1 // SPDX-License-Identifier: GPL-2.0 OR MIT
2 /*
3  * Copyright 2014-2022 Advanced Micro Devices, Inc.
4  *
5  * Permission is hereby granted, free of charge, to any person obtaining a
6  * copy of this software and associated documentation files (the "Software"),
7  * to deal in the Software without restriction, including without limitation
8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9  * and/or sell copies of the Software, and to permit persons to whom the
10  * Software is furnished to do so, subject to the following conditions:
11  *
12  * The above copyright notice and this permission notice shall be included in
13  * all copies or substantial portions of the Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21  * OTHER DEALINGS IN THE SOFTWARE.
22  */
23 
24 #include <linux/mutex.h>
25 #include <linux/log2.h>
26 #include <linux/sched.h>
27 #include <linux/sched/mm.h>
28 #include <linux/sched/task.h>
29 #include <linux/mmu_context.h>
30 #include <linux/slab.h>
31 #include <linux/notifier.h>
32 #include <linux/compat.h>
33 #include <linux/mman.h>
34 #include <linux/file.h>
35 #include <linux/pm_runtime.h>
36 #include "amdgpu_amdkfd.h"
37 #include "amdgpu.h"
38 
39 struct mm_struct;
40 
41 #include "kfd_priv.h"
42 #include "kfd_device_queue_manager.h"
43 #include "kfd_svm.h"
44 #include "kfd_smi_events.h"
45 #include "kfd_debug.h"
46 
47 /*
48  * List of struct kfd_process (field kfd_process).
49  * Unique/indexed by mm_struct*
50  */
51 DEFINE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
52 DEFINE_MUTEX(kfd_processes_mutex);
53 
54 DEFINE_SRCU(kfd_processes_srcu);
55 
56 /* For process termination handling */
57 static struct workqueue_struct *kfd_process_wq;
58 
59 /* Ordered, single-threaded workqueue for restoring evicted
60  * processes. Restoring multiple processes concurrently under memory
61  * pressure can lead to processes blocking each other from validating
62  * their BOs and result in a live-lock situation where processes
63  * remain evicted indefinitely.
64  */
65 static struct workqueue_struct *kfd_restore_wq;
66 
67 static struct kfd_process *find_process(const struct task_struct *thread,
68 					bool ref);
69 static void kfd_process_ref_release(struct kref *ref);
70 static struct kfd_process *create_process(const struct task_struct *thread);
71 
72 static void evict_process_worker(struct work_struct *work);
73 static void restore_process_worker(struct work_struct *work);
74 
75 static void kfd_process_device_destroy_cwsr_dgpu(struct kfd_process_device *pdd);
76 
77 struct kfd_procfs_tree {
78 	struct kobject *kobj;
79 };
80 
81 static struct kfd_procfs_tree procfs;
82 
83 /*
84  * Structure for SDMA activity tracking
85  */
86 struct kfd_sdma_activity_handler_workarea {
87 	struct work_struct sdma_activity_work;
88 	struct kfd_process_device *pdd;
89 	uint64_t sdma_activity_counter;
90 };
91 
92 struct temp_sdma_queue_list {
93 	uint64_t __user *rptr;
94 	uint64_t sdma_val;
95 	unsigned int queue_id;
96 	struct list_head list;
97 };
98 
99 static void kfd_sdma_activity_worker(struct work_struct *work)
100 {
101 	struct kfd_sdma_activity_handler_workarea *workarea;
102 	struct kfd_process_device *pdd;
103 	uint64_t val;
104 	struct mm_struct *mm;
105 	struct queue *q;
106 	struct qcm_process_device *qpd;
107 	struct device_queue_manager *dqm;
108 	int ret = 0;
109 	struct temp_sdma_queue_list sdma_q_list;
110 	struct temp_sdma_queue_list *sdma_q, *next;
111 
112 	workarea = container_of(work, struct kfd_sdma_activity_handler_workarea,
113 				sdma_activity_work);
114 
115 	pdd = workarea->pdd;
116 	if (!pdd)
117 		return;
118 	dqm = pdd->dev->dqm;
119 	qpd = &pdd->qpd;
120 	if (!dqm || !qpd)
121 		return;
122 	/*
123 	 * Total SDMA activity is current SDMA activity + past SDMA activity
124 	 * Past SDMA count is stored in pdd.
125 	 * To get the current activity counters for all active SDMA queues,
126 	 * we loop over all SDMA queues and get their counts from user-space.
127 	 *
128 	 * We cannot call get_user() with dqm_lock held as it can cause
129 	 * a circular lock dependency situation. To read the SDMA stats,
130 	 * we need to do the following:
131 	 *
132 	 * 1. Create a temporary list of SDMA queue nodes from the qpd->queues_list,
133 	 *    with dqm_lock/dqm_unlock().
134 	 * 2. Call get_user() for each node in temporary list without dqm_lock.
135 	 *    Save the SDMA count for each node and also add the count to the total
136 	 *    SDMA count counter.
137 	 *    Its possible, during this step, a few SDMA queue nodes got deleted
138 	 *    from the qpd->queues_list.
139 	 * 3. Do a second pass over qpd->queues_list to check if any nodes got deleted.
140 	 *    If any node got deleted, its SDMA count would be captured in the sdma
141 	 *    past activity counter. So subtract the SDMA counter stored in step 2
142 	 *    for this node from the total SDMA count.
143 	 */
144 	INIT_LIST_HEAD(&sdma_q_list.list);
145 
146 	/*
147 	 * Create the temp list of all SDMA queues
148 	 */
149 	dqm_lock(dqm);
150 
151 	list_for_each_entry(q, &qpd->queues_list, list) {
152 		if ((q->properties.type != KFD_QUEUE_TYPE_SDMA) &&
153 		    (q->properties.type != KFD_QUEUE_TYPE_SDMA_XGMI))
154 			continue;
155 
156 		sdma_q = kzalloc(sizeof(struct temp_sdma_queue_list), GFP_KERNEL);
157 		if (!sdma_q) {
158 			dqm_unlock(dqm);
159 			goto cleanup;
160 		}
161 
162 		INIT_LIST_HEAD(&sdma_q->list);
163 		sdma_q->rptr = (uint64_t __user *)q->properties.read_ptr;
164 		sdma_q->queue_id = q->properties.queue_id;
165 		list_add_tail(&sdma_q->list, &sdma_q_list.list);
166 	}
167 
168 	/*
169 	 * If the temp list is empty, then no SDMA queues nodes were found in
170 	 * qpd->queues_list. Return the past activity count as the total sdma
171 	 * count
172 	 */
173 	if (list_empty(&sdma_q_list.list)) {
174 		workarea->sdma_activity_counter = pdd->sdma_past_activity_counter;
175 		dqm_unlock(dqm);
176 		return;
177 	}
178 
179 	dqm_unlock(dqm);
180 
181 	/*
182 	 * Get the usage count for each SDMA queue in temp_list.
183 	 */
184 	mm = get_task_mm(pdd->process->lead_thread);
185 	if (!mm)
186 		goto cleanup;
187 
188 	kthread_use_mm(mm);
189 
190 	list_for_each_entry(sdma_q, &sdma_q_list.list, list) {
191 		val = 0;
192 		ret = read_sdma_queue_counter(sdma_q->rptr, &val);
193 		if (ret) {
194 			pr_debug("Failed to read SDMA queue active counter for queue id: %d",
195 				 sdma_q->queue_id);
196 		} else {
197 			sdma_q->sdma_val = val;
198 			workarea->sdma_activity_counter += val;
199 		}
200 	}
201 
202 	kthread_unuse_mm(mm);
203 	mmput(mm);
204 
205 	/*
206 	 * Do a second iteration over qpd_queues_list to check if any SDMA
207 	 * nodes got deleted while fetching SDMA counter.
208 	 */
209 	dqm_lock(dqm);
210 
211 	workarea->sdma_activity_counter += pdd->sdma_past_activity_counter;
212 
213 	list_for_each_entry(q, &qpd->queues_list, list) {
214 		if (list_empty(&sdma_q_list.list))
215 			break;
216 
217 		if ((q->properties.type != KFD_QUEUE_TYPE_SDMA) &&
218 		    (q->properties.type != KFD_QUEUE_TYPE_SDMA_XGMI))
219 			continue;
220 
221 		list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
222 			if (((uint64_t __user *)q->properties.read_ptr == sdma_q->rptr) &&
223 			     (sdma_q->queue_id == q->properties.queue_id)) {
224 				list_del(&sdma_q->list);
225 				kfree(sdma_q);
226 				break;
227 			}
228 		}
229 	}
230 
231 	dqm_unlock(dqm);
232 
233 	/*
234 	 * If temp list is not empty, it implies some queues got deleted
235 	 * from qpd->queues_list during SDMA usage read. Subtract the SDMA
236 	 * count for each node from the total SDMA count.
237 	 */
238 	list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
239 		workarea->sdma_activity_counter -= sdma_q->sdma_val;
240 		list_del(&sdma_q->list);
241 		kfree(sdma_q);
242 	}
243 
244 	return;
245 
246 cleanup:
247 	list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
248 		list_del(&sdma_q->list);
249 		kfree(sdma_q);
250 	}
251 }
252 
253 /**
254  * kfd_get_cu_occupancy - Collect number of waves in-flight on this device
255  * by current process. Translates acquired wave count into number of compute units
256  * that are occupied.
257  *
258  * @attr: Handle of attribute that allows reporting of wave count. The attribute
259  * handle encapsulates GPU device it is associated with, thereby allowing collection
260  * of waves in flight, etc
261  * @buffer: Handle of user provided buffer updated with wave count
262  *
263  * Return: Number of bytes written to user buffer or an error value
264  */
265 static int kfd_get_cu_occupancy(struct attribute *attr, char *buffer)
266 {
267 	int cu_cnt;
268 	int wave_cnt;
269 	int max_waves_per_cu;
270 	struct kfd_node *dev = NULL;
271 	struct kfd_process *proc = NULL;
272 	struct kfd_process_device *pdd = NULL;
273 
274 	pdd = container_of(attr, struct kfd_process_device, attr_cu_occupancy);
275 	dev = pdd->dev;
276 	if (dev->kfd2kgd->get_cu_occupancy == NULL)
277 		return -EINVAL;
278 
279 	cu_cnt = 0;
280 	proc = pdd->process;
281 	if (pdd->qpd.queue_count == 0) {
282 		pr_debug("Gpu-Id: %d has no active queues for process %d\n",
283 			 dev->id, proc->pasid);
284 		return snprintf(buffer, PAGE_SIZE, "%d\n", cu_cnt);
285 	}
286 
287 	/* Collect wave count from device if it supports */
288 	wave_cnt = 0;
289 	max_waves_per_cu = 0;
290 	dev->kfd2kgd->get_cu_occupancy(dev->adev, proc->pasid, &wave_cnt,
291 			&max_waves_per_cu, 0);
292 
293 	/* Translate wave count to number of compute units */
294 	cu_cnt = (wave_cnt + (max_waves_per_cu - 1)) / max_waves_per_cu;
295 	return snprintf(buffer, PAGE_SIZE, "%d\n", cu_cnt);
296 }
297 
298 static ssize_t kfd_procfs_show(struct kobject *kobj, struct attribute *attr,
299 			       char *buffer)
300 {
301 	if (strcmp(attr->name, "pasid") == 0) {
302 		struct kfd_process *p = container_of(attr, struct kfd_process,
303 						     attr_pasid);
304 
305 		return snprintf(buffer, PAGE_SIZE, "%d\n", p->pasid);
306 	} else if (strncmp(attr->name, "vram_", 5) == 0) {
307 		struct kfd_process_device *pdd = container_of(attr, struct kfd_process_device,
308 							      attr_vram);
309 		return snprintf(buffer, PAGE_SIZE, "%llu\n", READ_ONCE(pdd->vram_usage));
310 	} else if (strncmp(attr->name, "sdma_", 5) == 0) {
311 		struct kfd_process_device *pdd = container_of(attr, struct kfd_process_device,
312 							      attr_sdma);
313 		struct kfd_sdma_activity_handler_workarea sdma_activity_work_handler;
314 
315 		INIT_WORK(&sdma_activity_work_handler.sdma_activity_work,
316 					kfd_sdma_activity_worker);
317 
318 		sdma_activity_work_handler.pdd = pdd;
319 		sdma_activity_work_handler.sdma_activity_counter = 0;
320 
321 		schedule_work(&sdma_activity_work_handler.sdma_activity_work);
322 
323 		flush_work(&sdma_activity_work_handler.sdma_activity_work);
324 
325 		return snprintf(buffer, PAGE_SIZE, "%llu\n",
326 				(sdma_activity_work_handler.sdma_activity_counter)/
327 				 SDMA_ACTIVITY_DIVISOR);
328 	} else {
329 		pr_err("Invalid attribute");
330 		return -EINVAL;
331 	}
332 
333 	return 0;
334 }
335 
336 static void kfd_procfs_kobj_release(struct kobject *kobj)
337 {
338 	kfree(kobj);
339 }
340 
341 static const struct sysfs_ops kfd_procfs_ops = {
342 	.show = kfd_procfs_show,
343 };
344 
345 static const struct kobj_type procfs_type = {
346 	.release = kfd_procfs_kobj_release,
347 	.sysfs_ops = &kfd_procfs_ops,
348 };
349 
350 void kfd_procfs_init(void)
351 {
352 	int ret = 0;
353 
354 	procfs.kobj = kfd_alloc_struct(procfs.kobj);
355 	if (!procfs.kobj)
356 		return;
357 
358 	ret = kobject_init_and_add(procfs.kobj, &procfs_type,
359 				   &kfd_device->kobj, "proc");
360 	if (ret) {
361 		pr_warn("Could not create procfs proc folder");
362 		/* If we fail to create the procfs, clean up */
363 		kfd_procfs_shutdown();
364 	}
365 }
366 
367 void kfd_procfs_shutdown(void)
368 {
369 	if (procfs.kobj) {
370 		kobject_del(procfs.kobj);
371 		kobject_put(procfs.kobj);
372 		procfs.kobj = NULL;
373 	}
374 }
375 
376 static ssize_t kfd_procfs_queue_show(struct kobject *kobj,
377 				     struct attribute *attr, char *buffer)
378 {
379 	struct queue *q = container_of(kobj, struct queue, kobj);
380 
381 	if (!strcmp(attr->name, "size"))
382 		return snprintf(buffer, PAGE_SIZE, "%llu",
383 				q->properties.queue_size);
384 	else if (!strcmp(attr->name, "type"))
385 		return snprintf(buffer, PAGE_SIZE, "%d", q->properties.type);
386 	else if (!strcmp(attr->name, "gpuid"))
387 		return snprintf(buffer, PAGE_SIZE, "%u", q->device->id);
388 	else
389 		pr_err("Invalid attribute");
390 
391 	return 0;
392 }
393 
394 static ssize_t kfd_procfs_stats_show(struct kobject *kobj,
395 				     struct attribute *attr, char *buffer)
396 {
397 	if (strcmp(attr->name, "evicted_ms") == 0) {
398 		struct kfd_process_device *pdd = container_of(attr,
399 				struct kfd_process_device,
400 				attr_evict);
401 		uint64_t evict_jiffies;
402 
403 		evict_jiffies = atomic64_read(&pdd->evict_duration_counter);
404 
405 		return snprintf(buffer,
406 				PAGE_SIZE,
407 				"%llu\n",
408 				jiffies64_to_msecs(evict_jiffies));
409 
410 	/* Sysfs handle that gets CU occupancy is per device */
411 	} else if (strcmp(attr->name, "cu_occupancy") == 0) {
412 		return kfd_get_cu_occupancy(attr, buffer);
413 	} else {
414 		pr_err("Invalid attribute");
415 	}
416 
417 	return 0;
418 }
419 
420 static ssize_t kfd_sysfs_counters_show(struct kobject *kobj,
421 				       struct attribute *attr, char *buf)
422 {
423 	struct kfd_process_device *pdd;
424 
425 	if (!strcmp(attr->name, "faults")) {
426 		pdd = container_of(attr, struct kfd_process_device,
427 				   attr_faults);
428 		return sysfs_emit(buf, "%llu\n", READ_ONCE(pdd->faults));
429 	}
430 	if (!strcmp(attr->name, "page_in")) {
431 		pdd = container_of(attr, struct kfd_process_device,
432 				   attr_page_in);
433 		return sysfs_emit(buf, "%llu\n", READ_ONCE(pdd->page_in));
434 	}
435 	if (!strcmp(attr->name, "page_out")) {
436 		pdd = container_of(attr, struct kfd_process_device,
437 				   attr_page_out);
438 		return sysfs_emit(buf, "%llu\n", READ_ONCE(pdd->page_out));
439 	}
440 	return 0;
441 }
442 
443 static struct attribute attr_queue_size = {
444 	.name = "size",
445 	.mode = KFD_SYSFS_FILE_MODE
446 };
447 
448 static struct attribute attr_queue_type = {
449 	.name = "type",
450 	.mode = KFD_SYSFS_FILE_MODE
451 };
452 
453 static struct attribute attr_queue_gpuid = {
454 	.name = "gpuid",
455 	.mode = KFD_SYSFS_FILE_MODE
456 };
457 
458 static struct attribute *procfs_queue_attrs[] = {
459 	&attr_queue_size,
460 	&attr_queue_type,
461 	&attr_queue_gpuid,
462 	NULL
463 };
464 ATTRIBUTE_GROUPS(procfs_queue);
465 
466 static const struct sysfs_ops procfs_queue_ops = {
467 	.show = kfd_procfs_queue_show,
468 };
469 
470 static const struct kobj_type procfs_queue_type = {
471 	.sysfs_ops = &procfs_queue_ops,
472 	.default_groups = procfs_queue_groups,
473 };
474 
475 static const struct sysfs_ops procfs_stats_ops = {
476 	.show = kfd_procfs_stats_show,
477 };
478 
479 static const struct kobj_type procfs_stats_type = {
480 	.sysfs_ops = &procfs_stats_ops,
481 	.release = kfd_procfs_kobj_release,
482 };
483 
484 static const struct sysfs_ops sysfs_counters_ops = {
485 	.show = kfd_sysfs_counters_show,
486 };
487 
488 static const struct kobj_type sysfs_counters_type = {
489 	.sysfs_ops = &sysfs_counters_ops,
490 	.release = kfd_procfs_kobj_release,
491 };
492 
493 int kfd_procfs_add_queue(struct queue *q)
494 {
495 	struct kfd_process *proc;
496 	int ret;
497 
498 	if (!q || !q->process)
499 		return -EINVAL;
500 	proc = q->process;
501 
502 	/* Create proc/<pid>/queues/<queue id> folder */
503 	if (!proc->kobj_queues)
504 		return -EFAULT;
505 	ret = kobject_init_and_add(&q->kobj, &procfs_queue_type,
506 			proc->kobj_queues, "%u", q->properties.queue_id);
507 	if (ret < 0) {
508 		pr_warn("Creating proc/<pid>/queues/%u failed",
509 			q->properties.queue_id);
510 		kobject_put(&q->kobj);
511 		return ret;
512 	}
513 
514 	return 0;
515 }
516 
517 static void kfd_sysfs_create_file(struct kobject *kobj, struct attribute *attr,
518 				 char *name)
519 {
520 	int ret;
521 
522 	if (!kobj || !attr || !name)
523 		return;
524 
525 	attr->name = name;
526 	attr->mode = KFD_SYSFS_FILE_MODE;
527 	sysfs_attr_init(attr);
528 
529 	ret = sysfs_create_file(kobj, attr);
530 	if (ret)
531 		pr_warn("Create sysfs %s/%s failed %d", kobj->name, name, ret);
532 }
533 
534 static void kfd_procfs_add_sysfs_stats(struct kfd_process *p)
535 {
536 	int ret;
537 	int i;
538 	char stats_dir_filename[MAX_SYSFS_FILENAME_LEN];
539 
540 	if (!p || !p->kobj)
541 		return;
542 
543 	/*
544 	 * Create sysfs files for each GPU:
545 	 * - proc/<pid>/stats_<gpuid>/
546 	 * - proc/<pid>/stats_<gpuid>/evicted_ms
547 	 * - proc/<pid>/stats_<gpuid>/cu_occupancy
548 	 */
549 	for (i = 0; i < p->n_pdds; i++) {
550 		struct kfd_process_device *pdd = p->pdds[i];
551 
552 		snprintf(stats_dir_filename, MAX_SYSFS_FILENAME_LEN,
553 				"stats_%u", pdd->dev->id);
554 		pdd->kobj_stats = kfd_alloc_struct(pdd->kobj_stats);
555 		if (!pdd->kobj_stats)
556 			return;
557 
558 		ret = kobject_init_and_add(pdd->kobj_stats,
559 					   &procfs_stats_type,
560 					   p->kobj,
561 					   stats_dir_filename);
562 
563 		if (ret) {
564 			pr_warn("Creating KFD proc/stats_%s folder failed",
565 				stats_dir_filename);
566 			kobject_put(pdd->kobj_stats);
567 			pdd->kobj_stats = NULL;
568 			return;
569 		}
570 
571 		kfd_sysfs_create_file(pdd->kobj_stats, &pdd->attr_evict,
572 				      "evicted_ms");
573 		/* Add sysfs file to report compute unit occupancy */
574 		if (pdd->dev->kfd2kgd->get_cu_occupancy)
575 			kfd_sysfs_create_file(pdd->kobj_stats,
576 					      &pdd->attr_cu_occupancy,
577 					      "cu_occupancy");
578 	}
579 }
580 
581 static void kfd_procfs_add_sysfs_counters(struct kfd_process *p)
582 {
583 	int ret = 0;
584 	int i;
585 	char counters_dir_filename[MAX_SYSFS_FILENAME_LEN];
586 
587 	if (!p || !p->kobj)
588 		return;
589 
590 	/*
591 	 * Create sysfs files for each GPU which supports SVM
592 	 * - proc/<pid>/counters_<gpuid>/
593 	 * - proc/<pid>/counters_<gpuid>/faults
594 	 * - proc/<pid>/counters_<gpuid>/page_in
595 	 * - proc/<pid>/counters_<gpuid>/page_out
596 	 */
597 	for_each_set_bit(i, p->svms.bitmap_supported, p->n_pdds) {
598 		struct kfd_process_device *pdd = p->pdds[i];
599 		struct kobject *kobj_counters;
600 
601 		snprintf(counters_dir_filename, MAX_SYSFS_FILENAME_LEN,
602 			"counters_%u", pdd->dev->id);
603 		kobj_counters = kfd_alloc_struct(kobj_counters);
604 		if (!kobj_counters)
605 			return;
606 
607 		ret = kobject_init_and_add(kobj_counters, &sysfs_counters_type,
608 					   p->kobj, counters_dir_filename);
609 		if (ret) {
610 			pr_warn("Creating KFD proc/%s folder failed",
611 				counters_dir_filename);
612 			kobject_put(kobj_counters);
613 			return;
614 		}
615 
616 		pdd->kobj_counters = kobj_counters;
617 		kfd_sysfs_create_file(kobj_counters, &pdd->attr_faults,
618 				      "faults");
619 		kfd_sysfs_create_file(kobj_counters, &pdd->attr_page_in,
620 				      "page_in");
621 		kfd_sysfs_create_file(kobj_counters, &pdd->attr_page_out,
622 				      "page_out");
623 	}
624 }
625 
626 static void kfd_procfs_add_sysfs_files(struct kfd_process *p)
627 {
628 	int i;
629 
630 	if (!p || !p->kobj)
631 		return;
632 
633 	/*
634 	 * Create sysfs files for each GPU:
635 	 * - proc/<pid>/vram_<gpuid>
636 	 * - proc/<pid>/sdma_<gpuid>
637 	 */
638 	for (i = 0; i < p->n_pdds; i++) {
639 		struct kfd_process_device *pdd = p->pdds[i];
640 
641 		snprintf(pdd->vram_filename, MAX_SYSFS_FILENAME_LEN, "vram_%u",
642 			 pdd->dev->id);
643 		kfd_sysfs_create_file(p->kobj, &pdd->attr_vram,
644 				      pdd->vram_filename);
645 
646 		snprintf(pdd->sdma_filename, MAX_SYSFS_FILENAME_LEN, "sdma_%u",
647 			 pdd->dev->id);
648 		kfd_sysfs_create_file(p->kobj, &pdd->attr_sdma,
649 					    pdd->sdma_filename);
650 	}
651 }
652 
653 void kfd_procfs_del_queue(struct queue *q)
654 {
655 	if (!q)
656 		return;
657 
658 	kobject_del(&q->kobj);
659 	kobject_put(&q->kobj);
660 }
661 
662 int kfd_process_create_wq(void)
663 {
664 	if (!kfd_process_wq)
665 		kfd_process_wq = alloc_workqueue("kfd_process_wq", 0, 0);
666 	if (!kfd_restore_wq)
667 		kfd_restore_wq = alloc_ordered_workqueue("kfd_restore_wq", 0);
668 
669 	if (!kfd_process_wq || !kfd_restore_wq) {
670 		kfd_process_destroy_wq();
671 		return -ENOMEM;
672 	}
673 
674 	return 0;
675 }
676 
677 void kfd_process_destroy_wq(void)
678 {
679 	if (kfd_process_wq) {
680 		destroy_workqueue(kfd_process_wq);
681 		kfd_process_wq = NULL;
682 	}
683 	if (kfd_restore_wq) {
684 		destroy_workqueue(kfd_restore_wq);
685 		kfd_restore_wq = NULL;
686 	}
687 }
688 
689 static void kfd_process_free_gpuvm(struct kgd_mem *mem,
690 			struct kfd_process_device *pdd, void **kptr)
691 {
692 	struct kfd_node *dev = pdd->dev;
693 
694 	if (kptr && *kptr) {
695 		amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(mem);
696 		*kptr = NULL;
697 	}
698 
699 	amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(dev->adev, mem, pdd->drm_priv);
700 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->adev, mem, pdd->drm_priv,
701 					       NULL);
702 }
703 
704 /* kfd_process_alloc_gpuvm - Allocate GPU VM for the KFD process
705  *	This function should be only called right after the process
706  *	is created and when kfd_processes_mutex is still being held
707  *	to avoid concurrency. Because of that exclusiveness, we do
708  *	not need to take p->mutex.
709  */
710 static int kfd_process_alloc_gpuvm(struct kfd_process_device *pdd,
711 				   uint64_t gpu_va, uint32_t size,
712 				   uint32_t flags, struct kgd_mem **mem, void **kptr)
713 {
714 	struct kfd_node *kdev = pdd->dev;
715 	int err;
716 
717 	err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(kdev->adev, gpu_va, size,
718 						 pdd->drm_priv, mem, NULL,
719 						 flags, false);
720 	if (err)
721 		goto err_alloc_mem;
722 
723 	err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(kdev->adev, *mem,
724 			pdd->drm_priv);
725 	if (err)
726 		goto err_map_mem;
727 
728 	err = amdgpu_amdkfd_gpuvm_sync_memory(kdev->adev, *mem, true);
729 	if (err) {
730 		pr_debug("Sync memory failed, wait interrupted by user signal\n");
731 		goto sync_memory_failed;
732 	}
733 
734 	if (kptr) {
735 		err = amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(
736 				(struct kgd_mem *)*mem, kptr, NULL);
737 		if (err) {
738 			pr_debug("Map GTT BO to kernel failed\n");
739 			goto sync_memory_failed;
740 		}
741 	}
742 
743 	return err;
744 
745 sync_memory_failed:
746 	amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(kdev->adev, *mem, pdd->drm_priv);
747 
748 err_map_mem:
749 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(kdev->adev, *mem, pdd->drm_priv,
750 					       NULL);
751 err_alloc_mem:
752 	*mem = NULL;
753 	*kptr = NULL;
754 	return err;
755 }
756 
757 /* kfd_process_device_reserve_ib_mem - Reserve memory inside the
758  *	process for IB usage The memory reserved is for KFD to submit
759  *	IB to AMDGPU from kernel.  If the memory is reserved
760  *	successfully, ib_kaddr will have the CPU/kernel
761  *	address. Check ib_kaddr before accessing the memory.
762  */
763 static int kfd_process_device_reserve_ib_mem(struct kfd_process_device *pdd)
764 {
765 	struct qcm_process_device *qpd = &pdd->qpd;
766 	uint32_t flags = KFD_IOC_ALLOC_MEM_FLAGS_GTT |
767 			KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE |
768 			KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE |
769 			KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
770 	struct kgd_mem *mem;
771 	void *kaddr;
772 	int ret;
773 
774 	if (qpd->ib_kaddr || !qpd->ib_base)
775 		return 0;
776 
777 	/* ib_base is only set for dGPU */
778 	ret = kfd_process_alloc_gpuvm(pdd, qpd->ib_base, PAGE_SIZE, flags,
779 				      &mem, &kaddr);
780 	if (ret)
781 		return ret;
782 
783 	qpd->ib_mem = mem;
784 	qpd->ib_kaddr = kaddr;
785 
786 	return 0;
787 }
788 
789 static void kfd_process_device_destroy_ib_mem(struct kfd_process_device *pdd)
790 {
791 	struct qcm_process_device *qpd = &pdd->qpd;
792 
793 	if (!qpd->ib_kaddr || !qpd->ib_base)
794 		return;
795 
796 	kfd_process_free_gpuvm(qpd->ib_mem, pdd, &qpd->ib_kaddr);
797 }
798 
799 struct kfd_process *kfd_create_process(struct task_struct *thread)
800 {
801 	struct kfd_process *process;
802 	int ret;
803 
804 	if (!(thread->mm && mmget_not_zero(thread->mm)))
805 		return ERR_PTR(-EINVAL);
806 
807 	/* Only the pthreads threading model is supported. */
808 	if (thread->group_leader->mm != thread->mm) {
809 		mmput(thread->mm);
810 		return ERR_PTR(-EINVAL);
811 	}
812 
813 	/*
814 	 * take kfd processes mutex before starting of process creation
815 	 * so there won't be a case where two threads of the same process
816 	 * create two kfd_process structures
817 	 */
818 	mutex_lock(&kfd_processes_mutex);
819 
820 	if (kfd_is_locked()) {
821 		pr_debug("KFD is locked! Cannot create process");
822 		process = ERR_PTR(-EINVAL);
823 		goto out;
824 	}
825 
826 	/* A prior open of /dev/kfd could have already created the process. */
827 	process = find_process(thread, false);
828 	if (process) {
829 		pr_debug("Process already found\n");
830 	} else {
831 		/* If the process just called exec(3), it is possible that the
832 		 * cleanup of the kfd_process (following the release of the mm
833 		 * of the old process image) is still in the cleanup work queue.
834 		 * Make sure to drain any job before trying to recreate any
835 		 * resource for this process.
836 		 */
837 		flush_workqueue(kfd_process_wq);
838 
839 		process = create_process(thread);
840 		if (IS_ERR(process))
841 			goto out;
842 
843 		if (!procfs.kobj)
844 			goto out;
845 
846 		process->kobj = kfd_alloc_struct(process->kobj);
847 		if (!process->kobj) {
848 			pr_warn("Creating procfs kobject failed");
849 			goto out;
850 		}
851 		ret = kobject_init_and_add(process->kobj, &procfs_type,
852 					   procfs.kobj, "%d",
853 					   (int)process->lead_thread->pid);
854 		if (ret) {
855 			pr_warn("Creating procfs pid directory failed");
856 			kobject_put(process->kobj);
857 			goto out;
858 		}
859 
860 		kfd_sysfs_create_file(process->kobj, &process->attr_pasid,
861 				      "pasid");
862 
863 		process->kobj_queues = kobject_create_and_add("queues",
864 							process->kobj);
865 		if (!process->kobj_queues)
866 			pr_warn("Creating KFD proc/queues folder failed");
867 
868 		kfd_procfs_add_sysfs_stats(process);
869 		kfd_procfs_add_sysfs_files(process);
870 		kfd_procfs_add_sysfs_counters(process);
871 
872 		init_waitqueue_head(&process->wait_irq_drain);
873 	}
874 out:
875 	if (!IS_ERR(process))
876 		kref_get(&process->ref);
877 	mutex_unlock(&kfd_processes_mutex);
878 	mmput(thread->mm);
879 
880 	return process;
881 }
882 
883 struct kfd_process *kfd_get_process(const struct task_struct *thread)
884 {
885 	struct kfd_process *process;
886 
887 	if (!thread->mm)
888 		return ERR_PTR(-EINVAL);
889 
890 	/* Only the pthreads threading model is supported. */
891 	if (thread->group_leader->mm != thread->mm)
892 		return ERR_PTR(-EINVAL);
893 
894 	process = find_process(thread, false);
895 	if (!process)
896 		return ERR_PTR(-EINVAL);
897 
898 	return process;
899 }
900 
901 static struct kfd_process *find_process_by_mm(const struct mm_struct *mm)
902 {
903 	struct kfd_process *process;
904 
905 	hash_for_each_possible_rcu(kfd_processes_table, process,
906 					kfd_processes, (uintptr_t)mm)
907 		if (process->mm == mm)
908 			return process;
909 
910 	return NULL;
911 }
912 
913 static struct kfd_process *find_process(const struct task_struct *thread,
914 					bool ref)
915 {
916 	struct kfd_process *p;
917 	int idx;
918 
919 	idx = srcu_read_lock(&kfd_processes_srcu);
920 	p = find_process_by_mm(thread->mm);
921 	if (p && ref)
922 		kref_get(&p->ref);
923 	srcu_read_unlock(&kfd_processes_srcu, idx);
924 
925 	return p;
926 }
927 
928 void kfd_unref_process(struct kfd_process *p)
929 {
930 	kref_put(&p->ref, kfd_process_ref_release);
931 }
932 
933 /* This increments the process->ref counter. */
934 struct kfd_process *kfd_lookup_process_by_pid(struct pid *pid)
935 {
936 	struct task_struct *task = NULL;
937 	struct kfd_process *p    = NULL;
938 
939 	if (!pid) {
940 		task = current;
941 		get_task_struct(task);
942 	} else {
943 		task = get_pid_task(pid, PIDTYPE_PID);
944 	}
945 
946 	if (task) {
947 		p = find_process(task, true);
948 		put_task_struct(task);
949 	}
950 
951 	return p;
952 }
953 
954 static void kfd_process_device_free_bos(struct kfd_process_device *pdd)
955 {
956 	struct kfd_process *p = pdd->process;
957 	void *mem;
958 	int id;
959 	int i;
960 
961 	/*
962 	 * Remove all handles from idr and release appropriate
963 	 * local memory object
964 	 */
965 	idr_for_each_entry(&pdd->alloc_idr, mem, id) {
966 
967 		for (i = 0; i < p->n_pdds; i++) {
968 			struct kfd_process_device *peer_pdd = p->pdds[i];
969 
970 			if (!peer_pdd->drm_priv)
971 				continue;
972 			amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
973 				peer_pdd->dev->adev, mem, peer_pdd->drm_priv);
974 		}
975 
976 		amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, mem,
977 						       pdd->drm_priv, NULL);
978 		kfd_process_device_remove_obj_handle(pdd, id);
979 	}
980 }
981 
982 /*
983  * Just kunmap and unpin signal BO here. It will be freed in
984  * kfd_process_free_outstanding_kfd_bos()
985  */
986 static void kfd_process_kunmap_signal_bo(struct kfd_process *p)
987 {
988 	struct kfd_process_device *pdd;
989 	struct kfd_node *kdev;
990 	void *mem;
991 
992 	kdev = kfd_device_by_id(GET_GPU_ID(p->signal_handle));
993 	if (!kdev)
994 		return;
995 
996 	mutex_lock(&p->mutex);
997 
998 	pdd = kfd_get_process_device_data(kdev, p);
999 	if (!pdd)
1000 		goto out;
1001 
1002 	mem = kfd_process_device_translate_handle(
1003 		pdd, GET_IDR_HANDLE(p->signal_handle));
1004 	if (!mem)
1005 		goto out;
1006 
1007 	amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(mem);
1008 
1009 out:
1010 	mutex_unlock(&p->mutex);
1011 }
1012 
1013 static void kfd_process_free_outstanding_kfd_bos(struct kfd_process *p)
1014 {
1015 	int i;
1016 
1017 	for (i = 0; i < p->n_pdds; i++)
1018 		kfd_process_device_free_bos(p->pdds[i]);
1019 }
1020 
1021 static void kfd_process_destroy_pdds(struct kfd_process *p)
1022 {
1023 	int i;
1024 
1025 	for (i = 0; i < p->n_pdds; i++) {
1026 		struct kfd_process_device *pdd = p->pdds[i];
1027 
1028 		pr_debug("Releasing pdd (topology id %d) for process (pasid 0x%x)\n",
1029 				pdd->dev->id, p->pasid);
1030 
1031 		kfd_process_device_destroy_cwsr_dgpu(pdd);
1032 		kfd_process_device_destroy_ib_mem(pdd);
1033 
1034 		if (pdd->drm_file) {
1035 			amdgpu_amdkfd_gpuvm_release_process_vm(
1036 					pdd->dev->adev, pdd->drm_priv);
1037 			fput(pdd->drm_file);
1038 		}
1039 
1040 		if (pdd->qpd.cwsr_kaddr && !pdd->qpd.cwsr_base)
1041 			free_pages((unsigned long)pdd->qpd.cwsr_kaddr,
1042 				get_order(KFD_CWSR_TBA_TMA_SIZE));
1043 
1044 		idr_destroy(&pdd->alloc_idr);
1045 
1046 		kfd_free_process_doorbells(pdd->dev->kfd, pdd);
1047 
1048 		if (pdd->dev->kfd->shared_resources.enable_mes)
1049 			amdgpu_amdkfd_free_gtt_mem(pdd->dev->adev,
1050 						   pdd->proc_ctx_bo);
1051 		/*
1052 		 * before destroying pdd, make sure to report availability
1053 		 * for auto suspend
1054 		 */
1055 		if (pdd->runtime_inuse) {
1056 			pm_runtime_mark_last_busy(adev_to_drm(pdd->dev->adev)->dev);
1057 			pm_runtime_put_autosuspend(adev_to_drm(pdd->dev->adev)->dev);
1058 			pdd->runtime_inuse = false;
1059 		}
1060 
1061 		kfree(pdd);
1062 		p->pdds[i] = NULL;
1063 	}
1064 	p->n_pdds = 0;
1065 }
1066 
1067 static void kfd_process_remove_sysfs(struct kfd_process *p)
1068 {
1069 	struct kfd_process_device *pdd;
1070 	int i;
1071 
1072 	if (!p->kobj)
1073 		return;
1074 
1075 	sysfs_remove_file(p->kobj, &p->attr_pasid);
1076 	kobject_del(p->kobj_queues);
1077 	kobject_put(p->kobj_queues);
1078 	p->kobj_queues = NULL;
1079 
1080 	for (i = 0; i < p->n_pdds; i++) {
1081 		pdd = p->pdds[i];
1082 
1083 		sysfs_remove_file(p->kobj, &pdd->attr_vram);
1084 		sysfs_remove_file(p->kobj, &pdd->attr_sdma);
1085 
1086 		sysfs_remove_file(pdd->kobj_stats, &pdd->attr_evict);
1087 		if (pdd->dev->kfd2kgd->get_cu_occupancy)
1088 			sysfs_remove_file(pdd->kobj_stats,
1089 					  &pdd->attr_cu_occupancy);
1090 		kobject_del(pdd->kobj_stats);
1091 		kobject_put(pdd->kobj_stats);
1092 		pdd->kobj_stats = NULL;
1093 	}
1094 
1095 	for_each_set_bit(i, p->svms.bitmap_supported, p->n_pdds) {
1096 		pdd = p->pdds[i];
1097 
1098 		sysfs_remove_file(pdd->kobj_counters, &pdd->attr_faults);
1099 		sysfs_remove_file(pdd->kobj_counters, &pdd->attr_page_in);
1100 		sysfs_remove_file(pdd->kobj_counters, &pdd->attr_page_out);
1101 		kobject_del(pdd->kobj_counters);
1102 		kobject_put(pdd->kobj_counters);
1103 		pdd->kobj_counters = NULL;
1104 	}
1105 
1106 	kobject_del(p->kobj);
1107 	kobject_put(p->kobj);
1108 	p->kobj = NULL;
1109 }
1110 
1111 /* No process locking is needed in this function, because the process
1112  * is not findable any more. We must assume that no other thread is
1113  * using it any more, otherwise we couldn't safely free the process
1114  * structure in the end.
1115  */
1116 static void kfd_process_wq_release(struct work_struct *work)
1117 {
1118 	struct kfd_process *p = container_of(work, struct kfd_process,
1119 					     release_work);
1120 
1121 	kfd_process_dequeue_from_all_devices(p);
1122 	pqm_uninit(&p->pqm);
1123 
1124 	/* Signal the eviction fence after user mode queues are
1125 	 * destroyed. This allows any BOs to be freed without
1126 	 * triggering pointless evictions or waiting for fences.
1127 	 */
1128 	dma_fence_signal(p->ef);
1129 
1130 	kfd_process_remove_sysfs(p);
1131 
1132 	kfd_process_kunmap_signal_bo(p);
1133 	kfd_process_free_outstanding_kfd_bos(p);
1134 	svm_range_list_fini(p);
1135 
1136 	kfd_process_destroy_pdds(p);
1137 	dma_fence_put(p->ef);
1138 
1139 	kfd_event_free_process(p);
1140 
1141 	kfd_pasid_free(p->pasid);
1142 	mutex_destroy(&p->mutex);
1143 
1144 	put_task_struct(p->lead_thread);
1145 
1146 	kfree(p);
1147 }
1148 
1149 static void kfd_process_ref_release(struct kref *ref)
1150 {
1151 	struct kfd_process *p = container_of(ref, struct kfd_process, ref);
1152 
1153 	INIT_WORK(&p->release_work, kfd_process_wq_release);
1154 	queue_work(kfd_process_wq, &p->release_work);
1155 }
1156 
1157 static struct mmu_notifier *kfd_process_alloc_notifier(struct mm_struct *mm)
1158 {
1159 	int idx = srcu_read_lock(&kfd_processes_srcu);
1160 	struct kfd_process *p = find_process_by_mm(mm);
1161 
1162 	srcu_read_unlock(&kfd_processes_srcu, idx);
1163 
1164 	return p ? &p->mmu_notifier : ERR_PTR(-ESRCH);
1165 }
1166 
1167 static void kfd_process_free_notifier(struct mmu_notifier *mn)
1168 {
1169 	kfd_unref_process(container_of(mn, struct kfd_process, mmu_notifier));
1170 }
1171 
1172 static void kfd_process_notifier_release_internal(struct kfd_process *p)
1173 {
1174 	int i;
1175 
1176 	cancel_delayed_work_sync(&p->eviction_work);
1177 	cancel_delayed_work_sync(&p->restore_work);
1178 
1179 	for (i = 0; i < p->n_pdds; i++) {
1180 		struct kfd_process_device *pdd = p->pdds[i];
1181 
1182 		/* re-enable GFX OFF since runtime enable with ttmp setup disabled it. */
1183 		if (!kfd_dbg_is_rlc_restore_supported(pdd->dev) && p->runtime_info.ttmp_setup)
1184 			amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
1185 	}
1186 
1187 	/* Indicate to other users that MM is no longer valid */
1188 	p->mm = NULL;
1189 	kfd_dbg_trap_disable(p);
1190 
1191 	if (atomic_read(&p->debugged_process_count) > 0) {
1192 		struct kfd_process *target;
1193 		unsigned int temp;
1194 		int idx = srcu_read_lock(&kfd_processes_srcu);
1195 
1196 		hash_for_each_rcu(kfd_processes_table, temp, target, kfd_processes) {
1197 			if (target->debugger_process && target->debugger_process == p) {
1198 				mutex_lock_nested(&target->mutex, 1);
1199 				kfd_dbg_trap_disable(target);
1200 				mutex_unlock(&target->mutex);
1201 				if (atomic_read(&p->debugged_process_count) == 0)
1202 					break;
1203 			}
1204 		}
1205 
1206 		srcu_read_unlock(&kfd_processes_srcu, idx);
1207 	}
1208 
1209 	mmu_notifier_put(&p->mmu_notifier);
1210 }
1211 
1212 static void kfd_process_notifier_release(struct mmu_notifier *mn,
1213 					struct mm_struct *mm)
1214 {
1215 	struct kfd_process *p;
1216 
1217 	/*
1218 	 * The kfd_process structure can not be free because the
1219 	 * mmu_notifier srcu is read locked
1220 	 */
1221 	p = container_of(mn, struct kfd_process, mmu_notifier);
1222 	if (WARN_ON(p->mm != mm))
1223 		return;
1224 
1225 	mutex_lock(&kfd_processes_mutex);
1226 	/*
1227 	 * Do early return if table is empty.
1228 	 *
1229 	 * This could potentially happen if this function is called concurrently
1230 	 * by mmu_notifier and by kfd_cleanup_pocesses.
1231 	 *
1232 	 */
1233 	if (hash_empty(kfd_processes_table)) {
1234 		mutex_unlock(&kfd_processes_mutex);
1235 		return;
1236 	}
1237 	hash_del_rcu(&p->kfd_processes);
1238 	mutex_unlock(&kfd_processes_mutex);
1239 	synchronize_srcu(&kfd_processes_srcu);
1240 
1241 	kfd_process_notifier_release_internal(p);
1242 }
1243 
1244 static const struct mmu_notifier_ops kfd_process_mmu_notifier_ops = {
1245 	.release = kfd_process_notifier_release,
1246 	.alloc_notifier = kfd_process_alloc_notifier,
1247 	.free_notifier = kfd_process_free_notifier,
1248 };
1249 
1250 /*
1251  * This code handles the case when driver is being unloaded before all
1252  * mm_struct are released.  We need to safely free the kfd_process and
1253  * avoid race conditions with mmu_notifier that might try to free them.
1254  *
1255  */
1256 void kfd_cleanup_processes(void)
1257 {
1258 	struct kfd_process *p;
1259 	struct hlist_node *p_temp;
1260 	unsigned int temp;
1261 	HLIST_HEAD(cleanup_list);
1262 
1263 	/*
1264 	 * Move all remaining kfd_process from the process table to a
1265 	 * temp list for processing.   Once done, callback from mmu_notifier
1266 	 * release will not see the kfd_process in the table and do early return,
1267 	 * avoiding double free issues.
1268 	 */
1269 	mutex_lock(&kfd_processes_mutex);
1270 	hash_for_each_safe(kfd_processes_table, temp, p_temp, p, kfd_processes) {
1271 		hash_del_rcu(&p->kfd_processes);
1272 		synchronize_srcu(&kfd_processes_srcu);
1273 		hlist_add_head(&p->kfd_processes, &cleanup_list);
1274 	}
1275 	mutex_unlock(&kfd_processes_mutex);
1276 
1277 	hlist_for_each_entry_safe(p, p_temp, &cleanup_list, kfd_processes)
1278 		kfd_process_notifier_release_internal(p);
1279 
1280 	/*
1281 	 * Ensures that all outstanding free_notifier get called, triggering
1282 	 * the release of the kfd_process struct.
1283 	 */
1284 	mmu_notifier_synchronize();
1285 }
1286 
1287 int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep)
1288 {
1289 	unsigned long  offset;
1290 	int i;
1291 
1292 	if (p->has_cwsr)
1293 		return 0;
1294 
1295 	for (i = 0; i < p->n_pdds; i++) {
1296 		struct kfd_node *dev = p->pdds[i]->dev;
1297 		struct qcm_process_device *qpd = &p->pdds[i]->qpd;
1298 
1299 		if (!dev->kfd->cwsr_enabled || qpd->cwsr_kaddr || qpd->cwsr_base)
1300 			continue;
1301 
1302 		offset = KFD_MMAP_TYPE_RESERVED_MEM | KFD_MMAP_GPU_ID(dev->id);
1303 		qpd->tba_addr = (int64_t)vm_mmap(filep, 0,
1304 			KFD_CWSR_TBA_TMA_SIZE, PROT_READ | PROT_EXEC,
1305 			MAP_SHARED, offset);
1306 
1307 		if (IS_ERR_VALUE(qpd->tba_addr)) {
1308 			int err = qpd->tba_addr;
1309 
1310 			pr_err("Failure to set tba address. error %d.\n", err);
1311 			qpd->tba_addr = 0;
1312 			qpd->cwsr_kaddr = NULL;
1313 			return err;
1314 		}
1315 
1316 		memcpy(qpd->cwsr_kaddr, dev->kfd->cwsr_isa, dev->kfd->cwsr_isa_size);
1317 
1318 		kfd_process_set_trap_debug_flag(qpd, p->debug_trap_enabled);
1319 
1320 		qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET;
1321 		pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n",
1322 			qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr);
1323 	}
1324 
1325 	p->has_cwsr = true;
1326 
1327 	return 0;
1328 }
1329 
1330 static int kfd_process_device_init_cwsr_dgpu(struct kfd_process_device *pdd)
1331 {
1332 	struct kfd_node *dev = pdd->dev;
1333 	struct qcm_process_device *qpd = &pdd->qpd;
1334 	uint32_t flags = KFD_IOC_ALLOC_MEM_FLAGS_GTT
1335 			| KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE
1336 			| KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
1337 	struct kgd_mem *mem;
1338 	void *kaddr;
1339 	int ret;
1340 
1341 	if (!dev->kfd->cwsr_enabled || qpd->cwsr_kaddr || !qpd->cwsr_base)
1342 		return 0;
1343 
1344 	/* cwsr_base is only set for dGPU */
1345 	ret = kfd_process_alloc_gpuvm(pdd, qpd->cwsr_base,
1346 				      KFD_CWSR_TBA_TMA_SIZE, flags, &mem, &kaddr);
1347 	if (ret)
1348 		return ret;
1349 
1350 	qpd->cwsr_mem = mem;
1351 	qpd->cwsr_kaddr = kaddr;
1352 	qpd->tba_addr = qpd->cwsr_base;
1353 
1354 	memcpy(qpd->cwsr_kaddr, dev->kfd->cwsr_isa, dev->kfd->cwsr_isa_size);
1355 
1356 	kfd_process_set_trap_debug_flag(&pdd->qpd,
1357 					pdd->process->debug_trap_enabled);
1358 
1359 	qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET;
1360 	pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n",
1361 		 qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr);
1362 
1363 	return 0;
1364 }
1365 
1366 static void kfd_process_device_destroy_cwsr_dgpu(struct kfd_process_device *pdd)
1367 {
1368 	struct kfd_node *dev = pdd->dev;
1369 	struct qcm_process_device *qpd = &pdd->qpd;
1370 
1371 	if (!dev->kfd->cwsr_enabled || !qpd->cwsr_kaddr || !qpd->cwsr_base)
1372 		return;
1373 
1374 	kfd_process_free_gpuvm(qpd->cwsr_mem, pdd, &qpd->cwsr_kaddr);
1375 }
1376 
1377 void kfd_process_set_trap_handler(struct qcm_process_device *qpd,
1378 				  uint64_t tba_addr,
1379 				  uint64_t tma_addr)
1380 {
1381 	if (qpd->cwsr_kaddr) {
1382 		/* KFD trap handler is bound, record as second-level TBA/TMA
1383 		 * in first-level TMA. First-level trap will jump to second.
1384 		 */
1385 		uint64_t *tma =
1386 			(uint64_t *)(qpd->cwsr_kaddr + KFD_CWSR_TMA_OFFSET);
1387 		tma[0] = tba_addr;
1388 		tma[1] = tma_addr;
1389 	} else {
1390 		/* No trap handler bound, bind as first-level TBA/TMA. */
1391 		qpd->tba_addr = tba_addr;
1392 		qpd->tma_addr = tma_addr;
1393 	}
1394 }
1395 
1396 bool kfd_process_xnack_mode(struct kfd_process *p, bool supported)
1397 {
1398 	int i;
1399 
1400 	/* On most GFXv9 GPUs, the retry mode in the SQ must match the
1401 	 * boot time retry setting. Mixing processes with different
1402 	 * XNACK/retry settings can hang the GPU.
1403 	 *
1404 	 * Different GPUs can have different noretry settings depending
1405 	 * on HW bugs or limitations. We need to find at least one
1406 	 * XNACK mode for this process that's compatible with all GPUs.
1407 	 * Fortunately GPUs with retry enabled (noretry=0) can run code
1408 	 * built for XNACK-off. On GFXv9 it may perform slower.
1409 	 *
1410 	 * Therefore applications built for XNACK-off can always be
1411 	 * supported and will be our fallback if any GPU does not
1412 	 * support retry.
1413 	 */
1414 	for (i = 0; i < p->n_pdds; i++) {
1415 		struct kfd_node *dev = p->pdds[i]->dev;
1416 
1417 		/* Only consider GFXv9 and higher GPUs. Older GPUs don't
1418 		 * support the SVM APIs and don't need to be considered
1419 		 * for the XNACK mode selection.
1420 		 */
1421 		if (!KFD_IS_SOC15(dev))
1422 			continue;
1423 		/* Aldebaran can always support XNACK because it can support
1424 		 * per-process XNACK mode selection. But let the dev->noretry
1425 		 * setting still influence the default XNACK mode.
1426 		 */
1427 		if (supported && KFD_SUPPORT_XNACK_PER_PROCESS(dev))
1428 			continue;
1429 
1430 		/* GFXv10 and later GPUs do not support shader preemption
1431 		 * during page faults. This can lead to poor QoS for queue
1432 		 * management and memory-manager-related preemptions or
1433 		 * even deadlocks.
1434 		 */
1435 		if (KFD_GC_VERSION(dev) >= IP_VERSION(10, 1, 1))
1436 			return false;
1437 
1438 		if (dev->kfd->noretry)
1439 			return false;
1440 	}
1441 
1442 	return true;
1443 }
1444 
1445 void kfd_process_set_trap_debug_flag(struct qcm_process_device *qpd,
1446 				     bool enabled)
1447 {
1448 	if (qpd->cwsr_kaddr) {
1449 		uint64_t *tma =
1450 			(uint64_t *)(qpd->cwsr_kaddr + KFD_CWSR_TMA_OFFSET);
1451 		tma[2] = enabled;
1452 	}
1453 }
1454 
1455 /*
1456  * On return the kfd_process is fully operational and will be freed when the
1457  * mm is released
1458  */
1459 static struct kfd_process *create_process(const struct task_struct *thread)
1460 {
1461 	struct kfd_process *process;
1462 	struct mmu_notifier *mn;
1463 	int err = -ENOMEM;
1464 
1465 	process = kzalloc(sizeof(*process), GFP_KERNEL);
1466 	if (!process)
1467 		goto err_alloc_process;
1468 
1469 	kref_init(&process->ref);
1470 	mutex_init(&process->mutex);
1471 	process->mm = thread->mm;
1472 	process->lead_thread = thread->group_leader;
1473 	process->n_pdds = 0;
1474 	process->queues_paused = false;
1475 	INIT_DELAYED_WORK(&process->eviction_work, evict_process_worker);
1476 	INIT_DELAYED_WORK(&process->restore_work, restore_process_worker);
1477 	process->last_restore_timestamp = get_jiffies_64();
1478 	err = kfd_event_init_process(process);
1479 	if (err)
1480 		goto err_event_init;
1481 	process->is_32bit_user_mode = in_compat_syscall();
1482 	process->debug_trap_enabled = false;
1483 	process->debugger_process = NULL;
1484 	process->exception_enable_mask = 0;
1485 	atomic_set(&process->debugged_process_count, 0);
1486 	sema_init(&process->runtime_enable_sema, 0);
1487 
1488 	process->pasid = kfd_pasid_alloc();
1489 	if (process->pasid == 0) {
1490 		err = -ENOSPC;
1491 		goto err_alloc_pasid;
1492 	}
1493 
1494 	err = pqm_init(&process->pqm, process);
1495 	if (err != 0)
1496 		goto err_process_pqm_init;
1497 
1498 	/* init process apertures*/
1499 	err = kfd_init_apertures(process);
1500 	if (err != 0)
1501 		goto err_init_apertures;
1502 
1503 	/* Check XNACK support after PDDs are created in kfd_init_apertures */
1504 	process->xnack_enabled = kfd_process_xnack_mode(process, false);
1505 
1506 	err = svm_range_list_init(process);
1507 	if (err)
1508 		goto err_init_svm_range_list;
1509 
1510 	/* alloc_notifier needs to find the process in the hash table */
1511 	hash_add_rcu(kfd_processes_table, &process->kfd_processes,
1512 			(uintptr_t)process->mm);
1513 
1514 	/* Avoid free_notifier to start kfd_process_wq_release if
1515 	 * mmu_notifier_get failed because of pending signal.
1516 	 */
1517 	kref_get(&process->ref);
1518 
1519 	/* MMU notifier registration must be the last call that can fail
1520 	 * because after this point we cannot unwind the process creation.
1521 	 * After this point, mmu_notifier_put will trigger the cleanup by
1522 	 * dropping the last process reference in the free_notifier.
1523 	 */
1524 	mn = mmu_notifier_get(&kfd_process_mmu_notifier_ops, process->mm);
1525 	if (IS_ERR(mn)) {
1526 		err = PTR_ERR(mn);
1527 		goto err_register_notifier;
1528 	}
1529 	BUG_ON(mn != &process->mmu_notifier);
1530 
1531 	kfd_unref_process(process);
1532 	get_task_struct(process->lead_thread);
1533 
1534 	INIT_WORK(&process->debug_event_workarea, debug_event_write_work_handler);
1535 
1536 	return process;
1537 
1538 err_register_notifier:
1539 	hash_del_rcu(&process->kfd_processes);
1540 	svm_range_list_fini(process);
1541 err_init_svm_range_list:
1542 	kfd_process_free_outstanding_kfd_bos(process);
1543 	kfd_process_destroy_pdds(process);
1544 err_init_apertures:
1545 	pqm_uninit(&process->pqm);
1546 err_process_pqm_init:
1547 	kfd_pasid_free(process->pasid);
1548 err_alloc_pasid:
1549 	kfd_event_free_process(process);
1550 err_event_init:
1551 	mutex_destroy(&process->mutex);
1552 	kfree(process);
1553 err_alloc_process:
1554 	return ERR_PTR(err);
1555 }
1556 
1557 struct kfd_process_device *kfd_get_process_device_data(struct kfd_node *dev,
1558 							struct kfd_process *p)
1559 {
1560 	int i;
1561 
1562 	for (i = 0; i < p->n_pdds; i++)
1563 		if (p->pdds[i]->dev == dev)
1564 			return p->pdds[i];
1565 
1566 	return NULL;
1567 }
1568 
1569 struct kfd_process_device *kfd_create_process_device_data(struct kfd_node *dev,
1570 							struct kfd_process *p)
1571 {
1572 	struct kfd_process_device *pdd = NULL;
1573 	int retval = 0;
1574 
1575 	if (WARN_ON_ONCE(p->n_pdds >= MAX_GPU_INSTANCE))
1576 		return NULL;
1577 	pdd = kzalloc(sizeof(*pdd), GFP_KERNEL);
1578 	if (!pdd)
1579 		return NULL;
1580 
1581 	pdd->dev = dev;
1582 	INIT_LIST_HEAD(&pdd->qpd.queues_list);
1583 	INIT_LIST_HEAD(&pdd->qpd.priv_queue_list);
1584 	pdd->qpd.dqm = dev->dqm;
1585 	pdd->qpd.pqm = &p->pqm;
1586 	pdd->qpd.evicted = 0;
1587 	pdd->qpd.mapped_gws_queue = false;
1588 	pdd->process = p;
1589 	pdd->bound = PDD_UNBOUND;
1590 	pdd->already_dequeued = false;
1591 	pdd->runtime_inuse = false;
1592 	pdd->vram_usage = 0;
1593 	pdd->sdma_past_activity_counter = 0;
1594 	pdd->user_gpu_id = dev->id;
1595 	atomic64_set(&pdd->evict_duration_counter, 0);
1596 
1597 	if (dev->kfd->shared_resources.enable_mes) {
1598 		retval = amdgpu_amdkfd_alloc_gtt_mem(dev->adev,
1599 						AMDGPU_MES_PROC_CTX_SIZE,
1600 						&pdd->proc_ctx_bo,
1601 						&pdd->proc_ctx_gpu_addr,
1602 						&pdd->proc_ctx_cpu_ptr,
1603 						false);
1604 		if (retval) {
1605 			pr_err("failed to allocate process context bo\n");
1606 			goto err_free_pdd;
1607 		}
1608 		memset(pdd->proc_ctx_cpu_ptr, 0, AMDGPU_MES_PROC_CTX_SIZE);
1609 	}
1610 
1611 	p->pdds[p->n_pdds++] = pdd;
1612 	if (kfd_dbg_is_per_vmid_supported(pdd->dev))
1613 		pdd->spi_dbg_override = pdd->dev->kfd2kgd->disable_debug_trap(
1614 							pdd->dev->adev,
1615 							false,
1616 							0);
1617 
1618 	/* Init idr used for memory handle translation */
1619 	idr_init(&pdd->alloc_idr);
1620 
1621 	return pdd;
1622 
1623 err_free_pdd:
1624 	kfree(pdd);
1625 	return NULL;
1626 }
1627 
1628 /**
1629  * kfd_process_device_init_vm - Initialize a VM for a process-device
1630  *
1631  * @pdd: The process-device
1632  * @drm_file: Optional pointer to a DRM file descriptor
1633  *
1634  * If @drm_file is specified, it will be used to acquire the VM from
1635  * that file descriptor. If successful, the @pdd takes ownership of
1636  * the file descriptor.
1637  *
1638  * If @drm_file is NULL, a new VM is created.
1639  *
1640  * Returns 0 on success, -errno on failure.
1641  */
1642 int kfd_process_device_init_vm(struct kfd_process_device *pdd,
1643 			       struct file *drm_file)
1644 {
1645 	struct amdgpu_fpriv *drv_priv;
1646 	struct amdgpu_vm *avm;
1647 	struct kfd_process *p;
1648 	struct kfd_node *dev;
1649 	int ret;
1650 
1651 	if (!drm_file)
1652 		return -EINVAL;
1653 
1654 	if (pdd->drm_priv)
1655 		return -EBUSY;
1656 
1657 	ret = amdgpu_file_to_fpriv(drm_file, &drv_priv);
1658 	if (ret)
1659 		return ret;
1660 	avm = &drv_priv->vm;
1661 
1662 	p = pdd->process;
1663 	dev = pdd->dev;
1664 
1665 	ret = amdgpu_amdkfd_gpuvm_acquire_process_vm(dev->adev, avm,
1666 						     &p->kgd_process_info,
1667 						     &p->ef);
1668 	if (ret) {
1669 		pr_err("Failed to create process VM object\n");
1670 		return ret;
1671 	}
1672 	pdd->drm_priv = drm_file->private_data;
1673 	atomic64_set(&pdd->tlb_seq, 0);
1674 
1675 	ret = kfd_process_device_reserve_ib_mem(pdd);
1676 	if (ret)
1677 		goto err_reserve_ib_mem;
1678 	ret = kfd_process_device_init_cwsr_dgpu(pdd);
1679 	if (ret)
1680 		goto err_init_cwsr;
1681 
1682 	ret = amdgpu_amdkfd_gpuvm_set_vm_pasid(dev->adev, avm, p->pasid);
1683 	if (ret)
1684 		goto err_set_pasid;
1685 
1686 	pdd->drm_file = drm_file;
1687 
1688 	return 0;
1689 
1690 err_set_pasid:
1691 	kfd_process_device_destroy_cwsr_dgpu(pdd);
1692 err_init_cwsr:
1693 	kfd_process_device_destroy_ib_mem(pdd);
1694 err_reserve_ib_mem:
1695 	pdd->drm_priv = NULL;
1696 	amdgpu_amdkfd_gpuvm_destroy_cb(dev->adev, avm);
1697 
1698 	return ret;
1699 }
1700 
1701 /*
1702  * Direct the IOMMU to bind the process (specifically the pasid->mm)
1703  * to the device.
1704  * Unbinding occurs when the process dies or the device is removed.
1705  *
1706  * Assumes that the process lock is held.
1707  */
1708 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_node *dev,
1709 							struct kfd_process *p)
1710 {
1711 	struct kfd_process_device *pdd;
1712 	int err;
1713 
1714 	pdd = kfd_get_process_device_data(dev, p);
1715 	if (!pdd) {
1716 		pr_err("Process device data doesn't exist\n");
1717 		return ERR_PTR(-ENOMEM);
1718 	}
1719 
1720 	if (!pdd->drm_priv)
1721 		return ERR_PTR(-ENODEV);
1722 
1723 	/*
1724 	 * signal runtime-pm system to auto resume and prevent
1725 	 * further runtime suspend once device pdd is created until
1726 	 * pdd is destroyed.
1727 	 */
1728 	if (!pdd->runtime_inuse) {
1729 		err = pm_runtime_get_sync(adev_to_drm(dev->adev)->dev);
1730 		if (err < 0) {
1731 			pm_runtime_put_autosuspend(adev_to_drm(dev->adev)->dev);
1732 			return ERR_PTR(err);
1733 		}
1734 	}
1735 
1736 	/*
1737 	 * make sure that runtime_usage counter is incremented just once
1738 	 * per pdd
1739 	 */
1740 	pdd->runtime_inuse = true;
1741 
1742 	return pdd;
1743 }
1744 
1745 /* Create specific handle mapped to mem from process local memory idr
1746  * Assumes that the process lock is held.
1747  */
1748 int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
1749 					void *mem)
1750 {
1751 	return idr_alloc(&pdd->alloc_idr, mem, 0, 0, GFP_KERNEL);
1752 }
1753 
1754 /* Translate specific handle from process local memory idr
1755  * Assumes that the process lock is held.
1756  */
1757 void *kfd_process_device_translate_handle(struct kfd_process_device *pdd,
1758 					int handle)
1759 {
1760 	if (handle < 0)
1761 		return NULL;
1762 
1763 	return idr_find(&pdd->alloc_idr, handle);
1764 }
1765 
1766 /* Remove specific handle from process local memory idr
1767  * Assumes that the process lock is held.
1768  */
1769 void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,
1770 					int handle)
1771 {
1772 	if (handle >= 0)
1773 		idr_remove(&pdd->alloc_idr, handle);
1774 }
1775 
1776 /* This increments the process->ref counter. */
1777 struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid)
1778 {
1779 	struct kfd_process *p, *ret_p = NULL;
1780 	unsigned int temp;
1781 
1782 	int idx = srcu_read_lock(&kfd_processes_srcu);
1783 
1784 	hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1785 		if (p->pasid == pasid) {
1786 			kref_get(&p->ref);
1787 			ret_p = p;
1788 			break;
1789 		}
1790 	}
1791 
1792 	srcu_read_unlock(&kfd_processes_srcu, idx);
1793 
1794 	return ret_p;
1795 }
1796 
1797 /* This increments the process->ref counter. */
1798 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm)
1799 {
1800 	struct kfd_process *p;
1801 
1802 	int idx = srcu_read_lock(&kfd_processes_srcu);
1803 
1804 	p = find_process_by_mm(mm);
1805 	if (p)
1806 		kref_get(&p->ref);
1807 
1808 	srcu_read_unlock(&kfd_processes_srcu, idx);
1809 
1810 	return p;
1811 }
1812 
1813 /* kfd_process_evict_queues - Evict all user queues of a process
1814  *
1815  * Eviction is reference-counted per process-device. This means multiple
1816  * evictions from different sources can be nested safely.
1817  */
1818 int kfd_process_evict_queues(struct kfd_process *p, uint32_t trigger)
1819 {
1820 	int r = 0;
1821 	int i;
1822 	unsigned int n_evicted = 0;
1823 
1824 	for (i = 0; i < p->n_pdds; i++) {
1825 		struct kfd_process_device *pdd = p->pdds[i];
1826 
1827 		kfd_smi_event_queue_eviction(pdd->dev, p->lead_thread->pid,
1828 					     trigger);
1829 
1830 		r = pdd->dev->dqm->ops.evict_process_queues(pdd->dev->dqm,
1831 							    &pdd->qpd);
1832 		/* evict return -EIO if HWS is hang or asic is resetting, in this case
1833 		 * we would like to set all the queues to be in evicted state to prevent
1834 		 * them been add back since they actually not be saved right now.
1835 		 */
1836 		if (r && r != -EIO) {
1837 			pr_err("Failed to evict process queues\n");
1838 			goto fail;
1839 		}
1840 		n_evicted++;
1841 	}
1842 
1843 	return r;
1844 
1845 fail:
1846 	/* To keep state consistent, roll back partial eviction by
1847 	 * restoring queues
1848 	 */
1849 	for (i = 0; i < p->n_pdds; i++) {
1850 		struct kfd_process_device *pdd = p->pdds[i];
1851 
1852 		if (n_evicted == 0)
1853 			break;
1854 
1855 		kfd_smi_event_queue_restore(pdd->dev, p->lead_thread->pid);
1856 
1857 		if (pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm,
1858 							      &pdd->qpd))
1859 			pr_err("Failed to restore queues\n");
1860 
1861 		n_evicted--;
1862 	}
1863 
1864 	return r;
1865 }
1866 
1867 /* kfd_process_restore_queues - Restore all user queues of a process */
1868 int kfd_process_restore_queues(struct kfd_process *p)
1869 {
1870 	int r, ret = 0;
1871 	int i;
1872 
1873 	for (i = 0; i < p->n_pdds; i++) {
1874 		struct kfd_process_device *pdd = p->pdds[i];
1875 
1876 		kfd_smi_event_queue_restore(pdd->dev, p->lead_thread->pid);
1877 
1878 		r = pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm,
1879 							      &pdd->qpd);
1880 		if (r) {
1881 			pr_err("Failed to restore process queues\n");
1882 			if (!ret)
1883 				ret = r;
1884 		}
1885 	}
1886 
1887 	return ret;
1888 }
1889 
1890 int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id)
1891 {
1892 	int i;
1893 
1894 	for (i = 0; i < p->n_pdds; i++)
1895 		if (p->pdds[i] && gpu_id == p->pdds[i]->user_gpu_id)
1896 			return i;
1897 	return -EINVAL;
1898 }
1899 
1900 int
1901 kfd_process_gpuid_from_node(struct kfd_process *p, struct kfd_node *node,
1902 			    uint32_t *gpuid, uint32_t *gpuidx)
1903 {
1904 	int i;
1905 
1906 	for (i = 0; i < p->n_pdds; i++)
1907 		if (p->pdds[i] && p->pdds[i]->dev == node) {
1908 			*gpuid = p->pdds[i]->user_gpu_id;
1909 			*gpuidx = i;
1910 			return 0;
1911 		}
1912 	return -EINVAL;
1913 }
1914 
1915 static void evict_process_worker(struct work_struct *work)
1916 {
1917 	int ret;
1918 	struct kfd_process *p;
1919 	struct delayed_work *dwork;
1920 
1921 	dwork = to_delayed_work(work);
1922 
1923 	/* Process termination destroys this worker thread. So during the
1924 	 * lifetime of this thread, kfd_process p will be valid
1925 	 */
1926 	p = container_of(dwork, struct kfd_process, eviction_work);
1927 	WARN_ONCE(p->last_eviction_seqno != p->ef->seqno,
1928 		  "Eviction fence mismatch\n");
1929 
1930 	/* Narrow window of overlap between restore and evict work
1931 	 * item is possible. Once amdgpu_amdkfd_gpuvm_restore_process_bos
1932 	 * unreserves KFD BOs, it is possible to evicted again. But
1933 	 * restore has few more steps of finish. So lets wait for any
1934 	 * previous restore work to complete
1935 	 */
1936 	flush_delayed_work(&p->restore_work);
1937 
1938 	pr_debug("Started evicting pasid 0x%x\n", p->pasid);
1939 	ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_TRIGGER_TTM);
1940 	if (!ret) {
1941 		dma_fence_signal(p->ef);
1942 		dma_fence_put(p->ef);
1943 		p->ef = NULL;
1944 		queue_delayed_work(kfd_restore_wq, &p->restore_work,
1945 				msecs_to_jiffies(PROCESS_RESTORE_TIME_MS));
1946 
1947 		pr_debug("Finished evicting pasid 0x%x\n", p->pasid);
1948 	} else
1949 		pr_err("Failed to evict queues of pasid 0x%x\n", p->pasid);
1950 }
1951 
1952 static void restore_process_worker(struct work_struct *work)
1953 {
1954 	struct delayed_work *dwork;
1955 	struct kfd_process *p;
1956 	int ret = 0;
1957 
1958 	dwork = to_delayed_work(work);
1959 
1960 	/* Process termination destroys this worker thread. So during the
1961 	 * lifetime of this thread, kfd_process p will be valid
1962 	 */
1963 	p = container_of(dwork, struct kfd_process, restore_work);
1964 	pr_debug("Started restoring pasid 0x%x\n", p->pasid);
1965 
1966 	/* Setting last_restore_timestamp before successful restoration.
1967 	 * Otherwise this would have to be set by KGD (restore_process_bos)
1968 	 * before KFD BOs are unreserved. If not, the process can be evicted
1969 	 * again before the timestamp is set.
1970 	 * If restore fails, the timestamp will be set again in the next
1971 	 * attempt. This would mean that the minimum GPU quanta would be
1972 	 * PROCESS_ACTIVE_TIME_MS - (time to execute the following two
1973 	 * functions)
1974 	 */
1975 
1976 	p->last_restore_timestamp = get_jiffies_64();
1977 	/* VMs may not have been acquired yet during debugging. */
1978 	if (p->kgd_process_info)
1979 		ret = amdgpu_amdkfd_gpuvm_restore_process_bos(p->kgd_process_info,
1980 							     &p->ef);
1981 	if (ret) {
1982 		pr_debug("Failed to restore BOs of pasid 0x%x, retry after %d ms\n",
1983 			 p->pasid, PROCESS_BACK_OFF_TIME_MS);
1984 		ret = queue_delayed_work(kfd_restore_wq, &p->restore_work,
1985 				msecs_to_jiffies(PROCESS_BACK_OFF_TIME_MS));
1986 		WARN(!ret, "reschedule restore work failed\n");
1987 		return;
1988 	}
1989 
1990 	ret = kfd_process_restore_queues(p);
1991 	if (!ret)
1992 		pr_debug("Finished restoring pasid 0x%x\n", p->pasid);
1993 	else
1994 		pr_err("Failed to restore queues of pasid 0x%x\n", p->pasid);
1995 }
1996 
1997 void kfd_suspend_all_processes(void)
1998 {
1999 	struct kfd_process *p;
2000 	unsigned int temp;
2001 	int idx = srcu_read_lock(&kfd_processes_srcu);
2002 
2003 	WARN(debug_evictions, "Evicting all processes");
2004 	hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
2005 		cancel_delayed_work_sync(&p->eviction_work);
2006 		flush_delayed_work(&p->restore_work);
2007 
2008 		if (kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_TRIGGER_SUSPEND))
2009 			pr_err("Failed to suspend process 0x%x\n", p->pasid);
2010 		dma_fence_signal(p->ef);
2011 		dma_fence_put(p->ef);
2012 		p->ef = NULL;
2013 	}
2014 	srcu_read_unlock(&kfd_processes_srcu, idx);
2015 }
2016 
2017 int kfd_resume_all_processes(void)
2018 {
2019 	struct kfd_process *p;
2020 	unsigned int temp;
2021 	int ret = 0, idx = srcu_read_lock(&kfd_processes_srcu);
2022 
2023 	hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
2024 		if (!queue_delayed_work(kfd_restore_wq, &p->restore_work, 0)) {
2025 			pr_err("Restore process %d failed during resume\n",
2026 			       p->pasid);
2027 			ret = -EFAULT;
2028 		}
2029 	}
2030 	srcu_read_unlock(&kfd_processes_srcu, idx);
2031 	return ret;
2032 }
2033 
2034 int kfd_reserved_mem_mmap(struct kfd_node *dev, struct kfd_process *process,
2035 			  struct vm_area_struct *vma)
2036 {
2037 	struct kfd_process_device *pdd;
2038 	struct qcm_process_device *qpd;
2039 
2040 	if ((vma->vm_end - vma->vm_start) != KFD_CWSR_TBA_TMA_SIZE) {
2041 		pr_err("Incorrect CWSR mapping size.\n");
2042 		return -EINVAL;
2043 	}
2044 
2045 	pdd = kfd_get_process_device_data(dev, process);
2046 	if (!pdd)
2047 		return -EINVAL;
2048 	qpd = &pdd->qpd;
2049 
2050 	qpd->cwsr_kaddr = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
2051 					get_order(KFD_CWSR_TBA_TMA_SIZE));
2052 	if (!qpd->cwsr_kaddr) {
2053 		pr_err("Error allocating per process CWSR buffer.\n");
2054 		return -ENOMEM;
2055 	}
2056 
2057 	vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND
2058 		| VM_NORESERVE | VM_DONTDUMP | VM_PFNMAP);
2059 	/* Mapping pages to user process */
2060 	return remap_pfn_range(vma, vma->vm_start,
2061 			       PFN_DOWN(__pa(qpd->cwsr_kaddr)),
2062 			       KFD_CWSR_TBA_TMA_SIZE, vma->vm_page_prot);
2063 }
2064 
2065 void kfd_flush_tlb(struct kfd_process_device *pdd, enum TLB_FLUSH_TYPE type)
2066 {
2067 	struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv);
2068 	uint64_t tlb_seq = amdgpu_vm_tlb_seq(vm);
2069 	struct kfd_node *dev = pdd->dev;
2070 	uint32_t xcc_mask = dev->xcc_mask;
2071 	int xcc = 0;
2072 
2073 	/*
2074 	 * It can be that we race and lose here, but that is extremely unlikely
2075 	 * and the worst thing which could happen is that we flush the changes
2076 	 * into the TLB once more which is harmless.
2077 	 */
2078 	if (atomic64_xchg(&pdd->tlb_seq, tlb_seq) == tlb_seq)
2079 		return;
2080 
2081 	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
2082 		/* Nothing to flush until a VMID is assigned, which
2083 		 * only happens when the first queue is created.
2084 		 */
2085 		if (pdd->qpd.vmid)
2086 			amdgpu_amdkfd_flush_gpu_tlb_vmid(dev->adev,
2087 							pdd->qpd.vmid);
2088 	} else {
2089 		for_each_inst(xcc, xcc_mask)
2090 			amdgpu_amdkfd_flush_gpu_tlb_pasid(
2091 				dev->adev, pdd->process->pasid, type, xcc);
2092 	}
2093 }
2094 
2095 /* assumes caller holds process lock. */
2096 int kfd_process_drain_interrupts(struct kfd_process_device *pdd)
2097 {
2098 	uint32_t irq_drain_fence[8];
2099 	uint8_t node_id = 0;
2100 	int r = 0;
2101 
2102 	if (!KFD_IS_SOC15(pdd->dev))
2103 		return 0;
2104 
2105 	pdd->process->irq_drain_is_open = true;
2106 
2107 	memset(irq_drain_fence, 0, sizeof(irq_drain_fence));
2108 	irq_drain_fence[0] = (KFD_IRQ_FENCE_SOURCEID << 8) |
2109 							KFD_IRQ_FENCE_CLIENTID;
2110 	irq_drain_fence[3] = pdd->process->pasid;
2111 
2112 	/*
2113 	 * For GFX 9.4.3, send the NodeId also in IH cookie DW[3]
2114 	 */
2115 	if (KFD_GC_VERSION(pdd->dev->kfd) == IP_VERSION(9, 4, 3)) {
2116 		node_id = ffs(pdd->dev->interrupt_bitmap) - 1;
2117 		irq_drain_fence[3] |= node_id << 16;
2118 	}
2119 
2120 	/* ensure stale irqs scheduled KFD interrupts and send drain fence. */
2121 	if (amdgpu_amdkfd_send_close_event_drain_irq(pdd->dev->adev,
2122 						     irq_drain_fence)) {
2123 		pdd->process->irq_drain_is_open = false;
2124 		return 0;
2125 	}
2126 
2127 	r = wait_event_interruptible(pdd->process->wait_irq_drain,
2128 				     !READ_ONCE(pdd->process->irq_drain_is_open));
2129 	if (r)
2130 		pdd->process->irq_drain_is_open = false;
2131 
2132 	return r;
2133 }
2134 
2135 void kfd_process_close_interrupt_drain(unsigned int pasid)
2136 {
2137 	struct kfd_process *p;
2138 
2139 	p = kfd_lookup_process_by_pasid(pasid);
2140 
2141 	if (!p)
2142 		return;
2143 
2144 	WRITE_ONCE(p->irq_drain_is_open, false);
2145 	wake_up_all(&p->wait_irq_drain);
2146 	kfd_unref_process(p);
2147 }
2148 
2149 struct send_exception_work_handler_workarea {
2150 	struct work_struct work;
2151 	struct kfd_process *p;
2152 	unsigned int queue_id;
2153 	uint64_t error_reason;
2154 };
2155 
2156 static void send_exception_work_handler(struct work_struct *work)
2157 {
2158 	struct send_exception_work_handler_workarea *workarea;
2159 	struct kfd_process *p;
2160 	struct queue *q;
2161 	struct mm_struct *mm;
2162 	struct kfd_context_save_area_header __user *csa_header;
2163 	uint64_t __user *err_payload_ptr;
2164 	uint64_t cur_err;
2165 	uint32_t ev_id;
2166 
2167 	workarea = container_of(work,
2168 				struct send_exception_work_handler_workarea,
2169 				work);
2170 	p = workarea->p;
2171 
2172 	mm = get_task_mm(p->lead_thread);
2173 
2174 	if (!mm)
2175 		return;
2176 
2177 	kthread_use_mm(mm);
2178 
2179 	q = pqm_get_user_queue(&p->pqm, workarea->queue_id);
2180 
2181 	if (!q)
2182 		goto out;
2183 
2184 	csa_header = (void __user *)q->properties.ctx_save_restore_area_address;
2185 
2186 	get_user(err_payload_ptr, (uint64_t __user **)&csa_header->err_payload_addr);
2187 	get_user(cur_err, err_payload_ptr);
2188 	cur_err |= workarea->error_reason;
2189 	put_user(cur_err, err_payload_ptr);
2190 	get_user(ev_id, &csa_header->err_event_id);
2191 
2192 	kfd_set_event(p, ev_id);
2193 
2194 out:
2195 	kthread_unuse_mm(mm);
2196 	mmput(mm);
2197 }
2198 
2199 int kfd_send_exception_to_runtime(struct kfd_process *p,
2200 			unsigned int queue_id,
2201 			uint64_t error_reason)
2202 {
2203 	struct send_exception_work_handler_workarea worker;
2204 
2205 	INIT_WORK_ONSTACK(&worker.work, send_exception_work_handler);
2206 
2207 	worker.p = p;
2208 	worker.queue_id = queue_id;
2209 	worker.error_reason = error_reason;
2210 
2211 	schedule_work(&worker.work);
2212 	flush_work(&worker.work);
2213 	destroy_work_on_stack(&worker.work);
2214 
2215 	return 0;
2216 }
2217 
2218 struct kfd_process_device *kfd_process_device_data_by_id(struct kfd_process *p, uint32_t gpu_id)
2219 {
2220 	int i;
2221 
2222 	if (gpu_id) {
2223 		for (i = 0; i < p->n_pdds; i++) {
2224 			struct kfd_process_device *pdd = p->pdds[i];
2225 
2226 			if (pdd->user_gpu_id == gpu_id)
2227 				return pdd;
2228 		}
2229 	}
2230 	return NULL;
2231 }
2232 
2233 int kfd_process_get_user_gpu_id(struct kfd_process *p, uint32_t actual_gpu_id)
2234 {
2235 	int i;
2236 
2237 	if (!actual_gpu_id)
2238 		return 0;
2239 
2240 	for (i = 0; i < p->n_pdds; i++) {
2241 		struct kfd_process_device *pdd = p->pdds[i];
2242 
2243 		if (pdd->dev->id == actual_gpu_id)
2244 			return pdd->user_gpu_id;
2245 	}
2246 	return -EINVAL;
2247 }
2248 
2249 #if defined(CONFIG_DEBUG_FS)
2250 
2251 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data)
2252 {
2253 	struct kfd_process *p;
2254 	unsigned int temp;
2255 	int r = 0;
2256 
2257 	int idx = srcu_read_lock(&kfd_processes_srcu);
2258 
2259 	hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
2260 		seq_printf(m, "Process %d PASID 0x%x:\n",
2261 			   p->lead_thread->tgid, p->pasid);
2262 
2263 		mutex_lock(&p->mutex);
2264 		r = pqm_debugfs_mqds(m, &p->pqm);
2265 		mutex_unlock(&p->mutex);
2266 
2267 		if (r)
2268 			break;
2269 	}
2270 
2271 	srcu_read_unlock(&kfd_processes_srcu, idx);
2272 
2273 	return r;
2274 }
2275 
2276 #endif
2277