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 24 #include "aldebaran.h" 25 #include "amdgpu_reset.h" 26 #include "amdgpu_amdkfd.h" 27 #include "amdgpu_dpm.h" 28 #include "amdgpu_job.h" 29 #include "amdgpu_ring.h" 30 #include "amdgpu_ras.h" 31 #include "amdgpu_psp.h" 32 #include "amdgpu_xgmi.h" 33 34 static struct amdgpu_reset_handler * 35 aldebaran_get_reset_handler(struct amdgpu_reset_control *reset_ctl, 36 struct amdgpu_reset_context *reset_context) 37 { 38 struct amdgpu_reset_handler *handler; 39 struct amdgpu_device *adev = (struct amdgpu_device *)reset_ctl->handle; 40 41 if (reset_context->method != AMD_RESET_METHOD_NONE) { 42 dev_dbg(adev->dev, "Getting reset handler for method %d\n", 43 reset_context->method); 44 list_for_each_entry(handler, &reset_ctl->reset_handlers, 45 handler_list) { 46 if (handler->reset_method == reset_context->method) 47 return handler; 48 } 49 } 50 51 if (adev->gmc.xgmi.connected_to_cpu) { 52 list_for_each_entry(handler, &reset_ctl->reset_handlers, 53 handler_list) { 54 if (handler->reset_method == AMD_RESET_METHOD_MODE2) { 55 reset_context->method = AMD_RESET_METHOD_MODE2; 56 return handler; 57 } 58 } 59 } 60 61 dev_dbg(adev->dev, "Reset handler not found!\n"); 62 63 return NULL; 64 } 65 66 static int aldebaran_mode2_suspend_ip(struct amdgpu_device *adev) 67 { 68 int r, i; 69 70 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_UNGATE); 71 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_UNGATE); 72 73 for (i = adev->num_ip_blocks - 1; i >= 0; i--) { 74 if (!(adev->ip_blocks[i].version->type == 75 AMD_IP_BLOCK_TYPE_GFX || 76 adev->ip_blocks[i].version->type == 77 AMD_IP_BLOCK_TYPE_SDMA)) 78 continue; 79 80 r = adev->ip_blocks[i].version->funcs->suspend(adev); 81 82 if (r) { 83 dev_err(adev->dev, 84 "suspend of IP block <%s> failed %d\n", 85 adev->ip_blocks[i].version->funcs->name, r); 86 return r; 87 } 88 89 adev->ip_blocks[i].status.hw = false; 90 } 91 92 return r; 93 } 94 95 static int 96 aldebaran_mode2_prepare_hwcontext(struct amdgpu_reset_control *reset_ctl, 97 struct amdgpu_reset_context *reset_context) 98 { 99 int r = 0; 100 struct amdgpu_device *adev = (struct amdgpu_device *)reset_ctl->handle; 101 102 dev_dbg(adev->dev, "Aldebaran prepare hw context\n"); 103 /* Don't suspend on bare metal if we are not going to HW reset the ASIC */ 104 if (!amdgpu_sriov_vf(adev)) 105 r = aldebaran_mode2_suspend_ip(adev); 106 107 return r; 108 } 109 110 static void aldebaran_async_reset(struct work_struct *work) 111 { 112 struct amdgpu_reset_handler *handler; 113 struct amdgpu_reset_control *reset_ctl = 114 container_of(work, struct amdgpu_reset_control, reset_work); 115 struct amdgpu_device *adev = (struct amdgpu_device *)reset_ctl->handle; 116 117 list_for_each_entry(handler, &reset_ctl->reset_handlers, 118 handler_list) { 119 if (handler->reset_method == reset_ctl->active_reset) { 120 dev_dbg(adev->dev, "Resetting device\n"); 121 handler->do_reset(adev); 122 break; 123 } 124 } 125 } 126 127 static int aldebaran_mode2_reset(struct amdgpu_device *adev) 128 { 129 /* disable BM */ 130 pci_clear_master(adev->pdev); 131 adev->asic_reset_res = amdgpu_dpm_mode2_reset(adev); 132 return adev->asic_reset_res; 133 } 134 135 static int 136 aldebaran_mode2_perform_reset(struct amdgpu_reset_control *reset_ctl, 137 struct amdgpu_reset_context *reset_context) 138 { 139 struct amdgpu_device *tmp_adev = NULL; 140 struct amdgpu_device *adev = (struct amdgpu_device *)reset_ctl->handle; 141 int r = 0; 142 143 dev_dbg(adev->dev, "aldebaran perform hw reset\n"); 144 if (reset_context->hive == NULL) { 145 /* Wrong context, return error */ 146 return -EINVAL; 147 } 148 149 list_for_each_entry(tmp_adev, &reset_context->hive->device_list, 150 gmc.xgmi.head) { 151 mutex_lock(&tmp_adev->reset_cntl->reset_lock); 152 tmp_adev->reset_cntl->active_reset = AMD_RESET_METHOD_MODE2; 153 } 154 /* 155 * Mode2 reset doesn't need any sync between nodes in XGMI hive, instead launch 156 * them together so that they can be completed asynchronously on multiple nodes 157 */ 158 list_for_each_entry(tmp_adev, &reset_context->hive->device_list, 159 gmc.xgmi.head) { 160 /* For XGMI run all resets in parallel to speed up the process */ 161 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) { 162 if (!queue_work(system_unbound_wq, 163 &tmp_adev->reset_cntl->reset_work)) 164 r = -EALREADY; 165 } else 166 r = aldebaran_mode2_reset(tmp_adev); 167 if (r) { 168 dev_err(tmp_adev->dev, 169 "ASIC reset failed with error, %d for drm dev, %s", 170 r, adev_to_drm(tmp_adev)->unique); 171 break; 172 } 173 } 174 175 /* For XGMI wait for all resets to complete before proceed */ 176 if (!r) { 177 list_for_each_entry(tmp_adev, 178 &reset_context->hive->device_list, 179 gmc.xgmi.head) { 180 if (tmp_adev->gmc.xgmi.num_physical_nodes > 1) { 181 flush_work(&tmp_adev->reset_cntl->reset_work); 182 r = tmp_adev->asic_reset_res; 183 if (r) 184 break; 185 } 186 } 187 } 188 189 list_for_each_entry(tmp_adev, &reset_context->hive->device_list, 190 gmc.xgmi.head) { 191 mutex_unlock(&tmp_adev->reset_cntl->reset_lock); 192 tmp_adev->reset_cntl->active_reset = AMD_RESET_METHOD_NONE; 193 } 194 195 return r; 196 } 197 198 static int aldebaran_mode2_restore_ip(struct amdgpu_device *adev) 199 { 200 struct amdgpu_firmware_info *ucode_list[AMDGPU_UCODE_ID_MAXIMUM]; 201 struct amdgpu_firmware_info *ucode; 202 struct amdgpu_ip_block *cmn_block; 203 int ucode_count = 0; 204 int i, r; 205 206 dev_dbg(adev->dev, "Reloading ucodes after reset\n"); 207 for (i = 0; i < adev->firmware.max_ucodes; i++) { 208 ucode = &adev->firmware.ucode[i]; 209 if (!ucode->fw) 210 continue; 211 switch (ucode->ucode_id) { 212 case AMDGPU_UCODE_ID_SDMA0: 213 case AMDGPU_UCODE_ID_SDMA1: 214 case AMDGPU_UCODE_ID_SDMA2: 215 case AMDGPU_UCODE_ID_SDMA3: 216 case AMDGPU_UCODE_ID_SDMA4: 217 case AMDGPU_UCODE_ID_SDMA5: 218 case AMDGPU_UCODE_ID_SDMA6: 219 case AMDGPU_UCODE_ID_SDMA7: 220 case AMDGPU_UCODE_ID_CP_MEC1: 221 case AMDGPU_UCODE_ID_CP_MEC1_JT: 222 case AMDGPU_UCODE_ID_RLC_RESTORE_LIST_CNTL: 223 case AMDGPU_UCODE_ID_RLC_RESTORE_LIST_GPM_MEM: 224 case AMDGPU_UCODE_ID_RLC_RESTORE_LIST_SRM_MEM: 225 case AMDGPU_UCODE_ID_RLC_G: 226 ucode_list[ucode_count++] = ucode; 227 break; 228 default: 229 break; 230 } 231 } 232 233 /* Reinit NBIF block */ 234 cmn_block = 235 amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_COMMON); 236 if (unlikely(!cmn_block)) { 237 dev_err(adev->dev, "Failed to get BIF handle\n"); 238 return -EINVAL; 239 } 240 r = cmn_block->version->funcs->resume(adev); 241 if (r) 242 return r; 243 244 /* Reinit GFXHUB */ 245 adev->gfxhub.funcs->init(adev); 246 r = adev->gfxhub.funcs->gart_enable(adev); 247 if (r) { 248 dev_err(adev->dev, "GFXHUB gart reenable failed after reset\n"); 249 return r; 250 } 251 252 /* Reload GFX firmware */ 253 r = psp_load_fw_list(&adev->psp, ucode_list, ucode_count); 254 if (r) { 255 dev_err(adev->dev, "GFX ucode load failed after reset\n"); 256 return r; 257 } 258 259 /* Resume RLC, FW needs RLC alive to complete reset process */ 260 adev->gfx.rlc.funcs->resume(adev); 261 262 /* Wait for FW reset event complete */ 263 r = amdgpu_dpm_wait_for_event(adev, SMU_EVENT_RESET_COMPLETE, 0); 264 if (r) { 265 dev_err(adev->dev, 266 "Failed to get response from firmware after reset\n"); 267 return r; 268 } 269 270 for (i = 0; i < adev->num_ip_blocks; i++) { 271 if (!(adev->ip_blocks[i].version->type == 272 AMD_IP_BLOCK_TYPE_GFX || 273 adev->ip_blocks[i].version->type == 274 AMD_IP_BLOCK_TYPE_SDMA)) 275 continue; 276 r = adev->ip_blocks[i].version->funcs->resume(adev); 277 if (r) { 278 dev_err(adev->dev, 279 "resume of IP block <%s> failed %d\n", 280 adev->ip_blocks[i].version->funcs->name, r); 281 return r; 282 } 283 284 adev->ip_blocks[i].status.hw = true; 285 } 286 287 for (i = 0; i < adev->num_ip_blocks; i++) { 288 if (!(adev->ip_blocks[i].version->type == 289 AMD_IP_BLOCK_TYPE_GFX || 290 adev->ip_blocks[i].version->type == 291 AMD_IP_BLOCK_TYPE_SDMA || 292 adev->ip_blocks[i].version->type == 293 AMD_IP_BLOCK_TYPE_COMMON)) 294 continue; 295 296 if (adev->ip_blocks[i].version->funcs->late_init) { 297 r = adev->ip_blocks[i].version->funcs->late_init( 298 (void *)adev); 299 if (r) { 300 dev_err(adev->dev, 301 "late_init of IP block <%s> failed %d after reset\n", 302 adev->ip_blocks[i].version->funcs->name, 303 r); 304 return r; 305 } 306 } 307 adev->ip_blocks[i].status.late_initialized = true; 308 } 309 310 amdgpu_ras_set_error_query_ready(adev, true); 311 312 amdgpu_device_set_cg_state(adev, AMD_CG_STATE_GATE); 313 amdgpu_device_set_pg_state(adev, AMD_PG_STATE_GATE); 314 315 return r; 316 } 317 318 static int 319 aldebaran_mode2_restore_hwcontext(struct amdgpu_reset_control *reset_ctl, 320 struct amdgpu_reset_context *reset_context) 321 { 322 int r; 323 struct amdgpu_device *tmp_adev = NULL; 324 325 if (reset_context->hive == NULL) { 326 /* Wrong context, return error */ 327 return -EINVAL; 328 } 329 330 list_for_each_entry(tmp_adev, &reset_context->hive->device_list, 331 gmc.xgmi.head) { 332 dev_info(tmp_adev->dev, 333 "GPU reset succeeded, trying to resume\n"); 334 r = aldebaran_mode2_restore_ip(tmp_adev); 335 if (r) 336 goto end; 337 338 /* 339 * Add this ASIC as tracked as reset was already 340 * complete successfully. 341 */ 342 amdgpu_register_gpu_instance(tmp_adev); 343 344 /* Resume RAS */ 345 amdgpu_ras_resume(tmp_adev); 346 347 /* Update PSP FW topology after reset */ 348 if (reset_context->hive && 349 tmp_adev->gmc.xgmi.num_physical_nodes > 1) 350 r = amdgpu_xgmi_update_topology(reset_context->hive, 351 tmp_adev); 352 353 if (!r) { 354 amdgpu_irq_gpu_reset_resume_helper(tmp_adev); 355 356 r = amdgpu_ib_ring_tests(tmp_adev); 357 if (r) { 358 dev_err(tmp_adev->dev, 359 "ib ring test failed (%d).\n", r); 360 r = -EAGAIN; 361 tmp_adev->asic_reset_res = r; 362 goto end; 363 } 364 } 365 } 366 367 end: 368 return r; 369 } 370 371 static struct amdgpu_reset_handler aldebaran_mode2_handler = { 372 .reset_method = AMD_RESET_METHOD_MODE2, 373 .prepare_env = NULL, 374 .prepare_hwcontext = aldebaran_mode2_prepare_hwcontext, 375 .perform_reset = aldebaran_mode2_perform_reset, 376 .restore_hwcontext = aldebaran_mode2_restore_hwcontext, 377 .restore_env = NULL, 378 .do_reset = aldebaran_mode2_reset, 379 }; 380 381 int aldebaran_reset_init(struct amdgpu_device *adev) 382 { 383 struct amdgpu_reset_control *reset_ctl; 384 385 reset_ctl = kzalloc(sizeof(*reset_ctl), GFP_KERNEL); 386 if (!reset_ctl) 387 return -ENOMEM; 388 389 reset_ctl->handle = adev; 390 reset_ctl->async_reset = aldebaran_async_reset; 391 reset_ctl->active_reset = AMD_RESET_METHOD_NONE; 392 reset_ctl->get_reset_handler = aldebaran_get_reset_handler; 393 394 INIT_LIST_HEAD(&reset_ctl->reset_handlers); 395 INIT_WORK(&reset_ctl->reset_work, reset_ctl->async_reset); 396 /* Only mode2 is handled through reset control now */ 397 amdgpu_reset_add_handler(reset_ctl, &aldebaran_mode2_handler); 398 399 adev->reset_cntl = reset_ctl; 400 401 return 0; 402 } 403 404 int aldebaran_reset_fini(struct amdgpu_device *adev) 405 { 406 kfree(adev->reset_cntl); 407 adev->reset_cntl = NULL; 408 return 0; 409 } 410