 6447 for (j = 0; j < adev->gfx.me.num_pipe_per_me; j++) { 6448 cp_int_cntl_reg = gfx_v11_0_get_cpg_int_cntl(adev, i, j); 6449 6450 if (cp_int_cntl_reg) { 6451 cp_int_cntl = RREG32_SOC15_IP(GC, cp_int_cntl_reg); 6452 cp_int_cntl = REG_SET_FIELD(cp_int_cntl, CP_INT_CNTL_RING0, 6453 PRIV_INSTR_INT_ENABLE, 6454 state == AMDGPU_IRQ_STATE_ENABLE ? 1 : 0); 6455 WREG32_SOC15_IP(GC, cp_int_cntl_reg, cp_int_cntl); 6456 } 6457 } 6458 } 6459 break; 6460 default: 6461 break; 6462 } 6463 6464 return 0; 6465 } 6466 6467 static void gfx_v11_0_handle_priv_fault(struct amdgpu_device *adev, 6468 struct amdgpu_iv_entry *entry) 6469 { 6470 u8 me_id, pipe_id, queue_id; 6471 struct amdgpu_ring *ring; 6472 int i; 6473 6474 me_id = (entry->ring_id & 0x0c) >> 2; 6475 pipe_id = (entry->ring_id & 0x03) >> 0; 6476 queue_id = (entry->ring_id & 0x70) >> 4; 6477 6478 switch (me_id) { 6479 case 0: 6480 for (i = 0; i < adev->gfx.num_gfx_rings; i++) { 6481 ring = &adev->gfx.gfx_ring[i]; 6482 if (ring->me == me_id && ring->pipe == pipe_id && 6483 ring->queue == queue_id) 6484 drm_sched_fault(&ring->sched); 6485 } 6486 break; 6487 case 1: 6488 case 2: 6489 for (i = 0; i < adev->gfx.num_compute_rings; i++) { 6490 ring = &adev->gfx.compute_ring[i]; 6491 if (ring->me == me_id && ring->pipe == pipe_id && 6492 ring->queue == queue_id) 6493 drm_sched_fault(&ring->sched); 6494 } 6495 break; 6496 default: 6497 BUG(); 6498 break; 6499 } 6500 } 6501 6502 static int gfx_v11_0_priv_reg_irq(struct amdgpu_device *adev, 6503 struct amdgpu_irq_src *source, 6504 struct amdgpu_iv_entry *entry) 6505 { 6506 DRM_ERROR("Illegal register access in command stream\n"); 6507 gfx_v11_0_handle_priv_fault(adev, entry); 6508 return 0; 6509 } 6510 6511 static int gfx_v11_0_bad_op_irq(struct amdgpu_device *adev, 6512 struct amdgpu_irq_src *source, 6513 struct amdgpu_iv_entry *entry) 6514 { 6515 DRM_ERROR("Illegal opcode in command stream \n"); 6516 gfx_v11_0_handle_priv_fault(adev, entry); 6517 return 0; 6518 } 6519 6520 static int gfx_v11_0_priv_inst_irq(struct amdgpu_device *adev, 6521 struct amdgpu_irq_src *source, 6522 struct amdgpu_iv_entry *entry) 6523 { 6524 DRM_ERROR("Illegal instruction in command stream\n"); 6525 gfx_v11_0_handle_priv_fault(adev, entry); 6526 return 0; 6527 } 6528 6529 static int gfx_v11_0_rlc_gc_fed_irq(struct amdgpu_device *adev, 6530 struct amdgpu_irq_src *source, 6531 struct amdgpu_iv_entry *entry) 6532 { 6533 if (adev->gfx.ras && adev->gfx.ras->rlc_gc_fed_irq) 6534 return adev->gfx.ras->rlc_gc_fed_irq(adev, source, entry); 6535 6536 return 0; 6537 } 6538 6539 #if 0 6540 static int gfx_v11_0_kiq_set_interrupt_state(struct amdgpu_device *adev, 6541 struct amdgpu_irq_src *src, 6542 unsigned int type, 6543 enum amdgpu_interrupt_state state) 6544 { 6545 uint32_t tmp, target; 6546 struct amdgpu_ring *ring = &(adev->gfx.kiq[0].ring); 6547 6548 target = SOC15_REG_OFFSET(GC, 0, regCP_ME1_PIPE0_INT_CNTL); 6549 target += ring->pipe; 6550 6551 switch (type) { 6552 case AMDGPU_CP_KIQ_IRQ_DRIVER0: 6553 if (state == AMDGPU_IRQ_STATE_DISABLE) { 6554 tmp = RREG32_SOC15(GC, 0, regCPC_INT_CNTL); 6555 tmp = REG_SET_FIELD(tmp, CPC_INT_CNTL, 6556 GENERIC2_INT_ENABLE, 0); 6557 WREG32_SOC15(GC, 0, regCPC_INT_CNTL, tmp); 6558 6559 tmp = RREG32_SOC15_IP(GC, target); 6560 tmp = REG_SET_FIELD(tmp, CP_ME1_PIPE0_INT_CNTL, 6561 GENERIC2_INT_ENABLE, 0); 6562 WREG32_SOC15_IP(GC, target, tmp); 6563 } else { 6564 tmp = RREG32_SOC15(GC, 0, regCPC_INT_CNTL); 6565 tmp = REG_SET_FIELD(tmp, CPC_INT_CNTL, 6566 GENERIC2_INT_ENABLE, 1); 6567 WREG32_SOC15(GC, 0, regCPC_INT_CNTL, tmp); 6568 6569 tmp = RREG32_SOC15_IP(GC, target); 6570 tmp = REG_SET_FIELD(tmp, CP_ME1_PIPE0_INT_CNTL, 6571 GENERIC2_INT_ENABLE, 1); 6572 WREG32_SOC15_IP(GC, target, tmp); 6573 } 6574 break; 6575 default: 6576 BUG(); /* kiq only support GENERIC2_INT now */ 6577 break; 6578 } 6579 return 0; 6580 } 6581 #endif 6582 6583 static void gfx_v11_0_emit_mem_sync(struct amdgpu_ring *ring) 6584 { 6585 const unsigned int gcr_cntl = 6586 PACKET3_ACQUIRE_MEM_GCR_CNTL_GL2_INV(1) | 6587 PACKET3_ACQUIRE_MEM_GCR_CNTL_GL2_WB(1) | 6588 PACKET3_ACQUIRE_MEM_GCR_CNTL_GLM_INV(1) | 6589 PACKET3_ACQUIRE_MEM_GCR_CNTL_GLM_WB(1) | 6590 PACKET3_ACQUIRE_MEM_GCR_CNTL_GL1_INV(1) | 6591 PACKET3_ACQUIRE_MEM_GCR_CNTL_GLV_INV(1) | 6592 PACKET3_ACQUIRE_MEM_GCR_CNTL_GLK_INV(1) | 6593 PACKET3_ACQUIRE_MEM_GCR_CNTL_GLI_INV(1); 6594 6595 /* ACQUIRE_MEM - make one or more surfaces valid for use by the subsequent operations */ 6596 amdgpu_ring_write(ring, PACKET3(PACKET3_ACQUIRE_MEM, 6)); 6597 amdgpu_ring_write(ring, 0); /* CP_COHER_CNTL */ 6598 amdgpu_ring_write(ring, 0xffffffff); /* CP_COHER_SIZE */ 6599 amdgpu_ring_write(ring, 0xffffff); /* CP_COHER_SIZE_HI */ 6600 amdgpu_ring_write(ring, 0); /* CP_COHER_BASE */ 6601 amdgpu_ring_write(ring, 0); /* CP_COHER_BASE_HI */ 6602 amdgpu_ring_write(ring, 0x0000000A); /* POLL_INTERVAL */ 6603 amdgpu_ring_write(ring, gcr_cntl); /* GCR_CNTL */ 6604 } 6605 6606 static int gfx_v11_0_reset_kgq(struct amdgpu_ring *ring, unsigned int vmid) 6607 { 6608 struct amdgpu_device *adev = ring->adev; 6609 int r; 6610 6611 if (amdgpu_sriov_vf(adev)) 6612 return -EINVAL; 6613 6614 r = amdgpu_mes_reset_legacy_queue(ring->adev, ring, vmid, false); 6615 if (r) 6616 return r; 6617 6618 r = amdgpu_bo_reserve(ring->mqd_obj, false); 6619 if (unlikely(r != 0)) { 6620 dev_err(adev->dev, "fail to resv mqd_obj\n"); 6621 return r; 6622 } 6623 r = amdgpu_bo_kmap(ring->mqd_obj, (void **)&ring->mqd_ptr); 6624 if (!r) { 6625 r = gfx_v11_0_kgq_init_queue(ring, true); 6626 amdgpu_bo_kunmap(ring->mqd_obj); 6627 ring->mqd_ptr = NULL; 6628 } 6629 amdgpu_bo_unreserve(ring->mqd_obj); 6630 if (r) { 6631 dev_err(adev->dev, "fail to unresv mqd_obj\n"); 6632 return r; 6633 } 6634 6635 r = amdgpu_mes_map_legacy_queue(adev, ring); 6636 if (r) { 6637 dev_err(adev->dev, "failed to remap kgq\n"); 6638 return r; 6639 } 6640 6641 return amdgpu_ring_test_ring(ring); 6642 } 6643 6644 static int gfx_v11_0_reset_kcq(struct amdgpu_ring *ring, unsigned int vmid) 6645 { 6646 struct amdgpu_device *adev = ring->adev; 6647 int i, r = 0; 6648 6649 if (amdgpu_sriov_vf(adev)) 6650 return -EINVAL; 6651 6652 amdgpu_gfx_rlc_enter_safe_mode(adev, 0); 6653 mutex_lock(&adev->srbm_mutex); 6654 soc21_grbm_select(adev, ring->me, ring->pipe, ring->queue, 0); 6655 WREG32_SOC15(GC, 0, regCP_HQD_DEQUEUE_REQUEST, 0x2); 6656 WREG32_SOC15(GC, 0, regSPI_COMPUTE_QUEUE_RESET, 0x1); 6657 6658 /* make sure dequeue is complete*/ 6659 for (i = 0; i < adev->usec_timeout; i++) { 6660 if (!(RREG32_SOC15(GC, 0, regCP_HQD_ACTIVE) & 1)) 6661 break; 6662 udelay(1); 6663 } 6664 if (i >= adev->usec_timeout) 6665 r = -ETIMEDOUT; 6666 soc21_grbm_select(adev, 0, 0, 0, 0); 6667 mutex_unlock(&adev->srbm_mutex); 6668 amdgpu_gfx_rlc_exit_safe_mode(adev, 0); 6669 if (r) { 6670 dev_err(adev->dev, "fail to wait on hqd deactivate\n"); 6671 return r; 6672 } 6673 6674 r = amdgpu_bo_reserve(ring->mqd_obj, false); 6675 if (unlikely(r != 0)) { 6676 dev_err(adev->dev, "fail to resv mqd_obj\n"); 6677 return r; 6678 } 6679 r = amdgpu_bo_kmap(ring->mqd_obj, (void **)&ring->mqd_ptr); 6680 if (!r) { 6681 r = gfx_v11_0_kcq_init_queue(ring, true); 6682 amdgpu_bo_kunmap(ring->mqd_obj); 6683 ring->mqd_ptr = NULL; 6684 } 6685 amdgpu_bo_unreserve(ring->mqd_obj); 6686 if (r) { 6687 dev_err(adev->dev, "fail to unresv mqd_obj\n"); 6688 return r; 6689 } 6690 r = amdgpu_mes_map_legacy_queue(adev, ring); 6691 if (r) { 6692 dev_err(adev->dev, "failed to remap kcq\n"); 6693 return r; 6694 } 6695 6696 return amdgpu_ring_test_ring(ring); 6697 } 6698 6699 static void gfx_v11_ip_print(struct amdgpu_ip_block *ip_block, struct drm_printer *p) 6700 { 6701 struct amdgpu_device *adev = ip_block->adev; 6702 uint32_t i, j, k, reg, index = 0; 6703 uint32_t reg_count = ARRAY_SIZE(gc_reg_list_11_0); 6704 6705 if (!adev->gfx.ip_dump_core) 6706 return; 6707 6708 for (i = 0; i < reg_count; i++) 6709 drm_printf(p, "%-50s \t 0x%08x\n", 6710 gc_reg_list_11_0[i].reg_name, 6711 adev->gfx.ip_dump_core[i]); 6712 6713 /* print compute queue registers for all instances */ 6714 if (!adev->gfx.ip_dump_compute_queues) 6715 return; 6716 6717 reg_count = ARRAY_SIZE(gc_cp_reg_list_11); 6718 drm_printf(p, "\nnum_mec: %d num_pipe: %d num_queue: %d\n", 6719 adev->gfx.mec.num_mec, 6720 adev->gfx.mec.num_pipe_per_mec, 6721 adev->gfx.mec.num_queue_per_pipe); 6722 6723 for (i = 0; i < adev->gfx.mec.num_mec; i++) { 6724 for (j = 0; j < adev->gfx.mec.num_pipe_per_mec; j++) { 6725 for (k = 0; k < adev->gfx.mec.num_queue_per_pipe; k++) { 6726 drm_printf(p, "\nmec %d, pipe %d, queue %d\n", i, j, k); 6727 for (reg = 0; reg < reg_count; reg++) { 6728 drm_printf(p, "%-50s \t 0x%08x\n", 6729 gc_cp_reg_list_11[reg].reg_name, 6730 adev->gfx.ip_dump_compute_queues[index + reg]); 6731 } 6732 index += reg_count; 6733 } 6734 } 6735 } 6736 6737 /* print gfx queue registers for all instances */ 6738 if (!adev->gfx.ip_dump_gfx_queues) 6739 return; 6740 6741 index = 0; 6742 reg_count = ARRAY_SIZE(gc_gfx_queue_reg_list_11); 6743 drm_printf(p, "\nnum_me: %d num_pipe: %d num_queue: %d\n", 6744 adev->gfx.me.num_me, 6745 adev->gfx.me.num_pipe_per_me, 6746 adev->gfx.me.num_queue_per_pipe); 6747 6748 for (i = 0; i < adev->gfx.me.num_me; i++) { 6749 for (j = 0; j < adev->gfx.me.num_pipe_per_me; j++) { 6750 for (k = 0; k < adev->gfx.me.num_queue_per_pipe; k++) { 6751 drm_printf(p, "\nme %d, pipe %d, queue %d\n", i, j, k); 6752 for (reg = 0; reg < reg_count; reg++) { 6753 drm_printf(p, "%-50s \t 0x%08x\n", 6754 gc_gfx_queue_reg_list_11[reg].reg_name, 6755 adev->gfx.ip_dump_gfx_queues[index + reg]); 6756 } 6757 index += reg_count; 6758 } 6759 } 6760 } 6761 } 6762 6763 static void gfx_v11_ip_dump(struct amdgpu_ip_block *ip_block) 6764 { 6765 struct amdgpu_device *adev = ip_block->adev; 6766 uint32_t i, j, k, reg, index = 0; 6767 uint32_t reg_count = ARRAY_SIZE(gc_reg_list_11_0); 6768 6769 if (!adev->gfx.ip_dump_core) 6770 return; 6771 6772 amdgpu_gfx_off_ctrl(adev, false); 6773 for (i = 0; i < reg_count; i++) 6774 adev->gfx.ip_dump_core[i] = RREG32(SOC15_REG_ENTRY_OFFSET(gc_reg_list_11_0[i])); 6775 amdgpu_gfx_off_ctrl(adev, true); 6776 6777 /* dump compute queue registers for all instances */ 6778 if (!adev->gfx.ip_dump_compute_queues) 6779 return; 6780 6781 reg_count = ARRAY_SIZE(gc_cp_reg_list_11); 6782 amdgpu_gfx_off_ctrl(adev, false); 6783 mutex_lock(&adev->srbm_mutex); 6784 for (i = 0; i < adev->gfx.mec.num_mec; i++) { 6785 for (j = 0; j < adev->gfx.mec.num_pipe_per_mec; j++) { 6786 for (k = 0; k < adev->gfx.mec.num_queue_per_pipe; k++) { 6787 /* ME0 is for GFX so start from 1 for CP */ 6788 soc21_grbm_select(adev, adev->gfx.me.num_me + i, j, k, 0); 6789 for (reg = 0; reg < reg_count; reg++) { 6790 adev->gfx.ip_dump_compute_queues[index + reg] = 6791 RREG32(SOC15_REG_ENTRY_OFFSET( 6792 gc_cp_reg_list_11[reg])); 6793 } 6794 index += reg_count; 6795 } 6796 } 6797 } 6798 soc21_grbm_select(adev, 0, 0, 0, 0); 6799 mutex_unlock(&adev->srbm_mutex); 6800 amdgpu_gfx_off_ctrl(adev, true); 6801 6802 /* dump gfx queue registers for all instances */ 6803 if (!adev->gfx.ip_dump_gfx_queues) 6804 return; 6805 6806 index = 0; 6807 reg_count = ARRAY_SIZE(gc_gfx_queue_reg_list_11); 6808 amdgpu_gfx_off_ctrl(adev, false); 6809 mutex_lock(&adev->srbm_mutex); 6810 for (i = 0; i < adev->gfx.me.num_me; i++) { 6811 for (j = 0; j < adev->gfx.me.num_pipe_per_me; j++) { 6812 for (k = 0; k < adev->gfx.me.num_queue_per_pipe; k++) { 6813 soc21_grbm_select(adev, i, j, k, 0); 6814 6815 for (reg = 0; reg < reg_count; reg++) { 6816 adev->gfx.ip_dump_gfx_queues[index + reg] = 6817 RREG32(SOC15_REG_ENTRY_OFFSET( 6818 gc_gfx_queue_reg_list_11[reg])); 6819 } 6820 index += reg_count; 6821 } 6822 } 6823 } 6824 soc21_grbm_select(adev, 0, 0, 0, 0); 6825 mutex_unlock(&adev->srbm_mutex); 6826 amdgpu_gfx_off_ctrl(adev, true); 6827 } 6828 6829 static void gfx_v11_0_ring_emit_cleaner_shader(struct amdgpu_ring *ring) 6830 { 6831 /* Emit the cleaner shader */ 6832 amdgpu_ring_write(ring, PACKET3(PACKET3_RUN_CLEANER_SHADER, 0)); 6833 amdgpu_ring_write(ring, 0); /* RESERVED field, programmed to zero */ 6834 } 6835 6836 static const struct amd_ip_funcs gfx_v11_0_ip_funcs = { 6837 .name = "gfx_v11_0", 6838 .early_init = gfx_v11_0_early_init, 6839 .late_init = gfx_v11_0_late_init, 6840 .sw_init = gfx_v11_0_sw_init, 6841 .sw_fini = gfx_v11_0_sw_fini, 6842 .hw_init = gfx_v11_0_hw_init, 6843 .hw_fini = gfx_v11_0_hw_fini, 6844 .suspend = gfx_v11_0_suspend, 6845 .resume = gfx_v11_0_resume, 6846 .is_idle = gfx_v11_0_is_idle, 6847 .wait_for_idle = gfx_v11_0_wait_for_idle, 6848 .soft_reset = gfx_v11_0_soft_reset, 6849 .check_soft_reset = gfx_v11_0_check_soft_reset, 6850 .post_soft_reset = gfx_v11_0_post_soft_reset, 6851 .set_clockgating_state = gfx_v11_0_set_clockgating_state, 6852 .set_powergating_state = gfx_v11_0_set_powergating_state, 6853 .get_clockgating_state = gfx_v11_0_get_clockgating_state, 6854 .dump_ip_state = gfx_v11_ip_dump, 6855 .print_ip_state = gfx_v11_ip_print, 6856 }; 6857 6858 static const struct amdgpu_ring_funcs gfx_v11_0_ring_funcs_gfx = { 6859 .type = AMDGPU_RING_TYPE_GFX, 6860 .align_mask = 0xff, 6861 .nop = PACKET3(PACKET3_NOP, 0x3FFF), 6862 .support_64bit_ptrs = true, 6863 .secure_submission_supported = true, 6864 .get_rptr = gfx_v11_0_ring_get_rptr_gfx, 6865 .get_wptr = gfx_v11_0_ring_get_wptr_gfx, 6866 .set_wptr = gfx_v11_0_ring_set_wptr_gfx, 6867 .emit_frame_size = /* totally 247 maximum if 16 IBs */ 6868 5 + /* update_spm_vmid */ 6869 5 + /* COND_EXEC */ 6870 22 + /* SET_Q_PREEMPTION_MODE */ 6871 7 + /* PIPELINE_SYNC */ 6872 SOC15_FLUSH_GPU_TLB_NUM_WREG * 5 + 6873 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 7 + 6874 4 + /* VM_FLUSH */ 6875 8 + /* FENCE for VM_FLUSH */ 6876 20 + /* GDS switch */ 6877 5 + /* COND_EXEC */ 6878 7 + /* HDP_flush */ 6879 4 + /* VGT_flush */ 6880 31 + /* DE_META */ 6881 3 + /* CNTX_CTRL */ 6882 5 + /* HDP_INVL */ 6883 22 + /* SET_Q_PREEMPTION_MODE */ 6884 8 + 8 + /* FENCE x2 */ 6885 8 + /* gfx_v11_0_emit_mem_sync */ 6886 2, /* gfx_v11_0_ring_emit_cleaner_shader */ 6887 .emit_ib_size = 4, /* gfx_v11_0_ring_emit_ib_gfx */ 6888 .emit_ib = gfx_v11_0_ring_emit_ib_gfx, 6889 .emit_fence = gfx_v11_0_ring_emit_fence, 6890 .emit_pipeline_sync = gfx_v11_0_ring_emit_pipeline_sync, 6891 .emit_vm_flush = gfx_v11_0_ring_emit_vm_flush, 6892 .emit_gds_switch = gfx_v11_0_ring_emit_gds_switch, 6893 .emit_hdp_flush = gfx_v11_0_ring_emit_hdp_flush, 6894 .test_ring = gfx_v11_0_ring_test_ring, 6895 .test_ib = gfx_v11_0_ring_test_ib, 6896 .insert_nop = gfx_v11_ring_insert_nop, 6897 .pad_ib = amdgpu_ring_generic_pad_ib, 6898 .emit_cntxcntl = gfx_v11_0_ring_emit_cntxcntl, 6899 .emit_gfx_shadow = gfx_v11_0_ring_emit_gfx_shadow, 6900 .init_cond_exec = gfx_v11_0_ring_emit_init_cond_exec, 6901 .preempt_ib = gfx_v11_0_ring_preempt_ib, 6902 .emit_frame_cntl = gfx_v11_0_ring_emit_frame_cntl, 6903 .emit_wreg = gfx_v11_0_ring_emit_wreg, 6904 .emit_reg_wait = gfx_v11_0_ring_emit_reg_wait, 6905 .emit_reg_write_reg_wait = gfx_v11_0_ring_emit_reg_write_reg_wait, 6906 .soft_recovery = gfx_v11_0_ring_soft_recovery, 6907 .emit_mem_sync = gfx_v11_0_emit_mem_sync, 6908 .reset = gfx_v11_0_reset_kgq, 6909 .emit_cleaner_shader = gfx_v11_0_ring_emit_cleaner_shader, 6910 .begin_use = amdgpu_gfx_enforce_isolation_ring_begin_use, 6911 .end_use = amdgpu_gfx_enforce_isolation_ring_end_use, 6912 }; 6913 6914 static const struct amdgpu_ring_funcs gfx_v11_0_ring_funcs_compute = { 6915 .type = AMDGPU_RING_TYPE_COMPUTE, 6916 .align_mask = 0xff, 6917 .nop = PACKET3(PACKET3_NOP, 0x3FFF), 6918 .support_64bit_ptrs = true, 6919 .get_rptr = gfx_v11_0_ring_get_rptr_compute, 6920 .get_wptr = gfx_v11_0_ring_get_wptr_compute, 6921 .set_wptr = gfx_v11_0_ring_set_wptr_compute, 6922 .emit_frame_size = 6923 5 + /* update_spm_vmid */ 6924 20 + /* gfx_v11_0_ring_emit_gds_switch */ 6925 7 + /* gfx_v11_0_ring_emit_hdp_flush */ 6926 5 + /* hdp invalidate */ 6927 7 + /* gfx_v11_0_ring_emit_pipeline_sync */ 6928 SOC15_FLUSH_GPU_TLB_NUM_WREG * 5 + 6929 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 7 + 6930 2 + /* gfx_v11_0_ring_emit_vm_flush */ 6931 8 + 8 + 8 + /* gfx_v11_0_ring_emit_fence x3 for user fence, vm fence */ 6932 8 + /* gfx_v11_0_emit_mem_sync */ 6933 2, /* gfx_v11_0_ring_emit_cleaner_shader */ 6934 .emit_ib_size = 7, /* gfx_v11_0_ring_emit_ib_compute */ 6935 .emit_ib = gfx_v11_0_ring_emit_ib_compute, 6936 .emit_fence = gfx_v11_0_ring_emit_fence, 6937 .emit_pipeline_sync = gfx_v11_0_ring_emit_pipeline_sync, 6938 .emit_vm_flush = gfx_v11_0_ring_emit_vm_flush, 6939 .emit_gds_switch = gfx_v11_0_ring_emit_gds_switch, 6940 .emit_hdp_flush = gfx_v11_0_ring_emit_hdp_flush, 6941 .test_ring = gfx_v11_0_ring_test_ring, 6942 .test_ib = gfx_v11_0_ring_test_ib, 6943 .insert_nop = gfx_v11_ring_insert_nop, 6944 .pad_ib = amdgpu_ring_generic_pad_ib, 6945 .emit_wreg = gfx_v11_0_ring_emit_wreg, 6946 .emit_reg_wait = gfx_v11_0_ring_emit_reg_wait, 6947 .emit_reg_write_reg_wait = gfx_v11_0_ring_emit_reg_write_reg_wait, 6948 .soft_recovery = gfx_v11_0_ring_soft_recovery, 6949 .emit_mem_sync = gfx_v11_0_emit_mem_sync, 6950 .reset = gfx_v11_0_reset_kcq, 6951 .emit_cleaner_shader = gfx_v11_0_ring_emit_cleaner_shader, 6952 .begin_use = amdgpu_gfx_enforce_isolation_ring_begin_use, 6953 .end_use = amdgpu_gfx_enforce_isolation_ring_end_use, 6954 }; 6955 6956 static const struct amdgpu_ring_funcs gfx_v11_0_ring_funcs_kiq = { 6957 .type = AMDGPU_RING_TYPE_KIQ, 6958 .align_mask = 0xff, 6959 .nop = PACKET3(PACKET3_NOP, 0x3FFF), 6960 .support_64bit_ptrs = true, 6961 .get_rptr = gfx_v11_0_ring_get_rptr_compute, 6962 .get_wptr = gfx_v11_0_ring_get_wptr_compute, 6963 .set_wptr = gfx_v11_0_ring_set_wptr_compute, 6964 .emit_frame_size = 6965 20 + /* gfx_v11_0_ring_emit_gds_switch */ 6966 7 + /* gfx_v11_0_ring_emit_hdp_flush */ 6967 5 + /*hdp invalidate */ 6968 7 + /* gfx_v11_0_ring_emit_pipeline_sync */ 6969 SOC15_FLUSH_GPU_TLB_NUM_WREG * 5 + 6970 SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 7 + 6971 8 + 8 + 8, /* gfx_v11_0_ring_emit_fence_kiq x3 for user fence, vm fence */ 6972 .emit_ib_size = 7, /* gfx_v11_0_ring_emit_ib_compute */ 6973 .emit_ib = gfx_v11_0_ring_emit_ib_compute, 6974 .emit_fence = gfx_v11_0_ring_emit_fence_kiq, 6975 .test_ring = gfx_v11_0_ring_test_ring, 6976 .test_ib = gfx_v11_0_ring_test_ib, 6977 .insert_nop = amdgpu_ring_insert_nop, 6978 .pad_ib = amdgpu_ring_generic_pad_ib, 6979 .emit_rreg = gfx_v11_0_ring_emit_rreg, 6980 .emit_wreg = gfx_v11_0_ring_emit_wreg, 6981 .emit_reg_wait = gfx_v11_0_ring_emit_reg_wait, 6982 .emit_reg_write_reg_wait = gfx_v11_0_ring_emit_reg_write_reg_wait, 6983 }; 6984 6985 static void gfx_v11_0_set_ring_funcs(struct amdgpu_device *adev) 6986 { 6987 int i; 6988 6989 adev->gfx.kiq[0].ring.funcs = &gfx_v11_0_ring_funcs_kiq; 6990 6991 for (i = 0; i < adev->gfx.num_gfx_rings; i++) 6992 adev->gfx.gfx_ring[i].funcs = &gfx_v11_0_ring_funcs_gfx; 6993 6994 for (i = 0; i < adev->gfx.num_compute_rings; i++) 6995 adev->gfx.compute_ring[i].funcs = &gfx_v11_0_ring_funcs_compute; 6996 } 6997 6998 static const struct amdgpu_irq_src_funcs gfx_v11_0_eop_irq_funcs = { 6999 .set = gfx_v11_0_set_eop_interrupt_state, 7000 .process = gfx_v11_0_eop_irq, 7001 }; 7002 7003 static const struct amdgpu_irq_src_funcs gfx_v11_0_priv_reg_irq_funcs = { 7004 .set = gfx_v11_0_set_priv_reg_fault_state, 7005 .process = gfx_v11_0_priv_reg_irq, 7006 }; 7007 7008 static const struct amdgpu_irq_src_funcs gfx_v11_0_bad_op_irq_funcs = { 7009 .set = gfx_v11_0_set_bad_op_fault_state, 7010 .process = gfx_v11_0_bad_op_irq, 7011 }; 7012 7013 static const struct amdgpu_irq_src_funcs gfx_v11_0_priv_inst_irq_funcs = { 7014 .set = gfx_v11_0_set_priv_inst_fault_state, 7015 .process = gfx_v11_0_priv_inst_irq, 7016 }; 7017 7018 static const struct amdgpu_irq_src_funcs gfx_v11_0_rlc_gc_fed_irq_funcs = { 7019 .process = gfx_v11_0_rlc_gc_fed_irq, 7020 }; 7021 7022 static void gfx_v11_0_set_irq_funcs(struct amdgpu_device *adev) 7023 { 7024 adev->gfx.eop_irq.num_types = AMDGPU_CP_IRQ_LAST; 7025 adev->gfx.eop_irq.funcs = &gfx_v11_0_eop_irq_funcs; 7026 7027 adev->gfx.priv_reg_irq.num_types = 1; 7028 adev->gfx.priv_reg_irq.funcs = &gfx_v11_0_priv_reg_irq_funcs; 7029 7030 adev->gfx.bad_op_irq.num_types = 1; 7031 adev->gfx.bad_op_irq.funcs = &gfx_v11_0_bad_op_irq_funcs; 7032 7033 adev->gfx.priv_inst_irq.num_types = 1; 7034 adev->gfx.priv_inst_irq.funcs = &gfx_v11_0_priv_inst_irq_funcs; 7035 7036 adev->gfx.rlc_gc_fed_irq.num_types = 1; /* 0x80 FED error */ 7037 adev->gfx.rlc_gc_fed_irq.funcs = &gfx_v11_0_rlc_gc_fed_irq_funcs; 7038 7039 } 7040 7041 static void gfx_v11_0_set_imu_funcs(struct amdgpu_device *adev) 7042 { 7043 if (adev->flags & AMD_IS_APU) 7044 adev->gfx.imu.mode = MISSION_MODE; 7045 else 7046 adev->gfx.imu.mode = DEBUG_MODE; 7047 7048 adev->gfx.imu.funcs = &gfx_v11_0_imu_funcs; 7049 } 7050 7051 static void gfx_v11_0_set_rlc_funcs(struct amdgpu_device *adev) 7052 { 7053 adev->gfx.rlc.funcs = &gfx_v11_0_rlc_funcs; 7054 } 7055 7056 static void gfx_v11_0_set_gds_init(struct amdgpu_device *adev) 7057 { 7058 unsigned total_cu = adev->gfx.config.max_cu_per_sh * 7059 adev->gfx.config.max_sh_per_se * 7060 adev->gfx.config.max_shader_engines; 7061 7062 adev->gds.gds_size = 0x1000; 7063 adev->gds.gds_compute_max_wave_id = total_cu * 32 - 1; 7064 adev->gds.gws_size = 64; 7065 adev->gds.oa_size = 16; 7066 } 7067 7068 static void gfx_v11_0_set_mqd_funcs(struct amdgpu_device *adev) 7069 { 7070 /* set gfx eng mqd */ 7071 adev->mqds[AMDGPU_HW_IP_GFX].mqd_size = 7072 sizeof(struct v11_gfx_mqd); 7073 adev->mqds[AMDGPU_HW_IP_GFX].init_mqd = 7074 gfx_v11_0_gfx_mqd_init; 7075 /* set compute eng mqd */ 7076 adev->mqds[AMDGPU_HW_IP_COMPUTE].mqd_size = 7077 sizeof(struct v11_compute_mqd); 7078 adev->mqds[AMDGPU_HW_IP_COMPUTE].init_mqd = 7079 gfx_v11_0_compute_mqd_init; 7080 } 7081 7082 static void gfx_v11_0_set_user_wgp_inactive_bitmap_per_sh(struct amdgpu_device *adev, 7083 u32 bitmap) 7084 { 7085 u32 data; 7086 7087 if (!bitmap) 7088 return; 7089 7090 data = bitmap << GC_USER_SHADER_ARRAY_CONFIG__INACTIVE_WGPS__SHIFT; 7091 data &= GC_USER_SHADER_ARRAY_CONFIG__INACTIVE_WGPS_MASK; 7092 7093 WREG32_SOC15(GC, 0, regGC_USER_SHADER_ARRAY_CONFIG, data); 7094 } 7095 7096 static u32 gfx_v11_0_get_wgp_active_bitmap_per_sh(struct amdgpu_device *adev) 7097 { 7098 u32 data, wgp_bitmask; 7099 data = RREG32_SOC15(GC, 0, regCC_GC_SHADER_ARRAY_CONFIG); 7100 data |= RREG32_SOC15(GC, 0, regGC_USER_SHADER_ARRAY_CONFIG); 7101 7102 data &= CC_GC_SHADER_ARRAY_CONFIG__INACTIVE_WGPS_MASK; 7103 data >>= CC_GC_SHADER_ARRAY_CONFIG__INACTIVE_WGPS__SHIFT; 7104 7105 wgp_bitmask = 7106 amdgpu_gfx_create_bitmask(adev->gfx.config.max_cu_per_sh >> 1); 7107 7108 return (~data) & wgp_bitmask; 7109 } 7110 7111 static u32 gfx_v11_0_get_cu_active_bitmap_per_sh(struct amdgpu_device *adev) 7112 { 7113 u32 wgp_idx, wgp_active_bitmap; 7114 u32 cu_bitmap_per_wgp, cu_active_bitmap; 7115 7116 wgp_active_bitmap = gfx_v11_0_get_wgp_active_bitmap_per_sh(adev); 7117 cu_active_bitmap = 0; 7118 7119 for (wgp_idx = 0; wgp_idx < 16; wgp_idx++) { 7120 /* if there is one WGP enabled, it means 2 CUs will be enabled */ 7121 cu_bitmap_per_wgp = 3 << (2 * wgp_idx); 7122 if (wgp_active_bitmap & (1 << wgp_idx)) 7123 cu_active_bitmap |= cu_bitmap_per_wgp; 7124 } 7125 7126 return cu_active_bitmap; 7127 } 7128 7129 static int gfx_v11_0_get_cu_info(struct amdgpu_device *adev, 7130 struct amdgpu_cu_info *cu_info) 7131 { 7132 int i, j, k, counter, active_cu_number = 0; 7133 u32 mask, bitmap; 7134 unsigned disable_masks[8 * 2]; 7135 7136 if (!adev || !cu_info) 7137 return -EINVAL; 7138 7139 amdgpu_gfx_parse_disable_cu(disable_masks, 8, 2); 7140 7141 mutex_lock(&adev->grbm_idx_mutex); 7142 for (i = 0; i < adev->gfx.config.max_shader_engines; i++) { 7143 for (j = 0; j < adev->gfx.config.max_sh_per_se; j++) { 7144 bitmap = i * adev->gfx.config.max_sh_per_se + j; 7145 if (!((gfx_v11_0_get_sa_active_bitmap(adev) >> bitmap) & 1)) 7146 continue; 7147 mask = 1; 7148 counter = 0; 7149 gfx_v11_0_select_se_sh(adev, i, j, 0xffffffff, 0); 7150 if (i < 8 && j < 2) 7151 gfx_v11_0_set_user_wgp_inactive_bitmap_per_sh( 7152 adev, disable_masks[i * 2 + j]); 7153 bitmap = gfx_v11_0_get_cu_active_bitmap_per_sh(adev); 7154 7155 /** 7156 * GFX11 could support more than 4 SEs, while the bitmap 7157 * in cu_info struct is 4x4 and ioctl interface struct 7158 * drm_amdgpu_info_device should keep stable. 7159 * So we use last two columns of bitmap to store cu mask for 7160 * SEs 4 to 7, the layout of the bitmap is as below: 7161 * SE0: {SH0,SH1} --> {bitmap[0][0], bitmap[0][1]} 7162 * SE1: {SH0,SH1} --> {bitmap[1][0], bitmap[1][1]} 7163 * SE2: {SH0,SH1} --> {bitmap[2][0], bitmap[2][1]} 7164 * SE3: {SH0,SH1} --> {bitmap[3][0], bitmap[3][1]} 7165 * SE4: {SH0,SH1} --> {bitmap[0][2], bitmap[0][3]} 7166 * SE5: {SH0,SH1} --> {bitmap[1][2], bitmap[1][3]} 7167 * SE6: {SH0,SH1} --> {bitmap[2][2], bitmap[2][3]} 7168 * SE7: {SH0,SH1} --> {bitmap[3][2], bitmap[3][3]} 7169 */ 7170 cu_info->bitmap[0][i % 4][j + (i / 4) * 2] = bitmap; 7171 7172 for (k = 0; k < adev->gfx.config.max_cu_per_sh; k++) { 7173 if (bitmap & mask) 7174 counter++; 7175 7176 mask <<= 1; 7177 } 7178 active_cu_number += counter; 7179 } 7180 } 7181 gfx_v11_0_select_se_sh(adev, 0xffffffff, 0xffffffff, 0xffffffff, 0); 7182 mutex_unlock(&adev->grbm_idx_mutex); 7183 7184 cu_info->number = active_cu_number; 7185 cu_info->simd_per_cu = NUM_SIMD_PER_CU; 7186 7187 return 0; 7188 } 7189 7190 const struct amdgpu_ip_block_version gfx_v11_0_ip_block = 7191 { 7192 .type = AMD_IP_BLOCK_TYPE_GFX, 7193 .major = 11, 7194 .minor = 0, 7195 .rev = 0, 7196 .funcs = &gfx_v11_0_ip_funcs, 7197 }; 7198