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
kfd_dbg_ev_query_debug_event(struct kfd_process * process,unsigned int * queue_id,unsigned int * gpu_id,uint64_t exception_clear_mask,uint64_t * event_status)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
debug_event_write_work_handler(struct work_struct * work)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 */
kfd_dbg_ev_raise(uint64_t event_mask,struct kfd_process * process,struct kfd_node * dev,unsigned int source_id,bool use_worker,void * exception_data,size_t exception_data_size)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 */
kfd_set_dbg_ev_from_interrupt(struct kfd_node * dev,unsigned int pasid,uint32_t doorbell_id,uint64_t trap_mask,void * exception_data,size_t exception_data_size)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
kfd_dbg_send_exception_to_runtime(struct kfd_process * p,unsigned int dev_id,unsigned int queue_id,uint64_t error_reason)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
kfd_dbg_set_queue_workaround(struct queue * q,bool enable)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
kfd_dbg_set_workaround(struct kfd_process * target,bool enable)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
kfd_dbg_set_mes_debug_mode(struct kfd_process_device * pdd,bool sq_trap_en)347 int kfd_dbg_set_mes_debug_mode(struct kfd_process_device *pdd, bool sq_trap_en)
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
352 if (!kfd_dbg_is_per_vmid_supported(pdd->dev))
353 return 0;
354
355 return amdgpu_mes_set_shader_debugger(pdd->dev->adev, pdd->proc_ctx_gpu_addr, spi_dbg_cntl,
356 pdd->watch_points, flags, sq_trap_en);
357 }
358
359 #define KFD_DEBUGGER_INVALID_WATCH_POINT_ID -1
kfd_dbg_get_dev_watch_id(struct kfd_process_device * pdd,int * watch_id)360 static int kfd_dbg_get_dev_watch_id(struct kfd_process_device *pdd, int *watch_id)
361 {
362 int i;
363
364 *watch_id = KFD_DEBUGGER_INVALID_WATCH_POINT_ID;
365
366 spin_lock(&pdd->dev->kfd->watch_points_lock);
367
368 for (i = 0; i < MAX_WATCH_ADDRESSES; i++) {
369 /* device watchpoint in use so skip */
370 if ((pdd->dev->kfd->alloc_watch_ids >> i) & 0x1)
371 continue;
372
373 pdd->alloc_watch_ids |= 0x1 << i;
374 pdd->dev->kfd->alloc_watch_ids |= 0x1 << i;
375 *watch_id = i;
376 spin_unlock(&pdd->dev->kfd->watch_points_lock);
377 return 0;
378 }
379
380 spin_unlock(&pdd->dev->kfd->watch_points_lock);
381
382 return -ENOMEM;
383 }
384
kfd_dbg_clear_dev_watch_id(struct kfd_process_device * pdd,int watch_id)385 static void kfd_dbg_clear_dev_watch_id(struct kfd_process_device *pdd, int watch_id)
386 {
387 spin_lock(&pdd->dev->kfd->watch_points_lock);
388
389 /* process owns device watch point so safe to clear */
390 if ((pdd->alloc_watch_ids >> watch_id) & 0x1) {
391 pdd->alloc_watch_ids &= ~(0x1 << watch_id);
392 pdd->dev->kfd->alloc_watch_ids &= ~(0x1 << watch_id);
393 }
394
395 spin_unlock(&pdd->dev->kfd->watch_points_lock);
396 }
397
kfd_dbg_owns_dev_watch_id(struct kfd_process_device * pdd,int watch_id)398 static bool kfd_dbg_owns_dev_watch_id(struct kfd_process_device *pdd, int watch_id)
399 {
400 bool owns_watch_id = false;
401
402 spin_lock(&pdd->dev->kfd->watch_points_lock);
403 owns_watch_id = watch_id < MAX_WATCH_ADDRESSES &&
404 ((pdd->alloc_watch_ids >> watch_id) & 0x1);
405
406 spin_unlock(&pdd->dev->kfd->watch_points_lock);
407
408 return owns_watch_id;
409 }
410
kfd_dbg_trap_clear_dev_address_watch(struct kfd_process_device * pdd,uint32_t watch_id)411 int kfd_dbg_trap_clear_dev_address_watch(struct kfd_process_device *pdd,
412 uint32_t watch_id)
413 {
414 int r;
415
416 if (!kfd_dbg_owns_dev_watch_id(pdd, watch_id))
417 return -EINVAL;
418
419 if (!pdd->dev->kfd->shared_resources.enable_mes) {
420 r = debug_lock_and_unmap(pdd->dev->dqm);
421 if (r)
422 return r;
423 }
424
425 amdgpu_gfx_off_ctrl(pdd->dev->adev, false);
426 pdd->watch_points[watch_id] = pdd->dev->kfd2kgd->clear_address_watch(
427 pdd->dev->adev,
428 watch_id);
429 amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
430
431 if (!pdd->dev->kfd->shared_resources.enable_mes)
432 r = debug_map_and_unlock(pdd->dev->dqm);
433 else
434 r = kfd_dbg_set_mes_debug_mode(pdd, true);
435
436 kfd_dbg_clear_dev_watch_id(pdd, watch_id);
437
438 return r;
439 }
440
kfd_dbg_trap_set_dev_address_watch(struct kfd_process_device * pdd,uint64_t watch_address,uint32_t watch_address_mask,uint32_t * watch_id,uint32_t watch_mode)441 int kfd_dbg_trap_set_dev_address_watch(struct kfd_process_device *pdd,
442 uint64_t watch_address,
443 uint32_t watch_address_mask,
444 uint32_t *watch_id,
445 uint32_t watch_mode)
446 {
447 int xcc_id, r = kfd_dbg_get_dev_watch_id(pdd, watch_id);
448 uint32_t xcc_mask = pdd->dev->xcc_mask;
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 for_each_inst(xcc_id, xcc_mask)
463 pdd->watch_points[*watch_id] = pdd->dev->kfd2kgd->set_address_watch(
464 pdd->dev->adev,
465 watch_address,
466 watch_address_mask,
467 *watch_id,
468 watch_mode,
469 pdd->dev->vm_info.last_vmid_kfd,
470 xcc_id);
471 amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
472
473 if (!pdd->dev->kfd->shared_resources.enable_mes)
474 r = debug_map_and_unlock(pdd->dev->dqm);
475 else
476 r = kfd_dbg_set_mes_debug_mode(pdd, true);
477
478 /* HWS is broken so no point in HW rollback but release the watchpoint anyways */
479 if (r)
480 kfd_dbg_clear_dev_watch_id(pdd, *watch_id);
481
482 return 0;
483 }
484
kfd_dbg_clear_process_address_watch(struct kfd_process * target)485 static void kfd_dbg_clear_process_address_watch(struct kfd_process *target)
486 {
487 int i, j;
488
489 for (i = 0; i < target->n_pdds; i++)
490 for (j = 0; j < MAX_WATCH_ADDRESSES; j++)
491 kfd_dbg_trap_clear_dev_address_watch(target->pdds[i], j);
492 }
493
kfd_dbg_trap_set_flags(struct kfd_process * target,uint32_t * flags)494 int kfd_dbg_trap_set_flags(struct kfd_process *target, uint32_t *flags)
495 {
496 uint32_t prev_flags = target->dbg_flags;
497 int i, r = 0, rewind_count = 0;
498
499 for (i = 0; i < target->n_pdds; i++) {
500 if (!kfd_dbg_is_per_vmid_supported(target->pdds[i]->dev) &&
501 (*flags & KFD_DBG_TRAP_FLAG_SINGLE_MEM_OP)) {
502 *flags = prev_flags;
503 return -EACCES;
504 }
505 }
506
507 target->dbg_flags = *flags & KFD_DBG_TRAP_FLAG_SINGLE_MEM_OP;
508 *flags = prev_flags;
509 for (i = 0; i < target->n_pdds; i++) {
510 struct kfd_process_device *pdd = target->pdds[i];
511
512 if (!kfd_dbg_is_per_vmid_supported(pdd->dev))
513 continue;
514
515 if (!pdd->dev->kfd->shared_resources.enable_mes)
516 r = debug_refresh_runlist(pdd->dev->dqm);
517 else
518 r = kfd_dbg_set_mes_debug_mode(pdd, true);
519
520 if (r) {
521 target->dbg_flags = prev_flags;
522 break;
523 }
524
525 rewind_count++;
526 }
527
528 /* Rewind flags */
529 if (r) {
530 target->dbg_flags = prev_flags;
531
532 for (i = 0; i < rewind_count; i++) {
533 struct kfd_process_device *pdd = target->pdds[i];
534
535 if (!kfd_dbg_is_per_vmid_supported(pdd->dev))
536 continue;
537
538 if (!pdd->dev->kfd->shared_resources.enable_mes)
539 debug_refresh_runlist(pdd->dev->dqm);
540 else
541 kfd_dbg_set_mes_debug_mode(pdd, true);
542 }
543 }
544
545 return r;
546 }
547
548 /* kfd_dbg_trap_deactivate:
549 * target: target process
550 * unwind: If this is unwinding a failed kfd_dbg_trap_enable()
551 * unwind_count:
552 * If unwind == true, how far down the pdd list we need
553 * to unwind
554 * else: ignored
555 */
kfd_dbg_trap_deactivate(struct kfd_process * target,bool unwind,int unwind_count)556 void kfd_dbg_trap_deactivate(struct kfd_process *target, bool unwind, int unwind_count)
557 {
558 int i;
559
560 if (!unwind) {
561 uint32_t flags = 0;
562 int resume_count = resume_queues(target, 0, NULL);
563
564 if (resume_count)
565 pr_debug("Resumed %d queues\n", resume_count);
566
567 cancel_work_sync(&target->debug_event_workarea);
568 kfd_dbg_clear_process_address_watch(target);
569 kfd_dbg_trap_set_wave_launch_mode(target, 0);
570
571 kfd_dbg_trap_set_flags(target, &flags);
572 }
573
574 for (i = 0; i < target->n_pdds; i++) {
575 struct kfd_process_device *pdd = target->pdds[i];
576
577 /* If this is an unwind, and we have unwound the required
578 * enable calls on the pdd list, we need to stop now
579 * otherwise we may mess up another debugger session.
580 */
581 if (unwind && i == unwind_count)
582 break;
583
584 kfd_process_set_trap_debug_flag(&pdd->qpd, false);
585
586 /* GFX off is already disabled by debug activate if not RLC restore supported. */
587 if (kfd_dbg_is_rlc_restore_supported(pdd->dev))
588 amdgpu_gfx_off_ctrl(pdd->dev->adev, false);
589 pdd->spi_dbg_override =
590 pdd->dev->kfd2kgd->disable_debug_trap(
591 pdd->dev->adev,
592 target->runtime_info.ttmp_setup,
593 pdd->dev->vm_info.last_vmid_kfd);
594 amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
595
596 if (!kfd_dbg_is_per_vmid_supported(pdd->dev) &&
597 release_debug_trap_vmid(pdd->dev->dqm, &pdd->qpd))
598 pr_err("Failed to release debug vmid on [%i]\n", pdd->dev->id);
599
600 if (!pdd->dev->kfd->shared_resources.enable_mes)
601 debug_refresh_runlist(pdd->dev->dqm);
602 else
603 kfd_dbg_set_mes_debug_mode(pdd, !kfd_dbg_has_cwsr_workaround(pdd->dev));
604 }
605
606 kfd_dbg_set_workaround(target, false);
607 }
608
kfd_dbg_clean_exception_status(struct kfd_process * target)609 static void kfd_dbg_clean_exception_status(struct kfd_process *target)
610 {
611 struct process_queue_manager *pqm;
612 struct process_queue_node *pqn;
613 int i;
614
615 for (i = 0; i < target->n_pdds; i++) {
616 struct kfd_process_device *pdd = target->pdds[i];
617
618 kfd_process_drain_interrupts(pdd);
619
620 pdd->exception_status = 0;
621 }
622
623 pqm = &target->pqm;
624 list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
625 if (!pqn->q)
626 continue;
627
628 pqn->q->properties.exception_status = 0;
629 }
630
631 target->exception_status = 0;
632 }
633
kfd_dbg_trap_disable(struct kfd_process * target)634 int kfd_dbg_trap_disable(struct kfd_process *target)
635 {
636 if (!target->debug_trap_enabled)
637 return 0;
638
639 /*
640 * Defer deactivation to runtime if runtime not enabled otherwise reset
641 * attached running target runtime state to enable for re-attach.
642 */
643 if (target->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED)
644 kfd_dbg_trap_deactivate(target, false, 0);
645 else if (target->runtime_info.runtime_state != DEBUG_RUNTIME_STATE_DISABLED)
646 target->runtime_info.runtime_state = DEBUG_RUNTIME_STATE_ENABLED;
647
648 fput(target->dbg_ev_file);
649 target->dbg_ev_file = NULL;
650
651 if (target->debugger_process) {
652 atomic_dec(&target->debugger_process->debugged_process_count);
653 target->debugger_process = NULL;
654 }
655
656 target->debug_trap_enabled = false;
657 kfd_dbg_clean_exception_status(target);
658 kfd_unref_process(target);
659
660 return 0;
661 }
662
kfd_dbg_trap_activate(struct kfd_process * target)663 int kfd_dbg_trap_activate(struct kfd_process *target)
664 {
665 int i, r = 0;
666
667 r = kfd_dbg_set_workaround(target, true);
668 if (r)
669 return r;
670
671 for (i = 0; i < target->n_pdds; i++) {
672 struct kfd_process_device *pdd = target->pdds[i];
673
674 if (!kfd_dbg_is_per_vmid_supported(pdd->dev)) {
675 r = reserve_debug_trap_vmid(pdd->dev->dqm, &pdd->qpd);
676
677 if (r) {
678 target->runtime_info.runtime_state = (r == -EBUSY) ?
679 DEBUG_RUNTIME_STATE_ENABLED_BUSY :
680 DEBUG_RUNTIME_STATE_ENABLED_ERROR;
681
682 goto unwind_err;
683 }
684 }
685
686 /* Disable GFX OFF to prevent garbage read/writes to debug registers.
687 * If RLC restore of debug registers is not supported and runtime enable
688 * hasn't done so already on ttmp setup request, restore the trap config registers.
689 *
690 * If RLC restore of debug registers is not supported, keep gfx off disabled for
691 * the debug session.
692 */
693 amdgpu_gfx_off_ctrl(pdd->dev->adev, false);
694 if (!(kfd_dbg_is_rlc_restore_supported(pdd->dev) ||
695 target->runtime_info.ttmp_setup))
696 pdd->dev->kfd2kgd->enable_debug_trap(pdd->dev->adev, true,
697 pdd->dev->vm_info.last_vmid_kfd);
698
699 pdd->spi_dbg_override = pdd->dev->kfd2kgd->enable_debug_trap(
700 pdd->dev->adev,
701 false,
702 pdd->dev->vm_info.last_vmid_kfd);
703
704 if (kfd_dbg_is_rlc_restore_supported(pdd->dev))
705 amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
706
707 /*
708 * Setting the debug flag in the trap handler requires that the TMA has been
709 * allocated, which occurs during CWSR initialization.
710 * In the event that CWSR has not been initialized at this point, setting the
711 * flag will be called again during CWSR initialization if the target process
712 * is still debug enabled.
713 */
714 kfd_process_set_trap_debug_flag(&pdd->qpd, true);
715
716 if (!pdd->dev->kfd->shared_resources.enable_mes)
717 r = debug_refresh_runlist(pdd->dev->dqm);
718 else
719 r = kfd_dbg_set_mes_debug_mode(pdd, true);
720
721 if (r) {
722 target->runtime_info.runtime_state =
723 DEBUG_RUNTIME_STATE_ENABLED_ERROR;
724 goto unwind_err;
725 }
726 }
727
728 return 0;
729
730 unwind_err:
731 /* Enabling debug failed, we need to disable on
732 * all GPUs so the enable is all or nothing.
733 */
734 kfd_dbg_trap_deactivate(target, true, i);
735 return r;
736 }
737
kfd_dbg_trap_enable(struct kfd_process * target,uint32_t fd,void __user * runtime_info,uint32_t * runtime_size)738 int kfd_dbg_trap_enable(struct kfd_process *target, uint32_t fd,
739 void __user *runtime_info, uint32_t *runtime_size)
740 {
741 struct file *f;
742 uint32_t copy_size;
743 int i, r = 0;
744
745 if (target->debug_trap_enabled)
746 return -EALREADY;
747
748 /* Enable pre-checks */
749 for (i = 0; i < target->n_pdds; i++) {
750 struct kfd_process_device *pdd = target->pdds[i];
751
752 if (!KFD_IS_SOC15(pdd->dev))
753 return -ENODEV;
754
755 if (pdd->qpd.num_gws && (!kfd_dbg_has_gws_support(pdd->dev) ||
756 kfd_dbg_has_cwsr_workaround(pdd->dev)))
757 return -EBUSY;
758 }
759
760 copy_size = min((size_t)(*runtime_size), sizeof(target->runtime_info));
761
762 f = fget(fd);
763 if (!f) {
764 pr_err("Failed to get file for (%i)\n", fd);
765 return -EBADF;
766 }
767
768 target->dbg_ev_file = f;
769
770 /* defer activation to runtime if not runtime enabled */
771 if (target->runtime_info.runtime_state == DEBUG_RUNTIME_STATE_ENABLED)
772 kfd_dbg_trap_activate(target);
773
774 /* We already hold the process reference but hold another one for the
775 * debug session.
776 */
777 kref_get(&target->ref);
778 target->debug_trap_enabled = true;
779
780 if (target->debugger_process)
781 atomic_inc(&target->debugger_process->debugged_process_count);
782
783 if (copy_to_user(runtime_info, (void *)&target->runtime_info, copy_size)) {
784 kfd_dbg_trap_deactivate(target, false, 0);
785 r = -EFAULT;
786 }
787
788 *runtime_size = sizeof(target->runtime_info);
789
790 return r;
791 }
792
kfd_dbg_validate_trap_override_request(struct kfd_process * p,uint32_t trap_override,uint32_t trap_mask_request,uint32_t * trap_mask_supported)793 static int kfd_dbg_validate_trap_override_request(struct kfd_process *p,
794 uint32_t trap_override,
795 uint32_t trap_mask_request,
796 uint32_t *trap_mask_supported)
797 {
798 int i = 0;
799
800 *trap_mask_supported = 0xffffffff;
801
802 for (i = 0; i < p->n_pdds; i++) {
803 struct kfd_process_device *pdd = p->pdds[i];
804 int err = pdd->dev->kfd2kgd->validate_trap_override_request(
805 pdd->dev->adev,
806 trap_override,
807 trap_mask_supported);
808
809 if (err)
810 return err;
811 }
812
813 if (trap_mask_request & ~*trap_mask_supported)
814 return -EACCES;
815
816 return 0;
817 }
818
kfd_dbg_trap_set_wave_launch_override(struct kfd_process * target,uint32_t trap_override,uint32_t trap_mask_bits,uint32_t trap_mask_request,uint32_t * trap_mask_prev,uint32_t * trap_mask_supported)819 int kfd_dbg_trap_set_wave_launch_override(struct kfd_process *target,
820 uint32_t trap_override,
821 uint32_t trap_mask_bits,
822 uint32_t trap_mask_request,
823 uint32_t *trap_mask_prev,
824 uint32_t *trap_mask_supported)
825 {
826 int r = 0, i;
827
828 r = kfd_dbg_validate_trap_override_request(target,
829 trap_override,
830 trap_mask_request,
831 trap_mask_supported);
832
833 if (r)
834 return r;
835
836 for (i = 0; i < target->n_pdds; i++) {
837 struct kfd_process_device *pdd = target->pdds[i];
838
839 amdgpu_gfx_off_ctrl(pdd->dev->adev, false);
840 pdd->spi_dbg_override = pdd->dev->kfd2kgd->set_wave_launch_trap_override(
841 pdd->dev->adev,
842 pdd->dev->vm_info.last_vmid_kfd,
843 trap_override,
844 trap_mask_bits,
845 trap_mask_request,
846 trap_mask_prev,
847 pdd->spi_dbg_override);
848 amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
849
850 if (!pdd->dev->kfd->shared_resources.enable_mes)
851 r = debug_refresh_runlist(pdd->dev->dqm);
852 else
853 r = kfd_dbg_set_mes_debug_mode(pdd, true);
854
855 if (r)
856 break;
857 }
858
859 return r;
860 }
861
kfd_dbg_trap_set_wave_launch_mode(struct kfd_process * target,uint8_t wave_launch_mode)862 int kfd_dbg_trap_set_wave_launch_mode(struct kfd_process *target,
863 uint8_t wave_launch_mode)
864 {
865 int r = 0, i;
866
867 if (wave_launch_mode != KFD_DBG_TRAP_WAVE_LAUNCH_MODE_NORMAL &&
868 wave_launch_mode != KFD_DBG_TRAP_WAVE_LAUNCH_MODE_HALT &&
869 wave_launch_mode != KFD_DBG_TRAP_WAVE_LAUNCH_MODE_DEBUG)
870 return -EINVAL;
871
872 for (i = 0; i < target->n_pdds; i++) {
873 struct kfd_process_device *pdd = target->pdds[i];
874
875 amdgpu_gfx_off_ctrl(pdd->dev->adev, false);
876 pdd->spi_dbg_launch_mode = pdd->dev->kfd2kgd->set_wave_launch_mode(
877 pdd->dev->adev,
878 wave_launch_mode,
879 pdd->dev->vm_info.last_vmid_kfd);
880 amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
881
882 if (!pdd->dev->kfd->shared_resources.enable_mes)
883 r = debug_refresh_runlist(pdd->dev->dqm);
884 else
885 r = kfd_dbg_set_mes_debug_mode(pdd, true);
886
887 if (r)
888 break;
889 }
890
891 return r;
892 }
893
kfd_dbg_trap_query_exception_info(struct kfd_process * target,uint32_t source_id,uint32_t exception_code,bool clear_exception,void __user * info,uint32_t * info_size)894 int kfd_dbg_trap_query_exception_info(struct kfd_process *target,
895 uint32_t source_id,
896 uint32_t exception_code,
897 bool clear_exception,
898 void __user *info,
899 uint32_t *info_size)
900 {
901 bool found = false;
902 int r = 0;
903 uint32_t copy_size, actual_info_size = 0;
904 uint64_t *exception_status_ptr = NULL;
905
906 if (!target)
907 return -EINVAL;
908
909 if (!info || !info_size)
910 return -EINVAL;
911
912 mutex_lock(&target->event_mutex);
913
914 if (KFD_DBG_EC_TYPE_IS_QUEUE(exception_code)) {
915 /* Per queue exceptions */
916 struct queue *queue = NULL;
917 int i;
918
919 for (i = 0; i < target->n_pdds; i++) {
920 struct kfd_process_device *pdd = target->pdds[i];
921 struct qcm_process_device *qpd = &pdd->qpd;
922
923 list_for_each_entry(queue, &qpd->queues_list, list) {
924 if (!found && queue->properties.queue_id == source_id) {
925 found = true;
926 break;
927 }
928 }
929 if (found)
930 break;
931 }
932
933 if (!found) {
934 r = -EINVAL;
935 goto out;
936 }
937
938 if (!(queue->properties.exception_status & KFD_EC_MASK(exception_code))) {
939 r = -ENODATA;
940 goto out;
941 }
942 exception_status_ptr = &queue->properties.exception_status;
943 } else if (KFD_DBG_EC_TYPE_IS_DEVICE(exception_code)) {
944 /* Per device exceptions */
945 struct kfd_process_device *pdd = NULL;
946 int i;
947
948 for (i = 0; i < target->n_pdds; i++) {
949 pdd = target->pdds[i];
950 if (pdd->dev->id == source_id) {
951 found = true;
952 break;
953 }
954 }
955
956 if (!found) {
957 r = -EINVAL;
958 goto out;
959 }
960
961 if (!(pdd->exception_status & KFD_EC_MASK(exception_code))) {
962 r = -ENODATA;
963 goto out;
964 }
965
966 if (exception_code == EC_DEVICE_MEMORY_VIOLATION) {
967 copy_size = min((size_t)(*info_size), pdd->vm_fault_exc_data_size);
968
969 if (copy_to_user(info, pdd->vm_fault_exc_data, copy_size)) {
970 r = -EFAULT;
971 goto out;
972 }
973 actual_info_size = pdd->vm_fault_exc_data_size;
974 if (clear_exception) {
975 kfree(pdd->vm_fault_exc_data);
976 pdd->vm_fault_exc_data = NULL;
977 pdd->vm_fault_exc_data_size = 0;
978 }
979 }
980 exception_status_ptr = &pdd->exception_status;
981 } else if (KFD_DBG_EC_TYPE_IS_PROCESS(exception_code)) {
982 /* Per process exceptions */
983 if (!(target->exception_status & KFD_EC_MASK(exception_code))) {
984 r = -ENODATA;
985 goto out;
986 }
987
988 if (exception_code == EC_PROCESS_RUNTIME) {
989 copy_size = min((size_t)(*info_size), sizeof(target->runtime_info));
990
991 if (copy_to_user(info, (void *)&target->runtime_info, copy_size)) {
992 r = -EFAULT;
993 goto out;
994 }
995
996 actual_info_size = sizeof(target->runtime_info);
997 }
998
999 exception_status_ptr = &target->exception_status;
1000 } else {
1001 pr_debug("Bad exception type [%i]\n", exception_code);
1002 r = -EINVAL;
1003 goto out;
1004 }
1005
1006 *info_size = actual_info_size;
1007 if (clear_exception)
1008 *exception_status_ptr &= ~KFD_EC_MASK(exception_code);
1009 out:
1010 mutex_unlock(&target->event_mutex);
1011 return r;
1012 }
1013
kfd_dbg_trap_device_snapshot(struct kfd_process * target,uint64_t exception_clear_mask,void __user * user_info,uint32_t * number_of_device_infos,uint32_t * entry_size)1014 int kfd_dbg_trap_device_snapshot(struct kfd_process *target,
1015 uint64_t exception_clear_mask,
1016 void __user *user_info,
1017 uint32_t *number_of_device_infos,
1018 uint32_t *entry_size)
1019 {
1020 struct kfd_dbg_device_info_entry device_info;
1021 uint32_t tmp_entry_size = *entry_size, tmp_num_devices;
1022 int i, r = 0;
1023
1024 if (!(target && user_info && number_of_device_infos && entry_size))
1025 return -EINVAL;
1026
1027 tmp_num_devices = min_t(size_t, *number_of_device_infos, target->n_pdds);
1028 *number_of_device_infos = target->n_pdds;
1029 *entry_size = min_t(size_t, *entry_size, sizeof(device_info));
1030
1031 if (!tmp_num_devices)
1032 return 0;
1033
1034 memset(&device_info, 0, sizeof(device_info));
1035
1036 mutex_lock(&target->event_mutex);
1037
1038 /* Run over all pdd of the process */
1039 for (i = 0; i < tmp_num_devices; i++) {
1040 struct kfd_process_device *pdd = target->pdds[i];
1041 struct kfd_topology_device *topo_dev = kfd_topology_device_by_id(pdd->dev->id);
1042
1043 device_info.gpu_id = pdd->dev->id;
1044 device_info.exception_status = pdd->exception_status;
1045 device_info.lds_base = pdd->lds_base;
1046 device_info.lds_limit = pdd->lds_limit;
1047 device_info.scratch_base = pdd->scratch_base;
1048 device_info.scratch_limit = pdd->scratch_limit;
1049 device_info.gpuvm_base = pdd->gpuvm_base;
1050 device_info.gpuvm_limit = pdd->gpuvm_limit;
1051 device_info.location_id = topo_dev->node_props.location_id;
1052 device_info.vendor_id = topo_dev->node_props.vendor_id;
1053 device_info.device_id = topo_dev->node_props.device_id;
1054 device_info.revision_id = pdd->dev->adev->pdev->revision;
1055 device_info.subsystem_vendor_id = pdd->dev->adev->pdev->subsystem_vendor;
1056 device_info.subsystem_device_id = pdd->dev->adev->pdev->subsystem_device;
1057 device_info.fw_version = pdd->dev->kfd->mec_fw_version;
1058 device_info.gfx_target_version =
1059 topo_dev->node_props.gfx_target_version;
1060 device_info.simd_count = topo_dev->node_props.simd_count;
1061 device_info.max_waves_per_simd =
1062 topo_dev->node_props.max_waves_per_simd;
1063 device_info.array_count = topo_dev->node_props.array_count;
1064 device_info.simd_arrays_per_engine =
1065 topo_dev->node_props.simd_arrays_per_engine;
1066 device_info.num_xcc = NUM_XCC(pdd->dev->xcc_mask);
1067 device_info.capability = topo_dev->node_props.capability;
1068 device_info.debug_prop = topo_dev->node_props.debug_prop;
1069
1070 if (exception_clear_mask)
1071 pdd->exception_status &= ~exception_clear_mask;
1072
1073 if (copy_to_user(user_info, &device_info, *entry_size)) {
1074 r = -EFAULT;
1075 break;
1076 }
1077
1078 user_info += tmp_entry_size;
1079 }
1080
1081 mutex_unlock(&target->event_mutex);
1082
1083 return r;
1084 }
1085
kfd_dbg_set_enabled_debug_exception_mask(struct kfd_process * target,uint64_t exception_set_mask)1086 void kfd_dbg_set_enabled_debug_exception_mask(struct kfd_process *target,
1087 uint64_t exception_set_mask)
1088 {
1089 uint64_t found_mask = 0;
1090 struct process_queue_manager *pqm;
1091 struct process_queue_node *pqn;
1092 static const char write_data = '.';
1093 loff_t pos = 0;
1094 int i;
1095
1096 mutex_lock(&target->event_mutex);
1097
1098 found_mask |= target->exception_status;
1099
1100 pqm = &target->pqm;
1101 list_for_each_entry(pqn, &pqm->queues, process_queue_list) {
1102 if (!pqn->q)
1103 continue;
1104
1105 found_mask |= pqn->q->properties.exception_status;
1106 }
1107
1108 for (i = 0; i < target->n_pdds; i++) {
1109 struct kfd_process_device *pdd = target->pdds[i];
1110
1111 found_mask |= pdd->exception_status;
1112 }
1113
1114 if (exception_set_mask & found_mask)
1115 kernel_write(target->dbg_ev_file, &write_data, 1, &pos);
1116
1117 target->exception_enable_mask = exception_set_mask;
1118
1119 mutex_unlock(&target->event_mutex);
1120 }
1121
