1 /* 2 * Copyright 2018 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 <linux/list.h> 25 #include "amdgpu.h" 26 #include "amdgpu_xgmi.h" 27 #include "amdgpu_ras.h" 28 #include "soc15.h" 29 #include "df/df_3_6_offset.h" 30 #include "xgmi/xgmi_4_0_0_smn.h" 31 #include "xgmi/xgmi_4_0_0_sh_mask.h" 32 #include "wafl/wafl2_4_0_0_smn.h" 33 #include "wafl/wafl2_4_0_0_sh_mask.h" 34 35 static DEFINE_MUTEX(xgmi_mutex); 36 37 #define AMDGPU_MAX_XGMI_DEVICE_PER_HIVE 4 38 39 static LIST_HEAD(xgmi_hive_list); 40 41 static const int xgmi_pcs_err_status_reg_vg20[] = { 42 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS, 43 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x100000, 44 }; 45 46 static const int wafl_pcs_err_status_reg_vg20[] = { 47 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, 48 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS + 0x100000, 49 }; 50 51 static const int xgmi_pcs_err_status_reg_arct[] = { 52 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS, 53 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x100000, 54 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x500000, 55 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x600000, 56 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x700000, 57 smnXGMI0_PCS_GOPX16_PCS_ERROR_STATUS + 0x800000, 58 }; 59 60 /* same as vg20*/ 61 static const int wafl_pcs_err_status_reg_arct[] = { 62 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, 63 smnPCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS + 0x100000, 64 }; 65 66 static const struct amdgpu_pcs_ras_field xgmi_pcs_ras_fields[] = { 67 {"XGMI PCS DataLossErr", 68 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataLossErr)}, 69 {"XGMI PCS TrainingErr", 70 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, TrainingErr)}, 71 {"XGMI PCS CRCErr", 72 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, CRCErr)}, 73 {"XGMI PCS BERExceededErr", 74 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, BERExceededErr)}, 75 {"XGMI PCS TxMetaDataErr", 76 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, TxMetaDataErr)}, 77 {"XGMI PCS ReplayBufParityErr", 78 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayBufParityErr)}, 79 {"XGMI PCS DataParityErr", 80 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataParityErr)}, 81 {"XGMI PCS ReplayFifoOverflowErr", 82 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayFifoOverflowErr)}, 83 {"XGMI PCS ReplayFifoUnderflowErr", 84 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReplayFifoUnderflowErr)}, 85 {"XGMI PCS ElasticFifoOverflowErr", 86 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ElasticFifoOverflowErr)}, 87 {"XGMI PCS DeskewErr", 88 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DeskewErr)}, 89 {"XGMI PCS DataStartupLimitErr", 90 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, DataStartupLimitErr)}, 91 {"XGMI PCS FCInitTimeoutErr", 92 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, FCInitTimeoutErr)}, 93 {"XGMI PCS RecoveryTimeoutErr", 94 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryTimeoutErr)}, 95 {"XGMI PCS ReadySerialTimeoutErr", 96 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReadySerialTimeoutErr)}, 97 {"XGMI PCS ReadySerialAttemptErr", 98 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, ReadySerialAttemptErr)}, 99 {"XGMI PCS RecoveryAttemptErr", 100 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryAttemptErr)}, 101 {"XGMI PCS RecoveryRelockAttemptErr", 102 SOC15_REG_FIELD(XGMI0_PCS_GOPX16_PCS_ERROR_STATUS, RecoveryRelockAttemptErr)}, 103 }; 104 105 static const struct amdgpu_pcs_ras_field wafl_pcs_ras_fields[] = { 106 {"WAFL PCS DataLossErr", 107 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataLossErr)}, 108 {"WAFL PCS TrainingErr", 109 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, TrainingErr)}, 110 {"WAFL PCS CRCErr", 111 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, CRCErr)}, 112 {"WAFL PCS BERExceededErr", 113 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, BERExceededErr)}, 114 {"WAFL PCS TxMetaDataErr", 115 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, TxMetaDataErr)}, 116 {"WAFL PCS ReplayBufParityErr", 117 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayBufParityErr)}, 118 {"WAFL PCS DataParityErr", 119 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataParityErr)}, 120 {"WAFL PCS ReplayFifoOverflowErr", 121 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayFifoOverflowErr)}, 122 {"WAFL PCS ReplayFifoUnderflowErr", 123 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReplayFifoUnderflowErr)}, 124 {"WAFL PCS ElasticFifoOverflowErr", 125 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ElasticFifoOverflowErr)}, 126 {"WAFL PCS DeskewErr", 127 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DeskewErr)}, 128 {"WAFL PCS DataStartupLimitErr", 129 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, DataStartupLimitErr)}, 130 {"WAFL PCS FCInitTimeoutErr", 131 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, FCInitTimeoutErr)}, 132 {"WAFL PCS RecoveryTimeoutErr", 133 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryTimeoutErr)}, 134 {"WAFL PCS ReadySerialTimeoutErr", 135 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReadySerialTimeoutErr)}, 136 {"WAFL PCS ReadySerialAttemptErr", 137 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, ReadySerialAttemptErr)}, 138 {"WAFL PCS RecoveryAttemptErr", 139 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryAttemptErr)}, 140 {"WAFL PCS RecoveryRelockAttemptErr", 141 SOC15_REG_FIELD(PCS_GOPX1_0_PCS_GOPX1_PCS_ERROR_STATUS, RecoveryRelockAttemptErr)}, 142 }; 143 144 /** 145 * DOC: AMDGPU XGMI Support 146 * 147 * XGMI is a high speed interconnect that joins multiple GPU cards 148 * into a homogeneous memory space that is organized by a collective 149 * hive ID and individual node IDs, both of which are 64-bit numbers. 150 * 151 * The file xgmi_device_id contains the unique per GPU device ID and 152 * is stored in the /sys/class/drm/card${cardno}/device/ directory. 153 * 154 * Inside the device directory a sub-directory 'xgmi_hive_info' is 155 * created which contains the hive ID and the list of nodes. 156 * 157 * The hive ID is stored in: 158 * /sys/class/drm/card${cardno}/device/xgmi_hive_info/xgmi_hive_id 159 * 160 * The node information is stored in numbered directories: 161 * /sys/class/drm/card${cardno}/device/xgmi_hive_info/node${nodeno}/xgmi_device_id 162 * 163 * Each device has their own xgmi_hive_info direction with a mirror 164 * set of node sub-directories. 165 * 166 * The XGMI memory space is built by contiguously adding the power of 167 * two padded VRAM space from each node to each other. 168 * 169 */ 170 171 static struct attribute amdgpu_xgmi_hive_id = { 172 .name = "xgmi_hive_id", 173 .mode = S_IRUGO 174 }; 175 176 static struct attribute *amdgpu_xgmi_hive_attrs[] = { 177 &amdgpu_xgmi_hive_id, 178 NULL 179 }; 180 181 static ssize_t amdgpu_xgmi_show_attrs(struct kobject *kobj, 182 struct attribute *attr, char *buf) 183 { 184 struct amdgpu_hive_info *hive = container_of( 185 kobj, struct amdgpu_hive_info, kobj); 186 187 if (attr == &amdgpu_xgmi_hive_id) 188 return snprintf(buf, PAGE_SIZE, "%llu\n", hive->hive_id); 189 190 return 0; 191 } 192 193 static void amdgpu_xgmi_hive_release(struct kobject *kobj) 194 { 195 struct amdgpu_hive_info *hive = container_of( 196 kobj, struct amdgpu_hive_info, kobj); 197 198 mutex_destroy(&hive->hive_lock); 199 kfree(hive); 200 } 201 202 static const struct sysfs_ops amdgpu_xgmi_hive_ops = { 203 .show = amdgpu_xgmi_show_attrs, 204 }; 205 206 struct kobj_type amdgpu_xgmi_hive_type = { 207 .release = amdgpu_xgmi_hive_release, 208 .sysfs_ops = &amdgpu_xgmi_hive_ops, 209 .default_attrs = amdgpu_xgmi_hive_attrs, 210 }; 211 212 static ssize_t amdgpu_xgmi_show_device_id(struct device *dev, 213 struct device_attribute *attr, 214 char *buf) 215 { 216 struct drm_device *ddev = dev_get_drvdata(dev); 217 struct amdgpu_device *adev = drm_to_adev(ddev); 218 219 return sysfs_emit(buf, "%llu\n", adev->gmc.xgmi.node_id); 220 221 } 222 223 #define AMDGPU_XGMI_SET_FICAA(o) ((o) | 0x456801) 224 static ssize_t amdgpu_xgmi_show_error(struct device *dev, 225 struct device_attribute *attr, 226 char *buf) 227 { 228 struct drm_device *ddev = dev_get_drvdata(dev); 229 struct amdgpu_device *adev = drm_to_adev(ddev); 230 uint32_t ficaa_pie_ctl_in, ficaa_pie_status_in; 231 uint64_t fica_out; 232 unsigned int error_count = 0; 233 234 ficaa_pie_ctl_in = AMDGPU_XGMI_SET_FICAA(0x200); 235 ficaa_pie_status_in = AMDGPU_XGMI_SET_FICAA(0x208); 236 237 fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_ctl_in); 238 if (fica_out != 0x1f) 239 pr_err("xGMI error counters not enabled!\n"); 240 241 fica_out = adev->df.funcs->get_fica(adev, ficaa_pie_status_in); 242 243 if ((fica_out & 0xffff) == 2) 244 error_count = ((fica_out >> 62) & 0x1) + (fica_out >> 63); 245 246 adev->df.funcs->set_fica(adev, ficaa_pie_status_in, 0, 0); 247 248 return sysfs_emit(buf, "%u\n", error_count); 249 } 250 251 252 static DEVICE_ATTR(xgmi_device_id, S_IRUGO, amdgpu_xgmi_show_device_id, NULL); 253 static DEVICE_ATTR(xgmi_error, S_IRUGO, amdgpu_xgmi_show_error, NULL); 254 255 static int amdgpu_xgmi_sysfs_add_dev_info(struct amdgpu_device *adev, 256 struct amdgpu_hive_info *hive) 257 { 258 int ret = 0; 259 char node[10] = { 0 }; 260 261 /* Create xgmi device id file */ 262 ret = device_create_file(adev->dev, &dev_attr_xgmi_device_id); 263 if (ret) { 264 dev_err(adev->dev, "XGMI: Failed to create device file xgmi_device_id\n"); 265 return ret; 266 } 267 268 /* Create xgmi error file */ 269 ret = device_create_file(adev->dev, &dev_attr_xgmi_error); 270 if (ret) 271 pr_err("failed to create xgmi_error\n"); 272 273 274 /* Create sysfs link to hive info folder on the first device */ 275 if (hive->kobj.parent != (&adev->dev->kobj)) { 276 ret = sysfs_create_link(&adev->dev->kobj, &hive->kobj, 277 "xgmi_hive_info"); 278 if (ret) { 279 dev_err(adev->dev, "XGMI: Failed to create link to hive info"); 280 goto remove_file; 281 } 282 } 283 284 sprintf(node, "node%d", atomic_read(&hive->number_devices)); 285 /* Create sysfs link form the hive folder to yourself */ 286 ret = sysfs_create_link(&hive->kobj, &adev->dev->kobj, node); 287 if (ret) { 288 dev_err(adev->dev, "XGMI: Failed to create link from hive info"); 289 goto remove_link; 290 } 291 292 goto success; 293 294 295 remove_link: 296 sysfs_remove_link(&adev->dev->kobj, adev_to_drm(adev)->unique); 297 298 remove_file: 299 device_remove_file(adev->dev, &dev_attr_xgmi_device_id); 300 301 success: 302 return ret; 303 } 304 305 static void amdgpu_xgmi_sysfs_rem_dev_info(struct amdgpu_device *adev, 306 struct amdgpu_hive_info *hive) 307 { 308 char node[10]; 309 memset(node, 0, sizeof(node)); 310 311 device_remove_file(adev->dev, &dev_attr_xgmi_device_id); 312 device_remove_file(adev->dev, &dev_attr_xgmi_error); 313 314 if (hive->kobj.parent != (&adev->dev->kobj)) 315 sysfs_remove_link(&adev->dev->kobj,"xgmi_hive_info"); 316 317 sprintf(node, "node%d", atomic_read(&hive->number_devices)); 318 sysfs_remove_link(&hive->kobj, node); 319 320 } 321 322 323 324 struct amdgpu_hive_info *amdgpu_get_xgmi_hive(struct amdgpu_device *adev) 325 { 326 struct amdgpu_hive_info *hive = NULL; 327 int ret; 328 329 if (!adev->gmc.xgmi.hive_id) 330 return NULL; 331 332 if (adev->hive) { 333 kobject_get(&adev->hive->kobj); 334 return adev->hive; 335 } 336 337 mutex_lock(&xgmi_mutex); 338 339 list_for_each_entry(hive, &xgmi_hive_list, node) { 340 if (hive->hive_id == adev->gmc.xgmi.hive_id) 341 goto pro_end; 342 } 343 344 hive = kzalloc(sizeof(*hive), GFP_KERNEL); 345 if (!hive) { 346 dev_err(adev->dev, "XGMI: allocation failed\n"); 347 hive = NULL; 348 goto pro_end; 349 } 350 351 /* initialize new hive if not exist */ 352 ret = kobject_init_and_add(&hive->kobj, 353 &amdgpu_xgmi_hive_type, 354 &adev->dev->kobj, 355 "%s", "xgmi_hive_info"); 356 if (ret) { 357 dev_err(adev->dev, "XGMI: failed initializing kobject for xgmi hive\n"); 358 kfree(hive); 359 hive = NULL; 360 goto pro_end; 361 } 362 363 hive->hive_id = adev->gmc.xgmi.hive_id; 364 INIT_LIST_HEAD(&hive->device_list); 365 INIT_LIST_HEAD(&hive->node); 366 mutex_init(&hive->hive_lock); 367 atomic_set(&hive->in_reset, 0); 368 atomic_set(&hive->number_devices, 0); 369 task_barrier_init(&hive->tb); 370 hive->pstate = AMDGPU_XGMI_PSTATE_UNKNOWN; 371 hive->hi_req_gpu = NULL; 372 /* 373 * hive pstate on boot is high in vega20 so we have to go to low 374 * pstate on after boot. 375 */ 376 hive->hi_req_count = AMDGPU_MAX_XGMI_DEVICE_PER_HIVE; 377 list_add_tail(&hive->node, &xgmi_hive_list); 378 379 pro_end: 380 if (hive) 381 kobject_get(&hive->kobj); 382 mutex_unlock(&xgmi_mutex); 383 return hive; 384 } 385 386 void amdgpu_put_xgmi_hive(struct amdgpu_hive_info *hive) 387 { 388 if (hive) 389 kobject_put(&hive->kobj); 390 } 391 392 int amdgpu_xgmi_set_pstate(struct amdgpu_device *adev, int pstate) 393 { 394 int ret = 0; 395 struct amdgpu_hive_info *hive; 396 struct amdgpu_device *request_adev; 397 bool is_hi_req = pstate == AMDGPU_XGMI_PSTATE_MAX_VEGA20; 398 bool init_low; 399 400 hive = amdgpu_get_xgmi_hive(adev); 401 if (!hive) 402 return 0; 403 404 request_adev = hive->hi_req_gpu ? hive->hi_req_gpu : adev; 405 init_low = hive->pstate == AMDGPU_XGMI_PSTATE_UNKNOWN; 406 amdgpu_put_xgmi_hive(hive); 407 /* fw bug so temporarily disable pstate switching */ 408 return 0; 409 410 if (!hive || adev->asic_type != CHIP_VEGA20) 411 return 0; 412 413 mutex_lock(&hive->hive_lock); 414 415 if (is_hi_req) 416 hive->hi_req_count++; 417 else 418 hive->hi_req_count--; 419 420 /* 421 * Vega20 only needs single peer to request pstate high for the hive to 422 * go high but all peers must request pstate low for the hive to go low 423 */ 424 if (hive->pstate == pstate || 425 (!is_hi_req && hive->hi_req_count && !init_low)) 426 goto out; 427 428 dev_dbg(request_adev->dev, "Set xgmi pstate %d.\n", pstate); 429 430 ret = amdgpu_dpm_set_xgmi_pstate(request_adev, pstate); 431 if (ret) { 432 dev_err(request_adev->dev, 433 "XGMI: Set pstate failure on device %llx, hive %llx, ret %d", 434 request_adev->gmc.xgmi.node_id, 435 request_adev->gmc.xgmi.hive_id, ret); 436 goto out; 437 } 438 439 if (init_low) 440 hive->pstate = hive->hi_req_count ? 441 hive->pstate : AMDGPU_XGMI_PSTATE_MIN; 442 else { 443 hive->pstate = pstate; 444 hive->hi_req_gpu = pstate != AMDGPU_XGMI_PSTATE_MIN ? 445 adev : NULL; 446 } 447 out: 448 mutex_unlock(&hive->hive_lock); 449 return ret; 450 } 451 452 int amdgpu_xgmi_update_topology(struct amdgpu_hive_info *hive, struct amdgpu_device *adev) 453 { 454 int ret; 455 456 /* Each psp need to set the latest topology */ 457 ret = psp_xgmi_set_topology_info(&adev->psp, 458 atomic_read(&hive->number_devices), 459 &adev->psp.xgmi_context.top_info); 460 if (ret) 461 dev_err(adev->dev, 462 "XGMI: Set topology failure on device %llx, hive %llx, ret %d", 463 adev->gmc.xgmi.node_id, 464 adev->gmc.xgmi.hive_id, ret); 465 466 return ret; 467 } 468 469 470 /* 471 * NOTE psp_xgmi_node_info.num_hops layout is as follows: 472 * num_hops[7:6] = link type (0 = xGMI2, 1 = xGMI3, 2/3 = reserved) 473 * num_hops[5:3] = reserved 474 * num_hops[2:0] = number of hops 475 */ 476 int amdgpu_xgmi_get_hops_count(struct amdgpu_device *adev, 477 struct amdgpu_device *peer_adev) 478 { 479 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 480 uint8_t num_hops_mask = 0x7; 481 int i; 482 483 for (i = 0 ; i < top->num_nodes; ++i) 484 if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id) 485 return top->nodes[i].num_hops & num_hops_mask; 486 return -EINVAL; 487 } 488 489 int amdgpu_xgmi_get_num_links(struct amdgpu_device *adev, 490 struct amdgpu_device *peer_adev) 491 { 492 struct psp_xgmi_topology_info *top = &adev->psp.xgmi_context.top_info; 493 int i; 494 495 for (i = 0 ; i < top->num_nodes; ++i) 496 if (top->nodes[i].node_id == peer_adev->gmc.xgmi.node_id) 497 return top->nodes[i].num_links; 498 return -EINVAL; 499 } 500 501 int amdgpu_xgmi_add_device(struct amdgpu_device *adev) 502 { 503 struct psp_xgmi_topology_info *top_info; 504 struct amdgpu_hive_info *hive; 505 struct amdgpu_xgmi *entry; 506 struct amdgpu_device *tmp_adev = NULL; 507 508 int count = 0, ret = 0; 509 510 if (!adev->gmc.xgmi.supported) 511 return 0; 512 513 if (!adev->gmc.xgmi.pending_reset && 514 amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) { 515 ret = psp_xgmi_initialize(&adev->psp); 516 if (ret) { 517 dev_err(adev->dev, 518 "XGMI: Failed to initialize xgmi session\n"); 519 return ret; 520 } 521 522 ret = psp_xgmi_get_hive_id(&adev->psp, &adev->gmc.xgmi.hive_id); 523 if (ret) { 524 dev_err(adev->dev, 525 "XGMI: Failed to get hive id\n"); 526 return ret; 527 } 528 529 ret = psp_xgmi_get_node_id(&adev->psp, &adev->gmc.xgmi.node_id); 530 if (ret) { 531 dev_err(adev->dev, 532 "XGMI: Failed to get node id\n"); 533 return ret; 534 } 535 } else { 536 adev->gmc.xgmi.hive_id = 16; 537 adev->gmc.xgmi.node_id = adev->gmc.xgmi.physical_node_id + 16; 538 } 539 540 hive = amdgpu_get_xgmi_hive(adev); 541 if (!hive) { 542 ret = -EINVAL; 543 dev_err(adev->dev, 544 "XGMI: node 0x%llx, can not match hive 0x%llx in the hive list.\n", 545 adev->gmc.xgmi.node_id, adev->gmc.xgmi.hive_id); 546 goto exit; 547 } 548 mutex_lock(&hive->hive_lock); 549 550 top_info = &adev->psp.xgmi_context.top_info; 551 552 list_add_tail(&adev->gmc.xgmi.head, &hive->device_list); 553 list_for_each_entry(entry, &hive->device_list, head) 554 top_info->nodes[count++].node_id = entry->node_id; 555 top_info->num_nodes = count; 556 atomic_set(&hive->number_devices, count); 557 558 task_barrier_add_task(&hive->tb); 559 560 if (!adev->gmc.xgmi.pending_reset && 561 amdgpu_device_ip_get_ip_block(adev, AMD_IP_BLOCK_TYPE_PSP)) { 562 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 563 /* update node list for other device in the hive */ 564 if (tmp_adev != adev) { 565 top_info = &tmp_adev->psp.xgmi_context.top_info; 566 top_info->nodes[count - 1].node_id = 567 adev->gmc.xgmi.node_id; 568 top_info->num_nodes = count; 569 } 570 ret = amdgpu_xgmi_update_topology(hive, tmp_adev); 571 if (ret) 572 goto exit_unlock; 573 } 574 575 /* get latest topology info for each device from psp */ 576 list_for_each_entry(tmp_adev, &hive->device_list, gmc.xgmi.head) { 577 ret = psp_xgmi_get_topology_info(&tmp_adev->psp, count, 578 &tmp_adev->psp.xgmi_context.top_info); 579 if (ret) { 580 dev_err(tmp_adev->dev, 581 "XGMI: Get topology failure on device %llx, hive %llx, ret %d", 582 tmp_adev->gmc.xgmi.node_id, 583 tmp_adev->gmc.xgmi.hive_id, ret); 584 /* To do : continue with some node failed or disable the whole hive */ 585 goto exit_unlock; 586 } 587 } 588 } 589 590 if (!ret && !adev->gmc.xgmi.pending_reset) 591 ret = amdgpu_xgmi_sysfs_add_dev_info(adev, hive); 592 593 exit_unlock: 594 mutex_unlock(&hive->hive_lock); 595 exit: 596 if (!ret) { 597 adev->hive = hive; 598 dev_info(adev->dev, "XGMI: Add node %d, hive 0x%llx.\n", 599 adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id); 600 } else { 601 amdgpu_put_xgmi_hive(hive); 602 dev_err(adev->dev, "XGMI: Failed to add node %d, hive 0x%llx ret: %d\n", 603 adev->gmc.xgmi.physical_node_id, adev->gmc.xgmi.hive_id, 604 ret); 605 } 606 607 return ret; 608 } 609 610 int amdgpu_xgmi_remove_device(struct amdgpu_device *adev) 611 { 612 struct amdgpu_hive_info *hive = adev->hive; 613 614 if (!adev->gmc.xgmi.supported) 615 return -EINVAL; 616 617 if (!hive) 618 return -EINVAL; 619 620 mutex_lock(&hive->hive_lock); 621 task_barrier_rem_task(&hive->tb); 622 amdgpu_xgmi_sysfs_rem_dev_info(adev, hive); 623 if (hive->hi_req_gpu == adev) 624 hive->hi_req_gpu = NULL; 625 list_del(&adev->gmc.xgmi.head); 626 mutex_unlock(&hive->hive_lock); 627 628 amdgpu_put_xgmi_hive(hive); 629 adev->hive = NULL; 630 631 if (atomic_dec_return(&hive->number_devices) == 0) { 632 /* Remove the hive from global hive list */ 633 mutex_lock(&xgmi_mutex); 634 list_del(&hive->node); 635 mutex_unlock(&xgmi_mutex); 636 637 amdgpu_put_xgmi_hive(hive); 638 } 639 640 return psp_xgmi_terminate(&adev->psp); 641 } 642 643 static int amdgpu_xgmi_ras_late_init(struct amdgpu_device *adev) 644 { 645 int r; 646 struct ras_ih_if ih_info = { 647 .cb = NULL, 648 }; 649 struct ras_fs_if fs_info = { 650 .sysfs_name = "xgmi_wafl_err_count", 651 }; 652 653 if (!adev->gmc.xgmi.supported || 654 adev->gmc.xgmi.num_physical_nodes == 0) 655 return 0; 656 657 adev->gmc.xgmi.ras_funcs->reset_ras_error_count(adev); 658 659 if (!adev->gmc.xgmi.ras_if) { 660 adev->gmc.xgmi.ras_if = kmalloc(sizeof(struct ras_common_if), GFP_KERNEL); 661 if (!adev->gmc.xgmi.ras_if) 662 return -ENOMEM; 663 adev->gmc.xgmi.ras_if->block = AMDGPU_RAS_BLOCK__XGMI_WAFL; 664 adev->gmc.xgmi.ras_if->type = AMDGPU_RAS_ERROR__MULTI_UNCORRECTABLE; 665 adev->gmc.xgmi.ras_if->sub_block_index = 0; 666 strcpy(adev->gmc.xgmi.ras_if->name, "xgmi_wafl"); 667 } 668 ih_info.head = fs_info.head = *adev->gmc.xgmi.ras_if; 669 r = amdgpu_ras_late_init(adev, adev->gmc.xgmi.ras_if, 670 &fs_info, &ih_info); 671 if (r || !amdgpu_ras_is_supported(adev, adev->gmc.xgmi.ras_if->block)) { 672 kfree(adev->gmc.xgmi.ras_if); 673 adev->gmc.xgmi.ras_if = NULL; 674 } 675 676 return r; 677 } 678 679 static void amdgpu_xgmi_ras_fini(struct amdgpu_device *adev) 680 { 681 if (amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL) && 682 adev->gmc.xgmi.ras_if) { 683 struct ras_common_if *ras_if = adev->gmc.xgmi.ras_if; 684 struct ras_ih_if ih_info = { 685 .cb = NULL, 686 }; 687 688 amdgpu_ras_late_fini(adev, ras_if, &ih_info); 689 kfree(ras_if); 690 } 691 } 692 693 uint64_t amdgpu_xgmi_get_relative_phy_addr(struct amdgpu_device *adev, 694 uint64_t addr) 695 { 696 struct amdgpu_xgmi *xgmi = &adev->gmc.xgmi; 697 return (addr + xgmi->physical_node_id * xgmi->node_segment_size); 698 } 699 700 static void pcs_clear_status(struct amdgpu_device *adev, uint32_t pcs_status_reg) 701 { 702 WREG32_PCIE(pcs_status_reg, 0xFFFFFFFF); 703 WREG32_PCIE(pcs_status_reg, 0); 704 } 705 706 static void amdgpu_xgmi_reset_ras_error_count(struct amdgpu_device *adev) 707 { 708 uint32_t i; 709 710 switch (adev->asic_type) { 711 case CHIP_ARCTURUS: 712 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++) 713 pcs_clear_status(adev, 714 xgmi_pcs_err_status_reg_arct[i]); 715 break; 716 case CHIP_VEGA20: 717 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++) 718 pcs_clear_status(adev, 719 xgmi_pcs_err_status_reg_vg20[i]); 720 break; 721 default: 722 break; 723 } 724 } 725 726 static int amdgpu_xgmi_query_pcs_error_status(struct amdgpu_device *adev, 727 uint32_t value, 728 uint32_t *ue_count, 729 uint32_t *ce_count, 730 bool is_xgmi_pcs) 731 { 732 int i; 733 int ue_cnt; 734 735 if (is_xgmi_pcs) { 736 /* query xgmi pcs error status, 737 * only ue is supported */ 738 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_ras_fields); i ++) { 739 ue_cnt = (value & 740 xgmi_pcs_ras_fields[i].pcs_err_mask) >> 741 xgmi_pcs_ras_fields[i].pcs_err_shift; 742 if (ue_cnt) { 743 dev_info(adev->dev, "%s detected\n", 744 xgmi_pcs_ras_fields[i].err_name); 745 *ue_count += ue_cnt; 746 } 747 } 748 } else { 749 /* query wafl pcs error status, 750 * only ue is supported */ 751 for (i = 0; i < ARRAY_SIZE(wafl_pcs_ras_fields); i++) { 752 ue_cnt = (value & 753 wafl_pcs_ras_fields[i].pcs_err_mask) >> 754 wafl_pcs_ras_fields[i].pcs_err_shift; 755 if (ue_cnt) { 756 dev_info(adev->dev, "%s detected\n", 757 wafl_pcs_ras_fields[i].err_name); 758 *ue_count += ue_cnt; 759 } 760 } 761 } 762 763 return 0; 764 } 765 766 static int amdgpu_xgmi_query_ras_error_count(struct amdgpu_device *adev, 767 void *ras_error_status) 768 { 769 struct ras_err_data *err_data = (struct ras_err_data *)ras_error_status; 770 int i; 771 uint32_t data; 772 uint32_t ue_cnt = 0, ce_cnt = 0; 773 774 if (!amdgpu_ras_is_supported(adev, AMDGPU_RAS_BLOCK__XGMI_WAFL)) 775 return -EINVAL; 776 777 err_data->ue_count = 0; 778 err_data->ce_count = 0; 779 780 switch (adev->asic_type) { 781 case CHIP_ARCTURUS: 782 /* check xgmi pcs error */ 783 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_arct); i++) { 784 data = RREG32_PCIE(xgmi_pcs_err_status_reg_arct[i]); 785 if (data) 786 amdgpu_xgmi_query_pcs_error_status(adev, 787 data, &ue_cnt, &ce_cnt, true); 788 } 789 /* check wafl pcs error */ 790 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_arct); i++) { 791 data = RREG32_PCIE(wafl_pcs_err_status_reg_arct[i]); 792 if (data) 793 amdgpu_xgmi_query_pcs_error_status(adev, 794 data, &ue_cnt, &ce_cnt, false); 795 } 796 break; 797 case CHIP_VEGA20: 798 default: 799 /* check xgmi pcs error */ 800 for (i = 0; i < ARRAY_SIZE(xgmi_pcs_err_status_reg_vg20); i++) { 801 data = RREG32_PCIE(xgmi_pcs_err_status_reg_vg20[i]); 802 if (data) 803 amdgpu_xgmi_query_pcs_error_status(adev, 804 data, &ue_cnt, &ce_cnt, true); 805 } 806 /* check wafl pcs error */ 807 for (i = 0; i < ARRAY_SIZE(wafl_pcs_err_status_reg_vg20); i++) { 808 data = RREG32_PCIE(wafl_pcs_err_status_reg_vg20[i]); 809 if (data) 810 amdgpu_xgmi_query_pcs_error_status(adev, 811 data, &ue_cnt, &ce_cnt, false); 812 } 813 break; 814 } 815 816 adev->gmc.xgmi.ras_funcs->reset_ras_error_count(adev); 817 818 err_data->ue_count += ue_cnt; 819 err_data->ce_count += ce_cnt; 820 821 return 0; 822 } 823 824 const struct amdgpu_xgmi_ras_funcs xgmi_ras_funcs = { 825 .ras_late_init = amdgpu_xgmi_ras_late_init, 826 .ras_fini = amdgpu_xgmi_ras_fini, 827 .query_ras_error_count = amdgpu_xgmi_query_ras_error_count, 828 .reset_ras_error_count = amdgpu_xgmi_reset_ras_error_count, 829 }; 830