1 /*
2  * Copyright 2014 Advanced Micro Devices, Inc.
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice shall be included in
12  * all copies or substantial portions of the Software.
13  *
14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
20  * OTHER DEALINGS IN THE SOFTWARE.
21  */
22 
23 #include <linux/mutex.h>
24 #include <linux/log2.h>
25 #include <linux/sched.h>
26 #include <linux/sched/mm.h>
27 #include <linux/sched/task.h>
28 #include <linux/mmu_context.h>
29 #include <linux/slab.h>
30 #include <linux/amd-iommu.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_dbgmgr.h"
44 #include "kfd_iommu.h"
45 
46 /*
47  * List of struct kfd_process (field kfd_process).
48  * Unique/indexed by mm_struct*
49  */
50 DEFINE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
51 static DEFINE_MUTEX(kfd_processes_mutex);
52 
53 DEFINE_SRCU(kfd_processes_srcu);
54 
55 /* For process termination handling */
56 static struct workqueue_struct *kfd_process_wq;
57 
58 /* Ordered, single-threaded workqueue for restoring evicted
59  * processes. Restoring multiple processes concurrently under memory
60  * pressure can lead to processes blocking each other from validating
61  * their BOs and result in a live-lock situation where processes
62  * remain evicted indefinitely.
63  */
64 static struct workqueue_struct *kfd_restore_wq;
65 
66 static struct kfd_process *find_process(const struct task_struct *thread);
67 static void kfd_process_ref_release(struct kref *ref);
68 static struct kfd_process *create_process(const struct task_struct *thread);
69 static int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep);
70 
71 static void evict_process_worker(struct work_struct *work);
72 static void restore_process_worker(struct work_struct *work);
73 
74 struct kfd_procfs_tree {
75 	struct kobject *kobj;
76 };
77 
78 static struct kfd_procfs_tree procfs;
79 
80 /*
81  * Structure for SDMA activity tracking
82  */
83 struct kfd_sdma_activity_handler_workarea {
84 	struct work_struct sdma_activity_work;
85 	struct kfd_process_device *pdd;
86 	uint64_t sdma_activity_counter;
87 };
88 
89 struct temp_sdma_queue_list {
90 	uint64_t __user *rptr;
91 	uint64_t sdma_val;
92 	unsigned int queue_id;
93 	struct list_head list;
94 };
95 
96 static void kfd_sdma_activity_worker(struct work_struct *work)
97 {
98 	struct kfd_sdma_activity_handler_workarea *workarea;
99 	struct kfd_process_device *pdd;
100 	uint64_t val;
101 	struct mm_struct *mm;
102 	struct queue *q;
103 	struct qcm_process_device *qpd;
104 	struct device_queue_manager *dqm;
105 	int ret = 0;
106 	struct temp_sdma_queue_list sdma_q_list;
107 	struct temp_sdma_queue_list *sdma_q, *next;
108 
109 	workarea = container_of(work, struct kfd_sdma_activity_handler_workarea,
110 				sdma_activity_work);
111 	if (!workarea)
112 		return;
113 
114 	pdd = workarea->pdd;
115 	if (!pdd)
116 		return;
117 	dqm = pdd->dev->dqm;
118 	qpd = &pdd->qpd;
119 	if (!dqm || !qpd)
120 		return;
121 	/*
122 	 * Total SDMA activity is current SDMA activity + past SDMA activity
123 	 * Past SDMA count is stored in pdd.
124 	 * To get the current activity counters for all active SDMA queues,
125 	 * we loop over all SDMA queues and get their counts from user-space.
126 	 *
127 	 * We cannot call get_user() with dqm_lock held as it can cause
128 	 * a circular lock dependency situation. To read the SDMA stats,
129 	 * we need to do the following:
130 	 *
131 	 * 1. Create a temporary list of SDMA queue nodes from the qpd->queues_list,
132 	 *    with dqm_lock/dqm_unlock().
133 	 * 2. Call get_user() for each node in temporary list without dqm_lock.
134 	 *    Save the SDMA count for each node and also add the count to the total
135 	 *    SDMA count counter.
136 	 *    Its possible, during this step, a few SDMA queue nodes got deleted
137 	 *    from the qpd->queues_list.
138 	 * 3. Do a second pass over qpd->queues_list to check if any nodes got deleted.
139 	 *    If any node got deleted, its SDMA count would be captured in the sdma
140 	 *    past activity counter. So subtract the SDMA counter stored in step 2
141 	 *    for this node from the total SDMA count.
142 	 */
143 	INIT_LIST_HEAD(&sdma_q_list.list);
144 
145 	/*
146 	 * Create the temp list of all SDMA queues
147 	 */
148 	dqm_lock(dqm);
149 
150 	list_for_each_entry(q, &qpd->queues_list, list) {
151 		if ((q->properties.type != KFD_QUEUE_TYPE_SDMA) &&
152 		    (q->properties.type != KFD_QUEUE_TYPE_SDMA_XGMI))
153 			continue;
154 
155 		sdma_q = kzalloc(sizeof(struct temp_sdma_queue_list), GFP_KERNEL);
156 		if (!sdma_q) {
157 			dqm_unlock(dqm);
158 			goto cleanup;
159 		}
160 
161 		INIT_LIST_HEAD(&sdma_q->list);
162 		sdma_q->rptr = (uint64_t __user *)q->properties.read_ptr;
163 		sdma_q->queue_id = q->properties.queue_id;
164 		list_add_tail(&sdma_q->list, &sdma_q_list.list);
165 	}
166 
167 	/*
168 	 * If the temp list is empty, then no SDMA queues nodes were found in
169 	 * qpd->queues_list. Return the past activity count as the total sdma
170 	 * count
171 	 */
172 	if (list_empty(&sdma_q_list.list)) {
173 		workarea->sdma_activity_counter = pdd->sdma_past_activity_counter;
174 		dqm_unlock(dqm);
175 		return;
176 	}
177 
178 	dqm_unlock(dqm);
179 
180 	/*
181 	 * Get the usage count for each SDMA queue in temp_list.
182 	 */
183 	mm = get_task_mm(pdd->process->lead_thread);
184 	if (!mm)
185 		goto cleanup;
186 
187 	kthread_use_mm(mm);
188 
189 	list_for_each_entry(sdma_q, &sdma_q_list.list, list) {
190 		val = 0;
191 		ret = read_sdma_queue_counter(sdma_q->rptr, &val);
192 		if (ret) {
193 			pr_debug("Failed to read SDMA queue active counter for queue id: %d",
194 				 sdma_q->queue_id);
195 		} else {
196 			sdma_q->sdma_val = val;
197 			workarea->sdma_activity_counter += val;
198 		}
199 	}
200 
201 	kthread_unuse_mm(mm);
202 	mmput(mm);
203 
204 	/*
205 	 * Do a second iteration over qpd_queues_list to check if any SDMA
206 	 * nodes got deleted while fetching SDMA counter.
207 	 */
208 	dqm_lock(dqm);
209 
210 	workarea->sdma_activity_counter += pdd->sdma_past_activity_counter;
211 
212 	list_for_each_entry(q, &qpd->queues_list, list) {
213 		if (list_empty(&sdma_q_list.list))
214 			break;
215 
216 		if ((q->properties.type != KFD_QUEUE_TYPE_SDMA) &&
217 		    (q->properties.type != KFD_QUEUE_TYPE_SDMA_XGMI))
218 			continue;
219 
220 		list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
221 			if (((uint64_t __user *)q->properties.read_ptr == sdma_q->rptr) &&
222 			     (sdma_q->queue_id == q->properties.queue_id)) {
223 				list_del(&sdma_q->list);
224 				kfree(sdma_q);
225 				break;
226 			}
227 		}
228 	}
229 
230 	dqm_unlock(dqm);
231 
232 	/*
233 	 * If temp list is not empty, it implies some queues got deleted
234 	 * from qpd->queues_list during SDMA usage read. Subtract the SDMA
235 	 * count for each node from the total SDMA count.
236 	 */
237 	list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
238 		workarea->sdma_activity_counter -= sdma_q->sdma_val;
239 		list_del(&sdma_q->list);
240 		kfree(sdma_q);
241 	}
242 
243 	return;
244 
245 cleanup:
246 	list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
247 		list_del(&sdma_q->list);
248 		kfree(sdma_q);
249 	}
250 }
251 
252 /**
253  * @kfd_get_cu_occupancy() - Collect number of waves in-flight on this device
254  * by current process. Translates acquired wave count into number of compute units
255  * that are occupied.
256  *
257  * @atr: Handle of attribute that allows reporting of wave count. The attribute
258  * handle encapsulates GPU device it is associated with, thereby allowing collection
259  * of waves in flight, etc
260  *
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_dev *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->kgd, proc->pasid, &wave_cnt,
291 			&max_waves_per_cu);
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 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 struct attribute attr_queue_size = {
421 	.name = "size",
422 	.mode = KFD_SYSFS_FILE_MODE
423 };
424 
425 static struct attribute attr_queue_type = {
426 	.name = "type",
427 	.mode = KFD_SYSFS_FILE_MODE
428 };
429 
430 static struct attribute attr_queue_gpuid = {
431 	.name = "gpuid",
432 	.mode = KFD_SYSFS_FILE_MODE
433 };
434 
435 static struct attribute *procfs_queue_attrs[] = {
436 	&attr_queue_size,
437 	&attr_queue_type,
438 	&attr_queue_gpuid,
439 	NULL
440 };
441 
442 static const struct sysfs_ops procfs_queue_ops = {
443 	.show = kfd_procfs_queue_show,
444 };
445 
446 static struct kobj_type procfs_queue_type = {
447 	.sysfs_ops = &procfs_queue_ops,
448 	.default_attrs = procfs_queue_attrs,
449 };
450 
451 static const struct sysfs_ops procfs_stats_ops = {
452 	.show = kfd_procfs_stats_show,
453 };
454 
455 static struct attribute *procfs_stats_attrs[] = {
456 	NULL
457 };
458 
459 static struct kobj_type procfs_stats_type = {
460 	.sysfs_ops = &procfs_stats_ops,
461 	.default_attrs = procfs_stats_attrs,
462 };
463 
464 int kfd_procfs_add_queue(struct queue *q)
465 {
466 	struct kfd_process *proc;
467 	int ret;
468 
469 	if (!q || !q->process)
470 		return -EINVAL;
471 	proc = q->process;
472 
473 	/* Create proc/<pid>/queues/<queue id> folder */
474 	if (!proc->kobj_queues)
475 		return -EFAULT;
476 	ret = kobject_init_and_add(&q->kobj, &procfs_queue_type,
477 			proc->kobj_queues, "%u", q->properties.queue_id);
478 	if (ret < 0) {
479 		pr_warn("Creating proc/<pid>/queues/%u failed",
480 			q->properties.queue_id);
481 		kobject_put(&q->kobj);
482 		return ret;
483 	}
484 
485 	return 0;
486 }
487 
488 static int kfd_sysfs_create_file(struct kfd_process *p, struct attribute *attr,
489 				 char *name)
490 {
491 	int ret = 0;
492 
493 	if (!p || !attr || !name)
494 		return -EINVAL;
495 
496 	attr->name = name;
497 	attr->mode = KFD_SYSFS_FILE_MODE;
498 	sysfs_attr_init(attr);
499 
500 	ret = sysfs_create_file(p->kobj, attr);
501 
502 	return ret;
503 }
504 
505 static int kfd_procfs_add_sysfs_stats(struct kfd_process *p)
506 {
507 	int ret = 0;
508 	int i;
509 	char stats_dir_filename[MAX_SYSFS_FILENAME_LEN];
510 
511 	if (!p)
512 		return -EINVAL;
513 
514 	if (!p->kobj)
515 		return -EFAULT;
516 
517 	/*
518 	 * Create sysfs files for each GPU:
519 	 * - proc/<pid>/stats_<gpuid>/
520 	 * - proc/<pid>/stats_<gpuid>/evicted_ms
521 	 * - proc/<pid>/stats_<gpuid>/cu_occupancy
522 	 */
523 	for (i = 0; i < p->n_pdds; i++) {
524 		struct kfd_process_device *pdd = p->pdds[i];
525 		struct kobject *kobj_stats;
526 
527 		snprintf(stats_dir_filename, MAX_SYSFS_FILENAME_LEN,
528 				"stats_%u", pdd->dev->id);
529 		kobj_stats = kfd_alloc_struct(kobj_stats);
530 		if (!kobj_stats)
531 			return -ENOMEM;
532 
533 		ret = kobject_init_and_add(kobj_stats,
534 						&procfs_stats_type,
535 						p->kobj,
536 						stats_dir_filename);
537 
538 		if (ret) {
539 			pr_warn("Creating KFD proc/stats_%s folder failed",
540 					stats_dir_filename);
541 			kobject_put(kobj_stats);
542 			goto err;
543 		}
544 
545 		pdd->kobj_stats = kobj_stats;
546 		pdd->attr_evict.name = "evicted_ms";
547 		pdd->attr_evict.mode = KFD_SYSFS_FILE_MODE;
548 		sysfs_attr_init(&pdd->attr_evict);
549 		ret = sysfs_create_file(kobj_stats, &pdd->attr_evict);
550 		if (ret)
551 			pr_warn("Creating eviction stats for gpuid %d failed",
552 					(int)pdd->dev->id);
553 
554 		/* Add sysfs file to report compute unit occupancy */
555 		if (pdd->dev->kfd2kgd->get_cu_occupancy != NULL) {
556 			pdd->attr_cu_occupancy.name = "cu_occupancy";
557 			pdd->attr_cu_occupancy.mode = KFD_SYSFS_FILE_MODE;
558 			sysfs_attr_init(&pdd->attr_cu_occupancy);
559 			ret = sysfs_create_file(kobj_stats,
560 						&pdd->attr_cu_occupancy);
561 			if (ret)
562 				pr_warn("Creating %s failed for gpuid: %d",
563 					pdd->attr_cu_occupancy.name,
564 					(int)pdd->dev->id);
565 		}
566 	}
567 err:
568 	return ret;
569 }
570 
571 
572 static int kfd_procfs_add_sysfs_files(struct kfd_process *p)
573 {
574 	int ret = 0;
575 	int i;
576 
577 	if (!p)
578 		return -EINVAL;
579 
580 	if (!p->kobj)
581 		return -EFAULT;
582 
583 	/*
584 	 * Create sysfs files for each GPU:
585 	 * - proc/<pid>/vram_<gpuid>
586 	 * - proc/<pid>/sdma_<gpuid>
587 	 */
588 	for (i = 0; i < p->n_pdds; i++) {
589 		struct kfd_process_device *pdd = p->pdds[i];
590 
591 		snprintf(pdd->vram_filename, MAX_SYSFS_FILENAME_LEN, "vram_%u",
592 			 pdd->dev->id);
593 		ret = kfd_sysfs_create_file(p, &pdd->attr_vram, pdd->vram_filename);
594 		if (ret)
595 			pr_warn("Creating vram usage for gpu id %d failed",
596 				(int)pdd->dev->id);
597 
598 		snprintf(pdd->sdma_filename, MAX_SYSFS_FILENAME_LEN, "sdma_%u",
599 			 pdd->dev->id);
600 		ret = kfd_sysfs_create_file(p, &pdd->attr_sdma, pdd->sdma_filename);
601 		if (ret)
602 			pr_warn("Creating sdma usage for gpu id %d failed",
603 				(int)pdd->dev->id);
604 	}
605 
606 	return ret;
607 }
608 
609 void kfd_procfs_del_queue(struct queue *q)
610 {
611 	if (!q)
612 		return;
613 
614 	kobject_del(&q->kobj);
615 	kobject_put(&q->kobj);
616 }
617 
618 int kfd_process_create_wq(void)
619 {
620 	if (!kfd_process_wq)
621 		kfd_process_wq = alloc_workqueue("kfd_process_wq", 0, 0);
622 	if (!kfd_restore_wq)
623 		kfd_restore_wq = alloc_ordered_workqueue("kfd_restore_wq", 0);
624 
625 	if (!kfd_process_wq || !kfd_restore_wq) {
626 		kfd_process_destroy_wq();
627 		return -ENOMEM;
628 	}
629 
630 	return 0;
631 }
632 
633 void kfd_process_destroy_wq(void)
634 {
635 	if (kfd_process_wq) {
636 		destroy_workqueue(kfd_process_wq);
637 		kfd_process_wq = NULL;
638 	}
639 	if (kfd_restore_wq) {
640 		destroy_workqueue(kfd_restore_wq);
641 		kfd_restore_wq = NULL;
642 	}
643 }
644 
645 static void kfd_process_free_gpuvm(struct kgd_mem *mem,
646 			struct kfd_process_device *pdd)
647 {
648 	struct kfd_dev *dev = pdd->dev;
649 
650 	amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(dev->kgd, mem, pdd->vm);
651 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->kgd, mem, NULL);
652 }
653 
654 /* kfd_process_alloc_gpuvm - Allocate GPU VM for the KFD process
655  *	This function should be only called right after the process
656  *	is created and when kfd_processes_mutex is still being held
657  *	to avoid concurrency. Because of that exclusiveness, we do
658  *	not need to take p->mutex.
659  */
660 static int kfd_process_alloc_gpuvm(struct kfd_process_device *pdd,
661 				   uint64_t gpu_va, uint32_t size,
662 				   uint32_t flags, void **kptr)
663 {
664 	struct kfd_dev *kdev = pdd->dev;
665 	struct kgd_mem *mem = NULL;
666 	int handle;
667 	int err;
668 
669 	err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(kdev->kgd, gpu_va, size,
670 						 pdd->vm, &mem, NULL, flags);
671 	if (err)
672 		goto err_alloc_mem;
673 
674 	err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(kdev->kgd, mem, pdd->vm);
675 	if (err)
676 		goto err_map_mem;
677 
678 	err = amdgpu_amdkfd_gpuvm_sync_memory(kdev->kgd, mem, true);
679 	if (err) {
680 		pr_debug("Sync memory failed, wait interrupted by user signal\n");
681 		goto sync_memory_failed;
682 	}
683 
684 	/* Create an obj handle so kfd_process_device_remove_obj_handle
685 	 * will take care of the bo removal when the process finishes.
686 	 * We do not need to take p->mutex, because the process is just
687 	 * created and the ioctls have not had the chance to run.
688 	 */
689 	handle = kfd_process_device_create_obj_handle(pdd, mem);
690 
691 	if (handle < 0) {
692 		err = handle;
693 		goto free_gpuvm;
694 	}
695 
696 	if (kptr) {
697 		err = amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(kdev->kgd,
698 				(struct kgd_mem *)mem, kptr, NULL);
699 		if (err) {
700 			pr_debug("Map GTT BO to kernel failed\n");
701 			goto free_obj_handle;
702 		}
703 	}
704 
705 	return err;
706 
707 free_obj_handle:
708 	kfd_process_device_remove_obj_handle(pdd, handle);
709 free_gpuvm:
710 sync_memory_failed:
711 	kfd_process_free_gpuvm(mem, pdd);
712 	return err;
713 
714 err_map_mem:
715 	amdgpu_amdkfd_gpuvm_free_memory_of_gpu(kdev->kgd, mem, NULL);
716 err_alloc_mem:
717 	*kptr = NULL;
718 	return err;
719 }
720 
721 /* kfd_process_device_reserve_ib_mem - Reserve memory inside the
722  *	process for IB usage The memory reserved is for KFD to submit
723  *	IB to AMDGPU from kernel.  If the memory is reserved
724  *	successfully, ib_kaddr will have the CPU/kernel
725  *	address. Check ib_kaddr before accessing the memory.
726  */
727 static int kfd_process_device_reserve_ib_mem(struct kfd_process_device *pdd)
728 {
729 	struct qcm_process_device *qpd = &pdd->qpd;
730 	uint32_t flags = KFD_IOC_ALLOC_MEM_FLAGS_GTT |
731 			KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE |
732 			KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE |
733 			KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
734 	void *kaddr;
735 	int ret;
736 
737 	if (qpd->ib_kaddr || !qpd->ib_base)
738 		return 0;
739 
740 	/* ib_base is only set for dGPU */
741 	ret = kfd_process_alloc_gpuvm(pdd, qpd->ib_base, PAGE_SIZE, flags,
742 				      &kaddr);
743 	if (ret)
744 		return ret;
745 
746 	qpd->ib_kaddr = kaddr;
747 
748 	return 0;
749 }
750 
751 struct kfd_process *kfd_create_process(struct file *filep)
752 {
753 	struct kfd_process *process;
754 	struct task_struct *thread = current;
755 	int ret;
756 
757 	if (!thread->mm)
758 		return ERR_PTR(-EINVAL);
759 
760 	/* Only the pthreads threading model is supported. */
761 	if (thread->group_leader->mm != thread->mm)
762 		return ERR_PTR(-EINVAL);
763 
764 	/*
765 	 * take kfd processes mutex before starting of process creation
766 	 * so there won't be a case where two threads of the same process
767 	 * create two kfd_process structures
768 	 */
769 	mutex_lock(&kfd_processes_mutex);
770 
771 	/* A prior open of /dev/kfd could have already created the process. */
772 	process = find_process(thread);
773 	if (process) {
774 		pr_debug("Process already found\n");
775 	} else {
776 		process = create_process(thread);
777 		if (IS_ERR(process))
778 			goto out;
779 
780 		ret = kfd_process_init_cwsr_apu(process, filep);
781 		if (ret)
782 			goto out_destroy;
783 
784 		if (!procfs.kobj)
785 			goto out;
786 
787 		process->kobj = kfd_alloc_struct(process->kobj);
788 		if (!process->kobj) {
789 			pr_warn("Creating procfs kobject failed");
790 			goto out;
791 		}
792 		ret = kobject_init_and_add(process->kobj, &procfs_type,
793 					   procfs.kobj, "%d",
794 					   (int)process->lead_thread->pid);
795 		if (ret) {
796 			pr_warn("Creating procfs pid directory failed");
797 			kobject_put(process->kobj);
798 			goto out;
799 		}
800 
801 		process->attr_pasid.name = "pasid";
802 		process->attr_pasid.mode = KFD_SYSFS_FILE_MODE;
803 		sysfs_attr_init(&process->attr_pasid);
804 		ret = sysfs_create_file(process->kobj, &process->attr_pasid);
805 		if (ret)
806 			pr_warn("Creating pasid for pid %d failed",
807 					(int)process->lead_thread->pid);
808 
809 		process->kobj_queues = kobject_create_and_add("queues",
810 							process->kobj);
811 		if (!process->kobj_queues)
812 			pr_warn("Creating KFD proc/queues folder failed");
813 
814 		ret = kfd_procfs_add_sysfs_stats(process);
815 		if (ret)
816 			pr_warn("Creating sysfs stats dir for pid %d failed",
817 				(int)process->lead_thread->pid);
818 
819 		ret = kfd_procfs_add_sysfs_files(process);
820 		if (ret)
821 			pr_warn("Creating sysfs usage file for pid %d failed",
822 				(int)process->lead_thread->pid);
823 	}
824 out:
825 	if (!IS_ERR(process))
826 		kref_get(&process->ref);
827 	mutex_unlock(&kfd_processes_mutex);
828 
829 	return process;
830 
831 out_destroy:
832 	hash_del_rcu(&process->kfd_processes);
833 	mutex_unlock(&kfd_processes_mutex);
834 	synchronize_srcu(&kfd_processes_srcu);
835 	/* kfd_process_free_notifier will trigger the cleanup */
836 	mmu_notifier_put(&process->mmu_notifier);
837 	return ERR_PTR(ret);
838 }
839 
840 struct kfd_process *kfd_get_process(const struct task_struct *thread)
841 {
842 	struct kfd_process *process;
843 
844 	if (!thread->mm)
845 		return ERR_PTR(-EINVAL);
846 
847 	/* Only the pthreads threading model is supported. */
848 	if (thread->group_leader->mm != thread->mm)
849 		return ERR_PTR(-EINVAL);
850 
851 	process = find_process(thread);
852 	if (!process)
853 		return ERR_PTR(-EINVAL);
854 
855 	return process;
856 }
857 
858 static struct kfd_process *find_process_by_mm(const struct mm_struct *mm)
859 {
860 	struct kfd_process *process;
861 
862 	hash_for_each_possible_rcu(kfd_processes_table, process,
863 					kfd_processes, (uintptr_t)mm)
864 		if (process->mm == mm)
865 			return process;
866 
867 	return NULL;
868 }
869 
870 static struct kfd_process *find_process(const struct task_struct *thread)
871 {
872 	struct kfd_process *p;
873 	int idx;
874 
875 	idx = srcu_read_lock(&kfd_processes_srcu);
876 	p = find_process_by_mm(thread->mm);
877 	srcu_read_unlock(&kfd_processes_srcu, idx);
878 
879 	return p;
880 }
881 
882 void kfd_unref_process(struct kfd_process *p)
883 {
884 	kref_put(&p->ref, kfd_process_ref_release);
885 }
886 
887 
888 static void kfd_process_device_free_bos(struct kfd_process_device *pdd)
889 {
890 	struct kfd_process *p = pdd->process;
891 	void *mem;
892 	int id;
893 	int i;
894 
895 	/*
896 	 * Remove all handles from idr and release appropriate
897 	 * local memory object
898 	 */
899 	idr_for_each_entry(&pdd->alloc_idr, mem, id) {
900 
901 		for (i = 0; i < p->n_pdds; i++) {
902 			struct kfd_process_device *peer_pdd = p->pdds[i];
903 
904 			if (!peer_pdd->vm)
905 				continue;
906 			amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
907 				peer_pdd->dev->kgd, mem, peer_pdd->vm);
908 		}
909 
910 		amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->kgd, mem, NULL);
911 		kfd_process_device_remove_obj_handle(pdd, id);
912 	}
913 }
914 
915 static void kfd_process_free_outstanding_kfd_bos(struct kfd_process *p)
916 {
917 	int i;
918 
919 	for (i = 0; i < p->n_pdds; i++)
920 		kfd_process_device_free_bos(p->pdds[i]);
921 }
922 
923 static void kfd_process_destroy_pdds(struct kfd_process *p)
924 {
925 	int i;
926 
927 	for (i = 0; i < p->n_pdds; i++) {
928 		struct kfd_process_device *pdd = p->pdds[i];
929 
930 		pr_debug("Releasing pdd (topology id %d) for process (pasid 0x%x)\n",
931 				pdd->dev->id, p->pasid);
932 
933 		if (pdd->drm_file) {
934 			amdgpu_amdkfd_gpuvm_release_process_vm(
935 					pdd->dev->kgd, pdd->vm);
936 			fput(pdd->drm_file);
937 		}
938 
939 		if (pdd->qpd.cwsr_kaddr && !pdd->qpd.cwsr_base)
940 			free_pages((unsigned long)pdd->qpd.cwsr_kaddr,
941 				get_order(KFD_CWSR_TBA_TMA_SIZE));
942 
943 		kfree(pdd->qpd.doorbell_bitmap);
944 		idr_destroy(&pdd->alloc_idr);
945 
946 		kfd_free_process_doorbells(pdd->dev, pdd->doorbell_index);
947 
948 		/*
949 		 * before destroying pdd, make sure to report availability
950 		 * for auto suspend
951 		 */
952 		if (pdd->runtime_inuse) {
953 			pm_runtime_mark_last_busy(pdd->dev->ddev->dev);
954 			pm_runtime_put_autosuspend(pdd->dev->ddev->dev);
955 			pdd->runtime_inuse = false;
956 		}
957 
958 		kfree(pdd);
959 		p->pdds[i] = NULL;
960 	}
961 	p->n_pdds = 0;
962 }
963 
964 /* No process locking is needed in this function, because the process
965  * is not findable any more. We must assume that no other thread is
966  * using it any more, otherwise we couldn't safely free the process
967  * structure in the end.
968  */
969 static void kfd_process_wq_release(struct work_struct *work)
970 {
971 	struct kfd_process *p = container_of(work, struct kfd_process,
972 					     release_work);
973 	int i;
974 
975 	/* Remove the procfs files */
976 	if (p->kobj) {
977 		sysfs_remove_file(p->kobj, &p->attr_pasid);
978 		kobject_del(p->kobj_queues);
979 		kobject_put(p->kobj_queues);
980 		p->kobj_queues = NULL;
981 
982 		for (i = 0; i < p->n_pdds; i++) {
983 			struct kfd_process_device *pdd = p->pdds[i];
984 
985 			sysfs_remove_file(p->kobj, &pdd->attr_vram);
986 			sysfs_remove_file(p->kobj, &pdd->attr_sdma);
987 			sysfs_remove_file(p->kobj, &pdd->attr_evict);
988 			if (pdd->dev->kfd2kgd->get_cu_occupancy != NULL)
989 				sysfs_remove_file(p->kobj, &pdd->attr_cu_occupancy);
990 			kobject_del(pdd->kobj_stats);
991 			kobject_put(pdd->kobj_stats);
992 			pdd->kobj_stats = NULL;
993 		}
994 
995 		kobject_del(p->kobj);
996 		kobject_put(p->kobj);
997 		p->kobj = NULL;
998 	}
999 
1000 	kfd_iommu_unbind_process(p);
1001 
1002 	kfd_process_free_outstanding_kfd_bos(p);
1003 
1004 	kfd_process_destroy_pdds(p);
1005 	dma_fence_put(p->ef);
1006 
1007 	kfd_event_free_process(p);
1008 
1009 	kfd_pasid_free(p->pasid);
1010 	mutex_destroy(&p->mutex);
1011 
1012 	put_task_struct(p->lead_thread);
1013 
1014 	kfree(p);
1015 }
1016 
1017 static void kfd_process_ref_release(struct kref *ref)
1018 {
1019 	struct kfd_process *p = container_of(ref, struct kfd_process, ref);
1020 
1021 	INIT_WORK(&p->release_work, kfd_process_wq_release);
1022 	queue_work(kfd_process_wq, &p->release_work);
1023 }
1024 
1025 static struct mmu_notifier *kfd_process_alloc_notifier(struct mm_struct *mm)
1026 {
1027 	int idx = srcu_read_lock(&kfd_processes_srcu);
1028 	struct kfd_process *p = find_process_by_mm(mm);
1029 
1030 	srcu_read_unlock(&kfd_processes_srcu, idx);
1031 
1032 	return p ? &p->mmu_notifier : ERR_PTR(-ESRCH);
1033 }
1034 
1035 static void kfd_process_free_notifier(struct mmu_notifier *mn)
1036 {
1037 	kfd_unref_process(container_of(mn, struct kfd_process, mmu_notifier));
1038 }
1039 
1040 static void kfd_process_notifier_release(struct mmu_notifier *mn,
1041 					struct mm_struct *mm)
1042 {
1043 	struct kfd_process *p;
1044 	int i;
1045 
1046 	/*
1047 	 * The kfd_process structure can not be free because the
1048 	 * mmu_notifier srcu is read locked
1049 	 */
1050 	p = container_of(mn, struct kfd_process, mmu_notifier);
1051 	if (WARN_ON(p->mm != mm))
1052 		return;
1053 
1054 	mutex_lock(&kfd_processes_mutex);
1055 	hash_del_rcu(&p->kfd_processes);
1056 	mutex_unlock(&kfd_processes_mutex);
1057 	synchronize_srcu(&kfd_processes_srcu);
1058 
1059 	cancel_delayed_work_sync(&p->eviction_work);
1060 	cancel_delayed_work_sync(&p->restore_work);
1061 
1062 	mutex_lock(&p->mutex);
1063 
1064 	/* Iterate over all process device data structures and if the
1065 	 * pdd is in debug mode, we should first force unregistration,
1066 	 * then we will be able to destroy the queues
1067 	 */
1068 	for (i = 0; i < p->n_pdds; i++) {
1069 		struct kfd_dev *dev = p->pdds[i]->dev;
1070 
1071 		mutex_lock(kfd_get_dbgmgr_mutex());
1072 		if (dev && dev->dbgmgr && dev->dbgmgr->pasid == p->pasid) {
1073 			if (!kfd_dbgmgr_unregister(dev->dbgmgr, p)) {
1074 				kfd_dbgmgr_destroy(dev->dbgmgr);
1075 				dev->dbgmgr = NULL;
1076 			}
1077 		}
1078 		mutex_unlock(kfd_get_dbgmgr_mutex());
1079 	}
1080 
1081 	kfd_process_dequeue_from_all_devices(p);
1082 	pqm_uninit(&p->pqm);
1083 
1084 	/* Indicate to other users that MM is no longer valid */
1085 	p->mm = NULL;
1086 	/* Signal the eviction fence after user mode queues are
1087 	 * destroyed. This allows any BOs to be freed without
1088 	 * triggering pointless evictions or waiting for fences.
1089 	 */
1090 	dma_fence_signal(p->ef);
1091 
1092 	mutex_unlock(&p->mutex);
1093 
1094 	mmu_notifier_put(&p->mmu_notifier);
1095 }
1096 
1097 static const struct mmu_notifier_ops kfd_process_mmu_notifier_ops = {
1098 	.release = kfd_process_notifier_release,
1099 	.alloc_notifier = kfd_process_alloc_notifier,
1100 	.free_notifier = kfd_process_free_notifier,
1101 };
1102 
1103 static int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep)
1104 {
1105 	unsigned long  offset;
1106 	int i;
1107 
1108 	for (i = 0; i < p->n_pdds; i++) {
1109 		struct kfd_dev *dev = p->pdds[i]->dev;
1110 		struct qcm_process_device *qpd = &p->pdds[i]->qpd;
1111 
1112 		if (!dev->cwsr_enabled || qpd->cwsr_kaddr || qpd->cwsr_base)
1113 			continue;
1114 
1115 		offset = KFD_MMAP_TYPE_RESERVED_MEM | KFD_MMAP_GPU_ID(dev->id);
1116 		qpd->tba_addr = (int64_t)vm_mmap(filep, 0,
1117 			KFD_CWSR_TBA_TMA_SIZE, PROT_READ | PROT_EXEC,
1118 			MAP_SHARED, offset);
1119 
1120 		if (IS_ERR_VALUE(qpd->tba_addr)) {
1121 			int err = qpd->tba_addr;
1122 
1123 			pr_err("Failure to set tba address. error %d.\n", err);
1124 			qpd->tba_addr = 0;
1125 			qpd->cwsr_kaddr = NULL;
1126 			return err;
1127 		}
1128 
1129 		memcpy(qpd->cwsr_kaddr, dev->cwsr_isa, dev->cwsr_isa_size);
1130 
1131 		qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET;
1132 		pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n",
1133 			qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr);
1134 	}
1135 
1136 	return 0;
1137 }
1138 
1139 static int kfd_process_device_init_cwsr_dgpu(struct kfd_process_device *pdd)
1140 {
1141 	struct kfd_dev *dev = pdd->dev;
1142 	struct qcm_process_device *qpd = &pdd->qpd;
1143 	uint32_t flags = KFD_IOC_ALLOC_MEM_FLAGS_GTT
1144 			| KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE
1145 			| KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
1146 	void *kaddr;
1147 	int ret;
1148 
1149 	if (!dev->cwsr_enabled || qpd->cwsr_kaddr || !qpd->cwsr_base)
1150 		return 0;
1151 
1152 	/* cwsr_base is only set for dGPU */
1153 	ret = kfd_process_alloc_gpuvm(pdd, qpd->cwsr_base,
1154 				      KFD_CWSR_TBA_TMA_SIZE, flags, &kaddr);
1155 	if (ret)
1156 		return ret;
1157 
1158 	qpd->cwsr_kaddr = kaddr;
1159 	qpd->tba_addr = qpd->cwsr_base;
1160 
1161 	memcpy(qpd->cwsr_kaddr, dev->cwsr_isa, dev->cwsr_isa_size);
1162 
1163 	qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET;
1164 	pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n",
1165 		 qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr);
1166 
1167 	return 0;
1168 }
1169 
1170 void kfd_process_set_trap_handler(struct qcm_process_device *qpd,
1171 				  uint64_t tba_addr,
1172 				  uint64_t tma_addr)
1173 {
1174 	if (qpd->cwsr_kaddr) {
1175 		/* KFD trap handler is bound, record as second-level TBA/TMA
1176 		 * in first-level TMA. First-level trap will jump to second.
1177 		 */
1178 		uint64_t *tma =
1179 			(uint64_t *)(qpd->cwsr_kaddr + KFD_CWSR_TMA_OFFSET);
1180 		tma[0] = tba_addr;
1181 		tma[1] = tma_addr;
1182 	} else {
1183 		/* No trap handler bound, bind as first-level TBA/TMA. */
1184 		qpd->tba_addr = tba_addr;
1185 		qpd->tma_addr = tma_addr;
1186 	}
1187 }
1188 
1189 /*
1190  * On return the kfd_process is fully operational and will be freed when the
1191  * mm is released
1192  */
1193 static struct kfd_process *create_process(const struct task_struct *thread)
1194 {
1195 	struct kfd_process *process;
1196 	struct mmu_notifier *mn;
1197 	int err = -ENOMEM;
1198 
1199 	process = kzalloc(sizeof(*process), GFP_KERNEL);
1200 	if (!process)
1201 		goto err_alloc_process;
1202 
1203 	kref_init(&process->ref);
1204 	mutex_init(&process->mutex);
1205 	process->mm = thread->mm;
1206 	process->lead_thread = thread->group_leader;
1207 	process->n_pdds = 0;
1208 	INIT_DELAYED_WORK(&process->eviction_work, evict_process_worker);
1209 	INIT_DELAYED_WORK(&process->restore_work, restore_process_worker);
1210 	process->last_restore_timestamp = get_jiffies_64();
1211 	kfd_event_init_process(process);
1212 	process->is_32bit_user_mode = in_compat_syscall();
1213 
1214 	process->pasid = kfd_pasid_alloc();
1215 	if (process->pasid == 0)
1216 		goto err_alloc_pasid;
1217 
1218 	err = pqm_init(&process->pqm, process);
1219 	if (err != 0)
1220 		goto err_process_pqm_init;
1221 
1222 	/* init process apertures*/
1223 	err = kfd_init_apertures(process);
1224 	if (err != 0)
1225 		goto err_init_apertures;
1226 
1227 	/* alloc_notifier needs to find the process in the hash table */
1228 	hash_add_rcu(kfd_processes_table, &process->kfd_processes,
1229 			(uintptr_t)process->mm);
1230 
1231 	/* MMU notifier registration must be the last call that can fail
1232 	 * because after this point we cannot unwind the process creation.
1233 	 * After this point, mmu_notifier_put will trigger the cleanup by
1234 	 * dropping the last process reference in the free_notifier.
1235 	 */
1236 	mn = mmu_notifier_get(&kfd_process_mmu_notifier_ops, process->mm);
1237 	if (IS_ERR(mn)) {
1238 		err = PTR_ERR(mn);
1239 		goto err_register_notifier;
1240 	}
1241 	BUG_ON(mn != &process->mmu_notifier);
1242 
1243 	get_task_struct(process->lead_thread);
1244 
1245 	return process;
1246 
1247 err_register_notifier:
1248 	hash_del_rcu(&process->kfd_processes);
1249 	kfd_process_free_outstanding_kfd_bos(process);
1250 	kfd_process_destroy_pdds(process);
1251 err_init_apertures:
1252 	pqm_uninit(&process->pqm);
1253 err_process_pqm_init:
1254 	kfd_pasid_free(process->pasid);
1255 err_alloc_pasid:
1256 	mutex_destroy(&process->mutex);
1257 	kfree(process);
1258 err_alloc_process:
1259 	return ERR_PTR(err);
1260 }
1261 
1262 static int init_doorbell_bitmap(struct qcm_process_device *qpd,
1263 			struct kfd_dev *dev)
1264 {
1265 	unsigned int i;
1266 	int range_start = dev->shared_resources.non_cp_doorbells_start;
1267 	int range_end = dev->shared_resources.non_cp_doorbells_end;
1268 
1269 	if (!KFD_IS_SOC15(dev->device_info->asic_family))
1270 		return 0;
1271 
1272 	qpd->doorbell_bitmap =
1273 		kzalloc(DIV_ROUND_UP(KFD_MAX_NUM_OF_QUEUES_PER_PROCESS,
1274 				     BITS_PER_BYTE), GFP_KERNEL);
1275 	if (!qpd->doorbell_bitmap)
1276 		return -ENOMEM;
1277 
1278 	/* Mask out doorbells reserved for SDMA, IH, and VCN on SOC15. */
1279 	pr_debug("reserved doorbell 0x%03x - 0x%03x\n", range_start, range_end);
1280 	pr_debug("reserved doorbell 0x%03x - 0x%03x\n",
1281 			range_start + KFD_QUEUE_DOORBELL_MIRROR_OFFSET,
1282 			range_end + KFD_QUEUE_DOORBELL_MIRROR_OFFSET);
1283 
1284 	for (i = 0; i < KFD_MAX_NUM_OF_QUEUES_PER_PROCESS / 2; i++) {
1285 		if (i >= range_start && i <= range_end) {
1286 			set_bit(i, qpd->doorbell_bitmap);
1287 			set_bit(i + KFD_QUEUE_DOORBELL_MIRROR_OFFSET,
1288 				qpd->doorbell_bitmap);
1289 		}
1290 	}
1291 
1292 	return 0;
1293 }
1294 
1295 struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
1296 							struct kfd_process *p)
1297 {
1298 	int i;
1299 
1300 	for (i = 0; i < p->n_pdds; i++)
1301 		if (p->pdds[i]->dev == dev)
1302 			return p->pdds[i];
1303 
1304 	return NULL;
1305 }
1306 
1307 struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
1308 							struct kfd_process *p)
1309 {
1310 	struct kfd_process_device *pdd = NULL;
1311 
1312 	if (WARN_ON_ONCE(p->n_pdds >= MAX_GPU_INSTANCE))
1313 		return NULL;
1314 	pdd = kzalloc(sizeof(*pdd), GFP_KERNEL);
1315 	if (!pdd)
1316 		return NULL;
1317 
1318 	if (kfd_alloc_process_doorbells(dev, &pdd->doorbell_index) < 0) {
1319 		pr_err("Failed to alloc doorbell for pdd\n");
1320 		goto err_free_pdd;
1321 	}
1322 
1323 	if (init_doorbell_bitmap(&pdd->qpd, dev)) {
1324 		pr_err("Failed to init doorbell for process\n");
1325 		goto err_free_pdd;
1326 	}
1327 
1328 	pdd->dev = dev;
1329 	INIT_LIST_HEAD(&pdd->qpd.queues_list);
1330 	INIT_LIST_HEAD(&pdd->qpd.priv_queue_list);
1331 	pdd->qpd.dqm = dev->dqm;
1332 	pdd->qpd.pqm = &p->pqm;
1333 	pdd->qpd.evicted = 0;
1334 	pdd->qpd.mapped_gws_queue = false;
1335 	pdd->process = p;
1336 	pdd->bound = PDD_UNBOUND;
1337 	pdd->already_dequeued = false;
1338 	pdd->runtime_inuse = false;
1339 	pdd->vram_usage = 0;
1340 	pdd->sdma_past_activity_counter = 0;
1341 	atomic64_set(&pdd->evict_duration_counter, 0);
1342 	p->pdds[p->n_pdds++] = pdd;
1343 
1344 	/* Init idr used for memory handle translation */
1345 	idr_init(&pdd->alloc_idr);
1346 
1347 	return pdd;
1348 
1349 err_free_pdd:
1350 	kfree(pdd);
1351 	return NULL;
1352 }
1353 
1354 /**
1355  * kfd_process_device_init_vm - Initialize a VM for a process-device
1356  *
1357  * @pdd: The process-device
1358  * @drm_file: Optional pointer to a DRM file descriptor
1359  *
1360  * If @drm_file is specified, it will be used to acquire the VM from
1361  * that file descriptor. If successful, the @pdd takes ownership of
1362  * the file descriptor.
1363  *
1364  * If @drm_file is NULL, a new VM is created.
1365  *
1366  * Returns 0 on success, -errno on failure.
1367  */
1368 int kfd_process_device_init_vm(struct kfd_process_device *pdd,
1369 			       struct file *drm_file)
1370 {
1371 	struct kfd_process *p;
1372 	struct kfd_dev *dev;
1373 	int ret;
1374 
1375 	if (!drm_file)
1376 		return -EINVAL;
1377 
1378 	if (pdd->vm)
1379 		return -EBUSY;
1380 
1381 	p = pdd->process;
1382 	dev = pdd->dev;
1383 
1384 	ret = amdgpu_amdkfd_gpuvm_acquire_process_vm(
1385 		dev->kgd, drm_file, p->pasid,
1386 		&pdd->vm, &p->kgd_process_info, &p->ef);
1387 	if (ret) {
1388 		pr_err("Failed to create process VM object\n");
1389 		return ret;
1390 	}
1391 
1392 	amdgpu_vm_set_task_info(pdd->vm);
1393 
1394 	ret = kfd_process_device_reserve_ib_mem(pdd);
1395 	if (ret)
1396 		goto err_reserve_ib_mem;
1397 	ret = kfd_process_device_init_cwsr_dgpu(pdd);
1398 	if (ret)
1399 		goto err_init_cwsr;
1400 
1401 	pdd->drm_file = drm_file;
1402 
1403 	return 0;
1404 
1405 err_init_cwsr:
1406 err_reserve_ib_mem:
1407 	kfd_process_device_free_bos(pdd);
1408 	pdd->vm = NULL;
1409 
1410 	return ret;
1411 }
1412 
1413 /*
1414  * Direct the IOMMU to bind the process (specifically the pasid->mm)
1415  * to the device.
1416  * Unbinding occurs when the process dies or the device is removed.
1417  *
1418  * Assumes that the process lock is held.
1419  */
1420 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
1421 							struct kfd_process *p)
1422 {
1423 	struct kfd_process_device *pdd;
1424 	int err;
1425 
1426 	pdd = kfd_get_process_device_data(dev, p);
1427 	if (!pdd) {
1428 		pr_err("Process device data doesn't exist\n");
1429 		return ERR_PTR(-ENOMEM);
1430 	}
1431 
1432 	if (!pdd->vm)
1433 		return ERR_PTR(-ENODEV);
1434 
1435 	/*
1436 	 * signal runtime-pm system to auto resume and prevent
1437 	 * further runtime suspend once device pdd is created until
1438 	 * pdd is destroyed.
1439 	 */
1440 	if (!pdd->runtime_inuse) {
1441 		err = pm_runtime_get_sync(dev->ddev->dev);
1442 		if (err < 0) {
1443 			pm_runtime_put_autosuspend(dev->ddev->dev);
1444 			return ERR_PTR(err);
1445 		}
1446 	}
1447 
1448 	err = kfd_iommu_bind_process_to_device(pdd);
1449 	if (err)
1450 		goto out;
1451 
1452 	/*
1453 	 * make sure that runtime_usage counter is incremented just once
1454 	 * per pdd
1455 	 */
1456 	pdd->runtime_inuse = true;
1457 
1458 	return pdd;
1459 
1460 out:
1461 	/* balance runpm reference count and exit with error */
1462 	if (!pdd->runtime_inuse) {
1463 		pm_runtime_mark_last_busy(dev->ddev->dev);
1464 		pm_runtime_put_autosuspend(dev->ddev->dev);
1465 	}
1466 
1467 	return ERR_PTR(err);
1468 }
1469 
1470 /* Create specific handle mapped to mem from process local memory idr
1471  * Assumes that the process lock is held.
1472  */
1473 int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
1474 					void *mem)
1475 {
1476 	return idr_alloc(&pdd->alloc_idr, mem, 0, 0, GFP_KERNEL);
1477 }
1478 
1479 /* Translate specific handle from process local memory idr
1480  * Assumes that the process lock is held.
1481  */
1482 void *kfd_process_device_translate_handle(struct kfd_process_device *pdd,
1483 					int handle)
1484 {
1485 	if (handle < 0)
1486 		return NULL;
1487 
1488 	return idr_find(&pdd->alloc_idr, handle);
1489 }
1490 
1491 /* Remove specific handle from process local memory idr
1492  * Assumes that the process lock is held.
1493  */
1494 void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,
1495 					int handle)
1496 {
1497 	if (handle >= 0)
1498 		idr_remove(&pdd->alloc_idr, handle);
1499 }
1500 
1501 /* This increments the process->ref counter. */
1502 struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid)
1503 {
1504 	struct kfd_process *p, *ret_p = NULL;
1505 	unsigned int temp;
1506 
1507 	int idx = srcu_read_lock(&kfd_processes_srcu);
1508 
1509 	hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1510 		if (p->pasid == pasid) {
1511 			kref_get(&p->ref);
1512 			ret_p = p;
1513 			break;
1514 		}
1515 	}
1516 
1517 	srcu_read_unlock(&kfd_processes_srcu, idx);
1518 
1519 	return ret_p;
1520 }
1521 
1522 /* This increments the process->ref counter. */
1523 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm)
1524 {
1525 	struct kfd_process *p;
1526 
1527 	int idx = srcu_read_lock(&kfd_processes_srcu);
1528 
1529 	p = find_process_by_mm(mm);
1530 	if (p)
1531 		kref_get(&p->ref);
1532 
1533 	srcu_read_unlock(&kfd_processes_srcu, idx);
1534 
1535 	return p;
1536 }
1537 
1538 /* kfd_process_evict_queues - Evict all user queues of a process
1539  *
1540  * Eviction is reference-counted per process-device. This means multiple
1541  * evictions from different sources can be nested safely.
1542  */
1543 int kfd_process_evict_queues(struct kfd_process *p)
1544 {
1545 	int r = 0;
1546 	int i;
1547 	unsigned int n_evicted = 0;
1548 
1549 	for (i = 0; i < p->n_pdds; i++) {
1550 		struct kfd_process_device *pdd = p->pdds[i];
1551 
1552 		r = pdd->dev->dqm->ops.evict_process_queues(pdd->dev->dqm,
1553 							    &pdd->qpd);
1554 		if (r) {
1555 			pr_err("Failed to evict process queues\n");
1556 			goto fail;
1557 		}
1558 		n_evicted++;
1559 	}
1560 
1561 	return r;
1562 
1563 fail:
1564 	/* To keep state consistent, roll back partial eviction by
1565 	 * restoring queues
1566 	 */
1567 	for (i = 0; i < p->n_pdds; i++) {
1568 		struct kfd_process_device *pdd = p->pdds[i];
1569 
1570 		if (n_evicted == 0)
1571 			break;
1572 		if (pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm,
1573 							      &pdd->qpd))
1574 			pr_err("Failed to restore queues\n");
1575 
1576 		n_evicted--;
1577 	}
1578 
1579 	return r;
1580 }
1581 
1582 /* kfd_process_restore_queues - Restore all user queues of a process */
1583 int kfd_process_restore_queues(struct kfd_process *p)
1584 {
1585 	int r, ret = 0;
1586 	int i;
1587 
1588 	for (i = 0; i < p->n_pdds; i++) {
1589 		struct kfd_process_device *pdd = p->pdds[i];
1590 
1591 		r = pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm,
1592 							      &pdd->qpd);
1593 		if (r) {
1594 			pr_err("Failed to restore process queues\n");
1595 			if (!ret)
1596 				ret = r;
1597 		}
1598 	}
1599 
1600 	return ret;
1601 }
1602 
1603 static void evict_process_worker(struct work_struct *work)
1604 {
1605 	int ret;
1606 	struct kfd_process *p;
1607 	struct delayed_work *dwork;
1608 
1609 	dwork = to_delayed_work(work);
1610 
1611 	/* Process termination destroys this worker thread. So during the
1612 	 * lifetime of this thread, kfd_process p will be valid
1613 	 */
1614 	p = container_of(dwork, struct kfd_process, eviction_work);
1615 	WARN_ONCE(p->last_eviction_seqno != p->ef->seqno,
1616 		  "Eviction fence mismatch\n");
1617 
1618 	/* Narrow window of overlap between restore and evict work
1619 	 * item is possible. Once amdgpu_amdkfd_gpuvm_restore_process_bos
1620 	 * unreserves KFD BOs, it is possible to evicted again. But
1621 	 * restore has few more steps of finish. So lets wait for any
1622 	 * previous restore work to complete
1623 	 */
1624 	flush_delayed_work(&p->restore_work);
1625 
1626 	pr_debug("Started evicting pasid 0x%x\n", p->pasid);
1627 	ret = kfd_process_evict_queues(p);
1628 	if (!ret) {
1629 		dma_fence_signal(p->ef);
1630 		dma_fence_put(p->ef);
1631 		p->ef = NULL;
1632 		queue_delayed_work(kfd_restore_wq, &p->restore_work,
1633 				msecs_to_jiffies(PROCESS_RESTORE_TIME_MS));
1634 
1635 		pr_debug("Finished evicting pasid 0x%x\n", p->pasid);
1636 	} else
1637 		pr_err("Failed to evict queues of pasid 0x%x\n", p->pasid);
1638 }
1639 
1640 static void restore_process_worker(struct work_struct *work)
1641 {
1642 	struct delayed_work *dwork;
1643 	struct kfd_process *p;
1644 	int ret = 0;
1645 
1646 	dwork = to_delayed_work(work);
1647 
1648 	/* Process termination destroys this worker thread. So during the
1649 	 * lifetime of this thread, kfd_process p will be valid
1650 	 */
1651 	p = container_of(dwork, struct kfd_process, restore_work);
1652 	pr_debug("Started restoring pasid 0x%x\n", p->pasid);
1653 
1654 	/* Setting last_restore_timestamp before successful restoration.
1655 	 * Otherwise this would have to be set by KGD (restore_process_bos)
1656 	 * before KFD BOs are unreserved. If not, the process can be evicted
1657 	 * again before the timestamp is set.
1658 	 * If restore fails, the timestamp will be set again in the next
1659 	 * attempt. This would mean that the minimum GPU quanta would be
1660 	 * PROCESS_ACTIVE_TIME_MS - (time to execute the following two
1661 	 * functions)
1662 	 */
1663 
1664 	p->last_restore_timestamp = get_jiffies_64();
1665 	ret = amdgpu_amdkfd_gpuvm_restore_process_bos(p->kgd_process_info,
1666 						     &p->ef);
1667 	if (ret) {
1668 		pr_debug("Failed to restore BOs of pasid 0x%x, retry after %d ms\n",
1669 			 p->pasid, PROCESS_BACK_OFF_TIME_MS);
1670 		ret = queue_delayed_work(kfd_restore_wq, &p->restore_work,
1671 				msecs_to_jiffies(PROCESS_BACK_OFF_TIME_MS));
1672 		WARN(!ret, "reschedule restore work failed\n");
1673 		return;
1674 	}
1675 
1676 	ret = kfd_process_restore_queues(p);
1677 	if (!ret)
1678 		pr_debug("Finished restoring pasid 0x%x\n", p->pasid);
1679 	else
1680 		pr_err("Failed to restore queues of pasid 0x%x\n", p->pasid);
1681 }
1682 
1683 void kfd_suspend_all_processes(void)
1684 {
1685 	struct kfd_process *p;
1686 	unsigned int temp;
1687 	int idx = srcu_read_lock(&kfd_processes_srcu);
1688 
1689 	WARN(debug_evictions, "Evicting all processes");
1690 	hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1691 		cancel_delayed_work_sync(&p->eviction_work);
1692 		cancel_delayed_work_sync(&p->restore_work);
1693 
1694 		if (kfd_process_evict_queues(p))
1695 			pr_err("Failed to suspend process 0x%x\n", p->pasid);
1696 		dma_fence_signal(p->ef);
1697 		dma_fence_put(p->ef);
1698 		p->ef = NULL;
1699 	}
1700 	srcu_read_unlock(&kfd_processes_srcu, idx);
1701 }
1702 
1703 int kfd_resume_all_processes(void)
1704 {
1705 	struct kfd_process *p;
1706 	unsigned int temp;
1707 	int ret = 0, idx = srcu_read_lock(&kfd_processes_srcu);
1708 
1709 	hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1710 		if (!queue_delayed_work(kfd_restore_wq, &p->restore_work, 0)) {
1711 			pr_err("Restore process %d failed during resume\n",
1712 			       p->pasid);
1713 			ret = -EFAULT;
1714 		}
1715 	}
1716 	srcu_read_unlock(&kfd_processes_srcu, idx);
1717 	return ret;
1718 }
1719 
1720 int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process,
1721 			  struct vm_area_struct *vma)
1722 {
1723 	struct kfd_process_device *pdd;
1724 	struct qcm_process_device *qpd;
1725 
1726 	if ((vma->vm_end - vma->vm_start) != KFD_CWSR_TBA_TMA_SIZE) {
1727 		pr_err("Incorrect CWSR mapping size.\n");
1728 		return -EINVAL;
1729 	}
1730 
1731 	pdd = kfd_get_process_device_data(dev, process);
1732 	if (!pdd)
1733 		return -EINVAL;
1734 	qpd = &pdd->qpd;
1735 
1736 	qpd->cwsr_kaddr = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
1737 					get_order(KFD_CWSR_TBA_TMA_SIZE));
1738 	if (!qpd->cwsr_kaddr) {
1739 		pr_err("Error allocating per process CWSR buffer.\n");
1740 		return -ENOMEM;
1741 	}
1742 
1743 	vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND
1744 		| VM_NORESERVE | VM_DONTDUMP | VM_PFNMAP;
1745 	/* Mapping pages to user process */
1746 	return remap_pfn_range(vma, vma->vm_start,
1747 			       PFN_DOWN(__pa(qpd->cwsr_kaddr)),
1748 			       KFD_CWSR_TBA_TMA_SIZE, vma->vm_page_prot);
1749 }
1750 
1751 void kfd_flush_tlb(struct kfd_process_device *pdd)
1752 {
1753 	struct kfd_dev *dev = pdd->dev;
1754 
1755 	if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
1756 		/* Nothing to flush until a VMID is assigned, which
1757 		 * only happens when the first queue is created.
1758 		 */
1759 		if (pdd->qpd.vmid)
1760 			amdgpu_amdkfd_flush_gpu_tlb_vmid(dev->kgd,
1761 							pdd->qpd.vmid);
1762 	} else {
1763 		amdgpu_amdkfd_flush_gpu_tlb_pasid(dev->kgd,
1764 						pdd->process->pasid);
1765 	}
1766 }
1767 
1768 #if defined(CONFIG_DEBUG_FS)
1769 
1770 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data)
1771 {
1772 	struct kfd_process *p;
1773 	unsigned int temp;
1774 	int r = 0;
1775 
1776 	int idx = srcu_read_lock(&kfd_processes_srcu);
1777 
1778 	hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1779 		seq_printf(m, "Process %d PASID 0x%x:\n",
1780 			   p->lead_thread->tgid, p->pasid);
1781 
1782 		mutex_lock(&p->mutex);
1783 		r = pqm_debugfs_mqds(m, &p->pqm);
1784 		mutex_unlock(&p->mutex);
1785 
1786 		if (r)
1787 			break;
1788 	}
1789 
1790 	srcu_read_unlock(&kfd_processes_srcu, idx);
1791 
1792 	return r;
1793 }
1794 
1795 #endif
1796 
1797