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