1 /* 2 * Copyright 2023 Advanced Micro Devices, Inc. 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice shall be included in 12 * all copies or substantial portions of the Software. 13 * 14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 17 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR 18 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, 19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR 20 * OTHER DEALINGS IN THE SOFTWARE. 21 */ 22 23 #include "kfd_debug.h" 24 #include "kfd_device_queue_manager.h" 25 #include "kfd_topology.h" 26 #include <linux/file.h> 27 #include <uapi/linux/kfd_ioctl.h> 28 29 #define MAX_WATCH_ADDRESSES 4 30 31 int kfd_dbg_ev_query_debug_event(struct kfd_process *process, 32 unsigned int *queue_id, 33 unsigned int *gpu_id, 34 uint64_t exception_clear_mask, 35 uint64_t *event_status) 36 { 37 struct process_queue_manager *pqm; 38 struct process_queue_node *pqn; 39 int i; 40 41 if (!(process && process->debug_trap_enabled)) 42 return -ENODATA; 43 44 mutex_lock(&process->event_mutex); 45 *event_status = 0; 46 *queue_id = 0; 47 *gpu_id = 0; 48 49 /* find and report queue events */ 50 pqm = &process->pqm; 51 list_for_each_entry(pqn, &pqm->queues, process_queue_list) { 52 uint64_t tmp = process->exception_enable_mask; 53 54 if (!pqn->q) 55 continue; 56 57 tmp &= pqn->q->properties.exception_status; 58 59 if (!tmp) 60 continue; 61 62 *event_status = pqn->q->properties.exception_status; 63 *queue_id = pqn->q->properties.queue_id; 64 *gpu_id = pqn->q->device->id; 65 pqn->q->properties.exception_status &= ~exception_clear_mask; 66 goto out; 67 } 68 69 /* find and report device events */ 70 for (i = 0; i < process->n_pdds; i++) { 71 struct kfd_process_device *pdd = process->pdds[i]; 72 uint64_t tmp = process->exception_enable_mask 73 & pdd->exception_status; 74 75 if (!tmp) 76 continue; 77 78 *event_status = pdd->exception_status; 79 *gpu_id = pdd->dev->id; 80 pdd->exception_status &= ~exception_clear_mask; 81 goto out; 82 } 83 84 /* report process events */ 85 if (process->exception_enable_mask & process->exception_status) { 86 *event_status = process->exception_status; 87 process->exception_status &= ~exception_clear_mask; 88 } 89 90 out: 91 mutex_unlock(&process->event_mutex); 92 return *event_status ? 0 : -EAGAIN; 93 } 94 95 void debug_event_write_work_handler(struct work_struct *work) 96 { 97 struct kfd_process *process; 98 99 static const char write_data = '.'; 100 loff_t pos = 0; 101 102 process = container_of(work, 103 struct kfd_process, 104 debug_event_workarea); 105 106 kernel_write(process->dbg_ev_file, &write_data, 1, &pos); 107 } 108 109 /* update process/device/queue exception status, write to descriptor 110 * only if exception_status is enabled. 111 */ 112 bool kfd_dbg_ev_raise(uint64_t event_mask, 113 struct kfd_process *process, struct kfd_node *dev, 114 unsigned int source_id, bool use_worker, 115 void *exception_data, size_t exception_data_size) 116 { 117 struct process_queue_manager *pqm; 118 struct process_queue_node *pqn; 119 int i; 120 static const char write_data = '.'; 121 loff_t pos = 0; 122 bool is_subscribed = true; 123 124 if (!(process && process->debug_trap_enabled)) 125 return false; 126 127 mutex_lock(&process->event_mutex); 128 129 if (event_mask & KFD_EC_MASK_DEVICE) { 130 for (i = 0; i < process->n_pdds; i++) { 131 struct kfd_process_device *pdd = process->pdds[i]; 132 133 if (pdd->dev != dev) 134 continue; 135 136 pdd->exception_status |= event_mask & KFD_EC_MASK_DEVICE; 137 138 if (event_mask & KFD_EC_MASK(EC_DEVICE_MEMORY_VIOLATION)) { 139 if (!pdd->vm_fault_exc_data) { 140 pdd->vm_fault_exc_data = kmemdup( 141 exception_data, 142 exception_data_size, 143 GFP_KERNEL); 144 if (!pdd->vm_fault_exc_data) 145 pr_debug("Failed to allocate exception data memory"); 146 } else { 147 pr_debug("Debugger exception data not saved\n"); 148 print_hex_dump_bytes("exception data: ", 149 DUMP_PREFIX_OFFSET, 150 exception_data, 151 exception_data_size); 152 } 153 } 154 break; 155 } 156 } else if (event_mask & KFD_EC_MASK_PROCESS) { 157 process->exception_status |= event_mask & KFD_EC_MASK_PROCESS; 158 } else { 159 pqm = &process->pqm; 160 list_for_each_entry(pqn, &pqm->queues, 161 process_queue_list) { 162 int target_id; 163 164 if (!pqn->q) 165 continue; 166 167 target_id = event_mask & KFD_EC_MASK(EC_QUEUE_NEW) ? 168 pqn->q->properties.queue_id : 169 pqn->q->doorbell_id; 170 171 if (pqn->q->device != dev || target_id != source_id) 172 continue; 173 174 pqn->q->properties.exception_status |= event_mask; 175 break; 176 } 177 } 178 179 if (process->exception_enable_mask & event_mask) { 180 if (use_worker) 181 schedule_work(&process->debug_event_workarea); 182 else 183 kernel_write(process->dbg_ev_file, 184 &write_data, 185 1, 186 &pos); 187 } else { 188 is_subscribed = false; 189 } 190 191 mutex_unlock(&process->event_mutex); 192 193 return is_subscribed; 194 } 195 196 /* set pending event queue entry from ring entry */ 197 bool kfd_set_dbg_ev_from_interrupt(struct kfd_node *dev, 198 unsigned int pasid, 199 uint32_t doorbell_id, 200 uint64_t trap_mask, 201 void *exception_data, 202 size_t exception_data_size) 203 { 204 struct kfd_process *p; 205 bool signaled_to_debugger_or_runtime = false; 206 207 p = kfd_lookup_process_by_pasid(pasid); 208 209 if (!p) 210 return false; 211 212 if (!kfd_dbg_ev_raise(trap_mask, p, dev, doorbell_id, true, 213 exception_data, exception_data_size)) { 214 struct process_queue_manager *pqm; 215 struct process_queue_node *pqn; 216 217 if (!!(trap_mask & KFD_EC_MASK_QUEUE) && 218 p->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED) { 219 mutex_lock(&p->mutex); 220 221 pqm = &p->pqm; 222 list_for_each_entry(pqn, &pqm->queues, 223 process_queue_list) { 224 225 if (!(pqn->q && pqn->q->device == dev && 226 pqn->q->doorbell_id == doorbell_id)) 227 continue; 228 229 kfd_send_exception_to_runtime(p, pqn->q->properties.queue_id, 230 trap_mask); 231 232 signaled_to_debugger_or_runtime = true; 233 234 break; 235 } 236 237 mutex_unlock(&p->mutex); 238 } else if (trap_mask & KFD_EC_MASK(EC_DEVICE_MEMORY_VIOLATION)) { 239 kfd_dqm_evict_pasid(dev->dqm, p->pasid); 240 kfd_signal_vm_fault_event(dev, p->pasid, NULL, 241 exception_data); 242 243 signaled_to_debugger_or_runtime = true; 244 } 245 } else { 246 signaled_to_debugger_or_runtime = true; 247 } 248 249 kfd_unref_process(p); 250 251 return signaled_to_debugger_or_runtime; 252 } 253 254 int kfd_dbg_send_exception_to_runtime(struct kfd_process *p, 255 unsigned int dev_id, 256 unsigned int queue_id, 257 uint64_t error_reason) 258 { 259 if (error_reason & KFD_EC_MASK(EC_DEVICE_MEMORY_VIOLATION)) { 260 struct kfd_process_device *pdd = NULL; 261 struct kfd_hsa_memory_exception_data *data; 262 int i; 263 264 for (i = 0; i < p->n_pdds; i++) { 265 if (p->pdds[i]->dev->id == dev_id) { 266 pdd = p->pdds[i]; 267 break; 268 } 269 } 270 271 if (!pdd) 272 return -ENODEV; 273 274 data = (struct kfd_hsa_memory_exception_data *) 275 pdd->vm_fault_exc_data; 276 277 kfd_dqm_evict_pasid(pdd->dev->dqm, p->pasid); 278 kfd_signal_vm_fault_event(pdd->dev, p->pasid, NULL, data); 279 error_reason &= ~KFD_EC_MASK(EC_DEVICE_MEMORY_VIOLATION); 280 } 281 282 if (error_reason & (KFD_EC_MASK(EC_PROCESS_RUNTIME))) { 283 /* 284 * block should only happen after the debugger receives runtime 285 * enable notice. 286 */ 287 up(&p->runtime_enable_sema); 288 error_reason &= ~KFD_EC_MASK(EC_PROCESS_RUNTIME); 289 } 290 291 if (error_reason) 292 return kfd_send_exception_to_runtime(p, queue_id, error_reason); 293 294 return 0; 295 } 296 297 static int kfd_dbg_set_queue_workaround(struct queue *q, bool enable) 298 { 299 struct mqd_update_info minfo = {0}; 300 int err; 301 302 if (!q) 303 return 0; 304 305 if (!kfd_dbg_has_cwsr_workaround(q->device)) 306 return 0; 307 308 if (enable && q->properties.is_user_cu_masked) 309 return -EBUSY; 310 311 minfo.update_flag = enable ? UPDATE_FLAG_DBG_WA_ENABLE : UPDATE_FLAG_DBG_WA_DISABLE; 312 313 q->properties.is_dbg_wa = enable; 314 err = q->device->dqm->ops.update_queue(q->device->dqm, q, &minfo); 315 if (err) 316 q->properties.is_dbg_wa = false; 317 318 return err; 319 } 320 321 static int kfd_dbg_set_workaround(struct kfd_process *target, bool enable) 322 { 323 struct process_queue_manager *pqm = &target->pqm; 324 struct process_queue_node *pqn; 325 int r = 0; 326 327 list_for_each_entry(pqn, &pqm->queues, process_queue_list) { 328 r = kfd_dbg_set_queue_workaround(pqn->q, enable); 329 if (enable && r) 330 goto unwind; 331 } 332 333 return 0; 334 335 unwind: 336 list_for_each_entry(pqn, &pqm->queues, process_queue_list) 337 kfd_dbg_set_queue_workaround(pqn->q, false); 338 339 if (enable) 340 target->runtime_info.runtime_state = r == -EBUSY ? 341 DEBUG_RUNTIME_STATE_ENABLED_BUSY : 342 DEBUG_RUNTIME_STATE_ENABLED_ERROR; 343 344 return r; 345 } 346 347 int kfd_dbg_set_mes_debug_mode(struct kfd_process_device *pdd) 348 { 349 uint32_t spi_dbg_cntl = pdd->spi_dbg_override | pdd->spi_dbg_launch_mode; 350 uint32_t flags = pdd->process->dbg_flags; 351 bool sq_trap_en = !!spi_dbg_cntl || !kfd_dbg_has_cwsr_workaround(pdd->dev); 352 353 if (!kfd_dbg_is_per_vmid_supported(pdd->dev)) 354 return 0; 355 356 return amdgpu_mes_set_shader_debugger(pdd->dev->adev, pdd->proc_ctx_gpu_addr, spi_dbg_cntl, 357 pdd->watch_points, flags, sq_trap_en); 358 } 359 360 #define KFD_DEBUGGER_INVALID_WATCH_POINT_ID -1 361 static int kfd_dbg_get_dev_watch_id(struct kfd_process_device *pdd, int *watch_id) 362 { 363 int i; 364 365 *watch_id = KFD_DEBUGGER_INVALID_WATCH_POINT_ID; 366 367 spin_lock(&pdd->dev->kfd->watch_points_lock); 368 369 for (i = 0; i < MAX_WATCH_ADDRESSES; i++) { 370 /* device watchpoint in use so skip */ 371 if ((pdd->dev->kfd->alloc_watch_ids >> i) & 0x1) 372 continue; 373 374 pdd->alloc_watch_ids |= 0x1 << i; 375 pdd->dev->kfd->alloc_watch_ids |= 0x1 << i; 376 *watch_id = i; 377 spin_unlock(&pdd->dev->kfd->watch_points_lock); 378 return 0; 379 } 380 381 spin_unlock(&pdd->dev->kfd->watch_points_lock); 382 383 return -ENOMEM; 384 } 385 386 static void kfd_dbg_clear_dev_watch_id(struct kfd_process_device *pdd, int watch_id) 387 { 388 spin_lock(&pdd->dev->kfd->watch_points_lock); 389 390 /* process owns device watch point so safe to clear */ 391 if ((pdd->alloc_watch_ids >> watch_id) & 0x1) { 392 pdd->alloc_watch_ids &= ~(0x1 << watch_id); 393 pdd->dev->kfd->alloc_watch_ids &= ~(0x1 << watch_id); 394 } 395 396 spin_unlock(&pdd->dev->kfd->watch_points_lock); 397 } 398 399 static bool kfd_dbg_owns_dev_watch_id(struct kfd_process_device *pdd, int watch_id) 400 { 401 bool owns_watch_id = false; 402 403 spin_lock(&pdd->dev->kfd->watch_points_lock); 404 owns_watch_id = watch_id < MAX_WATCH_ADDRESSES && 405 ((pdd->alloc_watch_ids >> watch_id) & 0x1); 406 407 spin_unlock(&pdd->dev->kfd->watch_points_lock); 408 409 return owns_watch_id; 410 } 411 412 int kfd_dbg_trap_clear_dev_address_watch(struct kfd_process_device *pdd, 413 uint32_t watch_id) 414 { 415 int r; 416 417 if (!kfd_dbg_owns_dev_watch_id(pdd, watch_id)) 418 return -EINVAL; 419 420 if (!pdd->dev->kfd->shared_resources.enable_mes) { 421 r = debug_lock_and_unmap(pdd->dev->dqm); 422 if (r) 423 return r; 424 } 425 426 amdgpu_gfx_off_ctrl(pdd->dev->adev, false); 427 pdd->watch_points[watch_id] = pdd->dev->kfd2kgd->clear_address_watch( 428 pdd->dev->adev, 429 watch_id); 430 amdgpu_gfx_off_ctrl(pdd->dev->adev, true); 431 432 if (!pdd->dev->kfd->shared_resources.enable_mes) 433 r = debug_map_and_unlock(pdd->dev->dqm); 434 else 435 r = kfd_dbg_set_mes_debug_mode(pdd); 436 437 kfd_dbg_clear_dev_watch_id(pdd, watch_id); 438 439 return r; 440 } 441 442 int kfd_dbg_trap_set_dev_address_watch(struct kfd_process_device *pdd, 443 uint64_t watch_address, 444 uint32_t watch_address_mask, 445 uint32_t *watch_id, 446 uint32_t watch_mode) 447 { 448 int r = kfd_dbg_get_dev_watch_id(pdd, watch_id); 449 450 if (r) 451 return r; 452 453 if (!pdd->dev->kfd->shared_resources.enable_mes) { 454 r = debug_lock_and_unmap(pdd->dev->dqm); 455 if (r) { 456 kfd_dbg_clear_dev_watch_id(pdd, *watch_id); 457 return r; 458 } 459 } 460 461 amdgpu_gfx_off_ctrl(pdd->dev->adev, false); 462 pdd->watch_points[*watch_id] = pdd->dev->kfd2kgd->set_address_watch( 463 pdd->dev->adev, 464 watch_address, 465 watch_address_mask, 466 *watch_id, 467 watch_mode, 468 pdd->dev->vm_info.last_vmid_kfd); 469 amdgpu_gfx_off_ctrl(pdd->dev->adev, true); 470 471 if (!pdd->dev->kfd->shared_resources.enable_mes) 472 r = debug_map_and_unlock(pdd->dev->dqm); 473 else 474 r = kfd_dbg_set_mes_debug_mode(pdd); 475 476 /* HWS is broken so no point in HW rollback but release the watchpoint anyways */ 477 if (r) 478 kfd_dbg_clear_dev_watch_id(pdd, *watch_id); 479 480 return 0; 481 } 482 483 static void kfd_dbg_clear_process_address_watch(struct kfd_process *target) 484 { 485 int i, j; 486 487 for (i = 0; i < target->n_pdds; i++) 488 for (j = 0; j < MAX_WATCH_ADDRESSES; j++) 489 kfd_dbg_trap_clear_dev_address_watch(target->pdds[i], j); 490 } 491 492 int kfd_dbg_trap_set_flags(struct kfd_process *target, uint32_t *flags) 493 { 494 uint32_t prev_flags = target->dbg_flags; 495 int i, r = 0, rewind_count = 0; 496 497 for (i = 0; i < target->n_pdds; i++) { 498 if (!kfd_dbg_is_per_vmid_supported(target->pdds[i]->dev) && 499 (*flags & KFD_DBG_TRAP_FLAG_SINGLE_MEM_OP)) { 500 *flags = prev_flags; 501 return -EACCES; 502 } 503 } 504 505 target->dbg_flags = *flags & KFD_DBG_TRAP_FLAG_SINGLE_MEM_OP; 506 *flags = prev_flags; 507 for (i = 0; i < target->n_pdds; i++) { 508 struct kfd_process_device *pdd = target->pdds[i]; 509 510 if (!kfd_dbg_is_per_vmid_supported(pdd->dev)) 511 continue; 512 513 if (!pdd->dev->kfd->shared_resources.enable_mes) 514 r = debug_refresh_runlist(pdd->dev->dqm); 515 else 516 r = kfd_dbg_set_mes_debug_mode(pdd); 517 518 if (r) { 519 target->dbg_flags = prev_flags; 520 break; 521 } 522 523 rewind_count++; 524 } 525 526 /* Rewind flags */ 527 if (r) { 528 target->dbg_flags = prev_flags; 529 530 for (i = 0; i < rewind_count; i++) { 531 struct kfd_process_device *pdd = target->pdds[i]; 532 533 if (!kfd_dbg_is_per_vmid_supported(pdd->dev)) 534 continue; 535 536 if (!pdd->dev->kfd->shared_resources.enable_mes) 537 debug_refresh_runlist(pdd->dev->dqm); 538 else 539 kfd_dbg_set_mes_debug_mode(pdd); 540 } 541 } 542 543 return r; 544 } 545 546 /* kfd_dbg_trap_deactivate: 547 * target: target process 548 * unwind: If this is unwinding a failed kfd_dbg_trap_enable() 549 * unwind_count: 550 * If unwind == true, how far down the pdd list we need 551 * to unwind 552 * else: ignored 553 */ 554 void kfd_dbg_trap_deactivate(struct kfd_process *target, bool unwind, int unwind_count) 555 { 556 int i; 557 558 if (!unwind) { 559 uint32_t flags = 0; 560 int resume_count = resume_queues(target, 0, NULL); 561 562 if (resume_count) 563 pr_debug("Resumed %d queues\n", resume_count); 564 565 cancel_work_sync(&target->debug_event_workarea); 566 kfd_dbg_clear_process_address_watch(target); 567 kfd_dbg_trap_set_wave_launch_mode(target, 0); 568 569 kfd_dbg_trap_set_flags(target, &flags); 570 } 571 572 for (i = 0; i < target->n_pdds; i++) { 573 struct kfd_process_device *pdd = target->pdds[i]; 574 575 /* If this is an unwind, and we have unwound the required 576 * enable calls on the pdd list, we need to stop now 577 * otherwise we may mess up another debugger session. 578 */ 579 if (unwind && i == unwind_count) 580 break; 581 582 kfd_process_set_trap_debug_flag(&pdd->qpd, false); 583 584 /* GFX off is already disabled by debug activate if not RLC restore supported. */ 585 if (kfd_dbg_is_rlc_restore_supported(pdd->dev)) 586 amdgpu_gfx_off_ctrl(pdd->dev->adev, false); 587 pdd->spi_dbg_override = 588 pdd->dev->kfd2kgd->disable_debug_trap( 589 pdd->dev->adev, 590 target->runtime_info.ttmp_setup, 591 pdd->dev->vm_info.last_vmid_kfd); 592 amdgpu_gfx_off_ctrl(pdd->dev->adev, true); 593 594 if (!kfd_dbg_is_per_vmid_supported(pdd->dev) && 595 release_debug_trap_vmid(pdd->dev->dqm, &pdd->qpd)) 596 pr_err("Failed to release debug vmid on [%i]\n", pdd->dev->id); 597 598 if (!pdd->dev->kfd->shared_resources.enable_mes) 599 debug_refresh_runlist(pdd->dev->dqm); 600 else 601 kfd_dbg_set_mes_debug_mode(pdd); 602 } 603 604 kfd_dbg_set_workaround(target, false); 605 } 606 607 static void kfd_dbg_clean_exception_status(struct kfd_process *target) 608 { 609 struct process_queue_manager *pqm; 610 struct process_queue_node *pqn; 611 int i; 612 613 for (i = 0; i < target->n_pdds; i++) { 614 struct kfd_process_device *pdd = target->pdds[i]; 615 616 kfd_process_drain_interrupts(pdd); 617 618 pdd->exception_status = 0; 619 } 620 621 pqm = &target->pqm; 622 list_for_each_entry(pqn, &pqm->queues, process_queue_list) { 623 if (!pqn->q) 624 continue; 625 626 pqn->q->properties.exception_status = 0; 627 } 628 629 target->exception_status = 0; 630 } 631 632 int kfd_dbg_trap_disable(struct kfd_process *target) 633 { 634 if (!target->debug_trap_enabled) 635 return 0; 636 637 /* 638 * Defer deactivation to runtime if runtime not enabled otherwise reset 639 * attached running target runtime state to enable for re-attach. 640 */ 641 if (target->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED) 642 kfd_dbg_trap_deactivate(target, false, 0); 643 else if (target->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_DISABLED) 644 target->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_ENABLED; 645 646 fput(target->dbg_ev_file); 647 target->dbg_ev_file = NULL; 648 649 if (target->debugger_process) { 650 atomic_dec(&target->debugger_process->debugged_process_count); 651 target->debugger_process = NULL; 652 } 653 654 target->debug_trap_enabled = false; 655 kfd_dbg_clean_exception_status(target); 656 kfd_unref_process(target); 657 658 return 0; 659 } 660 661 int kfd_dbg_trap_activate(struct kfd_process *target) 662 { 663 int i, r = 0; 664 665 r = kfd_dbg_set_workaround(target, true); 666 if (r) 667 return r; 668 669 for (i = 0; i < target->n_pdds; i++) { 670 struct kfd_process_device *pdd = target->pdds[i]; 671 672 if (!kfd_dbg_is_per_vmid_supported(pdd->dev)) { 673 r = reserve_debug_trap_vmid(pdd->dev->dqm, &pdd->qpd); 674 675 if (r) { 676 target->runtime_info.runtime_state = (r == -EBUSY) ? 677 DEBUG_RUNTIME_STATE_ENABLED_BUSY : 678 DEBUG_RUNTIME_STATE_ENABLED_ERROR; 679 680 goto unwind_err; 681 } 682 } 683 684 /* Disable GFX OFF to prevent garbage read/writes to debug registers. 685 * If RLC restore of debug registers is not supported and runtime enable 686 * hasn't done so already on ttmp setup request, restore the trap config registers. 687 * 688 * If RLC restore of debug registers is not supported, keep gfx off disabled for 689 * the debug session. 690 */ 691 amdgpu_gfx_off_ctrl(pdd->dev->adev, false); 692 if (!(kfd_dbg_is_rlc_restore_supported(pdd->dev) || 693 target->runtime_info.ttmp_setup)) 694 pdd->dev->kfd2kgd->enable_debug_trap(pdd->dev->adev, true, 695 pdd->dev->vm_info.last_vmid_kfd); 696 697 pdd->spi_dbg_override = pdd->dev->kfd2kgd->enable_debug_trap( 698 pdd->dev->adev, 699 false, 700 pdd->dev->vm_info.last_vmid_kfd); 701 702 if (kfd_dbg_is_rlc_restore_supported(pdd->dev)) 703 amdgpu_gfx_off_ctrl(pdd->dev->adev, true); 704 705 /* 706 * Setting the debug flag in the trap handler requires that the TMA has been 707 * allocated, which occurs during CWSR initialization. 708 * In the event that CWSR has not been initialized at this point, setting the 709 * flag will be called again during CWSR initialization if the target process 710 * is still debug enabled. 711 */ 712 kfd_process_set_trap_debug_flag(&pdd->qpd, true); 713 714 if (!pdd->dev->kfd->shared_resources.enable_mes) 715 r = debug_refresh_runlist(pdd->dev->dqm); 716 else 717 r = kfd_dbg_set_mes_debug_mode(pdd); 718 719 if (r) { 720 target->runtime_info.runtime_state = 721 DEBUG_RUNTIME_STATE_ENABLED_ERROR; 722 goto unwind_err; 723 } 724 } 725 726 return 0; 727 728 unwind_err: 729 /* Enabling debug failed, we need to disable on 730 * all GPUs so the enable is all or nothing. 731 */ 732 kfd_dbg_trap_deactivate(target, true, i); 733 return r; 734 } 735 736 int kfd_dbg_trap_enable(struct kfd_process *target, uint32_t fd, 737 void __user *runtime_info, uint32_t *runtime_size) 738 { 739 struct file *f; 740 uint32_t copy_size; 741 int i, r = 0; 742 743 if (target->debug_trap_enabled) 744 return -EALREADY; 745 746 /* Enable pre-checks */ 747 for (i = 0; i < target->n_pdds; i++) { 748 struct kfd_process_device *pdd = target->pdds[i]; 749 750 if (!KFD_IS_SOC15(pdd->dev)) 751 return -ENODEV; 752 753 if (!kfd_dbg_has_gws_support(pdd->dev) && pdd->qpd.num_gws) 754 return -EBUSY; 755 } 756 757 copy_size = min((size_t)(*runtime_size), sizeof(target->runtime_info)); 758 759 f = fget(fd); 760 if (!f) { 761 pr_err("Failed to get file for (%i)\n", fd); 762 return -EBADF; 763 } 764 765 target->dbg_ev_file = f; 766 767 /* defer activation to runtime if not runtime enabled */ 768 if (target->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED) 769 kfd_dbg_trap_activate(target); 770 771 /* We already hold the process reference but hold another one for the 772 * debug session. 773 */ 774 kref_get(&target->ref); 775 target->debug_trap_enabled = true; 776 777 if (target->debugger_process) 778 atomic_inc(&target->debugger_process->debugged_process_count); 779 780 if (copy_to_user(runtime_info, (void *)&target->runtime_info, copy_size)) { 781 kfd_dbg_trap_deactivate(target, false, 0); 782 r = -EFAULT; 783 } 784 785 *runtime_size = sizeof(target->runtime_info); 786 787 return r; 788 } 789 790 static int kfd_dbg_validate_trap_override_request(struct kfd_process *p, 791 uint32_t trap_override, 792 uint32_t trap_mask_request, 793 uint32_t *trap_mask_supported) 794 { 795 int i = 0; 796 797 *trap_mask_supported = 0xffffffff; 798 799 for (i = 0; i < p->n_pdds; i++) { 800 struct kfd_process_device *pdd = p->pdds[i]; 801 int err = pdd->dev->kfd2kgd->validate_trap_override_request( 802 pdd->dev->adev, 803 trap_override, 804 trap_mask_supported); 805 806 if (err) 807 return err; 808 } 809 810 if (trap_mask_request & ~*trap_mask_supported) 811 return -EACCES; 812 813 return 0; 814 } 815 816 int kfd_dbg_trap_set_wave_launch_override(struct kfd_process *target, 817 uint32_t trap_override, 818 uint32_t trap_mask_bits, 819 uint32_t trap_mask_request, 820 uint32_t *trap_mask_prev, 821 uint32_t *trap_mask_supported) 822 { 823 int r = 0, i; 824 825 r = kfd_dbg_validate_trap_override_request(target, 826 trap_override, 827 trap_mask_request, 828 trap_mask_supported); 829 830 if (r) 831 return r; 832 833 for (i = 0; i < target->n_pdds; i++) { 834 struct kfd_process_device *pdd = target->pdds[i]; 835 836 amdgpu_gfx_off_ctrl(pdd->dev->adev, false); 837 pdd->spi_dbg_override = pdd->dev->kfd2kgd->set_wave_launch_trap_override( 838 pdd->dev->adev, 839 pdd->dev->vm_info.last_vmid_kfd, 840 trap_override, 841 trap_mask_bits, 842 trap_mask_request, 843 trap_mask_prev, 844 pdd->spi_dbg_override); 845 amdgpu_gfx_off_ctrl(pdd->dev->adev, true); 846 847 if (!pdd->dev->kfd->shared_resources.enable_mes) 848 r = debug_refresh_runlist(pdd->dev->dqm); 849 else 850 r = kfd_dbg_set_mes_debug_mode(pdd); 851 852 if (r) 853 break; 854 } 855 856 return r; 857 } 858 859 int kfd_dbg_trap_set_wave_launch_mode(struct kfd_process *target, 860 uint8_t wave_launch_mode) 861 { 862 int r = 0, i; 863 864 if (wave_launch_mode != KFD_DBG_TRAP_WAVE_LAUNCH_MODE_NORMAL && 865 wave_launch_mode != KFD_DBG_TRAP_WAVE_LAUNCH_MODE_HALT && 866 wave_launch_mode != KFD_DBG_TRAP_WAVE_LAUNCH_MODE_DEBUG) 867 return -EINVAL; 868 869 for (i = 0; i < target->n_pdds; i++) { 870 struct kfd_process_device *pdd = target->pdds[i]; 871 872 amdgpu_gfx_off_ctrl(pdd->dev->adev, false); 873 pdd->spi_dbg_launch_mode = pdd->dev->kfd2kgd->set_wave_launch_mode( 874 pdd->dev->adev, 875 wave_launch_mode, 876 pdd->dev->vm_info.last_vmid_kfd); 877 amdgpu_gfx_off_ctrl(pdd->dev->adev, true); 878 879 if (!pdd->dev->kfd->shared_resources.enable_mes) 880 r = debug_refresh_runlist(pdd->dev->dqm); 881 else 882 r = kfd_dbg_set_mes_debug_mode(pdd); 883 884 if (r) 885 break; 886 } 887 888 return r; 889 } 890 891 int kfd_dbg_trap_query_exception_info(struct kfd_process *target, 892 uint32_t source_id, 893 uint32_t exception_code, 894 bool clear_exception, 895 void __user *info, 896 uint32_t *info_size) 897 { 898 bool found = false; 899 int r = 0; 900 uint32_t copy_size, actual_info_size = 0; 901 uint64_t *exception_status_ptr = NULL; 902 903 if (!target) 904 return -EINVAL; 905 906 if (!info || !info_size) 907 return -EINVAL; 908 909 mutex_lock(&target->event_mutex); 910 911 if (KFD_DBG_EC_TYPE_IS_QUEUE(exception_code)) { 912 /* Per queue exceptions */ 913 struct queue *queue = NULL; 914 int i; 915 916 for (i = 0; i < target->n_pdds; i++) { 917 struct kfd_process_device *pdd = target->pdds[i]; 918 struct qcm_process_device *qpd = &pdd->qpd; 919 920 list_for_each_entry(queue, &qpd->queues_list, list) { 921 if (!found && queue->properties.queue_id == source_id) { 922 found = true; 923 break; 924 } 925 } 926 if (found) 927 break; 928 } 929 930 if (!found) { 931 r = -EINVAL; 932 goto out; 933 } 934 935 if (!(queue->properties.exception_status & KFD_EC_MASK(exception_code))) { 936 r = -ENODATA; 937 goto out; 938 } 939 exception_status_ptr = &queue->properties.exception_status; 940 } else if (KFD_DBG_EC_TYPE_IS_DEVICE(exception_code)) { 941 /* Per device exceptions */ 942 struct kfd_process_device *pdd = NULL; 943 int i; 944 945 for (i = 0; i < target->n_pdds; i++) { 946 pdd = target->pdds[i]; 947 if (pdd->dev->id == source_id) { 948 found = true; 949 break; 950 } 951 } 952 953 if (!found) { 954 r = -EINVAL; 955 goto out; 956 } 957 958 if (!(pdd->exception_status & KFD_EC_MASK(exception_code))) { 959 r = -ENODATA; 960 goto out; 961 } 962 963 if (exception_code == EC_DEVICE_MEMORY_VIOLATION) { 964 copy_size = min((size_t)(*info_size), pdd->vm_fault_exc_data_size); 965 966 if (copy_to_user(info, pdd->vm_fault_exc_data, copy_size)) { 967 r = -EFAULT; 968 goto out; 969 } 970 actual_info_size = pdd->vm_fault_exc_data_size; 971 if (clear_exception) { 972 kfree(pdd->vm_fault_exc_data); 973 pdd->vm_fault_exc_data = NULL; 974 pdd->vm_fault_exc_data_size = 0; 975 } 976 } 977 exception_status_ptr = &pdd->exception_status; 978 } else if (KFD_DBG_EC_TYPE_IS_PROCESS(exception_code)) { 979 /* Per process exceptions */ 980 if (!(target->exception_status & KFD_EC_MASK(exception_code))) { 981 r = -ENODATA; 982 goto out; 983 } 984 985 if (exception_code == EC_PROCESS_RUNTIME) { 986 copy_size = min((size_t)(*info_size), sizeof(target->runtime_info)); 987 988 if (copy_to_user(info, (void *)&target->runtime_info, copy_size)) { 989 r = -EFAULT; 990 goto out; 991 } 992 993 actual_info_size = sizeof(target->runtime_info); 994 } 995 996 exception_status_ptr = &target->exception_status; 997 } else { 998 pr_debug("Bad exception type [%i]\n", exception_code); 999 r = -EINVAL; 1000 goto out; 1001 } 1002 1003 *info_size = actual_info_size; 1004 if (clear_exception) 1005 *exception_status_ptr &= ~KFD_EC_MASK(exception_code); 1006 out: 1007 mutex_unlock(&target->event_mutex); 1008 return r; 1009 } 1010 1011 int kfd_dbg_trap_device_snapshot(struct kfd_process *target, 1012 uint64_t exception_clear_mask, 1013 void __user *user_info, 1014 uint32_t *number_of_device_infos, 1015 uint32_t *entry_size) 1016 { 1017 struct kfd_dbg_device_info_entry device_info; 1018 uint32_t tmp_entry_size = *entry_size, tmp_num_devices; 1019 int i, r = 0; 1020 1021 if (!(target && user_info && number_of_device_infos && entry_size)) 1022 return -EINVAL; 1023 1024 tmp_num_devices = min_t(size_t, *number_of_device_infos, target->n_pdds); 1025 *number_of_device_infos = target->n_pdds; 1026 *entry_size = min_t(size_t, *entry_size, sizeof(device_info)); 1027 1028 if (!tmp_num_devices) 1029 return 0; 1030 1031 memset(&device_info, 0, sizeof(device_info)); 1032 1033 mutex_lock(&target->event_mutex); 1034 1035 /* Run over all pdd of the process */ 1036 for (i = 0; i < tmp_num_devices; i++) { 1037 struct kfd_process_device *pdd = target->pdds[i]; 1038 struct kfd_topology_device *topo_dev = kfd_topology_device_by_id(pdd->dev->id); 1039 1040 device_info.gpu_id = pdd->dev->id; 1041 device_info.exception_status = pdd->exception_status; 1042 device_info.lds_base = pdd->lds_base; 1043 device_info.lds_limit = pdd->lds_limit; 1044 device_info.scratch_base = pdd->scratch_base; 1045 device_info.scratch_limit = pdd->scratch_limit; 1046 device_info.gpuvm_base = pdd->gpuvm_base; 1047 device_info.gpuvm_limit = pdd->gpuvm_limit; 1048 device_info.location_id = topo_dev->node_props.location_id; 1049 device_info.vendor_id = topo_dev->node_props.vendor_id; 1050 device_info.device_id = topo_dev->node_props.device_id; 1051 device_info.revision_id = pdd->dev->adev->pdev->revision; 1052 device_info.subsystem_vendor_id = pdd->dev->adev->pdev->subsystem_vendor; 1053 device_info.subsystem_device_id = pdd->dev->adev->pdev->subsystem_device; 1054 device_info.fw_version = pdd->dev->kfd->mec_fw_version; 1055 device_info.gfx_target_version = 1056 topo_dev->node_props.gfx_target_version; 1057 device_info.simd_count = topo_dev->node_props.simd_count; 1058 device_info.max_waves_per_simd = 1059 topo_dev->node_props.max_waves_per_simd; 1060 device_info.array_count = topo_dev->node_props.array_count; 1061 device_info.simd_arrays_per_engine = 1062 topo_dev->node_props.simd_arrays_per_engine; 1063 device_info.num_xcc = NUM_XCC(pdd->dev->xcc_mask); 1064 device_info.capability = topo_dev->node_props.capability; 1065 device_info.debug_prop = topo_dev->node_props.debug_prop; 1066 1067 if (exception_clear_mask) 1068 pdd->exception_status &= ~exception_clear_mask; 1069 1070 if (copy_to_user(user_info, &device_info, *entry_size)) { 1071 r = -EFAULT; 1072 break; 1073 } 1074 1075 user_info += tmp_entry_size; 1076 } 1077 1078 mutex_unlock(&target->event_mutex); 1079 1080 return r; 1081 } 1082 1083 void kfd_dbg_set_enabled_debug_exception_mask(struct kfd_process *target, 1084 uint64_t exception_set_mask) 1085 { 1086 uint64_t found_mask = 0; 1087 struct process_queue_manager *pqm; 1088 struct process_queue_node *pqn; 1089 static const char write_data = '.'; 1090 loff_t pos = 0; 1091 int i; 1092 1093 mutex_lock(&target->event_mutex); 1094 1095 found_mask |= target->exception_status; 1096 1097 pqm = &target->pqm; 1098 list_for_each_entry(pqn, &pqm->queues, process_queue_list) { 1099 if (!pqn->q) 1100 continue; 1101 1102 found_mask |= pqn->q->properties.exception_status; 1103 } 1104 1105 for (i = 0; i < target->n_pdds; i++) { 1106 struct kfd_process_device *pdd = target->pdds[i]; 1107 1108 found_mask |= pdd->exception_status; 1109 } 1110 1111 if (exception_set_mask & found_mask) 1112 kernel_write(target->dbg_ev_file, &write_data, 1, &pos); 1113 1114 target->exception_enable_mask = exception_set_mask; 1115 1116 mutex_unlock(&target->event_mutex); 1117 } 1118