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/amd-iommu.h>
32	#include <linux/notifier.h>
33	#include <linux/compat.h>
34	#include <linux/mman.h>
35	#include <linux/file.h>
36	#include <linux/pm_runtime.h>
37	#include "amdgpu_amdkfd.h"
38	#include "amdgpu.h"
39
40	struct mm_struct;
41
42	#include "kfd_priv.h"
43	#include "kfd_device_queue_manager.h"
44	#include "kfd_iommu.h"
45	#include "kfd_svm.h"
46	#include "kfd_smi_events.h"
47
48	/*
49	* List of struct kfd_process (field kfd_process).
50	* Unique/indexed by mm_struct*
51	*/
52	DEFINE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
53	static DEFINE_MUTEX(kfd_processes_mutex);
54
55	DEFINE_SRCU(kfd_processes_srcu);
56
57	/* For process termination handling */
58	static struct workqueue_struct *kfd_process_wq;
59
60	/* Ordered, single-threaded workqueue for restoring evicted
61	* processes. Restoring multiple processes concurrently under memory
62	* pressure can lead to processes blocking each other from validating
63	* their BOs and result in a live-lock situation where processes
64	* remain evicted indefinitely.
65	*/
66	static struct workqueue_struct *kfd_restore_wq;
67
68	static struct kfd_process *find_process(const struct task_struct *thread,
69	bool ref);
70	static void kfd_process_ref_release(struct kref *ref);
71	static struct kfd_process *create_process(const struct task_struct *thread);
72	static int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep);
73
74	static void evict_process_worker(struct work_struct *work);
75	static void restore_process_worker(struct work_struct *work);
76
77	static void kfd_process_device_destroy_cwsr_dgpu(struct kfd_process_device *pdd);
78
79	struct kfd_procfs_tree {
80	struct kobject *kobj;
81	};
82
83	static struct kfd_procfs_tree procfs;
84
85	/*
86	* Structure for SDMA activity tracking
87	*/
88	struct kfd_sdma_activity_handler_workarea {
89	struct work_struct sdma_activity_work;
90	struct kfd_process_device *pdd;
91	uint64_t sdma_activity_counter;
92	};
93
94	struct temp_sdma_queue_list {
95	uint64_t __user *rptr;
96	uint64_t sdma_val;
97	unsigned int queue_id;
98	struct list_head list;
99	};
100
101	static void kfd_sdma_activity_worker(struct work_struct *work)
102	{
103	struct kfd_sdma_activity_handler_workarea *workarea;
104	struct kfd_process_device *pdd;
105	uint64_t val;
106	struct mm_struct *mm;
107	struct queue *q;
108	struct qcm_process_device *qpd;
109	struct device_queue_manager *dqm;
110	int ret = 0;
111	struct temp_sdma_queue_list sdma_q_list;
112	struct temp_sdma_queue_list *sdma_q, *next;
113
114	workarea = container_of(work, struct kfd_sdma_activity_handler_workarea,
115	sdma_activity_work);
116
117	pdd = workarea->pdd;
118	if (!pdd)
119	return;
120	dqm = pdd->dev->dqm;
121	qpd = &pdd->qpd;
122	if (!dqm || !qpd)
123	return;
124	/*
125	* Total SDMA activity is current SDMA activity + past SDMA activity
126	* Past SDMA count is stored in pdd.
127	* To get the current activity counters for all active SDMA queues,
128	* we loop over all SDMA queues and get their counts from user-space.
129	*
130	* We cannot call get_user() with dqm_lock held as it can cause
131	* a circular lock dependency situation. To read the SDMA stats,
132	* we need to do the following:
133	*
134	* 1. Create a temporary list of SDMA queue nodes from the qpd->queues_list,
135	* with dqm_lock/dqm_unlock().
136	* 2. Call get_user() for each node in temporary list without dqm_lock.
137	* Save the SDMA count for each node and also add the count to the total
138	* SDMA count counter.
139	* Its possible, during this step, a few SDMA queue nodes got deleted
140	* from the qpd->queues_list.
141	* 3. Do a second pass over qpd->queues_list to check if any nodes got deleted.
142	* If any node got deleted, its SDMA count would be captured in the sdma
143	* past activity counter. So subtract the SDMA counter stored in step 2
144	* for this node from the total SDMA count.
145	*/
146	INIT_LIST_HEAD(list: &sdma_q_list.list);
147
148	/*
149	* Create the temp list of all SDMA queues
150	*/
151	dqm_lock(dqm);
152
153	list_for_each_entry(q, &qpd->queues_list, list) {
154	if ((q->properties.type != KFD_QUEUE_TYPE_SDMA) &&
155	(q->properties.type != KFD_QUEUE_TYPE_SDMA_XGMI))
156	continue;
157
158	sdma_q = kzalloc(size: sizeof(struct temp_sdma_queue_list), GFP_KERNEL);
159	if (!sdma_q) {
160	dqm_unlock(dqm);
161	goto cleanup;
162	}
163
164	INIT_LIST_HEAD(list: &sdma_q->list);
165	sdma_q->rptr = (uint64_t __user *)q->properties.read_ptr;
166	sdma_q->queue_id = q->properties.queue_id;
167	list_add_tail(new: &sdma_q->list, head: &sdma_q_list.list);
168	}
169
170	/*
171	* If the temp list is empty, then no SDMA queues nodes were found in
172	* qpd->queues_list. Return the past activity count as the total sdma
173	* count
174	*/
175	if (list_empty(head: &sdma_q_list.list)) {
176	workarea->sdma_activity_counter = pdd->sdma_past_activity_counter;
177	dqm_unlock(dqm);
178	return;
179	}
180
181	dqm_unlock(dqm);
182
183	/*
184	* Get the usage count for each SDMA queue in temp_list.
185	*/
186	mm = get_task_mm(task: pdd->process->lead_thread);
187	if (!mm)
188	goto cleanup;
189
190	kthread_use_mm(mm);
191
192	list_for_each_entry(sdma_q, &sdma_q_list.list, list) {
193	val = 0;
194	ret = read_sdma_queue_counter(q_rptr: sdma_q->rptr, val: &val);
195	if (ret) {
196	pr_debug("Failed to read SDMA queue active counter for queue id: %d",
197	sdma_q->queue_id);
198	} else {
199	sdma_q->sdma_val = val;
200	workarea->sdma_activity_counter += val;
201	}
202	}
203
204	kthread_unuse_mm(mm);
205	mmput(mm);
206
207	/*
208	* Do a second iteration over qpd_queues_list to check if any SDMA
209	* nodes got deleted while fetching SDMA counter.
210	*/
211	dqm_lock(dqm);
212
213	workarea->sdma_activity_counter += pdd->sdma_past_activity_counter;
214
215	list_for_each_entry(q, &qpd->queues_list, list) {
216	if (list_empty(head: &sdma_q_list.list))
217	break;
218
219	if ((q->properties.type != KFD_QUEUE_TYPE_SDMA) &&
220	(q->properties.type != KFD_QUEUE_TYPE_SDMA_XGMI))
221	continue;
222
223	list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
224	if (((uint64_t __user *)q->properties.read_ptr == sdma_q->rptr) &&
225	(sdma_q->queue_id == q->properties.queue_id)) {
226	list_del(entry: &sdma_q->list);
227	kfree(objp: sdma_q);
228	break;
229	}
230	}
231	}
232
233	dqm_unlock(dqm);
234
235	/*
236	* If temp list is not empty, it implies some queues got deleted
237	* from qpd->queues_list during SDMA usage read. Subtract the SDMA
238	* count for each node from the total SDMA count.
239	*/
240	list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
241	workarea->sdma_activity_counter -= sdma_q->sdma_val;
242	list_del(entry: &sdma_q->list);
243	kfree(objp: sdma_q);
244	}
245
246	return;
247
248	cleanup:
249	list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
250	list_del(entry: &sdma_q->list);
251	kfree(objp: sdma_q);
252	}
253	}
254
255	/**
256	* kfd_get_cu_occupancy - Collect number of waves in-flight on this device
257	* by current process. Translates acquired wave count into number of compute units
258	* that are occupied.
259	*
260	* @attr: Handle of attribute that allows reporting of wave count. The attribute
261	* handle encapsulates GPU device it is associated with, thereby allowing collection
262	* of waves in flight, etc
263	* @buffer: Handle of user provided buffer updated with wave count
264	*
265	* Return: Number of bytes written to user buffer or an error value
266	*/
267	static int kfd_get_cu_occupancy(struct attribute *attr, char *buffer)
268	{
269	int cu_cnt;
270	int wave_cnt;
271	int max_waves_per_cu;
272	struct kfd_dev *dev = NULL;
273	struct kfd_process *proc = NULL;
274	struct kfd_process_device *pdd = NULL;
275
276	pdd = container_of(attr, struct kfd_process_device, attr_cu_occupancy);
277	dev = pdd->dev;
278	if (dev->kfd2kgd->get_cu_occupancy == NULL)
279	return -EINVAL;
280
281	cu_cnt = 0;
282	proc = pdd->process;
283	if (pdd->qpd.queue_count == 0) {
284	pr_debug("Gpu-Id: %d has no active queues for process %d\n",
285	dev->id, proc->pasid);
286	return snprintf(buf: buffer, PAGE_SIZE, fmt: "%d\n", cu_cnt);
287	}
288
289	/* Collect wave count from device if it supports */
290	wave_cnt = 0;
291	max_waves_per_cu = 0;
292	dev->kfd2kgd->get_cu_occupancy(dev->adev, proc->pasid, &wave_cnt,
293	&max_waves_per_cu);
294
295	/* Translate wave count to number of compute units */
296	cu_cnt = (wave_cnt + (max_waves_per_cu - 1)) / max_waves_per_cu;
297	return snprintf(buf: buffer, PAGE_SIZE, fmt: "%d\n", cu_cnt);
298	}
299
300	static ssize_t kfd_procfs_show(struct kobject *kobj, struct attribute *attr,
301	char *buffer)
302	{
303	if (strcmp(attr->name, "pasid") == 0) {
304	struct kfd_process *p = container_of(attr, struct kfd_process,
305	attr_pasid);
306
307	return snprintf(buf: buffer, PAGE_SIZE, fmt: "%d\n", p->pasid);
308	} else if (strncmp(attr->name, "vram_", 5) == 0) {
309	struct kfd_process_device *pdd = container_of(attr, struct kfd_process_device,
310	attr_vram);
311	return snprintf(buf: buffer, PAGE_SIZE, fmt: "%llu\n", READ_ONCE(pdd->vram_usage));
312	} else if (strncmp(attr->name, "sdma_", 5) == 0) {
313	struct kfd_process_device *pdd = container_of(attr, struct kfd_process_device,
314	attr_sdma);
315	struct kfd_sdma_activity_handler_workarea sdma_activity_work_handler;
316
317	INIT_WORK(&sdma_activity_work_handler.sdma_activity_work,
318	kfd_sdma_activity_worker);
319
320	sdma_activity_work_handler.pdd = pdd;
321	sdma_activity_work_handler.sdma_activity_counter = 0;
322
323	schedule_work(work: &sdma_activity_work_handler.sdma_activity_work);
324
325	flush_work(work: &sdma_activity_work_handler.sdma_activity_work);
326
327	return snprintf(buf: buffer, PAGE_SIZE, fmt: "%llu\n",
328	(sdma_activity_work_handler.sdma_activity_counter)/
329	SDMA_ACTIVITY_DIVISOR);
330	} else {
331	pr_err("Invalid attribute");
332	return -EINVAL;
333	}
334
335	return 0;
336	}
337
338	static void kfd_procfs_kobj_release(struct kobject *kobj)
339	{
340	kfree(objp: kobj);
341	}
342
343	static const struct sysfs_ops kfd_procfs_ops = {
344	.show = kfd_procfs_show,
345	};
346
347	static const struct kobj_type procfs_type = {
348	.release = kfd_procfs_kobj_release,
349	.sysfs_ops = &kfd_procfs_ops,
350	};
351
352	void kfd_procfs_init(void)
353	{
354	int ret = 0;
355
356	procfs.kobj = kfd_alloc_struct(procfs.kobj);
357	if (!procfs.kobj)
358	return;
359
360	ret = kobject_init_and_add(kobj: procfs.kobj, ktype: &procfs_type,
361	parent: &kfd_device->kobj, fmt: "proc");
362	if (ret) {
363	pr_warn("Could not create procfs proc folder");
364	/* If we fail to create the procfs, clean up */
365	kfd_procfs_shutdown();
366	}
367	}
368
369	void kfd_procfs_shutdown(void)
370	{
371	if (procfs.kobj) {
372	kobject_del(kobj: procfs.kobj);
373	kobject_put(kobj: procfs.kobj);
374	procfs.kobj = NULL;
375	}
376	}
377
378	static ssize_t kfd_procfs_queue_show(struct kobject *kobj,
379	struct attribute *attr, char *buffer)
380	{
381	struct queue *q = container_of(kobj, struct queue, kobj);
382
383	if (!strcmp(attr->name, "size"))
384	return snprintf(buf: buffer, PAGE_SIZE, fmt: "%llu",
385	q->properties.queue_size);
386	else if (!strcmp(attr->name, "type"))
387	return snprintf(buf: buffer, PAGE_SIZE, fmt: "%d", q->properties.type);
388	else if (!strcmp(attr->name, "gpuid"))
389	return snprintf(buf: buffer, PAGE_SIZE, fmt: "%u", q->device->id);
390	else
391	pr_err("Invalid attribute");
392
393	return 0;
394	}
395
396	static ssize_t kfd_procfs_stats_show(struct kobject *kobj,
397	struct attribute *attr, char *buffer)
398	{
399	if (strcmp(attr->name, "evicted_ms") == 0) {
400	struct kfd_process_device *pdd = container_of(attr,
401	struct kfd_process_device,
402	attr_evict);
403	uint64_t evict_jiffies;
404
405	evict_jiffies = atomic64_read(&pdd->evict_duration_counter);
406
407	return snprintf(buffer,
408	PAGE_SIZE,
409	"%llu\n",
410	jiffies64_to_msecs(evict_jiffies));
411
412	/* Sysfs handle that gets CU occupancy is per device */
413	} else if (strcmp(attr->name, "cu_occupancy") == 0) {
414	return kfd_get_cu_occupancy(attr, buffer);
415	} else {
416	pr_err("Invalid attribute");
417	}
418
419	return 0;
420	}
421
422	static ssize_t kfd_sysfs_counters_show(struct kobject *kobj,
423	struct attribute *attr, char *buf)
424	{
425	struct kfd_process_device *pdd;
426
427	if (!strcmp(attr->name, "faults")) {
428	pdd = container_of(attr, struct kfd_process_device,
429	attr_faults);
430	return sysfs_emit(buf, fmt: "%llu\n", READ_ONCE(pdd->faults));
431	}
432	if (!strcmp(attr->name, "page_in")) {
433	pdd = container_of(attr, struct kfd_process_device,
434	attr_page_in);
435	return sysfs_emit(buf, fmt: "%llu\n", READ_ONCE(pdd->page_in));
436	}
437	if (!strcmp(attr->name, "page_out")) {
438	pdd = container_of(attr, struct kfd_process_device,
439	attr_page_out);
440	return sysfs_emit(buf, fmt: "%llu\n", READ_ONCE(pdd->page_out));
441	}
442	return 0;
443	}
444
445	static struct attribute attr_queue_size = {
446	.name = "size",
447	.mode = KFD_SYSFS_FILE_MODE
448	};
449
450	static struct attribute attr_queue_type = {
451	.name = "type",
452	.mode = KFD_SYSFS_FILE_MODE
453	};
454
455	static struct attribute attr_queue_gpuid = {
456	.name = "gpuid",
457	.mode = KFD_SYSFS_FILE_MODE
458	};
459
460	static struct attribute *procfs_queue_attrs[] = {
461	&attr_queue_size,
462	&attr_queue_type,
463	&attr_queue_gpuid,
464	NULL
465	};
466	ATTRIBUTE_GROUPS(procfs_queue);
467
468	static const struct sysfs_ops procfs_queue_ops = {
469	.show = kfd_procfs_queue_show,
470	};
471
472	static const struct kobj_type procfs_queue_type = {
473	.sysfs_ops = &procfs_queue_ops,
474	.default_groups = procfs_queue_groups,
475	};
476
477	static const struct sysfs_ops procfs_stats_ops = {
478	.show = kfd_procfs_stats_show,
479	};
480
481	static const struct kobj_type procfs_stats_type = {
482	.sysfs_ops = &procfs_stats_ops,
483	.release = kfd_procfs_kobj_release,
484	};
485
486	static const struct sysfs_ops sysfs_counters_ops = {
487	.show = kfd_sysfs_counters_show,
488	};
489
490	static const struct kobj_type sysfs_counters_type = {
491	.sysfs_ops = &sysfs_counters_ops,
492	.release = kfd_procfs_kobj_release,
493	};
494
495	int kfd_procfs_add_queue(struct queue *q)
496	{
497	struct kfd_process *proc;
498	int ret;
499
500	if (!q || !q->process)
501	return -EINVAL;
502	proc = q->process;
503
504	/* Create proc/<pid>/queues/<queue id> folder */
505	if (!proc->kobj_queues)
506	return -EFAULT;
507	ret = kobject_init_and_add(kobj: &q->kobj, ktype: &procfs_queue_type,
508	parent: proc->kobj_queues, fmt: "%u", q->properties.queue_id);
509	if (ret < 0) {
510	pr_warn("Creating proc/<pid>/queues/%u failed",
511	q->properties.queue_id);
512	kobject_put(kobj: &q->kobj);
513	return ret;
514	}
515
516	return 0;
517	}
518
519	static void kfd_sysfs_create_file(struct kobject *kobj, struct attribute *attr,
520	char *name)
521	{
522	int ret;
523
524	if (!kobj || !attr || !name)
525	return;
526
527	attr->name = name;
528	attr->mode = KFD_SYSFS_FILE_MODE;
529	sysfs_attr_init(attr);
530
531	ret = sysfs_create_file(kobj, attr);
532	if (ret)
533	pr_warn("Create sysfs %s/%s failed %d", kobj->name, name, ret);
534	}
535
536	static void kfd_procfs_add_sysfs_stats(struct kfd_process *p)
537	{
538	int ret;
539	int i;
540	char stats_dir_filename[MAX_SYSFS_FILENAME_LEN];
541
542	if (!p || !p->kobj)
543	return;
544
545	/*
546	* Create sysfs files for each GPU:
547	* - proc/<pid>/stats_<gpuid>/
548	* - proc/<pid>/stats_<gpuid>/evicted_ms
549	* - proc/<pid>/stats_<gpuid>/cu_occupancy
550	*/
551	for (i = 0; i < p->n_pdds; i++) {
552	struct kfd_process_device *pdd = p->pdds[i];
553
554	snprintf(buf: stats_dir_filename, MAX_SYSFS_FILENAME_LEN,
555	fmt: "stats_%u", pdd->dev->id);
556	pdd->kobj_stats = kfd_alloc_struct(pdd->kobj_stats);
557	if (!pdd->kobj_stats)
558	return;
559
560	ret = kobject_init_and_add(kobj: pdd->kobj_stats,
561	ktype: &procfs_stats_type,
562	parent: p->kobj,
563	fmt: stats_dir_filename);
564
565	if (ret) {
566	pr_warn("Creating KFD proc/stats_%s folder failed",
567	stats_dir_filename);
568	kobject_put(kobj: pdd->kobj_stats);
569	pdd->kobj_stats = NULL;
570	return;
571	}
572
573	kfd_sysfs_create_file(kobj: pdd->kobj_stats, attr: &pdd->attr_evict,
574	name: "evicted_ms");
575	/* Add sysfs file to report compute unit occupancy */
576	if (pdd->dev->kfd2kgd->get_cu_occupancy)
577	kfd_sysfs_create_file(kobj: pdd->kobj_stats,
578	attr: &pdd->attr_cu_occupancy,
579	name: "cu_occupancy");
580	}
581	}
582
583	static void kfd_procfs_add_sysfs_counters(struct kfd_process *p)
584	{
585	int ret = 0;
586	int i;
587	char counters_dir_filename[MAX_SYSFS_FILENAME_LEN];
588
589	if (!p || !p->kobj)
590	return;
591
592	/*
593	* Create sysfs files for each GPU which supports SVM
594	* - proc/<pid>/counters_<gpuid>/
595	* - proc/<pid>/counters_<gpuid>/faults
596	* - proc/<pid>/counters_<gpuid>/page_in
597	* - proc/<pid>/counters_<gpuid>/page_out
598	*/
599	for_each_set_bit(i, p->svms.bitmap_supported, p->n_pdds) {
600	struct kfd_process_device *pdd = p->pdds[i];
601	struct kobject *kobj_counters;
602
603	snprintf(buf: counters_dir_filename, MAX_SYSFS_FILENAME_LEN,
604	fmt: "counters_%u", pdd->dev->id);
605	kobj_counters = kfd_alloc_struct(kobj_counters);
606	if (!kobj_counters)
607	return;
608
609	ret = kobject_init_and_add(kobj: kobj_counters, ktype: &sysfs_counters_type,
610	parent: p->kobj, fmt: counters_dir_filename);
611	if (ret) {
612	pr_warn("Creating KFD proc/%s folder failed",
613	counters_dir_filename);
614	kobject_put(kobj: kobj_counters);
615	return;
616	}
617
618	pdd->kobj_counters = kobj_counters;
619	kfd_sysfs_create_file(kobj: kobj_counters, attr: &pdd->attr_faults,
620	name: "faults");
621	kfd_sysfs_create_file(kobj: kobj_counters, attr: &pdd->attr_page_in,
622	name: "page_in");
623	kfd_sysfs_create_file(kobj: kobj_counters, attr: &pdd->attr_page_out,
624	name: "page_out");
625	}
626	}
627
628	static void kfd_procfs_add_sysfs_files(struct kfd_process *p)
629	{
630	int i;
631
632	if (!p || !p->kobj)
633	return;
634
635	/*
636	* Create sysfs files for each GPU:
637	* - proc/<pid>/vram_<gpuid>
638	* - proc/<pid>/sdma_<gpuid>
639	*/
640	for (i = 0; i < p->n_pdds; i++) {
641	struct kfd_process_device *pdd = p->pdds[i];
642
643	snprintf(buf: pdd->vram_filename, MAX_SYSFS_FILENAME_LEN, fmt: "vram_%u",
644	pdd->dev->id);
645	kfd_sysfs_create_file(kobj: p->kobj, attr: &pdd->attr_vram,
646	name: pdd->vram_filename);
647
648	snprintf(buf: pdd->sdma_filename, MAX_SYSFS_FILENAME_LEN, fmt: "sdma_%u",
649	pdd->dev->id);
650	kfd_sysfs_create_file(kobj: p->kobj, attr: &pdd->attr_sdma,
651	name: pdd->sdma_filename);
652	}
653	}
654
655	void kfd_procfs_del_queue(struct queue *q)
656	{
657	if (!q)
658	return;
659
660	kobject_del(kobj: &q->kobj);
661	kobject_put(kobj: &q->kobj);
662	}
663
664	int kfd_process_create_wq(void)
665	{
666	if (!kfd_process_wq)
667	kfd_process_wq = alloc_workqueue(fmt: "kfd_process_wq", flags: 0, max_active: 0);
668	if (!kfd_restore_wq)
669	kfd_restore_wq = alloc_ordered_workqueue("kfd_restore_wq", 0);
670
671	if (!kfd_process_wq || !kfd_restore_wq) {
672	kfd_process_destroy_wq();
673	return -ENOMEM;
674	}
675
676	return 0;
677	}
678
679	void kfd_process_destroy_wq(void)
680	{
681	if (kfd_process_wq) {
682	destroy_workqueue(wq: kfd_process_wq);
683	kfd_process_wq = NULL;
684	}
685	if (kfd_restore_wq) {
686	destroy_workqueue(wq: kfd_restore_wq);
687	kfd_restore_wq = NULL;
688	}
689	}
690
691	static void kfd_process_free_gpuvm(struct kgd_mem *mem,
692	struct kfd_process_device *pdd, void **kptr)
693	{
694	struct kfd_dev *dev = pdd->dev;
695
696	if (kptr && *kptr) {
697	amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(mem);
698	*kptr = NULL;
699	}
700
701	amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(dev->adev, mem, pdd->drm_priv);
702	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->adev, mem, pdd->drm_priv,
703	NULL);
704	}
705
706	/* kfd_process_alloc_gpuvm - Allocate GPU VM for the KFD process
707	* This function should be only called right after the process
708	* is created and when kfd_processes_mutex is still being held
709	* to avoid concurrency. Because of that exclusiveness, we do
710	* not need to take p->mutex.
711	*/
712	static int kfd_process_alloc_gpuvm(struct kfd_process_device *pdd,
713	uint64_t gpu_va, uint32_t size,
714	uint32_t flags, struct kgd_mem **mem, void **kptr)
715	{
716	struct kfd_dev *kdev = pdd->dev;
717	int err;
718
719	err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(kdev->adev, gpu_va, size,
720	pdd->drm_priv, mem, NULL,
721	flags, false);
722	if (err)
723	goto err_alloc_mem;
724
725	err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(kdev->adev, *mem,
726	pdd->drm_priv);
727	if (err)
728	goto err_map_mem;
729
730	err = amdgpu_amdkfd_gpuvm_sync_memory(kdev->adev, *mem, true);
731	if (err) {
732	pr_debug("Sync memory failed, wait interrupted by user signal\n");
733	goto sync_memory_failed;
734	}
735
736	if (kptr) {
737	err = amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(
738	(struct kgd_mem *)*mem, kptr, NULL);
739	if (err) {
740	pr_debug("Map GTT BO to kernel failed\n");
741	goto sync_memory_failed;
742	}
743	}
744
745	return err;
746
747	sync_memory_failed:
748	amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(kdev->adev, *mem, pdd->drm_priv);
749
750	err_map_mem:
751	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(kdev->adev, *mem, pdd->drm_priv,
752	NULL);
753	err_alloc_mem:
754	*mem = NULL;
755	*kptr = NULL;
756	return err;
757	}
758
759	/* kfd_process_device_reserve_ib_mem - Reserve memory inside the
760	* process for IB usage The memory reserved is for KFD to submit
761	* IB to AMDGPU from kernel. If the memory is reserved
762	* successfully, ib_kaddr will have the CPU/kernel
763	* address. Check ib_kaddr before accessing the memory.
764	*/
765	static int kfd_process_device_reserve_ib_mem(struct kfd_process_device *pdd)
766	{
767	struct qcm_process_device *qpd = &pdd->qpd;
768	uint32_t flags = KFD_IOC_ALLOC_MEM_FLAGS_GTT |
769	KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE |
770	KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE |
771	KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
772	struct kgd_mem *mem;
773	void *kaddr;
774	int ret;
775
776	if (qpd->ib_kaddr || !qpd->ib_base)
777	return 0;
778
779	/* ib_base is only set for dGPU */
780	ret = kfd_process_alloc_gpuvm(pdd, gpu_va: qpd->ib_base, PAGE_SIZE, flags,
781	mem: &mem, kptr: &kaddr);
782	if (ret)
783	return ret;
784
785	qpd->ib_mem = mem;
786	qpd->ib_kaddr = kaddr;
787
788	return 0;
789	}
790
791	static void kfd_process_device_destroy_ib_mem(struct kfd_process_device *pdd)
792	{
793	struct qcm_process_device *qpd = &pdd->qpd;
794
795	if (!qpd->ib_kaddr || !qpd->ib_base)
796	return;
797
798	kfd_process_free_gpuvm(mem: qpd->ib_mem, pdd, kptr: &qpd->ib_kaddr);
799	}
800
801	struct kfd_process *kfd_create_process(struct file *filep)
802	{
803	struct kfd_process *process;
804	struct task_struct *thread = current;
805	int ret;
806
807	if (!thread->mm)
808	return ERR_PTR(-EINVAL);
809
810	/* Only the pthreads threading model is supported. */
811	if (thread->group_leader->mm != thread->mm)
812	return ERR_PTR(-EINVAL);
813
814	/*
815	* take kfd processes mutex before starting of process creation
816	* so there won't be a case where two threads of the same process
817	* create two kfd_process structures
818	*/
819	mutex_lock(lock: &kfd_processes_mutex);
820
821	/* A prior open of /dev/kfd could have already created the process. */
822	process = find_process(thread, ref: false);
823	if (process) {
824	pr_debug("Process already found\n");
825	} else {
826	process = create_process(thread);
827	if (IS_ERR(ptr: process))
828	goto out;
829
830	ret = kfd_process_init_cwsr_apu(p: process, filep);
831	if (ret)
832	goto out_destroy;
833
834	if (!procfs.kobj)
835	goto out;
836
837	process->kobj = kfd_alloc_struct(process->kobj);
838	if (!process->kobj) {
839	pr_warn("Creating procfs kobject failed");
840	goto out;
841	}
842	ret = kobject_init_and_add(kobj: process->kobj, ktype: &procfs_type,
843	parent: procfs.kobj, fmt: "%d",
844	(int)process->lead_thread->pid);
845	if (ret) {
846	pr_warn("Creating procfs pid directory failed");
847	kobject_put(kobj: process->kobj);
848	goto out;
849	}
850
851	kfd_sysfs_create_file(kobj: process->kobj, attr: &process->attr_pasid,
852	name: "pasid");
853
854	process->kobj_queues = kobject_create_and_add(name: "queues",
855	parent: process->kobj);
856	if (!process->kobj_queues)
857	pr_warn("Creating KFD proc/queues folder failed");
858
859	kfd_procfs_add_sysfs_stats(p: process);
860	kfd_procfs_add_sysfs_files(p: process);
861	kfd_procfs_add_sysfs_counters(p: process);
862	}
863	out:
864	if (!IS_ERR(ptr: process))
865	kref_get(kref: &process->ref);
866	mutex_unlock(lock: &kfd_processes_mutex);
867
868	return process;
869
870	out_destroy:
871	hash_del_rcu(node: &process->kfd_processes);
872	mutex_unlock(lock: &kfd_processes_mutex);
873	synchronize_srcu(ssp: &kfd_processes_srcu);
874	/* kfd_process_free_notifier will trigger the cleanup */
875	mmu_notifier_put(&process->mmu_notifier);
876	return ERR_PTR(error: ret);
877	}
878
879	struct kfd_process *kfd_get_process(const struct task_struct *thread)
880	{
881	struct kfd_process *process;
882
883	if (!thread->mm)
884	return ERR_PTR(-EINVAL);
885
886	/* Only the pthreads threading model is supported. */
887	if (thread->group_leader->mm != thread->mm)
888	return ERR_PTR(-EINVAL);
889
890	process = find_process(thread, ref: false);
891	if (!process)
892	return ERR_PTR(-EINVAL);
893
894	return process;
895	}
896
897	static struct kfd_process *find_process_by_mm(const struct mm_struct *mm)
898	{
899	struct kfd_process *process;
900
901	hash_for_each_possible_rcu(kfd_processes_table, process,
902	kfd_processes, (uintptr_t)mm)
903	if (process->mm == mm)
904	return process;
905
906	return NULL;
907	}
908
909	static struct kfd_process *find_process(const struct task_struct *thread,
910	bool ref)
911	{
912	struct kfd_process *p;
913	int idx;
914
915	idx = srcu_read_lock(ssp: &kfd_processes_srcu);
916	p = find_process_by_mm(mm: thread->mm);
917	if (p && ref)
918	kref_get(kref: &p->ref);
919	srcu_read_unlock(ssp: &kfd_processes_srcu, idx);
920
921	return p;
922	}
923
924	void kfd_unref_process(struct kfd_process *p)
925	{
926	kref_put(kref: &p->ref, release: kfd_process_ref_release);
927	}
928
929	/* This increments the process->ref counter. */
930	struct kfd_process *kfd_lookup_process_by_pid(struct pid *pid)
931	{
932	struct task_struct *task = NULL;
933	struct kfd_process *p = NULL;
934
935	if (!pid) {
936	task = current;
937	get_task_struct(t: task);
938	} else {
939	task = get_pid_task(pid, PIDTYPE_PID);
940	}
941
942	if (task) {
943	p = find_process(thread: task, ref: true);
944	put_task_struct(t: task);
945	}
946
947	return p;
948	}
949
950	static void kfd_process_device_free_bos(struct kfd_process_device *pdd)
951	{
952	struct kfd_process *p = pdd->process;
953	void *mem;
954	int id;
955	int i;
956
957	/*
958	* Remove all handles from idr and release appropriate
959	* local memory object
960	*/
961	idr_for_each_entry(&pdd->alloc_idr, mem, id) {
962
963	for (i = 0; i < p->n_pdds; i++) {
964	struct kfd_process_device *peer_pdd = p->pdds[i];
965
966	if (!peer_pdd->drm_priv)
967	continue;
968	amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
969	peer_pdd->dev->adev, mem, peer_pdd->drm_priv);
970	}
971
972	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, mem,
973	pdd->drm_priv, NULL);
974	kfd_process_device_remove_obj_handle(pdd, handle: id);
975	}
976	}
977
978	/*
979	* Just kunmap and unpin signal BO here. It will be freed in
980	* kfd_process_free_outstanding_kfd_bos()
981	*/
982	static void kfd_process_kunmap_signal_bo(struct kfd_process *p)
983	{
984	struct kfd_process_device *pdd;
985	struct kfd_dev *kdev;
986	void *mem;
987
988	kdev = kfd_device_by_id(GET_GPU_ID(p->signal_handle));
989	if (!kdev)
990	return;
991
992	mutex_lock(lock: &p->mutex);
993
994	pdd = kfd_get_process_device_data(dev: kdev, p);
995	if (!pdd)
996	goto out;
997
998	mem = kfd_process_device_translate_handle(
999	p: pdd, GET_IDR_HANDLE(p->signal_handle));
1000	if (!mem)
1001	goto out;
1002
1003	amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(mem);
1004
1005	out:
1006	mutex_unlock(lock: &p->mutex);
1007	}
1008
1009	static void kfd_process_free_outstanding_kfd_bos(struct kfd_process *p)
1010	{
1011	int i;
1012
1013	for (i = 0; i < p->n_pdds; i++)
1014	kfd_process_device_free_bos(pdd: p->pdds[i]);
1015	}
1016
1017	static void kfd_process_destroy_pdds(struct kfd_process *p)
1018	{
1019	int i;
1020
1021	for (i = 0; i < p->n_pdds; i++) {
1022	struct kfd_process_device *pdd = p->pdds[i];
1023
1024	pr_debug("Releasing pdd (topology id %d) for process (pasid 0x%x)\n",
1025	pdd->dev->id, p->pasid);
1026
1027	kfd_process_device_destroy_cwsr_dgpu(pdd);
1028	kfd_process_device_destroy_ib_mem(pdd);
1029
1030	if (pdd->drm_file) {
1031	amdgpu_amdkfd_gpuvm_release_process_vm(
1032	pdd->dev->adev, pdd->drm_priv);
1033	fput(pdd->drm_file);
1034	}
1035
1036	if (pdd->qpd.cwsr_kaddr && !pdd->qpd.cwsr_base)
1037	free_pages(addr: (unsigned long)pdd->qpd.cwsr_kaddr,
1038	order: get_order(KFD_CWSR_TBA_TMA_SIZE));
1039
1040	bitmap_free(bitmap: pdd->qpd.doorbell_bitmap);
1041	idr_destroy(&pdd->alloc_idr);
1042
1043	kfd_free_process_doorbells(kfd: pdd->dev, doorbell_index: pdd->doorbell_index);
1044
1045	if (pdd->dev->shared_resources.enable_mes)
1046	amdgpu_amdkfd_free_gtt_mem(pdd->dev->adev,
1047	pdd->proc_ctx_bo);
1048	/*
1049	* before destroying pdd, make sure to report availability
1050	* for auto suspend
1051	*/
1052	if (pdd->runtime_inuse) {
1053	pm_runtime_mark_last_busy(dev: adev_to_drm(pdd->dev->adev)->dev);
1054	pm_runtime_put_autosuspend(dev: adev_to_drm(pdd->dev->adev)->dev);
1055	pdd->runtime_inuse = false;
1056	}
1057
1058	kfree(objp: pdd);
1059	p->pdds[i] = NULL;
1060	}
1061	p->n_pdds = 0;
1062	}
1063
1064	static void kfd_process_remove_sysfs(struct kfd_process *p)
1065	{
1066	struct kfd_process_device *pdd;
1067	int i;
1068
1069	if (!p->kobj)
1070	return;
1071
1072	sysfs_remove_file(kobj: p->kobj, attr: &p->attr_pasid);
1073	kobject_del(kobj: p->kobj_queues);
1074	kobject_put(kobj: p->kobj_queues);
1075	p->kobj_queues = NULL;
1076
1077	for (i = 0; i < p->n_pdds; i++) {
1078	pdd = p->pdds[i];
1079
1080	sysfs_remove_file(kobj: p->kobj, attr: &pdd->attr_vram);
1081	sysfs_remove_file(kobj: p->kobj, attr: &pdd->attr_sdma);
1082
1083	sysfs_remove_file(kobj: pdd->kobj_stats, attr: &pdd->attr_evict);
1084	if (pdd->dev->kfd2kgd->get_cu_occupancy)
1085	sysfs_remove_file(kobj: pdd->kobj_stats,
1086	attr: &pdd->attr_cu_occupancy);
1087	kobject_del(kobj: pdd->kobj_stats);
1088	kobject_put(kobj: pdd->kobj_stats);
1089	pdd->kobj_stats = NULL;
1090	}
1091
1092	for_each_set_bit(i, p->svms.bitmap_supported, p->n_pdds) {
1093	pdd = p->pdds[i];
1094
1095	sysfs_remove_file(kobj: pdd->kobj_counters, attr: &pdd->attr_faults);
1096	sysfs_remove_file(kobj: pdd->kobj_counters, attr: &pdd->attr_page_in);
1097	sysfs_remove_file(kobj: pdd->kobj_counters, attr: &pdd->attr_page_out);
1098	kobject_del(kobj: pdd->kobj_counters);
1099	kobject_put(kobj: pdd->kobj_counters);
1100	pdd->kobj_counters = NULL;
1101	}
1102
1103	kobject_del(kobj: p->kobj);
1104	kobject_put(kobj: p->kobj);
1105	p->kobj = NULL;
1106	}
1107
1108	/* No process locking is needed in this function, because the process
1109	* is not findable any more. We must assume that no other thread is
1110	* using it any more, otherwise we couldn't safely free the process
1111	* structure in the end.
1112	*/
1113	static void kfd_process_wq_release(struct work_struct *work)
1114	{
1115	struct kfd_process *p = container_of(work, struct kfd_process,
1116	release_work);
1117
1118	kfd_process_dequeue_from_all_devices(p);
1119	pqm_uninit(pqm: &p->pqm);
1120
1121	/* Signal the eviction fence after user mode queues are
1122	* destroyed. This allows any BOs to be freed without
1123	* triggering pointless evictions or waiting for fences.
1124	*/
1125	dma_fence_signal(p->ef);
1126
1127	kfd_process_remove_sysfs(p);
1128	kfd_iommu_unbind_process(p);
1129
1130	kfd_process_kunmap_signal_bo(p);
1131	kfd_process_free_outstanding_kfd_bos(p);
1132	svm_range_list_fini(p);
1133
1134	kfd_process_destroy_pdds(p);
1135	dma_fence_put(p->ef);
1136
1137	kfd_event_free_process(p);
1138
1139	kfd_pasid_free(pasid: p->pasid);
1140	mutex_destroy(lock: &p->mutex);
1141
1142	put_task_struct(t: p->lead_thread);
1143
1144	kfree(objp: p);
1145	}
1146
1147	static void kfd_process_ref_release(struct kref *ref)
1148	{
1149	struct kfd_process *p = container_of(ref, struct kfd_process, ref);
1150
1151	INIT_WORK(&p->release_work, kfd_process_wq_release);
1152	queue_work(wq: kfd_process_wq, work: &p->release_work);
1153	}
1154
1155	static struct mmu_notifier *kfd_process_alloc_notifier(struct mm_struct *mm)
1156	{
1157	int idx = srcu_read_lock(ssp: &kfd_processes_srcu);
1158	struct kfd_process *p = find_process_by_mm(mm);
1159
1160	srcu_read_unlock(ssp: &kfd_processes_srcu, idx);
1161
1162	return p ? &p->mmu_notifier : ERR_PTR(-ESRCH);
1163	}
1164
1165	static void kfd_process_free_notifier(struct mmu_notifier *mn)
1166	{
1167	kfd_unref_process(container_of(mn, struct kfd_process, mmu_notifier));
1168	}
1169
1170	static void kfd_process_notifier_release_internal(struct kfd_process *p)
1171	{
1172	cancel_delayed_work_sync(dwork: &p->eviction_work);
1173	cancel_delayed_work_sync(dwork: &p->restore_work);
1174
1175	/* Indicate to other users that MM is no longer valid */
1176	p->mm = NULL;
1177
1178	mmu_notifier_put(&p->mmu_notifier);
1179	}
1180
1181	static void kfd_process_notifier_release(struct mmu_notifier *mn,
1182	struct mm_struct *mm)
1183	{
1184	struct kfd_process *p;
1185
1186	/*
1187	* The kfd_process structure can not be free because the
1188	* mmu_notifier srcu is read locked
1189	*/
1190	p = container_of(mn, struct kfd_process, mmu_notifier);
1191	if (WARN_ON(p->mm != mm))
1192	return;
1193
1194	mutex_lock(lock: &kfd_processes_mutex);
1195	/*
1196	* Do early return if table is empty.
1197	*
1198	* This could potentially happen if this function is called concurrently
1199	* by mmu_notifier and by kfd_cleanup_pocesses.
1200	*
1201	*/
1202	if (hash_empty(kfd_processes_table)) {
1203	mutex_unlock(lock: &kfd_processes_mutex);
1204	return;
1205	}
1206	hash_del_rcu(node: &p->kfd_processes);
1207	mutex_unlock(lock: &kfd_processes_mutex);
1208	synchronize_srcu(ssp: &kfd_processes_srcu);
1209
1210	kfd_process_notifier_release_internal(p);
1211	}
1212
1213	static const struct mmu_notifier_ops kfd_process_mmu_notifier_ops = {
1214	.release = kfd_process_notifier_release,
1215	.alloc_notifier = kfd_process_alloc_notifier,
1216	.free_notifier = kfd_process_free_notifier,
1217	};
1218
1219	/*
1220	* This code handles the case when driver is being unloaded before all
1221	* mm_struct are released. We need to safely free the kfd_process and
1222	* avoid race conditions with mmu_notifier that might try to free them.
1223	*
1224	*/
1225	void kfd_cleanup_processes(void)
1226	{
1227	struct kfd_process *p;
1228	struct hlist_node *p_temp;
1229	unsigned int temp;
1230	HLIST_HEAD(cleanup_list);
1231
1232	/*
1233	* Move all remaining kfd_process from the process table to a
1234	* temp list for processing. Once done, callback from mmu_notifier
1235	* release will not see the kfd_process in the table and do early return,
1236	* avoiding double free issues.
1237	*/
1238	mutex_lock(lock: &kfd_processes_mutex);
1239	hash_for_each_safe(kfd_processes_table, temp, p_temp, p, kfd_processes) {
1240	hash_del_rcu(node: &p->kfd_processes);
1241	synchronize_srcu(ssp: &kfd_processes_srcu);
1242	hlist_add_head(n: &p->kfd_processes, h: &cleanup_list);
1243	}
1244	mutex_unlock(lock: &kfd_processes_mutex);
1245
1246	hlist_for_each_entry_safe(p, p_temp, &cleanup_list, kfd_processes)
1247	kfd_process_notifier_release_internal(p);
1248
1249	/*
1250	* Ensures that all outstanding free_notifier get called, triggering
1251	* the release of the kfd_process struct.
1252	*/
1253	mmu_notifier_synchronize();
1254	}
1255
1256	static int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep)
1257	{
1258	unsigned long offset;
1259	int i;
1260
1261	for (i = 0; i < p->n_pdds; i++) {
1262	struct kfd_dev *dev = p->pdds[i]->dev;
1263	struct qcm_process_device *qpd = &p->pdds[i]->qpd;
1264
1265	if (!dev->cwsr_enabled || qpd->cwsr_kaddr || qpd->cwsr_base)
1266	continue;
1267
1268	offset = KFD_MMAP_TYPE_RESERVED_MEM | KFD_MMAP_GPU_ID(dev->id);
1269	qpd->tba_addr = (int64_t)vm_mmap(filep, 0,
1270	KFD_CWSR_TBA_TMA_SIZE, PROT_READ | PROT_EXEC,
1271	MAP_SHARED, offset);
1272
1273	if (IS_ERR_VALUE(qpd->tba_addr)) {
1274	int err = qpd->tba_addr;
1275
1276	pr_err("Failure to set tba address. error %d.\n", err);
1277	qpd->tba_addr = 0;
1278	qpd->cwsr_kaddr = NULL;
1279	return err;
1280	}
1281
1282	memcpy(to: qpd->cwsr_kaddr, from: dev->cwsr_isa, len: dev->cwsr_isa_size);
1283
1284	qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET;
1285	pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n",
1286	qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr);
1287	}
1288
1289	return 0;
1290	}
1291
1292	static int kfd_process_device_init_cwsr_dgpu(struct kfd_process_device *pdd)
1293	{
1294	struct kfd_dev *dev = pdd->dev;
1295	struct qcm_process_device *qpd = &pdd->qpd;
1296	uint32_t flags = KFD_IOC_ALLOC_MEM_FLAGS_GTT
1297	| KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE
1298	| KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
1299	struct kgd_mem *mem;
1300	void *kaddr;
1301	int ret;
1302
1303	if (!dev->cwsr_enabled || qpd->cwsr_kaddr || !qpd->cwsr_base)
1304	return 0;
1305
1306	/* cwsr_base is only set for dGPU */
1307	ret = kfd_process_alloc_gpuvm(pdd, gpu_va: qpd->cwsr_base,
1308	KFD_CWSR_TBA_TMA_SIZE, flags, mem: &mem, kptr: &kaddr);
1309	if (ret)
1310	return ret;
1311
1312	qpd->cwsr_mem = mem;
1313	qpd->cwsr_kaddr = kaddr;
1314	qpd->tba_addr = qpd->cwsr_base;
1315
1316	memcpy(to: qpd->cwsr_kaddr, from: dev->cwsr_isa, len: dev->cwsr_isa_size);
1317
1318	qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET;
1319	pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n",
1320	qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr);
1321
1322	return 0;
1323	}
1324
1325	static void kfd_process_device_destroy_cwsr_dgpu(struct kfd_process_device *pdd)
1326	{
1327	struct kfd_dev *dev = pdd->dev;
1328	struct qcm_process_device *qpd = &pdd->qpd;
1329
1330	if (!dev->cwsr_enabled || !qpd->cwsr_kaddr || !qpd->cwsr_base)
1331	return;
1332
1333	kfd_process_free_gpuvm(mem: qpd->cwsr_mem, pdd, kptr: &qpd->cwsr_kaddr);
1334	}
1335
1336	void kfd_process_set_trap_handler(struct qcm_process_device *qpd,
1337	uint64_t tba_addr,
1338	uint64_t tma_addr)
1339	{
1340	if (qpd->cwsr_kaddr) {
1341	/* KFD trap handler is bound, record as second-level TBA/TMA
1342	* in first-level TMA. First-level trap will jump to second.
1343	*/
1344	uint64_t *tma =
1345	(uint64_t *)(qpd->cwsr_kaddr + KFD_CWSR_TMA_OFFSET);
1346	tma[0] = tba_addr;
1347	tma[1] = tma_addr;
1348	} else {
1349	/* No trap handler bound, bind as first-level TBA/TMA. */
1350	qpd->tba_addr = tba_addr;
1351	qpd->tma_addr = tma_addr;
1352	}
1353	}
1354
1355	bool kfd_process_xnack_mode(struct kfd_process *p, bool supported)
1356	{
1357	int i;
1358
1359	/* On most GFXv9 GPUs, the retry mode in the SQ must match the
1360	* boot time retry setting. Mixing processes with different
1361	* XNACK/retry settings can hang the GPU.
1362	*
1363	* Different GPUs can have different noretry settings depending
1364	* on HW bugs or limitations. We need to find at least one
1365	* XNACK mode for this process that's compatible with all GPUs.
1366	* Fortunately GPUs with retry enabled (noretry=0) can run code
1367	* built for XNACK-off. On GFXv9 it may perform slower.
1368	*
1369	* Therefore applications built for XNACK-off can always be
1370	* supported and will be our fallback if any GPU does not
1371	* support retry.
1372	*/
1373	for (i = 0; i < p->n_pdds; i++) {
1374	struct kfd_dev *dev = p->pdds[i]->dev;
1375
1376	/* Only consider GFXv9 and higher GPUs. Older GPUs don't
1377	* support the SVM APIs and don't need to be considered
1378	* for the XNACK mode selection.
1379	*/
1380	if (!KFD_IS_SOC15(dev))
1381	continue;
1382	/* Aldebaran can always support XNACK because it can support
1383	* per-process XNACK mode selection. But let the dev->noretry
1384	* setting still influence the default XNACK mode.
1385	*/
1386	if (supported && KFD_SUPPORT_XNACK_PER_PROCESS(dev))
1387	continue;
1388
1389	/* GFXv10 and later GPUs do not support shader preemption
1390	* during page faults. This can lead to poor QoS for queue
1391	* management and memory-manager-related preemptions or
1392	* even deadlocks.
1393	*/
1394	if (KFD_GC_VERSION(dev) >= IP_VERSION(10, 1, 1))
1395	return false;
1396
1397	if (dev->noretry)
1398	return false;
1399	}
1400
1401	return true;
1402	}
1403
1404	/*
1405	* On return the kfd_process is fully operational and will be freed when the
1406	* mm is released
1407	*/
1408	static struct kfd_process *create_process(const struct task_struct *thread)
1409	{
1410	struct kfd_process *process;
1411	struct mmu_notifier *mn;
1412	int err = -ENOMEM;
1413
1414	process = kzalloc(size: sizeof(*process), GFP_KERNEL);
1415	if (!process)
1416	goto err_alloc_process;
1417
1418	kref_init(kref: &process->ref);
1419	mutex_init(&process->mutex);
1420	process->mm = thread->mm;
1421	process->lead_thread = thread->group_leader;
1422	process->n_pdds = 0;
1423	process->queues_paused = false;
1424	INIT_DELAYED_WORK(&process->eviction_work, evict_process_worker);
1425	INIT_DELAYED_WORK(&process->restore_work, restore_process_worker);
1426	process->last_restore_timestamp = get_jiffies_64();
1427	err = kfd_event_init_process(p: process);
1428	if (err)
1429	goto err_event_init;
1430	process->is_32bit_user_mode = in_compat_syscall();
1431
1432	process->pasid = kfd_pasid_alloc();
1433	if (process->pasid == 0) {
1434	err = -ENOSPC;
1435	goto err_alloc_pasid;
1436	}
1437
1438	err = pqm_init(pqm: &process->pqm, p: process);
1439	if (err != 0)
1440	goto err_process_pqm_init;
1441
1442	/* init process apertures*/
1443	err = kfd_init_apertures(process);
1444	if (err != 0)
1445	goto err_init_apertures;
1446
1447	/* Check XNACK support after PDDs are created in kfd_init_apertures */
1448	process->xnack_enabled = kfd_process_xnack_mode(p: process, supported: false);
1449
1450	err = svm_range_list_init(p: process);
1451	if (err)
1452	goto err_init_svm_range_list;
1453
1454	/* alloc_notifier needs to find the process in the hash table */
1455	hash_add_rcu(kfd_processes_table, &process->kfd_processes,
1456	(uintptr_t)process->mm);
1457
1458	/* Avoid free_notifier to start kfd_process_wq_release if
1459	* mmu_notifier_get failed because of pending signal.
1460	*/
1461	kref_get(kref: &process->ref);
1462
1463	/* MMU notifier registration must be the last call that can fail
1464	* because after this point we cannot unwind the process creation.
1465	* After this point, mmu_notifier_put will trigger the cleanup by
1466	* dropping the last process reference in the free_notifier.
1467	*/
1468	mn = mmu_notifier_get(&kfd_process_mmu_notifier_ops, process->mm);
1469	if (IS_ERR(ptr: mn)) {
1470	err = PTR_ERR(ptr: mn);
1471	goto err_register_notifier;
1472	}
1473	BUG_ON(mn != &process->mmu_notifier);
1474
1475	kfd_unref_process(p: process);
1476	get_task_struct(t: process->lead_thread);
1477
1478	return process;
1479
1480	err_register_notifier:
1481	hash_del_rcu(node: &process->kfd_processes);
1482	svm_range_list_fini(p: process);
1483	err_init_svm_range_list:
1484	kfd_process_free_outstanding_kfd_bos(p: process);
1485	kfd_process_destroy_pdds(p: process);
1486	err_init_apertures:
1487	pqm_uninit(pqm: &process->pqm);
1488	err_process_pqm_init:
1489	kfd_pasid_free(pasid: process->pasid);
1490	err_alloc_pasid:
1491	kfd_event_free_process(p: process);
1492	err_event_init:
1493	mutex_destroy(lock: &process->mutex);
1494	kfree(objp: process);
1495	err_alloc_process:
1496	return ERR_PTR(error: err);
1497	}
1498
1499	static int init_doorbell_bitmap(struct qcm_process_device *qpd,
1500	struct kfd_dev *dev)
1501	{
1502	unsigned int i;
1503	int range_start = dev->shared_resources.non_cp_doorbells_start;
1504	int range_end = dev->shared_resources.non_cp_doorbells_end;
1505
1506	if (!KFD_IS_SOC15(dev))
1507	return 0;
1508
1509	qpd->doorbell_bitmap = bitmap_zalloc(KFD_MAX_NUM_OF_QUEUES_PER_PROCESS,
1510	GFP_KERNEL);
1511	if (!qpd->doorbell_bitmap)
1512	return -ENOMEM;
1513
1514	/* Mask out doorbells reserved for SDMA, IH, and VCN on SOC15. */
1515	pr_debug("reserved doorbell 0x%03x - 0x%03x\n", range_start, range_end);
1516	pr_debug("reserved doorbell 0x%03x - 0x%03x\n",
1517	range_start + KFD_QUEUE_DOORBELL_MIRROR_OFFSET,
1518	range_end + KFD_QUEUE_DOORBELL_MIRROR_OFFSET);
1519
1520	for (i = 0; i < KFD_MAX_NUM_OF_QUEUES_PER_PROCESS / 2; i++) {
1521	if (i >= range_start && i <= range_end) {
1522	__set_bit(i, qpd->doorbell_bitmap);
1523	__set_bit(i + KFD_QUEUE_DOORBELL_MIRROR_OFFSET,
1524	qpd->doorbell_bitmap);
1525	}
1526	}
1527
1528	return 0;
1529	}
1530
1531	struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
1532	struct kfd_process *p)
1533	{
1534	int i;
1535
1536	for (i = 0; i < p->n_pdds; i++)
1537	if (p->pdds[i]->dev == dev)
1538	return p->pdds[i];
1539
1540	return NULL;
1541	}
1542
1543	struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
1544	struct kfd_process *p)
1545	{
1546	struct kfd_process_device *pdd = NULL;
1547	int retval = 0;
1548
1549	if (WARN_ON_ONCE(p->n_pdds >= MAX_GPU_INSTANCE))
1550	return NULL;
1551	pdd = kzalloc(size: sizeof(*pdd), GFP_KERNEL);
1552	if (!pdd)
1553	return NULL;
1554
1555	if (init_doorbell_bitmap(qpd: &pdd->qpd, dev)) {
1556	pr_err("Failed to init doorbell for process\n");
1557	goto err_free_pdd;
1558	}
1559
1560	pdd->dev = dev;
1561	INIT_LIST_HEAD(list: &pdd->qpd.queues_list);
1562	INIT_LIST_HEAD(list: &pdd->qpd.priv_queue_list);
1563	pdd->qpd.dqm = dev->dqm;
1564	pdd->qpd.pqm = &p->pqm;
1565	pdd->qpd.evicted = 0;
1566	pdd->qpd.mapped_gws_queue = false;
1567	pdd->process = p;
1568	pdd->bound = PDD_UNBOUND;
1569	pdd->already_dequeued = false;
1570	pdd->runtime_inuse = false;
1571	pdd->vram_usage = 0;
1572	pdd->sdma_past_activity_counter = 0;
1573	pdd->user_gpu_id = dev->id;
1574	atomic64_set(&pdd->evict_duration_counter, 0);
1575
1576	if (dev->shared_resources.enable_mes) {
1577	retval = amdgpu_amdkfd_alloc_gtt_mem(dev->adev,
1578	AMDGPU_MES_PROC_CTX_SIZE,
1579	&pdd->proc_ctx_bo,
1580	&pdd->proc_ctx_gpu_addr,
1581	&pdd->proc_ctx_cpu_ptr,
1582	false);
1583	if (retval) {
1584	pr_err("failed to allocate process context bo\n");
1585	goto err_free_pdd;
1586	}
1587	memset(pdd->proc_ctx_cpu_ptr, 0, AMDGPU_MES_PROC_CTX_SIZE);
1588	}
1589
1590	p->pdds[p->n_pdds++] = pdd;
1591
1592	/* Init idr used for memory handle translation */
1593	idr_init(idr: &pdd->alloc_idr);
1594
1595	return pdd;
1596
1597	err_free_pdd:
1598	kfree(objp: pdd);
1599	return NULL;
1600	}
1601
1602	/**
1603	* kfd_process_device_init_vm - Initialize a VM for a process-device
1604	*
1605	* @pdd: The process-device
1606	* @drm_file: Optional pointer to a DRM file descriptor
1607	*
1608	* If @drm_file is specified, it will be used to acquire the VM from
1609	* that file descriptor. If successful, the @pdd takes ownership of
1610	* the file descriptor.
1611	*
1612	* If @drm_file is NULL, a new VM is created.
1613	*
1614	* Returns 0 on success, -errno on failure.
1615	*/
1616	int kfd_process_device_init_vm(struct kfd_process_device *pdd,
1617	struct file *drm_file)
1618	{
1619	struct amdgpu_fpriv *drv_priv;
1620	struct amdgpu_vm *avm;
1621	struct kfd_process *p;
1622	struct kfd_dev *dev;
1623	int ret;
1624
1625	if (!drm_file)
1626	return -EINVAL;
1627
1628	if (pdd->drm_priv)
1629	return -EBUSY;
1630
1631	ret = amdgpu_file_to_fpriv(drm_file, &drv_priv);
1632	if (ret)
1633	return ret;
1634	avm = &drv_priv->vm;
1635
1636	p = pdd->process;
1637	dev = pdd->dev;
1638
1639	ret = amdgpu_amdkfd_gpuvm_acquire_process_vm(dev->adev, avm,
1640	&p->kgd_process_info,
1641	&p->ef);
1642	if (ret) {
1643	pr_err("Failed to create process VM object\n");
1644	return ret;
1645	}
1646	pdd->drm_priv = drm_file->private_data;
1647	atomic64_set(&pdd->tlb_seq, 0);
1648
1649	ret = kfd_process_device_reserve_ib_mem(pdd);
1650	if (ret)
1651	goto err_reserve_ib_mem;
1652	ret = kfd_process_device_init_cwsr_dgpu(pdd);
1653	if (ret)
1654	goto err_init_cwsr;
1655
1656	ret = amdgpu_amdkfd_gpuvm_set_vm_pasid(dev->adev, avm, p->pasid);
1657	if (ret)
1658	goto err_set_pasid;
1659
1660	pdd->drm_file = drm_file;
1661
1662	return 0;
1663
1664	err_set_pasid:
1665	kfd_process_device_destroy_cwsr_dgpu(pdd);
1666	err_init_cwsr:
1667	kfd_process_device_destroy_ib_mem(pdd);
1668	err_reserve_ib_mem:
1669	pdd->drm_priv = NULL;
1670	amdgpu_amdkfd_gpuvm_destroy_cb(dev->adev, avm);
1671
1672	return ret;
1673	}
1674
1675	/*
1676	* Direct the IOMMU to bind the process (specifically the pasid->mm)
1677	* to the device.
1678	* Unbinding occurs when the process dies or the device is removed.
1679	*
1680	* Assumes that the process lock is held.
1681	*/
1682	struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
1683	struct kfd_process *p)
1684	{
1685	struct kfd_process_device *pdd;
1686	int err;
1687
1688	pdd = kfd_get_process_device_data(dev, p);
1689	if (!pdd) {
1690	pr_err("Process device data doesn't exist\n");
1691	return ERR_PTR(-ENOMEM);
1692	}
1693
1694	if (!pdd->drm_priv)
1695	return ERR_PTR(-ENODEV);
1696
1697	/*
1698	* signal runtime-pm system to auto resume and prevent
1699	* further runtime suspend once device pdd is created until
1700	* pdd is destroyed.
1701	*/
1702	if (!pdd->runtime_inuse) {
1703	err = pm_runtime_get_sync(dev: adev_to_drm(dev->adev)->dev);
1704	if (err < 0) {
1705	pm_runtime_put_autosuspend(dev: adev_to_drm(dev->adev)->dev);
1706	return ERR_PTR(error: err);
1707	}
1708	}
1709
1710	err = kfd_iommu_bind_process_to_device(pdd);
1711	if (err)
1712	goto out;
1713
1714	/*
1715	* make sure that runtime_usage counter is incremented just once
1716	* per pdd
1717	*/
1718	pdd->runtime_inuse = true;
1719
1720	return pdd;
1721
1722	out:
1723	/* balance runpm reference count and exit with error */
1724	if (!pdd->runtime_inuse) {
1725	pm_runtime_mark_last_busy(dev: adev_to_drm(dev->adev)->dev);
1726	pm_runtime_put_autosuspend(dev: adev_to_drm(dev->adev)->dev);
1727	}
1728
1729	return ERR_PTR(error: err);
1730	}
1731
1732	/* Create specific handle mapped to mem from process local memory idr
1733	* Assumes that the process lock is held.
1734	*/
1735	int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
1736	void *mem)
1737	{
1738	return idr_alloc(&pdd->alloc_idr, ptr: mem, start: 0, end: 0, GFP_KERNEL);
1739	}
1740
1741	/* Translate specific handle from process local memory idr
1742	* Assumes that the process lock is held.
1743	*/
1744	void *kfd_process_device_translate_handle(struct kfd_process_device *pdd,
1745	int handle)
1746	{
1747	if (handle < 0)
1748	return NULL;
1749
1750	return idr_find(&pdd->alloc_idr, id: handle);
1751	}
1752
1753	/* Remove specific handle from process local memory idr
1754	* Assumes that the process lock is held.
1755	*/
1756	void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,
1757	int handle)
1758	{
1759	if (handle >= 0)
1760	idr_remove(&pdd->alloc_idr, id: handle);
1761	}
1762
1763	/* This increments the process->ref counter. */
1764	struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid)
1765	{
1766	struct kfd_process *p, *ret_p = NULL;
1767	unsigned int temp;
1768
1769	int idx = srcu_read_lock(ssp: &kfd_processes_srcu);
1770
1771	hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1772	if (p->pasid == pasid) {
1773	kref_get(kref: &p->ref);
1774	ret_p = p;
1775	break;
1776	}
1777	}
1778
1779	srcu_read_unlock(ssp: &kfd_processes_srcu, idx);
1780
1781	return ret_p;
1782	}
1783
1784	/* This increments the process->ref counter. */
1785	struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm)
1786	{
1787	struct kfd_process *p;
1788
1789	int idx = srcu_read_lock(ssp: &kfd_processes_srcu);
1790
1791	p = find_process_by_mm(mm);
1792	if (p)
1793	kref_get(kref: &p->ref);
1794
1795	srcu_read_unlock(ssp: &kfd_processes_srcu, idx);
1796
1797	return p;
1798	}
1799
1800	/* kfd_process_evict_queues - Evict all user queues of a process
1801	*
1802	* Eviction is reference-counted per process-device. This means multiple
1803	* evictions from different sources can be nested safely.
1804	*/
1805	int kfd_process_evict_queues(struct kfd_process *p, uint32_t trigger)
1806	{
1807	int r = 0;
1808	int i;
1809	unsigned int n_evicted = 0;
1810
1811	for (i = 0; i < p->n_pdds; i++) {
1812	struct kfd_process_device *pdd = p->pdds[i];
1813
1814	kfd_smi_event_queue_eviction(dev: pdd->dev, pid: p->lead_thread->pid,
1815	trigger);
1816
1817	r = pdd->dev->dqm->ops.evict_process_queues(pdd->dev->dqm,
1818	&pdd->qpd);
1819	/* evict return -EIO if HWS is hang or asic is resetting, in this case
1820	* we would like to set all the queues to be in evicted state to prevent
1821	* them been add back since they actually not be saved right now.
1822	*/
1823	if (r && r != -EIO) {
1824	pr_err("Failed to evict process queues\n");
1825	goto fail;
1826	}
1827	n_evicted++;
1828	}
1829
1830	return r;
1831
1832	fail:
1833	/* To keep state consistent, roll back partial eviction by
1834	* restoring queues
1835	*/
1836	for (i = 0; i < p->n_pdds; i++) {
1837	struct kfd_process_device *pdd = p->pdds[i];
1838
1839	if (n_evicted == 0)
1840	break;
1841
1842	kfd_smi_event_queue_restore(dev: pdd->dev, pid: p->lead_thread->pid);
1843
1844	if (pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm,
1845	&pdd->qpd))
1846	pr_err("Failed to restore queues\n");
1847
1848	n_evicted--;
1849	}
1850
1851	return r;
1852	}
1853
1854	/* kfd_process_restore_queues - Restore all user queues of a process */
1855	int kfd_process_restore_queues(struct kfd_process *p)
1856	{
1857	int r, ret = 0;
1858	int i;
1859
1860	for (i = 0; i < p->n_pdds; i++) {
1861	struct kfd_process_device *pdd = p->pdds[i];
1862
1863	kfd_smi_event_queue_restore(dev: pdd->dev, pid: p->lead_thread->pid);
1864
1865	r = pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm,
1866	&pdd->qpd);
1867	if (r) {
1868	pr_err("Failed to restore process queues\n");
1869	if (!ret)
1870	ret = r;
1871	}
1872	}
1873
1874	return ret;
1875	}
1876
1877	int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id)
1878	{
1879	int i;
1880
1881	for (i = 0; i < p->n_pdds; i++)
1882	if (p->pdds[i] && gpu_id == p->pdds[i]->user_gpu_id)
1883	return i;
1884	return -EINVAL;
1885	}
1886
1887	int
1888	kfd_process_gpuid_from_adev(struct kfd_process *p, struct amdgpu_device *adev,
1889	uint32_t *gpuid, uint32_t *gpuidx)
1890	{
1891	int i;
1892
1893	for (i = 0; i < p->n_pdds; i++)
1894	if (p->pdds[i] && p->pdds[i]->dev->adev == adev) {
1895	*gpuid = p->pdds[i]->user_gpu_id;
1896	*gpuidx = i;
1897	return 0;
1898	}
1899	return -EINVAL;
1900	}
1901
1902	static void evict_process_worker(struct work_struct *work)
1903	{
1904	int ret;
1905	struct kfd_process *p;
1906	struct delayed_work *dwork;
1907
1908	dwork = to_delayed_work(work);
1909
1910	/* Process termination destroys this worker thread. So during the
1911	* lifetime of this thread, kfd_process p will be valid
1912	*/
1913	p = container_of(dwork, struct kfd_process, eviction_work);
1914	WARN_ONCE(p->last_eviction_seqno != p->ef->seqno,
1915	"Eviction fence mismatch\n");
1916
1917	/* Narrow window of overlap between restore and evict work
1918	* item is possible. Once amdgpu_amdkfd_gpuvm_restore_process_bos
1919	* unreserves KFD BOs, it is possible to evicted again. But
1920	* restore has few more steps of finish. So lets wait for any
1921	* previous restore work to complete
1922	*/
1923	flush_delayed_work(dwork: &p->restore_work);
1924
1925	pr_debug("Started evicting pasid 0x%x\n", p->pasid);
1926	ret = kfd_process_evict_queues(p, trigger: KFD_QUEUE_EVICTION_TRIGGER_TTM);
1927	if (!ret) {
1928	dma_fence_signal(p->ef);
1929	dma_fence_put(p->ef);
1930	p->ef = NULL;
1931	queue_delayed_work(wq: kfd_restore_wq, dwork: &p->restore_work,
1932	delay: msecs_to_jiffies(PROCESS_RESTORE_TIME_MS));
1933
1934	pr_debug("Finished evicting pasid 0x%x\n", p->pasid);
1935	} else
1936	pr_err("Failed to evict queues of pasid 0x%x\n", p->pasid);
1937	}
1938
1939	static void restore_process_worker(struct work_struct *work)
1940	{
1941	struct delayed_work *dwork;
1942	struct kfd_process *p;
1943	int ret = 0;
1944
1945	dwork = to_delayed_work(work);
1946
1947	/* Process termination destroys this worker thread. So during the
1948	* lifetime of this thread, kfd_process p will be valid
1949	*/
1950	p = container_of(dwork, struct kfd_process, restore_work);
1951	pr_debug("Started restoring pasid 0x%x\n", p->pasid);
1952
1953	/* Setting last_restore_timestamp before successful restoration.
1954	* Otherwise this would have to be set by KGD (restore_process_bos)
1955	* before KFD BOs are unreserved. If not, the process can be evicted
1956	* again before the timestamp is set.
1957	* If restore fails, the timestamp will be set again in the next
1958	* attempt. This would mean that the minimum GPU quanta would be
1959	* PROCESS_ACTIVE_TIME_MS - (time to execute the following two
1960	* functions)
1961	*/
1962
1963	p->last_restore_timestamp = get_jiffies_64();
1964	ret = amdgpu_amdkfd_gpuvm_restore_process_bos(p->kgd_process_info,
1965	&p->ef);
1966	if (ret) {
1967	pr_debug("Failed to restore BOs of pasid 0x%x, retry after %d ms\n",
1968	p->pasid, PROCESS_BACK_OFF_TIME_MS);
1969	ret = queue_delayed_work(wq: kfd_restore_wq, dwork: &p->restore_work,
1970	delay: msecs_to_jiffies(PROCESS_BACK_OFF_TIME_MS));
1971	WARN(!ret, "reschedule restore work failed\n");
1972	return;
1973	}
1974
1975	ret = kfd_process_restore_queues(p);
1976	if (!ret)
1977	pr_debug("Finished restoring pasid 0x%x\n", p->pasid);
1978	else
1979	pr_err("Failed to restore queues of pasid 0x%x\n", p->pasid);
1980	}
1981
1982	void kfd_suspend_all_processes(void)
1983	{
1984	struct kfd_process *p;
1985	unsigned int temp;
1986	int idx = srcu_read_lock(ssp: &kfd_processes_srcu);
1987
1988	WARN(debug_evictions, "Evicting all processes");
1989	hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1990	cancel_delayed_work_sync(dwork: &p->eviction_work);
1991	cancel_delayed_work_sync(dwork: &p->restore_work);
1992
1993	if (kfd_process_evict_queues(p, trigger: KFD_QUEUE_EVICTION_TRIGGER_SUSPEND))
1994	pr_err("Failed to suspend process 0x%x\n", p->pasid);
1995	dma_fence_signal(p->ef);
1996	dma_fence_put(p->ef);
1997	p->ef = NULL;
1998	}
1999	srcu_read_unlock(ssp: &kfd_processes_srcu, idx);
2000	}
2001
2002	int kfd_resume_all_processes(void)
2003	{
2004	struct kfd_process *p;
2005	unsigned int temp;
2006	int ret = 0, idx = srcu_read_lock(ssp: &kfd_processes_srcu);
2007
2008	hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
2009	if (!queue_delayed_work(wq: kfd_restore_wq, dwork: &p->restore_work, delay: 0)) {
2010	pr_err("Restore process %d failed during resume\n",
2011	p->pasid);
2012	ret = -EFAULT;
2013	}
2014	}
2015	srcu_read_unlock(ssp: &kfd_processes_srcu, idx);
2016	return ret;
2017	}
2018
2019	int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process,
2020	struct vm_area_struct *vma)
2021	{
2022	struct kfd_process_device *pdd;
2023	struct qcm_process_device *qpd;
2024
2025	if ((vma->vm_end - vma->vm_start) != KFD_CWSR_TBA_TMA_SIZE) {
2026	pr_err("Incorrect CWSR mapping size.\n");
2027	return -EINVAL;
2028	}
2029
2030	pdd = kfd_get_process_device_data(dev, p: process);
2031	if (!pdd)
2032	return -EINVAL;
2033	qpd = &pdd->qpd;
2034
2035	qpd->cwsr_kaddr = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
2036	order: get_order(KFD_CWSR_TBA_TMA_SIZE));
2037	if (!qpd->cwsr_kaddr) {
2038	pr_err("Error allocating per process CWSR buffer.\n");
2039	return -ENOMEM;
2040	}
2041
2042	vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND
2043	| VM_NORESERVE | VM_DONTDUMP | VM_PFNMAP);
2044	/* Mapping pages to user process */
2045	return remap_pfn_range(vma, vma->vm_start,
2046	PFN_DOWN(__pa(qpd->cwsr_kaddr)),
2047	KFD_CWSR_TBA_TMA_SIZE, vma->vm_page_prot);
2048	}
2049
2050	void kfd_flush_tlb(struct kfd_process_device *pdd, enum TLB_FLUSH_TYPE type)
2051	{
2052	struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv);
2053	uint64_t tlb_seq = amdgpu_vm_tlb_seq(vm);
2054	struct kfd_dev *dev = pdd->dev;
2055
2056	/*
2057	* It can be that we race and lose here, but that is extremely unlikely
2058	* and the worst thing which could happen is that we flush the changes
2059	* into the TLB once more which is harmless.
2060	*/
2061	if (atomic64_xchg(&pdd->tlb_seq, tlb_seq) == tlb_seq)
2062	return;
2063
2064	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
2065	/* Nothing to flush until a VMID is assigned, which
2066	* only happens when the first queue is created.
2067	*/
2068	if (pdd->qpd.vmid)
2069	amdgpu_amdkfd_flush_gpu_tlb_vmid(dev->adev,
2070	pdd->qpd.vmid);
2071	} else {
2072	amdgpu_amdkfd_flush_gpu_tlb_pasid(dev->adev,
2073	pdd->process->pasid, type);
2074	}
2075	}
2076
2077	struct kfd_process_device *kfd_process_device_data_by_id(struct kfd_process *p, uint32_t gpu_id)
2078	{
2079	int i;
2080
2081	if (gpu_id) {
2082	for (i = 0; i < p->n_pdds; i++) {
2083	struct kfd_process_device *pdd = p->pdds[i];
2084
2085	if (pdd->user_gpu_id == gpu_id)
2086	return pdd;
2087	}
2088	}
2089	return NULL;
2090	}
2091
2092	int kfd_process_get_user_gpu_id(struct kfd_process *p, uint32_t actual_gpu_id)
2093	{
2094	int i;
2095
2096	if (!actual_gpu_id)
2097	return 0;
2098
2099	for (i = 0; i < p->n_pdds; i++) {
2100	struct kfd_process_device *pdd = p->pdds[i];
2101
2102	if (pdd->dev->id == actual_gpu_id)
2103	return pdd->user_gpu_id;
2104	}
2105	return -EINVAL;
2106	}
2107
2108	#if defined(CONFIG_DEBUG_FS)
2109
2110	int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data)
2111	{
2112	struct kfd_process *p;
2113	unsigned int temp;
2114	int r = 0;
2115
2116	int idx = srcu_read_lock(&kfd_processes_srcu);
2117
2118	hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
2119	seq_printf(m, "Process %d PASID 0x%x:\n",
2120	p->lead_thread->tgid, p->pasid);
2121
2122	mutex_lock(&p->mutex);
2123	r = pqm_debugfs_mqds(m, &p->pqm);
2124	mutex_unlock(&p->mutex);
2125
2126	if (r)
2127	break;
2128	}
2129
2130	srcu_read_unlock(&kfd_processes_srcu, idx);
2131
2132	return r;
2133	}
2134
2135	#endif
2136
2137