xref: /linux/drivers/gpu/drm/amd/amdkfd/kfd_debug.c (revision 52a5a22d8afe3bd195f7b470c7535c63717f5ff7)
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 #include <uapi/linux/kfd_sysfs.h>
29 
30 #define MAX_WATCH_ADDRESSES	4
31 
32 int kfd_dbg_ev_query_debug_event(struct kfd_process *process,
33 		      unsigned int *queue_id,
34 		      unsigned int *gpu_id,
35 		      uint64_t exception_clear_mask,
36 		      uint64_t *event_status)
37 {
38 	struct process_queue_manager *pqm;
39 	struct process_queue_node *pqn;
40 	int i;
41 
42 	if (!(process && process->debug_trap_enabled))
43 		return -ENODATA;
44 
45 	mutex_lock(&process->event_mutex);
46 	*event_status = 0;
47 	*queue_id = 0;
48 	*gpu_id = 0;
49 
50 	/* find and report queue events */
51 	pqm = &process->pqm;
52 	list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
53 		uint64_t tmp = process->exception_enable_mask;
54 
55 		if (!pqn->q)
56 			continue;
57 
58 		tmp &= pqn->q->properties.exception_status;
59 
60 		if (!tmp)
61 			continue;
62 
63 		*event_status = pqn->q->properties.exception_status;
64 		*queue_id = pqn->q->properties.queue_id;
65 		*gpu_id = pqn->q->device->id;
66 		pqn->q->properties.exception_status &= ~exception_clear_mask;
67 		goto out;
68 	}
69 
70 	/* find and report device events */
71 	for (i = 0; i < process->n_pdds; i++) {
72 		struct kfd_process_device *pdd = process->pdds[i];
73 		uint64_t tmp = process->exception_enable_mask
74 						& pdd->exception_status;
75 
76 		if (!tmp)
77 			continue;
78 
79 		*event_status = pdd->exception_status;
80 		*gpu_id = pdd->dev->id;
81 		pdd->exception_status &= ~exception_clear_mask;
82 		goto out;
83 	}
84 
85 	/* report process events */
86 	if (process->exception_enable_mask & process->exception_status) {
87 		*event_status = process->exception_status;
88 		process->exception_status &= ~exception_clear_mask;
89 	}
90 
91 out:
92 	mutex_unlock(&process->event_mutex);
93 	return *event_status ? 0 : -EAGAIN;
94 }
95 
96 void debug_event_write_work_handler(struct work_struct *work)
97 {
98 	struct kfd_process *process;
99 
100 	static const char write_data = '.';
101 	loff_t pos = 0;
102 
103 	process = container_of(work,
104 			struct kfd_process,
105 			debug_event_workarea);
106 
107 	if (process->debug_trap_enabled && process->dbg_ev_file)
108 		kernel_write(process->dbg_ev_file, &write_data, 1, &pos);
109 }
110 
111 /* update process/device/queue exception status, write to descriptor
112  * only if exception_status is enabled.
113  */
114 bool kfd_dbg_ev_raise(uint64_t event_mask,
115 			struct kfd_process *process, struct kfd_node *dev,
116 			unsigned int source_id, bool use_worker,
117 			void *exception_data, size_t exception_data_size)
118 {
119 	struct process_queue_manager *pqm;
120 	struct process_queue_node *pqn;
121 	int i;
122 	static const char write_data = '.';
123 	loff_t pos = 0;
124 	bool is_subscribed = true;
125 
126 	if (!(process && process->debug_trap_enabled))
127 		return false;
128 
129 	mutex_lock(&process->event_mutex);
130 
131 	if (event_mask & KFD_EC_MASK_DEVICE) {
132 		for (i = 0; i < process->n_pdds; i++) {
133 			struct kfd_process_device *pdd = process->pdds[i];
134 
135 			if (pdd->dev != dev)
136 				continue;
137 
138 			pdd->exception_status |= event_mask & KFD_EC_MASK_DEVICE;
139 
140 			if (event_mask & KFD_EC_MASK(EC_DEVICE_MEMORY_VIOLATION)) {
141 				if (!pdd->vm_fault_exc_data) {
142 					pdd->vm_fault_exc_data = kmemdup(
143 							exception_data,
144 							exception_data_size,
145 							GFP_KERNEL);
146 					if (!pdd->vm_fault_exc_data)
147 						pr_debug("Failed to allocate exception data memory");
148 				} else {
149 					pr_debug("Debugger exception data not saved\n");
150 					print_hex_dump_bytes("exception data: ",
151 							DUMP_PREFIX_OFFSET,
152 							exception_data,
153 							exception_data_size);
154 				}
155 			}
156 			break;
157 		}
158 	} else if (event_mask & KFD_EC_MASK_PROCESS) {
159 		process->exception_status |= event_mask & KFD_EC_MASK_PROCESS;
160 	} else {
161 		pqm = &process->pqm;
162 		list_for_each_entry(pqn, &pqm->queues,
163 				process_queue_list) {
164 			int target_id;
165 
166 			if (!pqn->q)
167 				continue;
168 
169 			target_id = event_mask & KFD_EC_MASK(EC_QUEUE_NEW) ?
170 					pqn->q->properties.queue_id :
171 							pqn->q->doorbell_id;
172 
173 			if (pqn->q->device != dev || target_id != source_id)
174 				continue;
175 
176 			pqn->q->properties.exception_status |= event_mask;
177 			break;
178 		}
179 	}
180 
181 	if (process->exception_enable_mask & event_mask) {
182 		if (use_worker)
183 			schedule_work(&process->debug_event_workarea);
184 		else
185 			kernel_write(process->dbg_ev_file,
186 					&write_data,
187 					1,
188 					&pos);
189 	} else {
190 		is_subscribed = false;
191 	}
192 
193 	mutex_unlock(&process->event_mutex);
194 
195 	return is_subscribed;
196 }
197 
198 /* set pending event queue entry from ring entry  */
199 bool kfd_set_dbg_ev_from_interrupt(struct kfd_node *dev,
200 				   unsigned int pasid,
201 				   uint32_t doorbell_id,
202 				   uint64_t trap_mask,
203 				   void *exception_data,
204 				   size_t exception_data_size)
205 {
206 	struct kfd_process *p;
207 	bool signaled_to_debugger_or_runtime = false;
208 
209 	p = kfd_lookup_process_by_pasid(pasid);
210 
211 	if (!p)
212 		return false;
213 
214 	if (!kfd_dbg_ev_raise(trap_mask, p, dev, doorbell_id, true,
215 			      exception_data, exception_data_size)) {
216 		struct process_queue_manager *pqm;
217 		struct process_queue_node *pqn;
218 
219 		if (!!(trap_mask & KFD_EC_MASK_QUEUE) &&
220 		       p->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED) {
221 			mutex_lock(&p->mutex);
222 
223 			pqm = &p->pqm;
224 			list_for_each_entry(pqn, &pqm->queues,
225 							process_queue_list) {
226 
227 				if (!(pqn->q && pqn->q->device == dev &&
228 				      pqn->q->doorbell_id == doorbell_id))
229 					continue;
230 
231 				kfd_send_exception_to_runtime(p, pqn->q->properties.queue_id,
232 							      trap_mask);
233 
234 				signaled_to_debugger_or_runtime = true;
235 
236 				break;
237 			}
238 
239 			mutex_unlock(&p->mutex);
240 		} else if (trap_mask & KFD_EC_MASK(EC_DEVICE_MEMORY_VIOLATION)) {
241 			kfd_dqm_evict_pasid(dev->dqm, p->pasid);
242 			kfd_signal_vm_fault_event(dev, p->pasid, NULL,
243 							exception_data);
244 
245 			signaled_to_debugger_or_runtime = true;
246 		}
247 	} else {
248 		signaled_to_debugger_or_runtime = true;
249 	}
250 
251 	kfd_unref_process(p);
252 
253 	return signaled_to_debugger_or_runtime;
254 }
255 
256 int kfd_dbg_send_exception_to_runtime(struct kfd_process *p,
257 					unsigned int dev_id,
258 					unsigned int queue_id,
259 					uint64_t error_reason)
260 {
261 	if (error_reason & KFD_EC_MASK(EC_DEVICE_MEMORY_VIOLATION)) {
262 		struct kfd_process_device *pdd = NULL;
263 		struct kfd_hsa_memory_exception_data *data;
264 		int i;
265 
266 		for (i = 0; i < p->n_pdds; i++) {
267 			if (p->pdds[i]->dev->id == dev_id) {
268 				pdd = p->pdds[i];
269 				break;
270 			}
271 		}
272 
273 		if (!pdd)
274 			return -ENODEV;
275 
276 		data = (struct kfd_hsa_memory_exception_data *)
277 						pdd->vm_fault_exc_data;
278 
279 		kfd_dqm_evict_pasid(pdd->dev->dqm, p->pasid);
280 		kfd_signal_vm_fault_event(pdd->dev, p->pasid, NULL, data);
281 		error_reason &= ~KFD_EC_MASK(EC_DEVICE_MEMORY_VIOLATION);
282 	}
283 
284 	if (error_reason & (KFD_EC_MASK(EC_PROCESS_RUNTIME))) {
285 		/*
286 		 * block should only happen after the debugger receives runtime
287 		 * enable notice.
288 		 */
289 		up(&p->runtime_enable_sema);
290 		error_reason &= ~KFD_EC_MASK(EC_PROCESS_RUNTIME);
291 	}
292 
293 	if (error_reason)
294 		return kfd_send_exception_to_runtime(p, queue_id, error_reason);
295 
296 	return 0;
297 }
298 
299 static int kfd_dbg_set_queue_workaround(struct queue *q, bool enable)
300 {
301 	struct mqd_update_info minfo = {0};
302 	int err;
303 
304 	if (!q)
305 		return 0;
306 
307 	if (!kfd_dbg_has_cwsr_workaround(q->device))
308 		return 0;
309 
310 	if (enable && q->properties.is_user_cu_masked)
311 		return -EBUSY;
312 
313 	minfo.update_flag = enable ? UPDATE_FLAG_DBG_WA_ENABLE : UPDATE_FLAG_DBG_WA_DISABLE;
314 
315 	q->properties.is_dbg_wa = enable;
316 	err = q->device->dqm->ops.update_queue(q->device->dqm, q, &minfo);
317 	if (err)
318 		q->properties.is_dbg_wa = false;
319 
320 	return err;
321 }
322 
323 static int kfd_dbg_set_workaround(struct kfd_process *target, bool enable)
324 {
325 	struct process_queue_manager *pqm = &target->pqm;
326 	struct process_queue_node *pqn;
327 	int r = 0;
328 
329 	list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
330 		r = kfd_dbg_set_queue_workaround(pqn->q, enable);
331 		if (enable && r)
332 			goto unwind;
333 	}
334 
335 	return 0;
336 
337 unwind:
338 	list_for_each_entry(pqn, &pqm->queues, process_queue_list)
339 		kfd_dbg_set_queue_workaround(pqn->q, false);
340 
341 	if (enable)
342 		target->runtime_info.runtime_state = r == -EBUSY ?
343 				DEBUG_RUNTIME_STATE_ENABLED_BUSY :
344 				DEBUG_RUNTIME_STATE_ENABLED_ERROR;
345 
346 	return r;
347 }
348 
349 int kfd_dbg_set_mes_debug_mode(struct kfd_process_device *pdd, bool sq_trap_en)
350 {
351 	uint32_t spi_dbg_cntl = pdd->spi_dbg_override | pdd->spi_dbg_launch_mode;
352 	uint32_t flags = pdd->process->dbg_flags;
353 	struct amdgpu_device *adev = pdd->dev->adev;
354 	int r;
355 
356 	if (!kfd_dbg_is_per_vmid_supported(pdd->dev))
357 		return 0;
358 
359 	if (!pdd->proc_ctx_cpu_ptr) {
360 			r = amdgpu_amdkfd_alloc_gtt_mem(adev,
361 				AMDGPU_MES_PROC_CTX_SIZE,
362 				&pdd->proc_ctx_bo,
363 				&pdd->proc_ctx_gpu_addr,
364 				&pdd->proc_ctx_cpu_ptr,
365 				false);
366 		if (r) {
367 			dev_err(adev->dev,
368 			"failed to allocate process context bo\n");
369 			return r;
370 		}
371 		memset(pdd->proc_ctx_cpu_ptr, 0, AMDGPU_MES_PROC_CTX_SIZE);
372 	}
373 
374 	return amdgpu_mes_set_shader_debugger(pdd->dev->adev, pdd->proc_ctx_gpu_addr, spi_dbg_cntl,
375 						pdd->watch_points, flags, sq_trap_en);
376 }
377 
378 #define KFD_DEBUGGER_INVALID_WATCH_POINT_ID -1
379 static int kfd_dbg_get_dev_watch_id(struct kfd_process_device *pdd, int *watch_id)
380 {
381 	int i;
382 
383 	*watch_id = KFD_DEBUGGER_INVALID_WATCH_POINT_ID;
384 
385 	spin_lock(&pdd->dev->watch_points_lock);
386 
387 	for (i = 0; i < MAX_WATCH_ADDRESSES; i++) {
388 		/* device watchpoint in use so skip */
389 		if ((pdd->dev->alloc_watch_ids >> i) & 0x1)
390 			continue;
391 
392 		pdd->alloc_watch_ids |= 0x1 << i;
393 		pdd->dev->alloc_watch_ids |= 0x1 << i;
394 		*watch_id = i;
395 		spin_unlock(&pdd->dev->watch_points_lock);
396 		return 0;
397 	}
398 
399 	spin_unlock(&pdd->dev->watch_points_lock);
400 
401 	return -ENOMEM;
402 }
403 
404 static void kfd_dbg_clear_dev_watch_id(struct kfd_process_device *pdd, int watch_id)
405 {
406 	spin_lock(&pdd->dev->watch_points_lock);
407 
408 	/* process owns device watch point so safe to clear */
409 	if ((pdd->alloc_watch_ids >> watch_id) & 0x1) {
410 		pdd->alloc_watch_ids &= ~(0x1 << watch_id);
411 		pdd->dev->alloc_watch_ids &= ~(0x1 << watch_id);
412 	}
413 
414 	spin_unlock(&pdd->dev->watch_points_lock);
415 }
416 
417 static bool kfd_dbg_owns_dev_watch_id(struct kfd_process_device *pdd, int watch_id)
418 {
419 	bool owns_watch_id = false;
420 
421 	spin_lock(&pdd->dev->watch_points_lock);
422 	owns_watch_id = watch_id < MAX_WATCH_ADDRESSES &&
423 			((pdd->alloc_watch_ids >> watch_id) & 0x1);
424 
425 	spin_unlock(&pdd->dev->watch_points_lock);
426 
427 	return owns_watch_id;
428 }
429 
430 int kfd_dbg_trap_clear_dev_address_watch(struct kfd_process_device *pdd,
431 					uint32_t watch_id)
432 {
433 	int r;
434 
435 	if (!kfd_dbg_owns_dev_watch_id(pdd, watch_id))
436 		return -EINVAL;
437 
438 	if (!pdd->dev->kfd->shared_resources.enable_mes) {
439 		r = debug_lock_and_unmap(pdd->dev->dqm);
440 		if (r)
441 			return r;
442 	}
443 
444 	amdgpu_gfx_off_ctrl(pdd->dev->adev, false);
445 	pdd->watch_points[watch_id] = pdd->dev->kfd2kgd->clear_address_watch(
446 							pdd->dev->adev,
447 							watch_id);
448 	amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
449 
450 	if (!pdd->dev->kfd->shared_resources.enable_mes)
451 		r = debug_map_and_unlock(pdd->dev->dqm);
452 	else
453 		r = kfd_dbg_set_mes_debug_mode(pdd, true);
454 
455 	kfd_dbg_clear_dev_watch_id(pdd, watch_id);
456 
457 	return r;
458 }
459 
460 int kfd_dbg_trap_set_dev_address_watch(struct kfd_process_device *pdd,
461 					uint64_t watch_address,
462 					uint32_t watch_address_mask,
463 					uint32_t *watch_id,
464 					uint32_t watch_mode)
465 {
466 	int xcc_id, r = kfd_dbg_get_dev_watch_id(pdd, watch_id);
467 	uint32_t xcc_mask = pdd->dev->xcc_mask;
468 
469 	if (r)
470 		return r;
471 
472 	if (!pdd->dev->kfd->shared_resources.enable_mes) {
473 		r = debug_lock_and_unmap(pdd->dev->dqm);
474 		if (r) {
475 			kfd_dbg_clear_dev_watch_id(pdd, *watch_id);
476 			return r;
477 		}
478 	}
479 
480 	amdgpu_gfx_off_ctrl(pdd->dev->adev, false);
481 	for_each_inst(xcc_id, xcc_mask)
482 		pdd->watch_points[*watch_id] = pdd->dev->kfd2kgd->set_address_watch(
483 				pdd->dev->adev,
484 				watch_address,
485 				watch_address_mask,
486 				*watch_id,
487 				watch_mode,
488 				pdd->dev->vm_info.last_vmid_kfd,
489 				xcc_id);
490 	amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
491 
492 	if (!pdd->dev->kfd->shared_resources.enable_mes)
493 		r = debug_map_and_unlock(pdd->dev->dqm);
494 	else
495 		r = kfd_dbg_set_mes_debug_mode(pdd, true);
496 
497 	/* HWS is broken so no point in HW rollback but release the watchpoint anyways */
498 	if (r)
499 		kfd_dbg_clear_dev_watch_id(pdd, *watch_id);
500 
501 	return 0;
502 }
503 
504 static void kfd_dbg_clear_process_address_watch(struct kfd_process *target)
505 {
506 	int i, j;
507 
508 	for (i = 0; i < target->n_pdds; i++)
509 		for (j = 0; j < MAX_WATCH_ADDRESSES; j++)
510 			kfd_dbg_trap_clear_dev_address_watch(target->pdds[i], j);
511 }
512 
513 int kfd_dbg_trap_set_flags(struct kfd_process *target, uint32_t *flags)
514 {
515 	uint32_t prev_flags = target->dbg_flags;
516 	int i, r = 0, rewind_count = 0;
517 
518 	for (i = 0; i < target->n_pdds; i++) {
519 		struct kfd_topology_device *topo_dev =
520 				kfd_topology_device_by_id(target->pdds[i]->dev->id);
521 		uint32_t caps = topo_dev->node_props.capability;
522 
523 		if (!(caps & HSA_CAP_TRAP_DEBUG_PRECISE_MEMORY_OPERATIONS_SUPPORTED) &&
524 			(*flags & KFD_DBG_TRAP_FLAG_SINGLE_MEM_OP)) {
525 			*flags = prev_flags;
526 			return -EACCES;
527 		}
528 
529 		if (!(caps & HSA_CAP_TRAP_DEBUG_PRECISE_ALU_OPERATIONS_SUPPORTED) &&
530 		    (*flags & KFD_DBG_TRAP_FLAG_SINGLE_ALU_OP)) {
531 			*flags = prev_flags;
532 			return -EACCES;
533 		}
534 	}
535 
536 	target->dbg_flags = *flags;
537 	*flags = prev_flags;
538 	for (i = 0; i < target->n_pdds; i++) {
539 		struct kfd_process_device *pdd = target->pdds[i];
540 
541 		if (!kfd_dbg_is_per_vmid_supported(pdd->dev))
542 			continue;
543 
544 		if (!pdd->dev->kfd->shared_resources.enable_mes)
545 			r = debug_refresh_runlist(pdd->dev->dqm);
546 		else
547 			r = kfd_dbg_set_mes_debug_mode(pdd, true);
548 
549 		if (r) {
550 			target->dbg_flags = prev_flags;
551 			break;
552 		}
553 
554 		rewind_count++;
555 	}
556 
557 	/* Rewind flags */
558 	if (r) {
559 		target->dbg_flags = prev_flags;
560 
561 		for (i = 0; i < rewind_count; i++) {
562 			struct kfd_process_device *pdd = target->pdds[i];
563 
564 			if (!kfd_dbg_is_per_vmid_supported(pdd->dev))
565 				continue;
566 
567 			if (!pdd->dev->kfd->shared_resources.enable_mes)
568 				debug_refresh_runlist(pdd->dev->dqm);
569 			else
570 				kfd_dbg_set_mes_debug_mode(pdd, true);
571 		}
572 	}
573 
574 	return r;
575 }
576 
577 /* kfd_dbg_trap_deactivate:
578  *	target: target process
579  *	unwind: If this is unwinding a failed kfd_dbg_trap_enable()
580  *	unwind_count:
581  *		If unwind == true, how far down the pdd list we need
582  *				to unwind
583  *		else: ignored
584  */
585 void kfd_dbg_trap_deactivate(struct kfd_process *target, bool unwind, int unwind_count)
586 {
587 	int i;
588 
589 	if (!unwind) {
590 		uint32_t flags = 0;
591 		int resume_count = resume_queues(target, 0, NULL);
592 
593 		if (resume_count)
594 			pr_debug("Resumed %d queues\n", resume_count);
595 
596 		cancel_work_sync(&target->debug_event_workarea);
597 		kfd_dbg_clear_process_address_watch(target);
598 		kfd_dbg_trap_set_wave_launch_mode(target, 0);
599 
600 		kfd_dbg_trap_set_flags(target, &flags);
601 	}
602 
603 	for (i = 0; i < target->n_pdds; i++) {
604 		struct kfd_process_device *pdd = target->pdds[i];
605 
606 		/* If this is an unwind, and we have unwound the required
607 		 * enable calls on the pdd list, we need to stop now
608 		 * otherwise we may mess up another debugger session.
609 		 */
610 		if (unwind && i == unwind_count)
611 			break;
612 
613 		kfd_process_set_trap_debug_flag(&pdd->qpd, false);
614 
615 		/* GFX off is already disabled by debug activate if not RLC restore supported. */
616 		if (kfd_dbg_is_rlc_restore_supported(pdd->dev))
617 			amdgpu_gfx_off_ctrl(pdd->dev->adev, false);
618 		pdd->spi_dbg_override =
619 				pdd->dev->kfd2kgd->disable_debug_trap(
620 				pdd->dev->adev,
621 				target->runtime_info.ttmp_setup,
622 				pdd->dev->vm_info.last_vmid_kfd);
623 		amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
624 
625 		if (!kfd_dbg_is_per_vmid_supported(pdd->dev) &&
626 				release_debug_trap_vmid(pdd->dev->dqm, &pdd->qpd))
627 			pr_err("Failed to release debug vmid on [%i]\n", pdd->dev->id);
628 
629 		if (!pdd->dev->kfd->shared_resources.enable_mes)
630 			debug_refresh_runlist(pdd->dev->dqm);
631 		else
632 			kfd_dbg_set_mes_debug_mode(pdd, !kfd_dbg_has_cwsr_workaround(pdd->dev));
633 	}
634 
635 	kfd_dbg_set_workaround(target, false);
636 }
637 
638 static void kfd_dbg_clean_exception_status(struct kfd_process *target)
639 {
640 	struct process_queue_manager *pqm;
641 	struct process_queue_node *pqn;
642 	int i;
643 
644 	for (i = 0; i < target->n_pdds; i++) {
645 		struct kfd_process_device *pdd = target->pdds[i];
646 
647 		kfd_process_drain_interrupts(pdd);
648 
649 		pdd->exception_status = 0;
650 	}
651 
652 	pqm = &target->pqm;
653 	list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
654 		if (!pqn->q)
655 			continue;
656 
657 		pqn->q->properties.exception_status = 0;
658 	}
659 
660 	target->exception_status = 0;
661 }
662 
663 int kfd_dbg_trap_disable(struct kfd_process *target)
664 {
665 	if (!target->debug_trap_enabled)
666 		return 0;
667 
668 	/*
669 	 * Defer deactivation to runtime if runtime not enabled otherwise reset
670 	 * attached running target runtime state to enable for re-attach.
671 	 */
672 	if (target->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED)
673 		kfd_dbg_trap_deactivate(target, false, 0);
674 	else if (target->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_DISABLED)
675 		target->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_ENABLED;
676 
677 	cancel_work_sync(&target->debug_event_workarea);
678 	fput(target->dbg_ev_file);
679 	target->dbg_ev_file = NULL;
680 
681 	if (target->debugger_process) {
682 		atomic_dec(&target->debugger_process->debugged_process_count);
683 		target->debugger_process = NULL;
684 	}
685 
686 	target->debug_trap_enabled = false;
687 	kfd_dbg_clean_exception_status(target);
688 	kfd_unref_process(target);
689 
690 	return 0;
691 }
692 
693 int kfd_dbg_trap_activate(struct kfd_process *target)
694 {
695 	int i, r = 0;
696 
697 	r = kfd_dbg_set_workaround(target, true);
698 	if (r)
699 		return r;
700 
701 	for (i = 0; i < target->n_pdds; i++) {
702 		struct kfd_process_device *pdd = target->pdds[i];
703 
704 		if (!kfd_dbg_is_per_vmid_supported(pdd->dev)) {
705 			r = reserve_debug_trap_vmid(pdd->dev->dqm, &pdd->qpd);
706 
707 			if (r) {
708 				target->runtime_info.runtime_state = (r == -EBUSY) ?
709 							DEBUG_RUNTIME_STATE_ENABLED_BUSY :
710 							DEBUG_RUNTIME_STATE_ENABLED_ERROR;
711 
712 				goto unwind_err;
713 			}
714 		}
715 
716 		/* Disable GFX OFF to prevent garbage read/writes to debug registers.
717 		 * If RLC restore of debug registers is not supported and runtime enable
718 		 * hasn't done so already on ttmp setup request, restore the trap config registers.
719 		 *
720 		 * If RLC restore of debug registers is not supported, keep gfx off disabled for
721 		 * the debug session.
722 		 */
723 		amdgpu_gfx_off_ctrl(pdd->dev->adev, false);
724 		if (!(kfd_dbg_is_rlc_restore_supported(pdd->dev) ||
725 						target->runtime_info.ttmp_setup))
726 			pdd->dev->kfd2kgd->enable_debug_trap(pdd->dev->adev, true,
727 								pdd->dev->vm_info.last_vmid_kfd);
728 
729 		pdd->spi_dbg_override = pdd->dev->kfd2kgd->enable_debug_trap(
730 					pdd->dev->adev,
731 					false,
732 					pdd->dev->vm_info.last_vmid_kfd);
733 
734 		if (kfd_dbg_is_rlc_restore_supported(pdd->dev))
735 			amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
736 
737 		/*
738 		 * Setting the debug flag in the trap handler requires that the TMA has been
739 		 * allocated, which occurs during CWSR initialization.
740 		 * In the event that CWSR has not been initialized at this point, setting the
741 		 * flag will be called again during CWSR initialization if the target process
742 		 * is still debug enabled.
743 		 */
744 		kfd_process_set_trap_debug_flag(&pdd->qpd, true);
745 
746 		if (!pdd->dev->kfd->shared_resources.enable_mes)
747 			r = debug_refresh_runlist(pdd->dev->dqm);
748 		else
749 			r = kfd_dbg_set_mes_debug_mode(pdd, true);
750 
751 		if (r) {
752 			target->runtime_info.runtime_state =
753 					DEBUG_RUNTIME_STATE_ENABLED_ERROR;
754 			goto unwind_err;
755 		}
756 	}
757 
758 	return 0;
759 
760 unwind_err:
761 	/* Enabling debug failed, we need to disable on
762 	 * all GPUs so the enable is all or nothing.
763 	 */
764 	kfd_dbg_trap_deactivate(target, true, i);
765 	return r;
766 }
767 
768 int kfd_dbg_trap_enable(struct kfd_process *target, uint32_t fd,
769 			void __user *runtime_info, uint32_t *runtime_size)
770 {
771 	struct file *f;
772 	uint32_t copy_size;
773 	int i, r = 0;
774 
775 	if (target->debug_trap_enabled)
776 		return -EALREADY;
777 
778 	/* Enable pre-checks */
779 	for (i = 0; i < target->n_pdds; i++) {
780 		struct kfd_process_device *pdd = target->pdds[i];
781 
782 		if (!KFD_IS_SOC15(pdd->dev))
783 			return -ENODEV;
784 
785 		if (pdd->qpd.num_gws && (!kfd_dbg_has_gws_support(pdd->dev) ||
786 					 kfd_dbg_has_cwsr_workaround(pdd->dev)))
787 			return -EBUSY;
788 	}
789 
790 	copy_size = min((size_t)(*runtime_size), sizeof(target->runtime_info));
791 
792 	f = fget(fd);
793 	if (!f) {
794 		pr_err("Failed to get file for (%i)\n", fd);
795 		return -EBADF;
796 	}
797 
798 	target->dbg_ev_file = f;
799 
800 	/* defer activation to runtime if not runtime enabled */
801 	if (target->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED)
802 		kfd_dbg_trap_activate(target);
803 
804 	/* We already hold the process reference but hold another one for the
805 	 * debug session.
806 	 */
807 	kref_get(&target->ref);
808 	target->debug_trap_enabled = true;
809 
810 	if (target->debugger_process)
811 		atomic_inc(&target->debugger_process->debugged_process_count);
812 
813 	if (copy_to_user(runtime_info, (void *)&target->runtime_info, copy_size)) {
814 		kfd_dbg_trap_deactivate(target, false, 0);
815 		r = -EFAULT;
816 	}
817 
818 	*runtime_size = sizeof(target->runtime_info);
819 
820 	return r;
821 }
822 
823 static int kfd_dbg_validate_trap_override_request(struct kfd_process *p,
824 						uint32_t trap_override,
825 						uint32_t trap_mask_request,
826 						uint32_t *trap_mask_supported)
827 {
828 	int i = 0;
829 
830 	*trap_mask_supported = 0xffffffff;
831 
832 	for (i = 0; i < p->n_pdds; i++) {
833 		struct kfd_process_device *pdd = p->pdds[i];
834 		int err = pdd->dev->kfd2kgd->validate_trap_override_request(
835 								pdd->dev->adev,
836 								trap_override,
837 								trap_mask_supported);
838 
839 		if (err)
840 			return err;
841 	}
842 
843 	if (trap_mask_request & ~*trap_mask_supported)
844 		return -EACCES;
845 
846 	return 0;
847 }
848 
849 int kfd_dbg_trap_set_wave_launch_override(struct kfd_process *target,
850 					uint32_t trap_override,
851 					uint32_t trap_mask_bits,
852 					uint32_t trap_mask_request,
853 					uint32_t *trap_mask_prev,
854 					uint32_t *trap_mask_supported)
855 {
856 	int r = 0, i;
857 
858 	r = kfd_dbg_validate_trap_override_request(target,
859 						trap_override,
860 						trap_mask_request,
861 						trap_mask_supported);
862 
863 	if (r)
864 		return r;
865 
866 	for (i = 0; i < target->n_pdds; i++) {
867 		struct kfd_process_device *pdd = target->pdds[i];
868 
869 		amdgpu_gfx_off_ctrl(pdd->dev->adev, false);
870 		pdd->spi_dbg_override = pdd->dev->kfd2kgd->set_wave_launch_trap_override(
871 				pdd->dev->adev,
872 				pdd->dev->vm_info.last_vmid_kfd,
873 				trap_override,
874 				trap_mask_bits,
875 				trap_mask_request,
876 				trap_mask_prev,
877 				pdd->spi_dbg_override);
878 		amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
879 
880 		if (!pdd->dev->kfd->shared_resources.enable_mes)
881 			r = debug_refresh_runlist(pdd->dev->dqm);
882 		else
883 			r = kfd_dbg_set_mes_debug_mode(pdd, true);
884 
885 		if (r)
886 			break;
887 	}
888 
889 	return r;
890 }
891 
892 int kfd_dbg_trap_set_wave_launch_mode(struct kfd_process *target,
893 					uint8_t wave_launch_mode)
894 {
895 	int r = 0, i;
896 
897 	if (wave_launch_mode != KFD_DBG_TRAP_WAVE_LAUNCH_MODE_NORMAL &&
898 			wave_launch_mode != KFD_DBG_TRAP_WAVE_LAUNCH_MODE_HALT &&
899 			wave_launch_mode != KFD_DBG_TRAP_WAVE_LAUNCH_MODE_DEBUG)
900 		return -EINVAL;
901 
902 	for (i = 0; i < target->n_pdds; i++) {
903 		struct kfd_process_device *pdd = target->pdds[i];
904 
905 		amdgpu_gfx_off_ctrl(pdd->dev->adev, false);
906 		pdd->spi_dbg_launch_mode = pdd->dev->kfd2kgd->set_wave_launch_mode(
907 				pdd->dev->adev,
908 				wave_launch_mode,
909 				pdd->dev->vm_info.last_vmid_kfd);
910 		amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
911 
912 		if (!pdd->dev->kfd->shared_resources.enable_mes)
913 			r = debug_refresh_runlist(pdd->dev->dqm);
914 		else
915 			r = kfd_dbg_set_mes_debug_mode(pdd, true);
916 
917 		if (r)
918 			break;
919 	}
920 
921 	return r;
922 }
923 
924 int kfd_dbg_trap_query_exception_info(struct kfd_process *target,
925 		uint32_t source_id,
926 		uint32_t exception_code,
927 		bool clear_exception,
928 		void __user *info,
929 		uint32_t *info_size)
930 {
931 	bool found = false;
932 	int r = 0;
933 	uint32_t copy_size, actual_info_size = 0;
934 	uint64_t *exception_status_ptr = NULL;
935 
936 	if (!target)
937 		return -EINVAL;
938 
939 	if (!info || !info_size)
940 		return -EINVAL;
941 
942 	mutex_lock(&target->event_mutex);
943 
944 	if (KFD_DBG_EC_TYPE_IS_QUEUE(exception_code)) {
945 		/* Per queue exceptions */
946 		struct queue *queue = NULL;
947 		int i;
948 
949 		for (i = 0; i < target->n_pdds; i++) {
950 			struct kfd_process_device *pdd = target->pdds[i];
951 			struct qcm_process_device *qpd = &pdd->qpd;
952 
953 			list_for_each_entry(queue, &qpd->queues_list, list) {
954 				if (!found && queue->properties.queue_id == source_id) {
955 					found = true;
956 					break;
957 				}
958 			}
959 			if (found)
960 				break;
961 		}
962 
963 		if (!found) {
964 			r = -EINVAL;
965 			goto out;
966 		}
967 
968 		if (!(queue->properties.exception_status & KFD_EC_MASK(exception_code))) {
969 			r = -ENODATA;
970 			goto out;
971 		}
972 		exception_status_ptr = &queue->properties.exception_status;
973 	} else if (KFD_DBG_EC_TYPE_IS_DEVICE(exception_code)) {
974 		/* Per device exceptions */
975 		struct kfd_process_device *pdd = NULL;
976 		int i;
977 
978 		for (i = 0; i < target->n_pdds; i++) {
979 			pdd = target->pdds[i];
980 			if (pdd->dev->id == source_id) {
981 				found = true;
982 				break;
983 			}
984 		}
985 
986 		if (!found) {
987 			r = -EINVAL;
988 			goto out;
989 		}
990 
991 		if (!(pdd->exception_status & KFD_EC_MASK(exception_code))) {
992 			r = -ENODATA;
993 			goto out;
994 		}
995 
996 		if (exception_code == EC_DEVICE_MEMORY_VIOLATION) {
997 			copy_size = min((size_t)(*info_size), pdd->vm_fault_exc_data_size);
998 
999 			if (copy_to_user(info, pdd->vm_fault_exc_data, copy_size)) {
1000 				r = -EFAULT;
1001 				goto out;
1002 			}
1003 			actual_info_size = pdd->vm_fault_exc_data_size;
1004 			if (clear_exception) {
1005 				kfree(pdd->vm_fault_exc_data);
1006 				pdd->vm_fault_exc_data = NULL;
1007 				pdd->vm_fault_exc_data_size = 0;
1008 			}
1009 		}
1010 		exception_status_ptr = &pdd->exception_status;
1011 	} else if (KFD_DBG_EC_TYPE_IS_PROCESS(exception_code)) {
1012 		/* Per process exceptions */
1013 		if (!(target->exception_status & KFD_EC_MASK(exception_code))) {
1014 			r = -ENODATA;
1015 			goto out;
1016 		}
1017 
1018 		if (exception_code == EC_PROCESS_RUNTIME) {
1019 			copy_size = min((size_t)(*info_size), sizeof(target->runtime_info));
1020 
1021 			if (copy_to_user(info, (void *)&target->runtime_info, copy_size)) {
1022 				r = -EFAULT;
1023 				goto out;
1024 			}
1025 
1026 			actual_info_size = sizeof(target->runtime_info);
1027 		}
1028 
1029 		exception_status_ptr = &target->exception_status;
1030 	} else {
1031 		pr_debug("Bad exception type [%i]\n", exception_code);
1032 		r = -EINVAL;
1033 		goto out;
1034 	}
1035 
1036 	*info_size = actual_info_size;
1037 	if (clear_exception)
1038 		*exception_status_ptr &= ~KFD_EC_MASK(exception_code);
1039 out:
1040 	mutex_unlock(&target->event_mutex);
1041 	return r;
1042 }
1043 
1044 int kfd_dbg_trap_device_snapshot(struct kfd_process *target,
1045 		uint64_t exception_clear_mask,
1046 		void __user *user_info,
1047 		uint32_t *number_of_device_infos,
1048 		uint32_t *entry_size)
1049 {
1050 	struct kfd_dbg_device_info_entry device_info;
1051 	uint32_t tmp_entry_size, tmp_num_devices;
1052 	int i, r = 0;
1053 
1054 	if (!(target && user_info && number_of_device_infos && entry_size))
1055 		return -EINVAL;
1056 
1057 	tmp_entry_size = *entry_size;
1058 
1059 	tmp_num_devices = min_t(size_t, *number_of_device_infos, target->n_pdds);
1060 	*number_of_device_infos = target->n_pdds;
1061 	*entry_size = min_t(size_t, *entry_size, sizeof(device_info));
1062 
1063 	if (!tmp_num_devices)
1064 		return 0;
1065 
1066 	memset(&device_info, 0, sizeof(device_info));
1067 
1068 	mutex_lock(&target->event_mutex);
1069 
1070 	/* Run over all pdd of the process */
1071 	for (i = 0; i < tmp_num_devices; i++) {
1072 		struct kfd_process_device *pdd = target->pdds[i];
1073 		struct kfd_topology_device *topo_dev = kfd_topology_device_by_id(pdd->dev->id);
1074 
1075 		device_info.gpu_id = pdd->dev->id;
1076 		device_info.exception_status = pdd->exception_status;
1077 		device_info.lds_base = pdd->lds_base;
1078 		device_info.lds_limit = pdd->lds_limit;
1079 		device_info.scratch_base = pdd->scratch_base;
1080 		device_info.scratch_limit = pdd->scratch_limit;
1081 		device_info.gpuvm_base = pdd->gpuvm_base;
1082 		device_info.gpuvm_limit = pdd->gpuvm_limit;
1083 		device_info.location_id = topo_dev->node_props.location_id;
1084 		device_info.vendor_id = topo_dev->node_props.vendor_id;
1085 		device_info.device_id = topo_dev->node_props.device_id;
1086 		device_info.revision_id = pdd->dev->adev->pdev->revision;
1087 		device_info.subsystem_vendor_id = pdd->dev->adev->pdev->subsystem_vendor;
1088 		device_info.subsystem_device_id = pdd->dev->adev->pdev->subsystem_device;
1089 		device_info.fw_version = pdd->dev->kfd->mec_fw_version;
1090 		device_info.gfx_target_version =
1091 			topo_dev->node_props.gfx_target_version;
1092 		device_info.simd_count = topo_dev->node_props.simd_count;
1093 		device_info.max_waves_per_simd =
1094 			topo_dev->node_props.max_waves_per_simd;
1095 		device_info.array_count = topo_dev->node_props.array_count;
1096 		device_info.simd_arrays_per_engine =
1097 			topo_dev->node_props.simd_arrays_per_engine;
1098 		device_info.num_xcc = NUM_XCC(pdd->dev->xcc_mask);
1099 		device_info.capability = topo_dev->node_props.capability;
1100 		device_info.debug_prop = topo_dev->node_props.debug_prop;
1101 
1102 		if (exception_clear_mask)
1103 			pdd->exception_status &= ~exception_clear_mask;
1104 
1105 		if (copy_to_user(user_info, &device_info, *entry_size)) {
1106 			r = -EFAULT;
1107 			break;
1108 		}
1109 
1110 		user_info += tmp_entry_size;
1111 	}
1112 
1113 	mutex_unlock(&target->event_mutex);
1114 
1115 	return r;
1116 }
1117 
1118 void kfd_dbg_set_enabled_debug_exception_mask(struct kfd_process *target,
1119 					uint64_t exception_set_mask)
1120 {
1121 	uint64_t found_mask = 0;
1122 	struct process_queue_manager *pqm;
1123 	struct process_queue_node *pqn;
1124 	static const char write_data = '.';
1125 	loff_t pos = 0;
1126 	int i;
1127 
1128 	mutex_lock(&target->event_mutex);
1129 
1130 	found_mask |= target->exception_status;
1131 
1132 	pqm = &target->pqm;
1133 	list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
1134 		if (!pqn->q)
1135 			continue;
1136 
1137 		found_mask |= pqn->q->properties.exception_status;
1138 	}
1139 
1140 	for (i = 0; i < target->n_pdds; i++) {
1141 		struct kfd_process_device *pdd = target->pdds[i];
1142 
1143 		found_mask |= pdd->exception_status;
1144 	}
1145 
1146 	if (exception_set_mask & found_mask)
1147 		kernel_write(target->dbg_ev_file, &write_data, 1, &pos);
1148 
1149 	target->exception_enable_mask = exception_set_mask;
1150 
1151 	mutex_unlock(&target->event_mutex);
1152 }
1153