Loading...
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
40struct 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 */
52DEFINE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
53static DEFINE_MUTEX(kfd_processes_mutex);
54
55DEFINE_SRCU(kfd_processes_srcu);
56
57/* For process termination handling */
58static 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 */
66static struct workqueue_struct *kfd_restore_wq;
67
68static struct kfd_process *find_process(const struct task_struct *thread,
69 bool ref);
70static void kfd_process_ref_release(struct kref *ref);
71static struct kfd_process *create_process(const struct task_struct *thread);
72static int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep);
73
74static void evict_process_worker(struct work_struct *work);
75static void restore_process_worker(struct work_struct *work);
76
77static void kfd_process_device_destroy_cwsr_dgpu(struct kfd_process_device *pdd);
78
79struct kfd_procfs_tree {
80 struct kobject *kobj;
81};
82
83static struct kfd_procfs_tree procfs;
84
85/*
86 * Structure for SDMA activity tracking
87 */
88struct 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
94struct 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
101static 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(&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(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(&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(&sdma_q->list, &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(&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(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(sdma_q->rptr, &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(&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(&sdma_q->list);
227 kfree(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(&sdma_q->list);
243 kfree(sdma_q);
244 }
245
246 return;
247
248cleanup:
249 list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
250 list_del(&sdma_q->list);
251 kfree(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 */
267static 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(buffer, PAGE_SIZE, "%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(buffer, PAGE_SIZE, "%d\n", cu_cnt);
298}
299
300static 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(buffer, PAGE_SIZE, "%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(buffer, PAGE_SIZE, "%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(&sdma_activity_work_handler.sdma_activity_work);
324
325 flush_work(&sdma_activity_work_handler.sdma_activity_work);
326
327 return snprintf(buffer, PAGE_SIZE, "%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
338static void kfd_procfs_kobj_release(struct kobject *kobj)
339{
340 kfree(kobj);
341}
342
343static const struct sysfs_ops kfd_procfs_ops = {
344 .show = kfd_procfs_show,
345};
346
347static struct kobj_type procfs_type = {
348 .release = kfd_procfs_kobj_release,
349 .sysfs_ops = &kfd_procfs_ops,
350};
351
352void 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(procfs.kobj, &procfs_type,
361 &kfd_device->kobj, "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
369void kfd_procfs_shutdown(void)
370{
371 if (procfs.kobj) {
372 kobject_del(procfs.kobj);
373 kobject_put(procfs.kobj);
374 procfs.kobj = NULL;
375 }
376}
377
378static 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(buffer, PAGE_SIZE, "%llu",
385 q->properties.queue_size);
386 else if (!strcmp(attr->name, "type"))
387 return snprintf(buffer, PAGE_SIZE, "%d", q->properties.type);
388 else if (!strcmp(attr->name, "gpuid"))
389 return snprintf(buffer, PAGE_SIZE, "%u", q->device->id);
390 else
391 pr_err("Invalid attribute");
392
393 return 0;
394}
395
396static 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
422static 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, "%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, "%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, "%llu\n", READ_ONCE(pdd->page_out));
441 }
442 return 0;
443}
444
445static struct attribute attr_queue_size = {
446 .name = "size",
447 .mode = KFD_SYSFS_FILE_MODE
448};
449
450static struct attribute attr_queue_type = {
451 .name = "type",
452 .mode = KFD_SYSFS_FILE_MODE
453};
454
455static struct attribute attr_queue_gpuid = {
456 .name = "gpuid",
457 .mode = KFD_SYSFS_FILE_MODE
458};
459
460static struct attribute *procfs_queue_attrs[] = {
461 &attr_queue_size,
462 &attr_queue_type,
463 &attr_queue_gpuid,
464 NULL
465};
466ATTRIBUTE_GROUPS(procfs_queue);
467
468static const struct sysfs_ops procfs_queue_ops = {
469 .show = kfd_procfs_queue_show,
470};
471
472static struct kobj_type procfs_queue_type = {
473 .sysfs_ops = &procfs_queue_ops,
474 .default_groups = procfs_queue_groups,
475};
476
477static const struct sysfs_ops procfs_stats_ops = {
478 .show = kfd_procfs_stats_show,
479};
480
481static struct kobj_type procfs_stats_type = {
482 .sysfs_ops = &procfs_stats_ops,
483 .release = kfd_procfs_kobj_release,
484};
485
486static const struct sysfs_ops sysfs_counters_ops = {
487 .show = kfd_sysfs_counters_show,
488};
489
490static struct kobj_type sysfs_counters_type = {
491 .sysfs_ops = &sysfs_counters_ops,
492 .release = kfd_procfs_kobj_release,
493};
494
495int 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(&q->kobj, &procfs_queue_type,
508 proc->kobj_queues, "%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(&q->kobj);
513 return ret;
514 }
515
516 return 0;
517}
518
519static 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
536static 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(stats_dir_filename, MAX_SYSFS_FILENAME_LEN,
555 "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(pdd->kobj_stats,
561 &procfs_stats_type,
562 p->kobj,
563 stats_dir_filename);
564
565 if (ret) {
566 pr_warn("Creating KFD proc/stats_%s folder failed",
567 stats_dir_filename);
568 kobject_put(pdd->kobj_stats);
569 pdd->kobj_stats = NULL;
570 return;
571 }
572
573 kfd_sysfs_create_file(pdd->kobj_stats, &pdd->attr_evict,
574 "evicted_ms");
575 /* Add sysfs file to report compute unit occupancy */
576 if (pdd->dev->kfd2kgd->get_cu_occupancy)
577 kfd_sysfs_create_file(pdd->kobj_stats,
578 &pdd->attr_cu_occupancy,
579 "cu_occupancy");
580 }
581}
582
583static 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(counters_dir_filename, MAX_SYSFS_FILENAME_LEN,
604 "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_counters, &sysfs_counters_type,
610 p->kobj, counters_dir_filename);
611 if (ret) {
612 pr_warn("Creating KFD proc/%s folder failed",
613 counters_dir_filename);
614 kobject_put(kobj_counters);
615 return;
616 }
617
618 pdd->kobj_counters = kobj_counters;
619 kfd_sysfs_create_file(kobj_counters, &pdd->attr_faults,
620 "faults");
621 kfd_sysfs_create_file(kobj_counters, &pdd->attr_page_in,
622 "page_in");
623 kfd_sysfs_create_file(kobj_counters, &pdd->attr_page_out,
624 "page_out");
625 }
626}
627
628static 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(pdd->vram_filename, MAX_SYSFS_FILENAME_LEN, "vram_%u",
644 pdd->dev->id);
645 kfd_sysfs_create_file(p->kobj, &pdd->attr_vram,
646 pdd->vram_filename);
647
648 snprintf(pdd->sdma_filename, MAX_SYSFS_FILENAME_LEN, "sdma_%u",
649 pdd->dev->id);
650 kfd_sysfs_create_file(p->kobj, &pdd->attr_sdma,
651 pdd->sdma_filename);
652 }
653}
654
655void kfd_procfs_del_queue(struct queue *q)
656{
657 if (!q)
658 return;
659
660 kobject_del(&q->kobj);
661 kobject_put(&q->kobj);
662}
663
664int kfd_process_create_wq(void)
665{
666 if (!kfd_process_wq)
667 kfd_process_wq = alloc_workqueue("kfd_process_wq", 0, 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
679void kfd_process_destroy_wq(void)
680{
681 if (kfd_process_wq) {
682 destroy_workqueue(kfd_process_wq);
683 kfd_process_wq = NULL;
684 }
685 if (kfd_restore_wq) {
686 destroy_workqueue(kfd_restore_wq);
687 kfd_restore_wq = NULL;
688 }
689}
690
691static 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 */
712static 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
747sync_memory_failed:
748 amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(kdev->adev, *mem, pdd->drm_priv);
749
750err_map_mem:
751 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(kdev->adev, *mem, pdd->drm_priv,
752 NULL);
753err_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 */
765static 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, qpd->ib_base, PAGE_SIZE, flags,
781 &mem, &kaddr);
782 if (ret)
783 return ret;
784
785 qpd->ib_mem = mem;
786 qpd->ib_kaddr = kaddr;
787
788 return 0;
789}
790
791static 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(qpd->ib_mem, pdd, &qpd->ib_kaddr);
799}
800
801struct 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(&kfd_processes_mutex);
820
821 /* A prior open of /dev/kfd could have already created the process. */
822 process = find_process(thread, false);
823 if (process) {
824 pr_debug("Process already found\n");
825 } else {
826 process = create_process(thread);
827 if (IS_ERR(process))
828 goto out;
829
830 ret = kfd_process_init_cwsr_apu(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(process->kobj, &procfs_type,
843 procfs.kobj, "%d",
844 (int)process->lead_thread->pid);
845 if (ret) {
846 pr_warn("Creating procfs pid directory failed");
847 kobject_put(process->kobj);
848 goto out;
849 }
850
851 kfd_sysfs_create_file(process->kobj, &process->attr_pasid,
852 "pasid");
853
854 process->kobj_queues = kobject_create_and_add("queues",
855 process->kobj);
856 if (!process->kobj_queues)
857 pr_warn("Creating KFD proc/queues folder failed");
858
859 kfd_procfs_add_sysfs_stats(process);
860 kfd_procfs_add_sysfs_files(process);
861 kfd_procfs_add_sysfs_counters(process);
862 }
863out:
864 if (!IS_ERR(process))
865 kref_get(&process->ref);
866 mutex_unlock(&kfd_processes_mutex);
867
868 return process;
869
870out_destroy:
871 hash_del_rcu(&process->kfd_processes);
872 mutex_unlock(&kfd_processes_mutex);
873 synchronize_srcu(&kfd_processes_srcu);
874 /* kfd_process_free_notifier will trigger the cleanup */
875 mmu_notifier_put(&process->mmu_notifier);
876 return ERR_PTR(ret);
877}
878
879struct 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, false);
891 if (!process)
892 return ERR_PTR(-EINVAL);
893
894 return process;
895}
896
897static 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
909static 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(&kfd_processes_srcu);
916 p = find_process_by_mm(thread->mm);
917 if (p && ref)
918 kref_get(&p->ref);
919 srcu_read_unlock(&kfd_processes_srcu, idx);
920
921 return p;
922}
923
924void kfd_unref_process(struct kfd_process *p)
925{
926 kref_put(&p->ref, kfd_process_ref_release);
927}
928
929/* This increments the process->ref counter. */
930struct 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(task);
938 } else {
939 task = get_pid_task(pid, PIDTYPE_PID);
940 }
941
942 if (task) {
943 p = find_process(task, true);
944 put_task_struct(task);
945 }
946
947 return p;
948}
949
950static 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, 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 */
982static 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(&p->mutex);
993
994 pdd = kfd_get_process_device_data(kdev, p);
995 if (!pdd)
996 goto out;
997
998 mem = kfd_process_device_translate_handle(
999 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
1005out:
1006 mutex_unlock(&p->mutex);
1007}
1008
1009static 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(p->pdds[i]);
1015}
1016
1017static 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((unsigned long)pdd->qpd.cwsr_kaddr,
1038 get_order(KFD_CWSR_TBA_TMA_SIZE));
1039
1040 bitmap_free(pdd->qpd.doorbell_bitmap);
1041 idr_destroy(&pdd->alloc_idr);
1042
1043 kfd_free_process_doorbells(pdd->dev, 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(adev_to_drm(pdd->dev->adev)->dev);
1054 pm_runtime_put_autosuspend(adev_to_drm(pdd->dev->adev)->dev);
1055 pdd->runtime_inuse = false;
1056 }
1057
1058 kfree(pdd);
1059 p->pdds[i] = NULL;
1060 }
1061 p->n_pdds = 0;
1062}
1063
1064static 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(p->kobj, &p->attr_pasid);
1073 kobject_del(p->kobj_queues);
1074 kobject_put(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(p->kobj, &pdd->attr_vram);
1081 sysfs_remove_file(p->kobj, &pdd->attr_sdma);
1082
1083 sysfs_remove_file(pdd->kobj_stats, &pdd->attr_evict);
1084 if (pdd->dev->kfd2kgd->get_cu_occupancy)
1085 sysfs_remove_file(pdd->kobj_stats,
1086 &pdd->attr_cu_occupancy);
1087 kobject_del(pdd->kobj_stats);
1088 kobject_put(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(pdd->kobj_counters, &pdd->attr_faults);
1096 sysfs_remove_file(pdd->kobj_counters, &pdd->attr_page_in);
1097 sysfs_remove_file(pdd->kobj_counters, &pdd->attr_page_out);
1098 kobject_del(pdd->kobj_counters);
1099 kobject_put(pdd->kobj_counters);
1100 pdd->kobj_counters = NULL;
1101 }
1102
1103 kobject_del(p->kobj);
1104 kobject_put(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 */
1113static 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(&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(p->pasid);
1140 mutex_destroy(&p->mutex);
1141
1142 put_task_struct(p->lead_thread);
1143
1144 kfree(p);
1145}
1146
1147static 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(kfd_process_wq, &p->release_work);
1153}
1154
1155static struct mmu_notifier *kfd_process_alloc_notifier(struct mm_struct *mm)
1156{
1157 int idx = srcu_read_lock(&kfd_processes_srcu);
1158 struct kfd_process *p = find_process_by_mm(mm);
1159
1160 srcu_read_unlock(&kfd_processes_srcu, idx);
1161
1162 return p ? &p->mmu_notifier : ERR_PTR(-ESRCH);
1163}
1164
1165static void kfd_process_free_notifier(struct mmu_notifier *mn)
1166{
1167 kfd_unref_process(container_of(mn, struct kfd_process, mmu_notifier));
1168}
1169
1170static void kfd_process_notifier_release(struct mmu_notifier *mn,
1171 struct mm_struct *mm)
1172{
1173 struct kfd_process *p;
1174
1175 /*
1176 * The kfd_process structure can not be free because the
1177 * mmu_notifier srcu is read locked
1178 */
1179 p = container_of(mn, struct kfd_process, mmu_notifier);
1180 if (WARN_ON(p->mm != mm))
1181 return;
1182
1183 mutex_lock(&kfd_processes_mutex);
1184 hash_del_rcu(&p->kfd_processes);
1185 mutex_unlock(&kfd_processes_mutex);
1186 synchronize_srcu(&kfd_processes_srcu);
1187
1188 cancel_delayed_work_sync(&p->eviction_work);
1189 cancel_delayed_work_sync(&p->restore_work);
1190
1191 /* Indicate to other users that MM is no longer valid */
1192 p->mm = NULL;
1193
1194 mmu_notifier_put(&p->mmu_notifier);
1195}
1196
1197static const struct mmu_notifier_ops kfd_process_mmu_notifier_ops = {
1198 .release = kfd_process_notifier_release,
1199 .alloc_notifier = kfd_process_alloc_notifier,
1200 .free_notifier = kfd_process_free_notifier,
1201};
1202
1203static int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep)
1204{
1205 unsigned long offset;
1206 int i;
1207
1208 for (i = 0; i < p->n_pdds; i++) {
1209 struct kfd_dev *dev = p->pdds[i]->dev;
1210 struct qcm_process_device *qpd = &p->pdds[i]->qpd;
1211
1212 if (!dev->cwsr_enabled || qpd->cwsr_kaddr || qpd->cwsr_base)
1213 continue;
1214
1215 offset = KFD_MMAP_TYPE_RESERVED_MEM | KFD_MMAP_GPU_ID(dev->id);
1216 qpd->tba_addr = (int64_t)vm_mmap(filep, 0,
1217 KFD_CWSR_TBA_TMA_SIZE, PROT_READ | PROT_EXEC,
1218 MAP_SHARED, offset);
1219
1220 if (IS_ERR_VALUE(qpd->tba_addr)) {
1221 int err = qpd->tba_addr;
1222
1223 pr_err("Failure to set tba address. error %d.\n", err);
1224 qpd->tba_addr = 0;
1225 qpd->cwsr_kaddr = NULL;
1226 return err;
1227 }
1228
1229 memcpy(qpd->cwsr_kaddr, dev->cwsr_isa, dev->cwsr_isa_size);
1230
1231 qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET;
1232 pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n",
1233 qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr);
1234 }
1235
1236 return 0;
1237}
1238
1239static int kfd_process_device_init_cwsr_dgpu(struct kfd_process_device *pdd)
1240{
1241 struct kfd_dev *dev = pdd->dev;
1242 struct qcm_process_device *qpd = &pdd->qpd;
1243 uint32_t flags = KFD_IOC_ALLOC_MEM_FLAGS_GTT
1244 | KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE
1245 | KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
1246 struct kgd_mem *mem;
1247 void *kaddr;
1248 int ret;
1249
1250 if (!dev->cwsr_enabled || qpd->cwsr_kaddr || !qpd->cwsr_base)
1251 return 0;
1252
1253 /* cwsr_base is only set for dGPU */
1254 ret = kfd_process_alloc_gpuvm(pdd, qpd->cwsr_base,
1255 KFD_CWSR_TBA_TMA_SIZE, flags, &mem, &kaddr);
1256 if (ret)
1257 return ret;
1258
1259 qpd->cwsr_mem = mem;
1260 qpd->cwsr_kaddr = kaddr;
1261 qpd->tba_addr = qpd->cwsr_base;
1262
1263 memcpy(qpd->cwsr_kaddr, dev->cwsr_isa, dev->cwsr_isa_size);
1264
1265 qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET;
1266 pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n",
1267 qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr);
1268
1269 return 0;
1270}
1271
1272static void kfd_process_device_destroy_cwsr_dgpu(struct kfd_process_device *pdd)
1273{
1274 struct kfd_dev *dev = pdd->dev;
1275 struct qcm_process_device *qpd = &pdd->qpd;
1276
1277 if (!dev->cwsr_enabled || !qpd->cwsr_kaddr || !qpd->cwsr_base)
1278 return;
1279
1280 kfd_process_free_gpuvm(qpd->cwsr_mem, pdd, &qpd->cwsr_kaddr);
1281}
1282
1283void kfd_process_set_trap_handler(struct qcm_process_device *qpd,
1284 uint64_t tba_addr,
1285 uint64_t tma_addr)
1286{
1287 if (qpd->cwsr_kaddr) {
1288 /* KFD trap handler is bound, record as second-level TBA/TMA
1289 * in first-level TMA. First-level trap will jump to second.
1290 */
1291 uint64_t *tma =
1292 (uint64_t *)(qpd->cwsr_kaddr + KFD_CWSR_TMA_OFFSET);
1293 tma[0] = tba_addr;
1294 tma[1] = tma_addr;
1295 } else {
1296 /* No trap handler bound, bind as first-level TBA/TMA. */
1297 qpd->tba_addr = tba_addr;
1298 qpd->tma_addr = tma_addr;
1299 }
1300}
1301
1302bool kfd_process_xnack_mode(struct kfd_process *p, bool supported)
1303{
1304 int i;
1305
1306 /* On most GFXv9 GPUs, the retry mode in the SQ must match the
1307 * boot time retry setting. Mixing processes with different
1308 * XNACK/retry settings can hang the GPU.
1309 *
1310 * Different GPUs can have different noretry settings depending
1311 * on HW bugs or limitations. We need to find at least one
1312 * XNACK mode for this process that's compatible with all GPUs.
1313 * Fortunately GPUs with retry enabled (noretry=0) can run code
1314 * built for XNACK-off. On GFXv9 it may perform slower.
1315 *
1316 * Therefore applications built for XNACK-off can always be
1317 * supported and will be our fallback if any GPU does not
1318 * support retry.
1319 */
1320 for (i = 0; i < p->n_pdds; i++) {
1321 struct kfd_dev *dev = p->pdds[i]->dev;
1322
1323 /* Only consider GFXv9 and higher GPUs. Older GPUs don't
1324 * support the SVM APIs and don't need to be considered
1325 * for the XNACK mode selection.
1326 */
1327 if (!KFD_IS_SOC15(dev))
1328 continue;
1329 /* Aldebaran can always support XNACK because it can support
1330 * per-process XNACK mode selection. But let the dev->noretry
1331 * setting still influence the default XNACK mode.
1332 */
1333 if (supported && KFD_GC_VERSION(dev) == IP_VERSION(9, 4, 2))
1334 continue;
1335
1336 /* GFXv10 and later GPUs do not support shader preemption
1337 * during page faults. This can lead to poor QoS for queue
1338 * management and memory-manager-related preemptions or
1339 * even deadlocks.
1340 */
1341 if (KFD_GC_VERSION(dev) >= IP_VERSION(10, 1, 1))
1342 return false;
1343
1344 if (dev->noretry)
1345 return false;
1346 }
1347
1348 return true;
1349}
1350
1351/*
1352 * On return the kfd_process is fully operational and will be freed when the
1353 * mm is released
1354 */
1355static struct kfd_process *create_process(const struct task_struct *thread)
1356{
1357 struct kfd_process *process;
1358 struct mmu_notifier *mn;
1359 int err = -ENOMEM;
1360
1361 process = kzalloc(sizeof(*process), GFP_KERNEL);
1362 if (!process)
1363 goto err_alloc_process;
1364
1365 kref_init(&process->ref);
1366 mutex_init(&process->mutex);
1367 process->mm = thread->mm;
1368 process->lead_thread = thread->group_leader;
1369 process->n_pdds = 0;
1370 process->queues_paused = false;
1371 INIT_DELAYED_WORK(&process->eviction_work, evict_process_worker);
1372 INIT_DELAYED_WORK(&process->restore_work, restore_process_worker);
1373 process->last_restore_timestamp = get_jiffies_64();
1374 err = kfd_event_init_process(process);
1375 if (err)
1376 goto err_event_init;
1377 process->is_32bit_user_mode = in_compat_syscall();
1378
1379 process->pasid = kfd_pasid_alloc();
1380 if (process->pasid == 0) {
1381 err = -ENOSPC;
1382 goto err_alloc_pasid;
1383 }
1384
1385 err = pqm_init(&process->pqm, process);
1386 if (err != 0)
1387 goto err_process_pqm_init;
1388
1389 /* init process apertures*/
1390 err = kfd_init_apertures(process);
1391 if (err != 0)
1392 goto err_init_apertures;
1393
1394 /* Check XNACK support after PDDs are created in kfd_init_apertures */
1395 process->xnack_enabled = kfd_process_xnack_mode(process, false);
1396
1397 err = svm_range_list_init(process);
1398 if (err)
1399 goto err_init_svm_range_list;
1400
1401 /* alloc_notifier needs to find the process in the hash table */
1402 hash_add_rcu(kfd_processes_table, &process->kfd_processes,
1403 (uintptr_t)process->mm);
1404
1405 /* Avoid free_notifier to start kfd_process_wq_release if
1406 * mmu_notifier_get failed because of pending signal.
1407 */
1408 kref_get(&process->ref);
1409
1410 /* MMU notifier registration must be the last call that can fail
1411 * because after this point we cannot unwind the process creation.
1412 * After this point, mmu_notifier_put will trigger the cleanup by
1413 * dropping the last process reference in the free_notifier.
1414 */
1415 mn = mmu_notifier_get(&kfd_process_mmu_notifier_ops, process->mm);
1416 if (IS_ERR(mn)) {
1417 err = PTR_ERR(mn);
1418 goto err_register_notifier;
1419 }
1420 BUG_ON(mn != &process->mmu_notifier);
1421
1422 kfd_unref_process(process);
1423 get_task_struct(process->lead_thread);
1424
1425 return process;
1426
1427err_register_notifier:
1428 hash_del_rcu(&process->kfd_processes);
1429 svm_range_list_fini(process);
1430err_init_svm_range_list:
1431 kfd_process_free_outstanding_kfd_bos(process);
1432 kfd_process_destroy_pdds(process);
1433err_init_apertures:
1434 pqm_uninit(&process->pqm);
1435err_process_pqm_init:
1436 kfd_pasid_free(process->pasid);
1437err_alloc_pasid:
1438 kfd_event_free_process(process);
1439err_event_init:
1440 mutex_destroy(&process->mutex);
1441 kfree(process);
1442err_alloc_process:
1443 return ERR_PTR(err);
1444}
1445
1446static int init_doorbell_bitmap(struct qcm_process_device *qpd,
1447 struct kfd_dev *dev)
1448{
1449 unsigned int i;
1450 int range_start = dev->shared_resources.non_cp_doorbells_start;
1451 int range_end = dev->shared_resources.non_cp_doorbells_end;
1452
1453 if (!KFD_IS_SOC15(dev))
1454 return 0;
1455
1456 qpd->doorbell_bitmap = bitmap_zalloc(KFD_MAX_NUM_OF_QUEUES_PER_PROCESS,
1457 GFP_KERNEL);
1458 if (!qpd->doorbell_bitmap)
1459 return -ENOMEM;
1460
1461 /* Mask out doorbells reserved for SDMA, IH, and VCN on SOC15. */
1462 pr_debug("reserved doorbell 0x%03x - 0x%03x\n", range_start, range_end);
1463 pr_debug("reserved doorbell 0x%03x - 0x%03x\n",
1464 range_start + KFD_QUEUE_DOORBELL_MIRROR_OFFSET,
1465 range_end + KFD_QUEUE_DOORBELL_MIRROR_OFFSET);
1466
1467 for (i = 0; i < KFD_MAX_NUM_OF_QUEUES_PER_PROCESS / 2; i++) {
1468 if (i >= range_start && i <= range_end) {
1469 __set_bit(i, qpd->doorbell_bitmap);
1470 __set_bit(i + KFD_QUEUE_DOORBELL_MIRROR_OFFSET,
1471 qpd->doorbell_bitmap);
1472 }
1473 }
1474
1475 return 0;
1476}
1477
1478struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
1479 struct kfd_process *p)
1480{
1481 int i;
1482
1483 for (i = 0; i < p->n_pdds; i++)
1484 if (p->pdds[i]->dev == dev)
1485 return p->pdds[i];
1486
1487 return NULL;
1488}
1489
1490struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
1491 struct kfd_process *p)
1492{
1493 struct kfd_process_device *pdd = NULL;
1494 int retval = 0;
1495
1496 if (WARN_ON_ONCE(p->n_pdds >= MAX_GPU_INSTANCE))
1497 return NULL;
1498 pdd = kzalloc(sizeof(*pdd), GFP_KERNEL);
1499 if (!pdd)
1500 return NULL;
1501
1502 if (init_doorbell_bitmap(&pdd->qpd, dev)) {
1503 pr_err("Failed to init doorbell for process\n");
1504 goto err_free_pdd;
1505 }
1506
1507 pdd->dev = dev;
1508 INIT_LIST_HEAD(&pdd->qpd.queues_list);
1509 INIT_LIST_HEAD(&pdd->qpd.priv_queue_list);
1510 pdd->qpd.dqm = dev->dqm;
1511 pdd->qpd.pqm = &p->pqm;
1512 pdd->qpd.evicted = 0;
1513 pdd->qpd.mapped_gws_queue = false;
1514 pdd->process = p;
1515 pdd->bound = PDD_UNBOUND;
1516 pdd->already_dequeued = false;
1517 pdd->runtime_inuse = false;
1518 pdd->vram_usage = 0;
1519 pdd->sdma_past_activity_counter = 0;
1520 pdd->user_gpu_id = dev->id;
1521 atomic64_set(&pdd->evict_duration_counter, 0);
1522
1523 if (dev->shared_resources.enable_mes) {
1524 retval = amdgpu_amdkfd_alloc_gtt_mem(dev->adev,
1525 AMDGPU_MES_PROC_CTX_SIZE,
1526 &pdd->proc_ctx_bo,
1527 &pdd->proc_ctx_gpu_addr,
1528 &pdd->proc_ctx_cpu_ptr,
1529 false);
1530 if (retval) {
1531 pr_err("failed to allocate process context bo\n");
1532 goto err_free_pdd;
1533 }
1534 memset(pdd->proc_ctx_cpu_ptr, 0, AMDGPU_MES_PROC_CTX_SIZE);
1535 }
1536
1537 p->pdds[p->n_pdds++] = pdd;
1538
1539 /* Init idr used for memory handle translation */
1540 idr_init(&pdd->alloc_idr);
1541
1542 return pdd;
1543
1544err_free_pdd:
1545 kfree(pdd);
1546 return NULL;
1547}
1548
1549/**
1550 * kfd_process_device_init_vm - Initialize a VM for a process-device
1551 *
1552 * @pdd: The process-device
1553 * @drm_file: Optional pointer to a DRM file descriptor
1554 *
1555 * If @drm_file is specified, it will be used to acquire the VM from
1556 * that file descriptor. If successful, the @pdd takes ownership of
1557 * the file descriptor.
1558 *
1559 * If @drm_file is NULL, a new VM is created.
1560 *
1561 * Returns 0 on success, -errno on failure.
1562 */
1563int kfd_process_device_init_vm(struct kfd_process_device *pdd,
1564 struct file *drm_file)
1565{
1566 struct kfd_process *p;
1567 struct kfd_dev *dev;
1568 int ret;
1569
1570 if (!drm_file)
1571 return -EINVAL;
1572
1573 if (pdd->drm_priv)
1574 return -EBUSY;
1575
1576 p = pdd->process;
1577 dev = pdd->dev;
1578
1579 ret = amdgpu_amdkfd_gpuvm_acquire_process_vm(dev->adev, drm_file,
1580 &p->kgd_process_info,
1581 &p->ef);
1582 if (ret) {
1583 pr_err("Failed to create process VM object\n");
1584 return ret;
1585 }
1586 pdd->drm_priv = drm_file->private_data;
1587 atomic64_set(&pdd->tlb_seq, 0);
1588
1589 ret = kfd_process_device_reserve_ib_mem(pdd);
1590 if (ret)
1591 goto err_reserve_ib_mem;
1592 ret = kfd_process_device_init_cwsr_dgpu(pdd);
1593 if (ret)
1594 goto err_init_cwsr;
1595
1596 ret = amdgpu_amdkfd_gpuvm_set_vm_pasid(dev->adev, drm_file, p->pasid);
1597 if (ret)
1598 goto err_set_pasid;
1599
1600 pdd->drm_file = drm_file;
1601
1602 return 0;
1603
1604err_set_pasid:
1605 kfd_process_device_destroy_cwsr_dgpu(pdd);
1606err_init_cwsr:
1607 kfd_process_device_destroy_ib_mem(pdd);
1608err_reserve_ib_mem:
1609 pdd->drm_priv = NULL;
1610
1611 return ret;
1612}
1613
1614/*
1615 * Direct the IOMMU to bind the process (specifically the pasid->mm)
1616 * to the device.
1617 * Unbinding occurs when the process dies or the device is removed.
1618 *
1619 * Assumes that the process lock is held.
1620 */
1621struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
1622 struct kfd_process *p)
1623{
1624 struct kfd_process_device *pdd;
1625 int err;
1626
1627 pdd = kfd_get_process_device_data(dev, p);
1628 if (!pdd) {
1629 pr_err("Process device data doesn't exist\n");
1630 return ERR_PTR(-ENOMEM);
1631 }
1632
1633 if (!pdd->drm_priv)
1634 return ERR_PTR(-ENODEV);
1635
1636 /*
1637 * signal runtime-pm system to auto resume and prevent
1638 * further runtime suspend once device pdd is created until
1639 * pdd is destroyed.
1640 */
1641 if (!pdd->runtime_inuse) {
1642 err = pm_runtime_get_sync(adev_to_drm(dev->adev)->dev);
1643 if (err < 0) {
1644 pm_runtime_put_autosuspend(adev_to_drm(dev->adev)->dev);
1645 return ERR_PTR(err);
1646 }
1647 }
1648
1649 err = kfd_iommu_bind_process_to_device(pdd);
1650 if (err)
1651 goto out;
1652
1653 /*
1654 * make sure that runtime_usage counter is incremented just once
1655 * per pdd
1656 */
1657 pdd->runtime_inuse = true;
1658
1659 return pdd;
1660
1661out:
1662 /* balance runpm reference count and exit with error */
1663 if (!pdd->runtime_inuse) {
1664 pm_runtime_mark_last_busy(adev_to_drm(dev->adev)->dev);
1665 pm_runtime_put_autosuspend(adev_to_drm(dev->adev)->dev);
1666 }
1667
1668 return ERR_PTR(err);
1669}
1670
1671/* Create specific handle mapped to mem from process local memory idr
1672 * Assumes that the process lock is held.
1673 */
1674int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
1675 void *mem)
1676{
1677 return idr_alloc(&pdd->alloc_idr, mem, 0, 0, GFP_KERNEL);
1678}
1679
1680/* Translate specific handle from process local memory idr
1681 * Assumes that the process lock is held.
1682 */
1683void *kfd_process_device_translate_handle(struct kfd_process_device *pdd,
1684 int handle)
1685{
1686 if (handle < 0)
1687 return NULL;
1688
1689 return idr_find(&pdd->alloc_idr, handle);
1690}
1691
1692/* Remove specific handle from process local memory idr
1693 * Assumes that the process lock is held.
1694 */
1695void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,
1696 int handle)
1697{
1698 if (handle >= 0)
1699 idr_remove(&pdd->alloc_idr, handle);
1700}
1701
1702/* This increments the process->ref counter. */
1703struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid)
1704{
1705 struct kfd_process *p, *ret_p = NULL;
1706 unsigned int temp;
1707
1708 int idx = srcu_read_lock(&kfd_processes_srcu);
1709
1710 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1711 if (p->pasid == pasid) {
1712 kref_get(&p->ref);
1713 ret_p = p;
1714 break;
1715 }
1716 }
1717
1718 srcu_read_unlock(&kfd_processes_srcu, idx);
1719
1720 return ret_p;
1721}
1722
1723/* This increments the process->ref counter. */
1724struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm)
1725{
1726 struct kfd_process *p;
1727
1728 int idx = srcu_read_lock(&kfd_processes_srcu);
1729
1730 p = find_process_by_mm(mm);
1731 if (p)
1732 kref_get(&p->ref);
1733
1734 srcu_read_unlock(&kfd_processes_srcu, idx);
1735
1736 return p;
1737}
1738
1739/* kfd_process_evict_queues - Evict all user queues of a process
1740 *
1741 * Eviction is reference-counted per process-device. This means multiple
1742 * evictions from different sources can be nested safely.
1743 */
1744int kfd_process_evict_queues(struct kfd_process *p, uint32_t trigger)
1745{
1746 int r = 0;
1747 int i;
1748 unsigned int n_evicted = 0;
1749
1750 for (i = 0; i < p->n_pdds; i++) {
1751 struct kfd_process_device *pdd = p->pdds[i];
1752
1753 kfd_smi_event_queue_eviction(pdd->dev, p->lead_thread->pid,
1754 trigger);
1755
1756 r = pdd->dev->dqm->ops.evict_process_queues(pdd->dev->dqm,
1757 &pdd->qpd);
1758 /* evict return -EIO if HWS is hang or asic is resetting, in this case
1759 * we would like to set all the queues to be in evicted state to prevent
1760 * them been add back since they actually not be saved right now.
1761 */
1762 if (r && r != -EIO) {
1763 pr_err("Failed to evict process queues\n");
1764 goto fail;
1765 }
1766 n_evicted++;
1767 }
1768
1769 return r;
1770
1771fail:
1772 /* To keep state consistent, roll back partial eviction by
1773 * restoring queues
1774 */
1775 for (i = 0; i < p->n_pdds; i++) {
1776 struct kfd_process_device *pdd = p->pdds[i];
1777
1778 if (n_evicted == 0)
1779 break;
1780
1781 kfd_smi_event_queue_restore(pdd->dev, p->lead_thread->pid);
1782
1783 if (pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm,
1784 &pdd->qpd))
1785 pr_err("Failed to restore queues\n");
1786
1787 n_evicted--;
1788 }
1789
1790 return r;
1791}
1792
1793/* kfd_process_restore_queues - Restore all user queues of a process */
1794int kfd_process_restore_queues(struct kfd_process *p)
1795{
1796 int r, ret = 0;
1797 int i;
1798
1799 for (i = 0; i < p->n_pdds; i++) {
1800 struct kfd_process_device *pdd = p->pdds[i];
1801
1802 kfd_smi_event_queue_restore(pdd->dev, p->lead_thread->pid);
1803
1804 r = pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm,
1805 &pdd->qpd);
1806 if (r) {
1807 pr_err("Failed to restore process queues\n");
1808 if (!ret)
1809 ret = r;
1810 }
1811 }
1812
1813 return ret;
1814}
1815
1816int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id)
1817{
1818 int i;
1819
1820 for (i = 0; i < p->n_pdds; i++)
1821 if (p->pdds[i] && gpu_id == p->pdds[i]->user_gpu_id)
1822 return i;
1823 return -EINVAL;
1824}
1825
1826int
1827kfd_process_gpuid_from_adev(struct kfd_process *p, struct amdgpu_device *adev,
1828 uint32_t *gpuid, uint32_t *gpuidx)
1829{
1830 int i;
1831
1832 for (i = 0; i < p->n_pdds; i++)
1833 if (p->pdds[i] && p->pdds[i]->dev->adev == adev) {
1834 *gpuid = p->pdds[i]->user_gpu_id;
1835 *gpuidx = i;
1836 return 0;
1837 }
1838 return -EINVAL;
1839}
1840
1841static void evict_process_worker(struct work_struct *work)
1842{
1843 int ret;
1844 struct kfd_process *p;
1845 struct delayed_work *dwork;
1846
1847 dwork = to_delayed_work(work);
1848
1849 /* Process termination destroys this worker thread. So during the
1850 * lifetime of this thread, kfd_process p will be valid
1851 */
1852 p = container_of(dwork, struct kfd_process, eviction_work);
1853 WARN_ONCE(p->last_eviction_seqno != p->ef->seqno,
1854 "Eviction fence mismatch\n");
1855
1856 /* Narrow window of overlap between restore and evict work
1857 * item is possible. Once amdgpu_amdkfd_gpuvm_restore_process_bos
1858 * unreserves KFD BOs, it is possible to evicted again. But
1859 * restore has few more steps of finish. So lets wait for any
1860 * previous restore work to complete
1861 */
1862 flush_delayed_work(&p->restore_work);
1863
1864 pr_debug("Started evicting pasid 0x%x\n", p->pasid);
1865 ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_TRIGGER_TTM);
1866 if (!ret) {
1867 dma_fence_signal(p->ef);
1868 dma_fence_put(p->ef);
1869 p->ef = NULL;
1870 queue_delayed_work(kfd_restore_wq, &p->restore_work,
1871 msecs_to_jiffies(PROCESS_RESTORE_TIME_MS));
1872
1873 pr_debug("Finished evicting pasid 0x%x\n", p->pasid);
1874 } else
1875 pr_err("Failed to evict queues of pasid 0x%x\n", p->pasid);
1876}
1877
1878static void restore_process_worker(struct work_struct *work)
1879{
1880 struct delayed_work *dwork;
1881 struct kfd_process *p;
1882 int ret = 0;
1883
1884 dwork = to_delayed_work(work);
1885
1886 /* Process termination destroys this worker thread. So during the
1887 * lifetime of this thread, kfd_process p will be valid
1888 */
1889 p = container_of(dwork, struct kfd_process, restore_work);
1890 pr_debug("Started restoring pasid 0x%x\n", p->pasid);
1891
1892 /* Setting last_restore_timestamp before successful restoration.
1893 * Otherwise this would have to be set by KGD (restore_process_bos)
1894 * before KFD BOs are unreserved. If not, the process can be evicted
1895 * again before the timestamp is set.
1896 * If restore fails, the timestamp will be set again in the next
1897 * attempt. This would mean that the minimum GPU quanta would be
1898 * PROCESS_ACTIVE_TIME_MS - (time to execute the following two
1899 * functions)
1900 */
1901
1902 p->last_restore_timestamp = get_jiffies_64();
1903 ret = amdgpu_amdkfd_gpuvm_restore_process_bos(p->kgd_process_info,
1904 &p->ef);
1905 if (ret) {
1906 pr_debug("Failed to restore BOs of pasid 0x%x, retry after %d ms\n",
1907 p->pasid, PROCESS_BACK_OFF_TIME_MS);
1908 ret = queue_delayed_work(kfd_restore_wq, &p->restore_work,
1909 msecs_to_jiffies(PROCESS_BACK_OFF_TIME_MS));
1910 WARN(!ret, "reschedule restore work failed\n");
1911 return;
1912 }
1913
1914 ret = kfd_process_restore_queues(p);
1915 if (!ret)
1916 pr_debug("Finished restoring pasid 0x%x\n", p->pasid);
1917 else
1918 pr_err("Failed to restore queues of pasid 0x%x\n", p->pasid);
1919}
1920
1921void kfd_suspend_all_processes(void)
1922{
1923 struct kfd_process *p;
1924 unsigned int temp;
1925 int idx = srcu_read_lock(&kfd_processes_srcu);
1926
1927 WARN(debug_evictions, "Evicting all processes");
1928 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1929 cancel_delayed_work_sync(&p->eviction_work);
1930 cancel_delayed_work_sync(&p->restore_work);
1931
1932 if (kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_TRIGGER_SUSPEND))
1933 pr_err("Failed to suspend process 0x%x\n", p->pasid);
1934 dma_fence_signal(p->ef);
1935 dma_fence_put(p->ef);
1936 p->ef = NULL;
1937 }
1938 srcu_read_unlock(&kfd_processes_srcu, idx);
1939}
1940
1941int kfd_resume_all_processes(void)
1942{
1943 struct kfd_process *p;
1944 unsigned int temp;
1945 int ret = 0, idx = srcu_read_lock(&kfd_processes_srcu);
1946
1947 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1948 if (!queue_delayed_work(kfd_restore_wq, &p->restore_work, 0)) {
1949 pr_err("Restore process %d failed during resume\n",
1950 p->pasid);
1951 ret = -EFAULT;
1952 }
1953 }
1954 srcu_read_unlock(&kfd_processes_srcu, idx);
1955 return ret;
1956}
1957
1958int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process,
1959 struct vm_area_struct *vma)
1960{
1961 struct kfd_process_device *pdd;
1962 struct qcm_process_device *qpd;
1963
1964 if ((vma->vm_end - vma->vm_start) != KFD_CWSR_TBA_TMA_SIZE) {
1965 pr_err("Incorrect CWSR mapping size.\n");
1966 return -EINVAL;
1967 }
1968
1969 pdd = kfd_get_process_device_data(dev, process);
1970 if (!pdd)
1971 return -EINVAL;
1972 qpd = &pdd->qpd;
1973
1974 qpd->cwsr_kaddr = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
1975 get_order(KFD_CWSR_TBA_TMA_SIZE));
1976 if (!qpd->cwsr_kaddr) {
1977 pr_err("Error allocating per process CWSR buffer.\n");
1978 return -ENOMEM;
1979 }
1980
1981 vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND
1982 | VM_NORESERVE | VM_DONTDUMP | VM_PFNMAP;
1983 /* Mapping pages to user process */
1984 return remap_pfn_range(vma, vma->vm_start,
1985 PFN_DOWN(__pa(qpd->cwsr_kaddr)),
1986 KFD_CWSR_TBA_TMA_SIZE, vma->vm_page_prot);
1987}
1988
1989void kfd_flush_tlb(struct kfd_process_device *pdd, enum TLB_FLUSH_TYPE type)
1990{
1991 struct amdgpu_vm *vm = drm_priv_to_vm(pdd->drm_priv);
1992 uint64_t tlb_seq = amdgpu_vm_tlb_seq(vm);
1993 struct kfd_dev *dev = pdd->dev;
1994
1995 /*
1996 * It can be that we race and lose here, but that is extremely unlikely
1997 * and the worst thing which could happen is that we flush the changes
1998 * into the TLB once more which is harmless.
1999 */
2000 if (atomic64_xchg(&pdd->tlb_seq, tlb_seq) == tlb_seq)
2001 return;
2002
2003 if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
2004 /* Nothing to flush until a VMID is assigned, which
2005 * only happens when the first queue is created.
2006 */
2007 if (pdd->qpd.vmid)
2008 amdgpu_amdkfd_flush_gpu_tlb_vmid(dev->adev,
2009 pdd->qpd.vmid);
2010 } else {
2011 amdgpu_amdkfd_flush_gpu_tlb_pasid(dev->adev,
2012 pdd->process->pasid, type);
2013 }
2014}
2015
2016struct kfd_process_device *kfd_process_device_data_by_id(struct kfd_process *p, uint32_t gpu_id)
2017{
2018 int i;
2019
2020 if (gpu_id) {
2021 for (i = 0; i < p->n_pdds; i++) {
2022 struct kfd_process_device *pdd = p->pdds[i];
2023
2024 if (pdd->user_gpu_id == gpu_id)
2025 return pdd;
2026 }
2027 }
2028 return NULL;
2029}
2030
2031int kfd_process_get_user_gpu_id(struct kfd_process *p, uint32_t actual_gpu_id)
2032{
2033 int i;
2034
2035 if (!actual_gpu_id)
2036 return 0;
2037
2038 for (i = 0; i < p->n_pdds; i++) {
2039 struct kfd_process_device *pdd = p->pdds[i];
2040
2041 if (pdd->dev->id == actual_gpu_id)
2042 return pdd->user_gpu_id;
2043 }
2044 return -EINVAL;
2045}
2046
2047#if defined(CONFIG_DEBUG_FS)
2048
2049int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data)
2050{
2051 struct kfd_process *p;
2052 unsigned int temp;
2053 int r = 0;
2054
2055 int idx = srcu_read_lock(&kfd_processes_srcu);
2056
2057 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
2058 seq_printf(m, "Process %d PASID 0x%x:\n",
2059 p->lead_thread->tgid, p->pasid);
2060
2061 mutex_lock(&p->mutex);
2062 r = pqm_debugfs_mqds(m, &p->pqm);
2063 mutex_unlock(&p->mutex);
2064
2065 if (r)
2066 break;
2067 }
2068
2069 srcu_read_unlock(&kfd_processes_srcu, idx);
2070
2071 return r;
2072}
2073
2074#endif
2075
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
39struct 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#include "kfd_svm.h"
46
47/*
48 * List of struct kfd_process (field kfd_process).
49 * Unique/indexed by mm_struct*
50 */
51DEFINE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
52static DEFINE_MUTEX(kfd_processes_mutex);
53
54DEFINE_SRCU(kfd_processes_srcu);
55
56/* For process termination handling */
57static struct workqueue_struct *kfd_process_wq;
58
59/* Ordered, single-threaded workqueue for restoring evicted
60 * processes. Restoring multiple processes concurrently under memory
61 * pressure can lead to processes blocking each other from validating
62 * their BOs and result in a live-lock situation where processes
63 * remain evicted indefinitely.
64 */
65static struct workqueue_struct *kfd_restore_wq;
66
67static struct kfd_process *find_process(const struct task_struct *thread);
68static void kfd_process_ref_release(struct kref *ref);
69static struct kfd_process *create_process(const struct task_struct *thread);
70static int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep);
71
72static void evict_process_worker(struct work_struct *work);
73static void restore_process_worker(struct work_struct *work);
74
75struct kfd_procfs_tree {
76 struct kobject *kobj;
77};
78
79static struct kfd_procfs_tree procfs;
80
81/*
82 * Structure for SDMA activity tracking
83 */
84struct kfd_sdma_activity_handler_workarea {
85 struct work_struct sdma_activity_work;
86 struct kfd_process_device *pdd;
87 uint64_t sdma_activity_counter;
88};
89
90struct temp_sdma_queue_list {
91 uint64_t __user *rptr;
92 uint64_t sdma_val;
93 unsigned int queue_id;
94 struct list_head list;
95};
96
97static void kfd_sdma_activity_worker(struct work_struct *work)
98{
99 struct kfd_sdma_activity_handler_workarea *workarea;
100 struct kfd_process_device *pdd;
101 uint64_t val;
102 struct mm_struct *mm;
103 struct queue *q;
104 struct qcm_process_device *qpd;
105 struct device_queue_manager *dqm;
106 int ret = 0;
107 struct temp_sdma_queue_list sdma_q_list;
108 struct temp_sdma_queue_list *sdma_q, *next;
109
110 workarea = container_of(work, struct kfd_sdma_activity_handler_workarea,
111 sdma_activity_work);
112
113 pdd = workarea->pdd;
114 if (!pdd)
115 return;
116 dqm = pdd->dev->dqm;
117 qpd = &pdd->qpd;
118 if (!dqm || !qpd)
119 return;
120 /*
121 * Total SDMA activity is current SDMA activity + past SDMA activity
122 * Past SDMA count is stored in pdd.
123 * To get the current activity counters for all active SDMA queues,
124 * we loop over all SDMA queues and get their counts from user-space.
125 *
126 * We cannot call get_user() with dqm_lock held as it can cause
127 * a circular lock dependency situation. To read the SDMA stats,
128 * we need to do the following:
129 *
130 * 1. Create a temporary list of SDMA queue nodes from the qpd->queues_list,
131 * with dqm_lock/dqm_unlock().
132 * 2. Call get_user() for each node in temporary list without dqm_lock.
133 * Save the SDMA count for each node and also add the count to the total
134 * SDMA count counter.
135 * Its possible, during this step, a few SDMA queue nodes got deleted
136 * from the qpd->queues_list.
137 * 3. Do a second pass over qpd->queues_list to check if any nodes got deleted.
138 * If any node got deleted, its SDMA count would be captured in the sdma
139 * past activity counter. So subtract the SDMA counter stored in step 2
140 * for this node from the total SDMA count.
141 */
142 INIT_LIST_HEAD(&sdma_q_list.list);
143
144 /*
145 * Create the temp list of all SDMA queues
146 */
147 dqm_lock(dqm);
148
149 list_for_each_entry(q, &qpd->queues_list, list) {
150 if ((q->properties.type != KFD_QUEUE_TYPE_SDMA) &&
151 (q->properties.type != KFD_QUEUE_TYPE_SDMA_XGMI))
152 continue;
153
154 sdma_q = kzalloc(sizeof(struct temp_sdma_queue_list), GFP_KERNEL);
155 if (!sdma_q) {
156 dqm_unlock(dqm);
157 goto cleanup;
158 }
159
160 INIT_LIST_HEAD(&sdma_q->list);
161 sdma_q->rptr = (uint64_t __user *)q->properties.read_ptr;
162 sdma_q->queue_id = q->properties.queue_id;
163 list_add_tail(&sdma_q->list, &sdma_q_list.list);
164 }
165
166 /*
167 * If the temp list is empty, then no SDMA queues nodes were found in
168 * qpd->queues_list. Return the past activity count as the total sdma
169 * count
170 */
171 if (list_empty(&sdma_q_list.list)) {
172 workarea->sdma_activity_counter = pdd->sdma_past_activity_counter;
173 dqm_unlock(dqm);
174 return;
175 }
176
177 dqm_unlock(dqm);
178
179 /*
180 * Get the usage count for each SDMA queue in temp_list.
181 */
182 mm = get_task_mm(pdd->process->lead_thread);
183 if (!mm)
184 goto cleanup;
185
186 kthread_use_mm(mm);
187
188 list_for_each_entry(sdma_q, &sdma_q_list.list, list) {
189 val = 0;
190 ret = read_sdma_queue_counter(sdma_q->rptr, &val);
191 if (ret) {
192 pr_debug("Failed to read SDMA queue active counter for queue id: %d",
193 sdma_q->queue_id);
194 } else {
195 sdma_q->sdma_val = val;
196 workarea->sdma_activity_counter += val;
197 }
198 }
199
200 kthread_unuse_mm(mm);
201 mmput(mm);
202
203 /*
204 * Do a second iteration over qpd_queues_list to check if any SDMA
205 * nodes got deleted while fetching SDMA counter.
206 */
207 dqm_lock(dqm);
208
209 workarea->sdma_activity_counter += pdd->sdma_past_activity_counter;
210
211 list_for_each_entry(q, &qpd->queues_list, list) {
212 if (list_empty(&sdma_q_list.list))
213 break;
214
215 if ((q->properties.type != KFD_QUEUE_TYPE_SDMA) &&
216 (q->properties.type != KFD_QUEUE_TYPE_SDMA_XGMI))
217 continue;
218
219 list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
220 if (((uint64_t __user *)q->properties.read_ptr == sdma_q->rptr) &&
221 (sdma_q->queue_id == q->properties.queue_id)) {
222 list_del(&sdma_q->list);
223 kfree(sdma_q);
224 break;
225 }
226 }
227 }
228
229 dqm_unlock(dqm);
230
231 /*
232 * If temp list is not empty, it implies some queues got deleted
233 * from qpd->queues_list during SDMA usage read. Subtract the SDMA
234 * count for each node from the total SDMA count.
235 */
236 list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
237 workarea->sdma_activity_counter -= sdma_q->sdma_val;
238 list_del(&sdma_q->list);
239 kfree(sdma_q);
240 }
241
242 return;
243
244cleanup:
245 list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
246 list_del(&sdma_q->list);
247 kfree(sdma_q);
248 }
249}
250
251/**
252 * @kfd_get_cu_occupancy - Collect number of waves in-flight on this device
253 * by current process. Translates acquired wave count into number of compute units
254 * that are occupied.
255 *
256 * @atr: Handle of attribute that allows reporting of wave count. The attribute
257 * handle encapsulates GPU device it is associated with, thereby allowing collection
258 * of waves in flight, etc
259 *
260 * @buffer: Handle of user provided buffer updated with wave count
261 *
262 * Return: Number of bytes written to user buffer or an error value
263 */
264static int kfd_get_cu_occupancy(struct attribute *attr, char *buffer)
265{
266 int cu_cnt;
267 int wave_cnt;
268 int max_waves_per_cu;
269 struct kfd_dev *dev = NULL;
270 struct kfd_process *proc = NULL;
271 struct kfd_process_device *pdd = NULL;
272
273 pdd = container_of(attr, struct kfd_process_device, attr_cu_occupancy);
274 dev = pdd->dev;
275 if (dev->kfd2kgd->get_cu_occupancy == NULL)
276 return -EINVAL;
277
278 cu_cnt = 0;
279 proc = pdd->process;
280 if (pdd->qpd.queue_count == 0) {
281 pr_debug("Gpu-Id: %d has no active queues for process %d\n",
282 dev->id, proc->pasid);
283 return snprintf(buffer, PAGE_SIZE, "%d\n", cu_cnt);
284 }
285
286 /* Collect wave count from device if it supports */
287 wave_cnt = 0;
288 max_waves_per_cu = 0;
289 dev->kfd2kgd->get_cu_occupancy(dev->kgd, proc->pasid, &wave_cnt,
290 &max_waves_per_cu);
291
292 /* Translate wave count to number of compute units */
293 cu_cnt = (wave_cnt + (max_waves_per_cu - 1)) / max_waves_per_cu;
294 return snprintf(buffer, PAGE_SIZE, "%d\n", cu_cnt);
295}
296
297static ssize_t kfd_procfs_show(struct kobject *kobj, struct attribute *attr,
298 char *buffer)
299{
300 if (strcmp(attr->name, "pasid") == 0) {
301 struct kfd_process *p = container_of(attr, struct kfd_process,
302 attr_pasid);
303
304 return snprintf(buffer, PAGE_SIZE, "%d\n", p->pasid);
305 } else if (strncmp(attr->name, "vram_", 5) == 0) {
306 struct kfd_process_device *pdd = container_of(attr, struct kfd_process_device,
307 attr_vram);
308 return snprintf(buffer, PAGE_SIZE, "%llu\n", READ_ONCE(pdd->vram_usage));
309 } else if (strncmp(attr->name, "sdma_", 5) == 0) {
310 struct kfd_process_device *pdd = container_of(attr, struct kfd_process_device,
311 attr_sdma);
312 struct kfd_sdma_activity_handler_workarea sdma_activity_work_handler;
313
314 INIT_WORK(&sdma_activity_work_handler.sdma_activity_work,
315 kfd_sdma_activity_worker);
316
317 sdma_activity_work_handler.pdd = pdd;
318 sdma_activity_work_handler.sdma_activity_counter = 0;
319
320 schedule_work(&sdma_activity_work_handler.sdma_activity_work);
321
322 flush_work(&sdma_activity_work_handler.sdma_activity_work);
323
324 return snprintf(buffer, PAGE_SIZE, "%llu\n",
325 (sdma_activity_work_handler.sdma_activity_counter)/
326 SDMA_ACTIVITY_DIVISOR);
327 } else {
328 pr_err("Invalid attribute");
329 return -EINVAL;
330 }
331
332 return 0;
333}
334
335static void kfd_procfs_kobj_release(struct kobject *kobj)
336{
337 kfree(kobj);
338}
339
340static const struct sysfs_ops kfd_procfs_ops = {
341 .show = kfd_procfs_show,
342};
343
344static struct kobj_type procfs_type = {
345 .release = kfd_procfs_kobj_release,
346 .sysfs_ops = &kfd_procfs_ops,
347};
348
349void kfd_procfs_init(void)
350{
351 int ret = 0;
352
353 procfs.kobj = kfd_alloc_struct(procfs.kobj);
354 if (!procfs.kobj)
355 return;
356
357 ret = kobject_init_and_add(procfs.kobj, &procfs_type,
358 &kfd_device->kobj, "proc");
359 if (ret) {
360 pr_warn("Could not create procfs proc folder");
361 /* If we fail to create the procfs, clean up */
362 kfd_procfs_shutdown();
363 }
364}
365
366void kfd_procfs_shutdown(void)
367{
368 if (procfs.kobj) {
369 kobject_del(procfs.kobj);
370 kobject_put(procfs.kobj);
371 procfs.kobj = NULL;
372 }
373}
374
375static ssize_t kfd_procfs_queue_show(struct kobject *kobj,
376 struct attribute *attr, char *buffer)
377{
378 struct queue *q = container_of(kobj, struct queue, kobj);
379
380 if (!strcmp(attr->name, "size"))
381 return snprintf(buffer, PAGE_SIZE, "%llu",
382 q->properties.queue_size);
383 else if (!strcmp(attr->name, "type"))
384 return snprintf(buffer, PAGE_SIZE, "%d", q->properties.type);
385 else if (!strcmp(attr->name, "gpuid"))
386 return snprintf(buffer, PAGE_SIZE, "%u", q->device->id);
387 else
388 pr_err("Invalid attribute");
389
390 return 0;
391}
392
393static ssize_t kfd_procfs_stats_show(struct kobject *kobj,
394 struct attribute *attr, char *buffer)
395{
396 if (strcmp(attr->name, "evicted_ms") == 0) {
397 struct kfd_process_device *pdd = container_of(attr,
398 struct kfd_process_device,
399 attr_evict);
400 uint64_t evict_jiffies;
401
402 evict_jiffies = atomic64_read(&pdd->evict_duration_counter);
403
404 return snprintf(buffer,
405 PAGE_SIZE,
406 "%llu\n",
407 jiffies64_to_msecs(evict_jiffies));
408
409 /* Sysfs handle that gets CU occupancy is per device */
410 } else if (strcmp(attr->name, "cu_occupancy") == 0) {
411 return kfd_get_cu_occupancy(attr, buffer);
412 } else {
413 pr_err("Invalid attribute");
414 }
415
416 return 0;
417}
418
419static ssize_t kfd_sysfs_counters_show(struct kobject *kobj,
420 struct attribute *attr, char *buf)
421{
422 struct kfd_process_device *pdd;
423
424 if (!strcmp(attr->name, "faults")) {
425 pdd = container_of(attr, struct kfd_process_device,
426 attr_faults);
427 return sysfs_emit(buf, "%llu\n", READ_ONCE(pdd->faults));
428 }
429 if (!strcmp(attr->name, "page_in")) {
430 pdd = container_of(attr, struct kfd_process_device,
431 attr_page_in);
432 return sysfs_emit(buf, "%llu\n", READ_ONCE(pdd->page_in));
433 }
434 if (!strcmp(attr->name, "page_out")) {
435 pdd = container_of(attr, struct kfd_process_device,
436 attr_page_out);
437 return sysfs_emit(buf, "%llu\n", READ_ONCE(pdd->page_out));
438 }
439 return 0;
440}
441
442static struct attribute attr_queue_size = {
443 .name = "size",
444 .mode = KFD_SYSFS_FILE_MODE
445};
446
447static struct attribute attr_queue_type = {
448 .name = "type",
449 .mode = KFD_SYSFS_FILE_MODE
450};
451
452static struct attribute attr_queue_gpuid = {
453 .name = "gpuid",
454 .mode = KFD_SYSFS_FILE_MODE
455};
456
457static struct attribute *procfs_queue_attrs[] = {
458 &attr_queue_size,
459 &attr_queue_type,
460 &attr_queue_gpuid,
461 NULL
462};
463
464static const struct sysfs_ops procfs_queue_ops = {
465 .show = kfd_procfs_queue_show,
466};
467
468static struct kobj_type procfs_queue_type = {
469 .sysfs_ops = &procfs_queue_ops,
470 .default_attrs = procfs_queue_attrs,
471};
472
473static const struct sysfs_ops procfs_stats_ops = {
474 .show = kfd_procfs_stats_show,
475};
476
477static struct kobj_type procfs_stats_type = {
478 .sysfs_ops = &procfs_stats_ops,
479 .release = kfd_procfs_kobj_release,
480};
481
482static const struct sysfs_ops sysfs_counters_ops = {
483 .show = kfd_sysfs_counters_show,
484};
485
486static struct kobj_type sysfs_counters_type = {
487 .sysfs_ops = &sysfs_counters_ops,
488 .release = kfd_procfs_kobj_release,
489};
490
491int kfd_procfs_add_queue(struct queue *q)
492{
493 struct kfd_process *proc;
494 int ret;
495
496 if (!q || !q->process)
497 return -EINVAL;
498 proc = q->process;
499
500 /* Create proc/<pid>/queues/<queue id> folder */
501 if (!proc->kobj_queues)
502 return -EFAULT;
503 ret = kobject_init_and_add(&q->kobj, &procfs_queue_type,
504 proc->kobj_queues, "%u", q->properties.queue_id);
505 if (ret < 0) {
506 pr_warn("Creating proc/<pid>/queues/%u failed",
507 q->properties.queue_id);
508 kobject_put(&q->kobj);
509 return ret;
510 }
511
512 return 0;
513}
514
515static void kfd_sysfs_create_file(struct kobject *kobj, struct attribute *attr,
516 char *name)
517{
518 int ret;
519
520 if (!kobj || !attr || !name)
521 return;
522
523 attr->name = name;
524 attr->mode = KFD_SYSFS_FILE_MODE;
525 sysfs_attr_init(attr);
526
527 ret = sysfs_create_file(kobj, attr);
528 if (ret)
529 pr_warn("Create sysfs %s/%s failed %d", kobj->name, name, ret);
530}
531
532static void kfd_procfs_add_sysfs_stats(struct kfd_process *p)
533{
534 int ret;
535 int i;
536 char stats_dir_filename[MAX_SYSFS_FILENAME_LEN];
537
538 if (!p || !p->kobj)
539 return;
540
541 /*
542 * Create sysfs files for each GPU:
543 * - proc/<pid>/stats_<gpuid>/
544 * - proc/<pid>/stats_<gpuid>/evicted_ms
545 * - proc/<pid>/stats_<gpuid>/cu_occupancy
546 */
547 for (i = 0; i < p->n_pdds; i++) {
548 struct kfd_process_device *pdd = p->pdds[i];
549
550 snprintf(stats_dir_filename, MAX_SYSFS_FILENAME_LEN,
551 "stats_%u", pdd->dev->id);
552 pdd->kobj_stats = kfd_alloc_struct(pdd->kobj_stats);
553 if (!pdd->kobj_stats)
554 return;
555
556 ret = kobject_init_and_add(pdd->kobj_stats,
557 &procfs_stats_type,
558 p->kobj,
559 stats_dir_filename);
560
561 if (ret) {
562 pr_warn("Creating KFD proc/stats_%s folder failed",
563 stats_dir_filename);
564 kobject_put(pdd->kobj_stats);
565 pdd->kobj_stats = NULL;
566 return;
567 }
568
569 kfd_sysfs_create_file(pdd->kobj_stats, &pdd->attr_evict,
570 "evicted_ms");
571 /* Add sysfs file to report compute unit occupancy */
572 if (pdd->dev->kfd2kgd->get_cu_occupancy)
573 kfd_sysfs_create_file(pdd->kobj_stats,
574 &pdd->attr_cu_occupancy,
575 "cu_occupancy");
576 }
577}
578
579static void kfd_procfs_add_sysfs_counters(struct kfd_process *p)
580{
581 int ret = 0;
582 int i;
583 char counters_dir_filename[MAX_SYSFS_FILENAME_LEN];
584
585 if (!p || !p->kobj)
586 return;
587
588 /*
589 * Create sysfs files for each GPU which supports SVM
590 * - proc/<pid>/counters_<gpuid>/
591 * - proc/<pid>/counters_<gpuid>/faults
592 * - proc/<pid>/counters_<gpuid>/page_in
593 * - proc/<pid>/counters_<gpuid>/page_out
594 */
595 for_each_set_bit(i, p->svms.bitmap_supported, p->n_pdds) {
596 struct kfd_process_device *pdd = p->pdds[i];
597 struct kobject *kobj_counters;
598
599 snprintf(counters_dir_filename, MAX_SYSFS_FILENAME_LEN,
600 "counters_%u", pdd->dev->id);
601 kobj_counters = kfd_alloc_struct(kobj_counters);
602 if (!kobj_counters)
603 return;
604
605 ret = kobject_init_and_add(kobj_counters, &sysfs_counters_type,
606 p->kobj, counters_dir_filename);
607 if (ret) {
608 pr_warn("Creating KFD proc/%s folder failed",
609 counters_dir_filename);
610 kobject_put(kobj_counters);
611 return;
612 }
613
614 pdd->kobj_counters = kobj_counters;
615 kfd_sysfs_create_file(kobj_counters, &pdd->attr_faults,
616 "faults");
617 kfd_sysfs_create_file(kobj_counters, &pdd->attr_page_in,
618 "page_in");
619 kfd_sysfs_create_file(kobj_counters, &pdd->attr_page_out,
620 "page_out");
621 }
622}
623
624static void kfd_procfs_add_sysfs_files(struct kfd_process *p)
625{
626 int i;
627
628 if (!p || !p->kobj)
629 return;
630
631 /*
632 * Create sysfs files for each GPU:
633 * - proc/<pid>/vram_<gpuid>
634 * - proc/<pid>/sdma_<gpuid>
635 */
636 for (i = 0; i < p->n_pdds; i++) {
637 struct kfd_process_device *pdd = p->pdds[i];
638
639 snprintf(pdd->vram_filename, MAX_SYSFS_FILENAME_LEN, "vram_%u",
640 pdd->dev->id);
641 kfd_sysfs_create_file(p->kobj, &pdd->attr_vram,
642 pdd->vram_filename);
643
644 snprintf(pdd->sdma_filename, MAX_SYSFS_FILENAME_LEN, "sdma_%u",
645 pdd->dev->id);
646 kfd_sysfs_create_file(p->kobj, &pdd->attr_sdma,
647 pdd->sdma_filename);
648 }
649}
650
651void kfd_procfs_del_queue(struct queue *q)
652{
653 if (!q)
654 return;
655
656 kobject_del(&q->kobj);
657 kobject_put(&q->kobj);
658}
659
660int kfd_process_create_wq(void)
661{
662 if (!kfd_process_wq)
663 kfd_process_wq = alloc_workqueue("kfd_process_wq", 0, 0);
664 if (!kfd_restore_wq)
665 kfd_restore_wq = alloc_ordered_workqueue("kfd_restore_wq", 0);
666
667 if (!kfd_process_wq || !kfd_restore_wq) {
668 kfd_process_destroy_wq();
669 return -ENOMEM;
670 }
671
672 return 0;
673}
674
675void kfd_process_destroy_wq(void)
676{
677 if (kfd_process_wq) {
678 destroy_workqueue(kfd_process_wq);
679 kfd_process_wq = NULL;
680 }
681 if (kfd_restore_wq) {
682 destroy_workqueue(kfd_restore_wq);
683 kfd_restore_wq = NULL;
684 }
685}
686
687static void kfd_process_free_gpuvm(struct kgd_mem *mem,
688 struct kfd_process_device *pdd)
689{
690 struct kfd_dev *dev = pdd->dev;
691
692 amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(dev->kgd, mem, pdd->drm_priv);
693 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->kgd, mem, pdd->drm_priv,
694 NULL);
695}
696
697/* kfd_process_alloc_gpuvm - Allocate GPU VM for the KFD process
698 * This function should be only called right after the process
699 * is created and when kfd_processes_mutex is still being held
700 * to avoid concurrency. Because of that exclusiveness, we do
701 * not need to take p->mutex.
702 */
703static int kfd_process_alloc_gpuvm(struct kfd_process_device *pdd,
704 uint64_t gpu_va, uint32_t size,
705 uint32_t flags, void **kptr)
706{
707 struct kfd_dev *kdev = pdd->dev;
708 struct kgd_mem *mem = NULL;
709 int handle;
710 int err;
711
712 err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(kdev->kgd, gpu_va, size,
713 pdd->drm_priv, &mem, NULL, flags);
714 if (err)
715 goto err_alloc_mem;
716
717 err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(kdev->kgd, mem, pdd->drm_priv);
718 if (err)
719 goto err_map_mem;
720
721 err = amdgpu_amdkfd_gpuvm_sync_memory(kdev->kgd, mem, true);
722 if (err) {
723 pr_debug("Sync memory failed, wait interrupted by user signal\n");
724 goto sync_memory_failed;
725 }
726
727 /* Create an obj handle so kfd_process_device_remove_obj_handle
728 * will take care of the bo removal when the process finishes.
729 * We do not need to take p->mutex, because the process is just
730 * created and the ioctls have not had the chance to run.
731 */
732 handle = kfd_process_device_create_obj_handle(pdd, mem);
733
734 if (handle < 0) {
735 err = handle;
736 goto free_gpuvm;
737 }
738
739 if (kptr) {
740 err = amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(kdev->kgd,
741 (struct kgd_mem *)mem, kptr, NULL);
742 if (err) {
743 pr_debug("Map GTT BO to kernel failed\n");
744 goto free_obj_handle;
745 }
746 }
747
748 return err;
749
750free_obj_handle:
751 kfd_process_device_remove_obj_handle(pdd, handle);
752free_gpuvm:
753sync_memory_failed:
754 kfd_process_free_gpuvm(mem, pdd);
755 return err;
756
757err_map_mem:
758 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(kdev->kgd, mem, pdd->drm_priv,
759 NULL);
760err_alloc_mem:
761 *kptr = NULL;
762 return err;
763}
764
765/* kfd_process_device_reserve_ib_mem - Reserve memory inside the
766 * process for IB usage The memory reserved is for KFD to submit
767 * IB to AMDGPU from kernel. If the memory is reserved
768 * successfully, ib_kaddr will have the CPU/kernel
769 * address. Check ib_kaddr before accessing the memory.
770 */
771static int kfd_process_device_reserve_ib_mem(struct kfd_process_device *pdd)
772{
773 struct qcm_process_device *qpd = &pdd->qpd;
774 uint32_t flags = KFD_IOC_ALLOC_MEM_FLAGS_GTT |
775 KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE |
776 KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE |
777 KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
778 void *kaddr;
779 int ret;
780
781 if (qpd->ib_kaddr || !qpd->ib_base)
782 return 0;
783
784 /* ib_base is only set for dGPU */
785 ret = kfd_process_alloc_gpuvm(pdd, qpd->ib_base, PAGE_SIZE, flags,
786 &kaddr);
787 if (ret)
788 return ret;
789
790 qpd->ib_kaddr = kaddr;
791
792 return 0;
793}
794
795struct kfd_process *kfd_create_process(struct file *filep)
796{
797 struct kfd_process *process;
798 struct task_struct *thread = current;
799 int ret;
800
801 if (!thread->mm)
802 return ERR_PTR(-EINVAL);
803
804 /* Only the pthreads threading model is supported. */
805 if (thread->group_leader->mm != thread->mm)
806 return ERR_PTR(-EINVAL);
807
808 /*
809 * take kfd processes mutex before starting of process creation
810 * so there won't be a case where two threads of the same process
811 * create two kfd_process structures
812 */
813 mutex_lock(&kfd_processes_mutex);
814
815 /* A prior open of /dev/kfd could have already created the process. */
816 process = find_process(thread);
817 if (process) {
818 pr_debug("Process already found\n");
819 } else {
820 process = create_process(thread);
821 if (IS_ERR(process))
822 goto out;
823
824 ret = kfd_process_init_cwsr_apu(process, filep);
825 if (ret)
826 goto out_destroy;
827
828 if (!procfs.kobj)
829 goto out;
830
831 process->kobj = kfd_alloc_struct(process->kobj);
832 if (!process->kobj) {
833 pr_warn("Creating procfs kobject failed");
834 goto out;
835 }
836 ret = kobject_init_and_add(process->kobj, &procfs_type,
837 procfs.kobj, "%d",
838 (int)process->lead_thread->pid);
839 if (ret) {
840 pr_warn("Creating procfs pid directory failed");
841 kobject_put(process->kobj);
842 goto out;
843 }
844
845 kfd_sysfs_create_file(process->kobj, &process->attr_pasid,
846 "pasid");
847
848 process->kobj_queues = kobject_create_and_add("queues",
849 process->kobj);
850 if (!process->kobj_queues)
851 pr_warn("Creating KFD proc/queues folder failed");
852
853 kfd_procfs_add_sysfs_stats(process);
854 kfd_procfs_add_sysfs_files(process);
855 kfd_procfs_add_sysfs_counters(process);
856 }
857out:
858 if (!IS_ERR(process))
859 kref_get(&process->ref);
860 mutex_unlock(&kfd_processes_mutex);
861
862 return process;
863
864out_destroy:
865 hash_del_rcu(&process->kfd_processes);
866 mutex_unlock(&kfd_processes_mutex);
867 synchronize_srcu(&kfd_processes_srcu);
868 /* kfd_process_free_notifier will trigger the cleanup */
869 mmu_notifier_put(&process->mmu_notifier);
870 return ERR_PTR(ret);
871}
872
873struct kfd_process *kfd_get_process(const struct task_struct *thread)
874{
875 struct kfd_process *process;
876
877 if (!thread->mm)
878 return ERR_PTR(-EINVAL);
879
880 /* Only the pthreads threading model is supported. */
881 if (thread->group_leader->mm != thread->mm)
882 return ERR_PTR(-EINVAL);
883
884 process = find_process(thread);
885 if (!process)
886 return ERR_PTR(-EINVAL);
887
888 return process;
889}
890
891static struct kfd_process *find_process_by_mm(const struct mm_struct *mm)
892{
893 struct kfd_process *process;
894
895 hash_for_each_possible_rcu(kfd_processes_table, process,
896 kfd_processes, (uintptr_t)mm)
897 if (process->mm == mm)
898 return process;
899
900 return NULL;
901}
902
903static struct kfd_process *find_process(const struct task_struct *thread)
904{
905 struct kfd_process *p;
906 int idx;
907
908 idx = srcu_read_lock(&kfd_processes_srcu);
909 p = find_process_by_mm(thread->mm);
910 srcu_read_unlock(&kfd_processes_srcu, idx);
911
912 return p;
913}
914
915void kfd_unref_process(struct kfd_process *p)
916{
917 kref_put(&p->ref, kfd_process_ref_release);
918}
919
920
921static void kfd_process_device_free_bos(struct kfd_process_device *pdd)
922{
923 struct kfd_process *p = pdd->process;
924 void *mem;
925 int id;
926 int i;
927
928 /*
929 * Remove all handles from idr and release appropriate
930 * local memory object
931 */
932 idr_for_each_entry(&pdd->alloc_idr, mem, id) {
933
934 for (i = 0; i < p->n_pdds; i++) {
935 struct kfd_process_device *peer_pdd = p->pdds[i];
936
937 if (!peer_pdd->drm_priv)
938 continue;
939 amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
940 peer_pdd->dev->kgd, mem, peer_pdd->drm_priv);
941 }
942
943 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->kgd, mem,
944 pdd->drm_priv, NULL);
945 kfd_process_device_remove_obj_handle(pdd, id);
946 }
947}
948
949static void kfd_process_free_outstanding_kfd_bos(struct kfd_process *p)
950{
951 int i;
952
953 for (i = 0; i < p->n_pdds; i++)
954 kfd_process_device_free_bos(p->pdds[i]);
955}
956
957static void kfd_process_destroy_pdds(struct kfd_process *p)
958{
959 int i;
960
961 for (i = 0; i < p->n_pdds; i++) {
962 struct kfd_process_device *pdd = p->pdds[i];
963
964 pr_debug("Releasing pdd (topology id %d) for process (pasid 0x%x)\n",
965 pdd->dev->id, p->pasid);
966
967 if (pdd->drm_file) {
968 amdgpu_amdkfd_gpuvm_release_process_vm(
969 pdd->dev->kgd, pdd->drm_priv);
970 fput(pdd->drm_file);
971 }
972
973 if (pdd->qpd.cwsr_kaddr && !pdd->qpd.cwsr_base)
974 free_pages((unsigned long)pdd->qpd.cwsr_kaddr,
975 get_order(KFD_CWSR_TBA_TMA_SIZE));
976
977 kfree(pdd->qpd.doorbell_bitmap);
978 idr_destroy(&pdd->alloc_idr);
979
980 kfd_free_process_doorbells(pdd->dev, pdd->doorbell_index);
981
982 /*
983 * before destroying pdd, make sure to report availability
984 * for auto suspend
985 */
986 if (pdd->runtime_inuse) {
987 pm_runtime_mark_last_busy(pdd->dev->ddev->dev);
988 pm_runtime_put_autosuspend(pdd->dev->ddev->dev);
989 pdd->runtime_inuse = false;
990 }
991
992 kfree(pdd);
993 p->pdds[i] = NULL;
994 }
995 p->n_pdds = 0;
996}
997
998static void kfd_process_remove_sysfs(struct kfd_process *p)
999{
1000 struct kfd_process_device *pdd;
1001 int i;
1002
1003 if (!p->kobj)
1004 return;
1005
1006 sysfs_remove_file(p->kobj, &p->attr_pasid);
1007 kobject_del(p->kobj_queues);
1008 kobject_put(p->kobj_queues);
1009 p->kobj_queues = NULL;
1010
1011 for (i = 0; i < p->n_pdds; i++) {
1012 pdd = p->pdds[i];
1013
1014 sysfs_remove_file(p->kobj, &pdd->attr_vram);
1015 sysfs_remove_file(p->kobj, &pdd->attr_sdma);
1016
1017 sysfs_remove_file(pdd->kobj_stats, &pdd->attr_evict);
1018 if (pdd->dev->kfd2kgd->get_cu_occupancy)
1019 sysfs_remove_file(pdd->kobj_stats,
1020 &pdd->attr_cu_occupancy);
1021 kobject_del(pdd->kobj_stats);
1022 kobject_put(pdd->kobj_stats);
1023 pdd->kobj_stats = NULL;
1024 }
1025
1026 for_each_set_bit(i, p->svms.bitmap_supported, p->n_pdds) {
1027 pdd = p->pdds[i];
1028
1029 sysfs_remove_file(pdd->kobj_counters, &pdd->attr_faults);
1030 sysfs_remove_file(pdd->kobj_counters, &pdd->attr_page_in);
1031 sysfs_remove_file(pdd->kobj_counters, &pdd->attr_page_out);
1032 kobject_del(pdd->kobj_counters);
1033 kobject_put(pdd->kobj_counters);
1034 pdd->kobj_counters = NULL;
1035 }
1036
1037 kobject_del(p->kobj);
1038 kobject_put(p->kobj);
1039 p->kobj = NULL;
1040}
1041
1042/* No process locking is needed in this function, because the process
1043 * is not findable any more. We must assume that no other thread is
1044 * using it any more, otherwise we couldn't safely free the process
1045 * structure in the end.
1046 */
1047static void kfd_process_wq_release(struct work_struct *work)
1048{
1049 struct kfd_process *p = container_of(work, struct kfd_process,
1050 release_work);
1051 kfd_process_remove_sysfs(p);
1052 kfd_iommu_unbind_process(p);
1053
1054 kfd_process_free_outstanding_kfd_bos(p);
1055 svm_range_list_fini(p);
1056
1057 kfd_process_destroy_pdds(p);
1058 dma_fence_put(p->ef);
1059
1060 kfd_event_free_process(p);
1061
1062 kfd_pasid_free(p->pasid);
1063 mutex_destroy(&p->mutex);
1064
1065 put_task_struct(p->lead_thread);
1066
1067 kfree(p);
1068}
1069
1070static void kfd_process_ref_release(struct kref *ref)
1071{
1072 struct kfd_process *p = container_of(ref, struct kfd_process, ref);
1073
1074 INIT_WORK(&p->release_work, kfd_process_wq_release);
1075 queue_work(kfd_process_wq, &p->release_work);
1076}
1077
1078static struct mmu_notifier *kfd_process_alloc_notifier(struct mm_struct *mm)
1079{
1080 int idx = srcu_read_lock(&kfd_processes_srcu);
1081 struct kfd_process *p = find_process_by_mm(mm);
1082
1083 srcu_read_unlock(&kfd_processes_srcu, idx);
1084
1085 return p ? &p->mmu_notifier : ERR_PTR(-ESRCH);
1086}
1087
1088static void kfd_process_free_notifier(struct mmu_notifier *mn)
1089{
1090 kfd_unref_process(container_of(mn, struct kfd_process, mmu_notifier));
1091}
1092
1093static void kfd_process_notifier_release(struct mmu_notifier *mn,
1094 struct mm_struct *mm)
1095{
1096 struct kfd_process *p;
1097 int i;
1098
1099 /*
1100 * The kfd_process structure can not be free because the
1101 * mmu_notifier srcu is read locked
1102 */
1103 p = container_of(mn, struct kfd_process, mmu_notifier);
1104 if (WARN_ON(p->mm != mm))
1105 return;
1106
1107 mutex_lock(&kfd_processes_mutex);
1108 hash_del_rcu(&p->kfd_processes);
1109 mutex_unlock(&kfd_processes_mutex);
1110 synchronize_srcu(&kfd_processes_srcu);
1111
1112 cancel_delayed_work_sync(&p->eviction_work);
1113 cancel_delayed_work_sync(&p->restore_work);
1114 cancel_delayed_work_sync(&p->svms.restore_work);
1115
1116 mutex_lock(&p->mutex);
1117
1118 /* Iterate over all process device data structures and if the
1119 * pdd is in debug mode, we should first force unregistration,
1120 * then we will be able to destroy the queues
1121 */
1122 for (i = 0; i < p->n_pdds; i++) {
1123 struct kfd_dev *dev = p->pdds[i]->dev;
1124
1125 mutex_lock(kfd_get_dbgmgr_mutex());
1126 if (dev && dev->dbgmgr && dev->dbgmgr->pasid == p->pasid) {
1127 if (!kfd_dbgmgr_unregister(dev->dbgmgr, p)) {
1128 kfd_dbgmgr_destroy(dev->dbgmgr);
1129 dev->dbgmgr = NULL;
1130 }
1131 }
1132 mutex_unlock(kfd_get_dbgmgr_mutex());
1133 }
1134
1135 kfd_process_dequeue_from_all_devices(p);
1136 pqm_uninit(&p->pqm);
1137
1138 /* Indicate to other users that MM is no longer valid */
1139 p->mm = NULL;
1140 /* Signal the eviction fence after user mode queues are
1141 * destroyed. This allows any BOs to be freed without
1142 * triggering pointless evictions or waiting for fences.
1143 */
1144 dma_fence_signal(p->ef);
1145
1146 mutex_unlock(&p->mutex);
1147
1148 mmu_notifier_put(&p->mmu_notifier);
1149}
1150
1151static const struct mmu_notifier_ops kfd_process_mmu_notifier_ops = {
1152 .release = kfd_process_notifier_release,
1153 .alloc_notifier = kfd_process_alloc_notifier,
1154 .free_notifier = kfd_process_free_notifier,
1155};
1156
1157static int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep)
1158{
1159 unsigned long offset;
1160 int i;
1161
1162 for (i = 0; i < p->n_pdds; i++) {
1163 struct kfd_dev *dev = p->pdds[i]->dev;
1164 struct qcm_process_device *qpd = &p->pdds[i]->qpd;
1165
1166 if (!dev->cwsr_enabled || qpd->cwsr_kaddr || qpd->cwsr_base)
1167 continue;
1168
1169 offset = KFD_MMAP_TYPE_RESERVED_MEM | KFD_MMAP_GPU_ID(dev->id);
1170 qpd->tba_addr = (int64_t)vm_mmap(filep, 0,
1171 KFD_CWSR_TBA_TMA_SIZE, PROT_READ | PROT_EXEC,
1172 MAP_SHARED, offset);
1173
1174 if (IS_ERR_VALUE(qpd->tba_addr)) {
1175 int err = qpd->tba_addr;
1176
1177 pr_err("Failure to set tba address. error %d.\n", err);
1178 qpd->tba_addr = 0;
1179 qpd->cwsr_kaddr = NULL;
1180 return err;
1181 }
1182
1183 memcpy(qpd->cwsr_kaddr, dev->cwsr_isa, dev->cwsr_isa_size);
1184
1185 qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET;
1186 pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n",
1187 qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr);
1188 }
1189
1190 return 0;
1191}
1192
1193static int kfd_process_device_init_cwsr_dgpu(struct kfd_process_device *pdd)
1194{
1195 struct kfd_dev *dev = pdd->dev;
1196 struct qcm_process_device *qpd = &pdd->qpd;
1197 uint32_t flags = KFD_IOC_ALLOC_MEM_FLAGS_GTT
1198 | KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE
1199 | KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
1200 void *kaddr;
1201 int ret;
1202
1203 if (!dev->cwsr_enabled || qpd->cwsr_kaddr || !qpd->cwsr_base)
1204 return 0;
1205
1206 /* cwsr_base is only set for dGPU */
1207 ret = kfd_process_alloc_gpuvm(pdd, qpd->cwsr_base,
1208 KFD_CWSR_TBA_TMA_SIZE, flags, &kaddr);
1209 if (ret)
1210 return ret;
1211
1212 qpd->cwsr_kaddr = kaddr;
1213 qpd->tba_addr = qpd->cwsr_base;
1214
1215 memcpy(qpd->cwsr_kaddr, dev->cwsr_isa, dev->cwsr_isa_size);
1216
1217 qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET;
1218 pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n",
1219 qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr);
1220
1221 return 0;
1222}
1223
1224void kfd_process_set_trap_handler(struct qcm_process_device *qpd,
1225 uint64_t tba_addr,
1226 uint64_t tma_addr)
1227{
1228 if (qpd->cwsr_kaddr) {
1229 /* KFD trap handler is bound, record as second-level TBA/TMA
1230 * in first-level TMA. First-level trap will jump to second.
1231 */
1232 uint64_t *tma =
1233 (uint64_t *)(qpd->cwsr_kaddr + KFD_CWSR_TMA_OFFSET);
1234 tma[0] = tba_addr;
1235 tma[1] = tma_addr;
1236 } else {
1237 /* No trap handler bound, bind as first-level TBA/TMA. */
1238 qpd->tba_addr = tba_addr;
1239 qpd->tma_addr = tma_addr;
1240 }
1241}
1242
1243bool kfd_process_xnack_mode(struct kfd_process *p, bool supported)
1244{
1245 int i;
1246
1247 /* On most GFXv9 GPUs, the retry mode in the SQ must match the
1248 * boot time retry setting. Mixing processes with different
1249 * XNACK/retry settings can hang the GPU.
1250 *
1251 * Different GPUs can have different noretry settings depending
1252 * on HW bugs or limitations. We need to find at least one
1253 * XNACK mode for this process that's compatible with all GPUs.
1254 * Fortunately GPUs with retry enabled (noretry=0) can run code
1255 * built for XNACK-off. On GFXv9 it may perform slower.
1256 *
1257 * Therefore applications built for XNACK-off can always be
1258 * supported and will be our fallback if any GPU does not
1259 * support retry.
1260 */
1261 for (i = 0; i < p->n_pdds; i++) {
1262 struct kfd_dev *dev = p->pdds[i]->dev;
1263
1264 /* Only consider GFXv9 and higher GPUs. Older GPUs don't
1265 * support the SVM APIs and don't need to be considered
1266 * for the XNACK mode selection.
1267 */
1268 if (dev->device_info->asic_family < CHIP_VEGA10)
1269 continue;
1270 /* Aldebaran can always support XNACK because it can support
1271 * per-process XNACK mode selection. But let the dev->noretry
1272 * setting still influence the default XNACK mode.
1273 */
1274 if (supported &&
1275 dev->device_info->asic_family == CHIP_ALDEBARAN)
1276 continue;
1277
1278 /* GFXv10 and later GPUs do not support shader preemption
1279 * during page faults. This can lead to poor QoS for queue
1280 * management and memory-manager-related preemptions or
1281 * even deadlocks.
1282 */
1283 if (dev->device_info->asic_family >= CHIP_NAVI10)
1284 return false;
1285
1286 if (dev->noretry)
1287 return false;
1288 }
1289
1290 return true;
1291}
1292
1293/*
1294 * On return the kfd_process is fully operational and will be freed when the
1295 * mm is released
1296 */
1297static struct kfd_process *create_process(const struct task_struct *thread)
1298{
1299 struct kfd_process *process;
1300 struct mmu_notifier *mn;
1301 int err = -ENOMEM;
1302
1303 process = kzalloc(sizeof(*process), GFP_KERNEL);
1304 if (!process)
1305 goto err_alloc_process;
1306
1307 kref_init(&process->ref);
1308 mutex_init(&process->mutex);
1309 process->mm = thread->mm;
1310 process->lead_thread = thread->group_leader;
1311 process->n_pdds = 0;
1312 INIT_DELAYED_WORK(&process->eviction_work, evict_process_worker);
1313 INIT_DELAYED_WORK(&process->restore_work, restore_process_worker);
1314 process->last_restore_timestamp = get_jiffies_64();
1315 kfd_event_init_process(process);
1316 process->is_32bit_user_mode = in_compat_syscall();
1317
1318 process->pasid = kfd_pasid_alloc();
1319 if (process->pasid == 0)
1320 goto err_alloc_pasid;
1321
1322 err = pqm_init(&process->pqm, process);
1323 if (err != 0)
1324 goto err_process_pqm_init;
1325
1326 /* init process apertures*/
1327 err = kfd_init_apertures(process);
1328 if (err != 0)
1329 goto err_init_apertures;
1330
1331 /* Check XNACK support after PDDs are created in kfd_init_apertures */
1332 process->xnack_enabled = kfd_process_xnack_mode(process, false);
1333
1334 err = svm_range_list_init(process);
1335 if (err)
1336 goto err_init_svm_range_list;
1337
1338 /* alloc_notifier needs to find the process in the hash table */
1339 hash_add_rcu(kfd_processes_table, &process->kfd_processes,
1340 (uintptr_t)process->mm);
1341
1342 /* MMU notifier registration must be the last call that can fail
1343 * because after this point we cannot unwind the process creation.
1344 * After this point, mmu_notifier_put will trigger the cleanup by
1345 * dropping the last process reference in the free_notifier.
1346 */
1347 mn = mmu_notifier_get(&kfd_process_mmu_notifier_ops, process->mm);
1348 if (IS_ERR(mn)) {
1349 err = PTR_ERR(mn);
1350 goto err_register_notifier;
1351 }
1352 BUG_ON(mn != &process->mmu_notifier);
1353
1354 get_task_struct(process->lead_thread);
1355
1356 return process;
1357
1358err_register_notifier:
1359 hash_del_rcu(&process->kfd_processes);
1360 svm_range_list_fini(process);
1361err_init_svm_range_list:
1362 kfd_process_free_outstanding_kfd_bos(process);
1363 kfd_process_destroy_pdds(process);
1364err_init_apertures:
1365 pqm_uninit(&process->pqm);
1366err_process_pqm_init:
1367 kfd_pasid_free(process->pasid);
1368err_alloc_pasid:
1369 mutex_destroy(&process->mutex);
1370 kfree(process);
1371err_alloc_process:
1372 return ERR_PTR(err);
1373}
1374
1375static int init_doorbell_bitmap(struct qcm_process_device *qpd,
1376 struct kfd_dev *dev)
1377{
1378 unsigned int i;
1379 int range_start = dev->shared_resources.non_cp_doorbells_start;
1380 int range_end = dev->shared_resources.non_cp_doorbells_end;
1381
1382 if (!KFD_IS_SOC15(dev->device_info->asic_family))
1383 return 0;
1384
1385 qpd->doorbell_bitmap =
1386 kzalloc(DIV_ROUND_UP(KFD_MAX_NUM_OF_QUEUES_PER_PROCESS,
1387 BITS_PER_BYTE), GFP_KERNEL);
1388 if (!qpd->doorbell_bitmap)
1389 return -ENOMEM;
1390
1391 /* Mask out doorbells reserved for SDMA, IH, and VCN on SOC15. */
1392 pr_debug("reserved doorbell 0x%03x - 0x%03x\n", range_start, range_end);
1393 pr_debug("reserved doorbell 0x%03x - 0x%03x\n",
1394 range_start + KFD_QUEUE_DOORBELL_MIRROR_OFFSET,
1395 range_end + KFD_QUEUE_DOORBELL_MIRROR_OFFSET);
1396
1397 for (i = 0; i < KFD_MAX_NUM_OF_QUEUES_PER_PROCESS / 2; i++) {
1398 if (i >= range_start && i <= range_end) {
1399 set_bit(i, qpd->doorbell_bitmap);
1400 set_bit(i + KFD_QUEUE_DOORBELL_MIRROR_OFFSET,
1401 qpd->doorbell_bitmap);
1402 }
1403 }
1404
1405 return 0;
1406}
1407
1408struct kfd_process_device *kfd_get_process_device_data(struct kfd_dev *dev,
1409 struct kfd_process *p)
1410{
1411 int i;
1412
1413 for (i = 0; i < p->n_pdds; i++)
1414 if (p->pdds[i]->dev == dev)
1415 return p->pdds[i];
1416
1417 return NULL;
1418}
1419
1420struct kfd_process_device *kfd_create_process_device_data(struct kfd_dev *dev,
1421 struct kfd_process *p)
1422{
1423 struct kfd_process_device *pdd = NULL;
1424
1425 if (WARN_ON_ONCE(p->n_pdds >= MAX_GPU_INSTANCE))
1426 return NULL;
1427 pdd = kzalloc(sizeof(*pdd), GFP_KERNEL);
1428 if (!pdd)
1429 return NULL;
1430
1431 if (kfd_alloc_process_doorbells(dev, &pdd->doorbell_index) < 0) {
1432 pr_err("Failed to alloc doorbell for pdd\n");
1433 goto err_free_pdd;
1434 }
1435
1436 if (init_doorbell_bitmap(&pdd->qpd, dev)) {
1437 pr_err("Failed to init doorbell for process\n");
1438 goto err_free_pdd;
1439 }
1440
1441 pdd->dev = dev;
1442 INIT_LIST_HEAD(&pdd->qpd.queues_list);
1443 INIT_LIST_HEAD(&pdd->qpd.priv_queue_list);
1444 pdd->qpd.dqm = dev->dqm;
1445 pdd->qpd.pqm = &p->pqm;
1446 pdd->qpd.evicted = 0;
1447 pdd->qpd.mapped_gws_queue = false;
1448 pdd->process = p;
1449 pdd->bound = PDD_UNBOUND;
1450 pdd->already_dequeued = false;
1451 pdd->runtime_inuse = false;
1452 pdd->vram_usage = 0;
1453 pdd->sdma_past_activity_counter = 0;
1454 atomic64_set(&pdd->evict_duration_counter, 0);
1455 p->pdds[p->n_pdds++] = pdd;
1456
1457 /* Init idr used for memory handle translation */
1458 idr_init(&pdd->alloc_idr);
1459
1460 return pdd;
1461
1462err_free_pdd:
1463 kfree(pdd);
1464 return NULL;
1465}
1466
1467/**
1468 * kfd_process_device_init_vm - Initialize a VM for a process-device
1469 *
1470 * @pdd: The process-device
1471 * @drm_file: Optional pointer to a DRM file descriptor
1472 *
1473 * If @drm_file is specified, it will be used to acquire the VM from
1474 * that file descriptor. If successful, the @pdd takes ownership of
1475 * the file descriptor.
1476 *
1477 * If @drm_file is NULL, a new VM is created.
1478 *
1479 * Returns 0 on success, -errno on failure.
1480 */
1481int kfd_process_device_init_vm(struct kfd_process_device *pdd,
1482 struct file *drm_file)
1483{
1484 struct kfd_process *p;
1485 struct kfd_dev *dev;
1486 int ret;
1487
1488 if (!drm_file)
1489 return -EINVAL;
1490
1491 if (pdd->drm_priv)
1492 return -EBUSY;
1493
1494 p = pdd->process;
1495 dev = pdd->dev;
1496
1497 ret = amdgpu_amdkfd_gpuvm_acquire_process_vm(
1498 dev->kgd, drm_file, p->pasid,
1499 &p->kgd_process_info, &p->ef);
1500 if (ret) {
1501 pr_err("Failed to create process VM object\n");
1502 return ret;
1503 }
1504 pdd->drm_priv = drm_file->private_data;
1505
1506 ret = kfd_process_device_reserve_ib_mem(pdd);
1507 if (ret)
1508 goto err_reserve_ib_mem;
1509 ret = kfd_process_device_init_cwsr_dgpu(pdd);
1510 if (ret)
1511 goto err_init_cwsr;
1512
1513 pdd->drm_file = drm_file;
1514
1515 return 0;
1516
1517err_init_cwsr:
1518err_reserve_ib_mem:
1519 kfd_process_device_free_bos(pdd);
1520 pdd->drm_priv = NULL;
1521
1522 return ret;
1523}
1524
1525/*
1526 * Direct the IOMMU to bind the process (specifically the pasid->mm)
1527 * to the device.
1528 * Unbinding occurs when the process dies or the device is removed.
1529 *
1530 * Assumes that the process lock is held.
1531 */
1532struct kfd_process_device *kfd_bind_process_to_device(struct kfd_dev *dev,
1533 struct kfd_process *p)
1534{
1535 struct kfd_process_device *pdd;
1536 int err;
1537
1538 pdd = kfd_get_process_device_data(dev, p);
1539 if (!pdd) {
1540 pr_err("Process device data doesn't exist\n");
1541 return ERR_PTR(-ENOMEM);
1542 }
1543
1544 if (!pdd->drm_priv)
1545 return ERR_PTR(-ENODEV);
1546
1547 /*
1548 * signal runtime-pm system to auto resume and prevent
1549 * further runtime suspend once device pdd is created until
1550 * pdd is destroyed.
1551 */
1552 if (!pdd->runtime_inuse) {
1553 err = pm_runtime_get_sync(dev->ddev->dev);
1554 if (err < 0) {
1555 pm_runtime_put_autosuspend(dev->ddev->dev);
1556 return ERR_PTR(err);
1557 }
1558 }
1559
1560 err = kfd_iommu_bind_process_to_device(pdd);
1561 if (err)
1562 goto out;
1563
1564 /*
1565 * make sure that runtime_usage counter is incremented just once
1566 * per pdd
1567 */
1568 pdd->runtime_inuse = true;
1569
1570 return pdd;
1571
1572out:
1573 /* balance runpm reference count and exit with error */
1574 if (!pdd->runtime_inuse) {
1575 pm_runtime_mark_last_busy(dev->ddev->dev);
1576 pm_runtime_put_autosuspend(dev->ddev->dev);
1577 }
1578
1579 return ERR_PTR(err);
1580}
1581
1582/* Create specific handle mapped to mem from process local memory idr
1583 * Assumes that the process lock is held.
1584 */
1585int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
1586 void *mem)
1587{
1588 return idr_alloc(&pdd->alloc_idr, mem, 0, 0, GFP_KERNEL);
1589}
1590
1591/* Translate specific handle from process local memory idr
1592 * Assumes that the process lock is held.
1593 */
1594void *kfd_process_device_translate_handle(struct kfd_process_device *pdd,
1595 int handle)
1596{
1597 if (handle < 0)
1598 return NULL;
1599
1600 return idr_find(&pdd->alloc_idr, handle);
1601}
1602
1603/* Remove specific handle from process local memory idr
1604 * Assumes that the process lock is held.
1605 */
1606void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,
1607 int handle)
1608{
1609 if (handle >= 0)
1610 idr_remove(&pdd->alloc_idr, handle);
1611}
1612
1613/* This increments the process->ref counter. */
1614struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid)
1615{
1616 struct kfd_process *p, *ret_p = NULL;
1617 unsigned int temp;
1618
1619 int idx = srcu_read_lock(&kfd_processes_srcu);
1620
1621 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1622 if (p->pasid == pasid) {
1623 kref_get(&p->ref);
1624 ret_p = p;
1625 break;
1626 }
1627 }
1628
1629 srcu_read_unlock(&kfd_processes_srcu, idx);
1630
1631 return ret_p;
1632}
1633
1634/* This increments the process->ref counter. */
1635struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm)
1636{
1637 struct kfd_process *p;
1638
1639 int idx = srcu_read_lock(&kfd_processes_srcu);
1640
1641 p = find_process_by_mm(mm);
1642 if (p)
1643 kref_get(&p->ref);
1644
1645 srcu_read_unlock(&kfd_processes_srcu, idx);
1646
1647 return p;
1648}
1649
1650/* kfd_process_evict_queues - Evict all user queues of a process
1651 *
1652 * Eviction is reference-counted per process-device. This means multiple
1653 * evictions from different sources can be nested safely.
1654 */
1655int kfd_process_evict_queues(struct kfd_process *p)
1656{
1657 int r = 0;
1658 int i;
1659 unsigned int n_evicted = 0;
1660
1661 for (i = 0; i < p->n_pdds; i++) {
1662 struct kfd_process_device *pdd = p->pdds[i];
1663
1664 r = pdd->dev->dqm->ops.evict_process_queues(pdd->dev->dqm,
1665 &pdd->qpd);
1666 if (r) {
1667 pr_err("Failed to evict process queues\n");
1668 goto fail;
1669 }
1670 n_evicted++;
1671 }
1672
1673 return r;
1674
1675fail:
1676 /* To keep state consistent, roll back partial eviction by
1677 * restoring queues
1678 */
1679 for (i = 0; i < p->n_pdds; i++) {
1680 struct kfd_process_device *pdd = p->pdds[i];
1681
1682 if (n_evicted == 0)
1683 break;
1684 if (pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm,
1685 &pdd->qpd))
1686 pr_err("Failed to restore queues\n");
1687
1688 n_evicted--;
1689 }
1690
1691 return r;
1692}
1693
1694/* kfd_process_restore_queues - Restore all user queues of a process */
1695int kfd_process_restore_queues(struct kfd_process *p)
1696{
1697 int r, ret = 0;
1698 int i;
1699
1700 for (i = 0; i < p->n_pdds; i++) {
1701 struct kfd_process_device *pdd = p->pdds[i];
1702
1703 r = pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm,
1704 &pdd->qpd);
1705 if (r) {
1706 pr_err("Failed to restore process queues\n");
1707 if (!ret)
1708 ret = r;
1709 }
1710 }
1711
1712 return ret;
1713}
1714
1715int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id)
1716{
1717 int i;
1718
1719 for (i = 0; i < p->n_pdds; i++)
1720 if (p->pdds[i] && gpu_id == p->pdds[i]->dev->id)
1721 return i;
1722 return -EINVAL;
1723}
1724
1725int
1726kfd_process_gpuid_from_kgd(struct kfd_process *p, struct amdgpu_device *adev,
1727 uint32_t *gpuid, uint32_t *gpuidx)
1728{
1729 struct kgd_dev *kgd = (struct kgd_dev *)adev;
1730 int i;
1731
1732 for (i = 0; i < p->n_pdds; i++)
1733 if (p->pdds[i] && p->pdds[i]->dev->kgd == kgd) {
1734 *gpuid = p->pdds[i]->dev->id;
1735 *gpuidx = i;
1736 return 0;
1737 }
1738 return -EINVAL;
1739}
1740
1741static void evict_process_worker(struct work_struct *work)
1742{
1743 int ret;
1744 struct kfd_process *p;
1745 struct delayed_work *dwork;
1746
1747 dwork = to_delayed_work(work);
1748
1749 /* Process termination destroys this worker thread. So during the
1750 * lifetime of this thread, kfd_process p will be valid
1751 */
1752 p = container_of(dwork, struct kfd_process, eviction_work);
1753 WARN_ONCE(p->last_eviction_seqno != p->ef->seqno,
1754 "Eviction fence mismatch\n");
1755
1756 /* Narrow window of overlap between restore and evict work
1757 * item is possible. Once amdgpu_amdkfd_gpuvm_restore_process_bos
1758 * unreserves KFD BOs, it is possible to evicted again. But
1759 * restore has few more steps of finish. So lets wait for any
1760 * previous restore work to complete
1761 */
1762 flush_delayed_work(&p->restore_work);
1763
1764 pr_debug("Started evicting pasid 0x%x\n", p->pasid);
1765 ret = kfd_process_evict_queues(p);
1766 if (!ret) {
1767 dma_fence_signal(p->ef);
1768 dma_fence_put(p->ef);
1769 p->ef = NULL;
1770 queue_delayed_work(kfd_restore_wq, &p->restore_work,
1771 msecs_to_jiffies(PROCESS_RESTORE_TIME_MS));
1772
1773 pr_debug("Finished evicting pasid 0x%x\n", p->pasid);
1774 } else
1775 pr_err("Failed to evict queues of pasid 0x%x\n", p->pasid);
1776}
1777
1778static void restore_process_worker(struct work_struct *work)
1779{
1780 struct delayed_work *dwork;
1781 struct kfd_process *p;
1782 int ret = 0;
1783
1784 dwork = to_delayed_work(work);
1785
1786 /* Process termination destroys this worker thread. So during the
1787 * lifetime of this thread, kfd_process p will be valid
1788 */
1789 p = container_of(dwork, struct kfd_process, restore_work);
1790 pr_debug("Started restoring pasid 0x%x\n", p->pasid);
1791
1792 /* Setting last_restore_timestamp before successful restoration.
1793 * Otherwise this would have to be set by KGD (restore_process_bos)
1794 * before KFD BOs are unreserved. If not, the process can be evicted
1795 * again before the timestamp is set.
1796 * If restore fails, the timestamp will be set again in the next
1797 * attempt. This would mean that the minimum GPU quanta would be
1798 * PROCESS_ACTIVE_TIME_MS - (time to execute the following two
1799 * functions)
1800 */
1801
1802 p->last_restore_timestamp = get_jiffies_64();
1803 ret = amdgpu_amdkfd_gpuvm_restore_process_bos(p->kgd_process_info,
1804 &p->ef);
1805 if (ret) {
1806 pr_debug("Failed to restore BOs of pasid 0x%x, retry after %d ms\n",
1807 p->pasid, PROCESS_BACK_OFF_TIME_MS);
1808 ret = queue_delayed_work(kfd_restore_wq, &p->restore_work,
1809 msecs_to_jiffies(PROCESS_BACK_OFF_TIME_MS));
1810 WARN(!ret, "reschedule restore work failed\n");
1811 return;
1812 }
1813
1814 ret = kfd_process_restore_queues(p);
1815 if (!ret)
1816 pr_debug("Finished restoring pasid 0x%x\n", p->pasid);
1817 else
1818 pr_err("Failed to restore queues of pasid 0x%x\n", p->pasid);
1819}
1820
1821void kfd_suspend_all_processes(void)
1822{
1823 struct kfd_process *p;
1824 unsigned int temp;
1825 int idx = srcu_read_lock(&kfd_processes_srcu);
1826
1827 WARN(debug_evictions, "Evicting all processes");
1828 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1829 cancel_delayed_work_sync(&p->eviction_work);
1830 cancel_delayed_work_sync(&p->restore_work);
1831
1832 if (kfd_process_evict_queues(p))
1833 pr_err("Failed to suspend process 0x%x\n", p->pasid);
1834 dma_fence_signal(p->ef);
1835 dma_fence_put(p->ef);
1836 p->ef = NULL;
1837 }
1838 srcu_read_unlock(&kfd_processes_srcu, idx);
1839}
1840
1841int kfd_resume_all_processes(void)
1842{
1843 struct kfd_process *p;
1844 unsigned int temp;
1845 int ret = 0, idx = srcu_read_lock(&kfd_processes_srcu);
1846
1847 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1848 if (!queue_delayed_work(kfd_restore_wq, &p->restore_work, 0)) {
1849 pr_err("Restore process %d failed during resume\n",
1850 p->pasid);
1851 ret = -EFAULT;
1852 }
1853 }
1854 srcu_read_unlock(&kfd_processes_srcu, idx);
1855 return ret;
1856}
1857
1858int kfd_reserved_mem_mmap(struct kfd_dev *dev, struct kfd_process *process,
1859 struct vm_area_struct *vma)
1860{
1861 struct kfd_process_device *pdd;
1862 struct qcm_process_device *qpd;
1863
1864 if ((vma->vm_end - vma->vm_start) != KFD_CWSR_TBA_TMA_SIZE) {
1865 pr_err("Incorrect CWSR mapping size.\n");
1866 return -EINVAL;
1867 }
1868
1869 pdd = kfd_get_process_device_data(dev, process);
1870 if (!pdd)
1871 return -EINVAL;
1872 qpd = &pdd->qpd;
1873
1874 qpd->cwsr_kaddr = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
1875 get_order(KFD_CWSR_TBA_TMA_SIZE));
1876 if (!qpd->cwsr_kaddr) {
1877 pr_err("Error allocating per process CWSR buffer.\n");
1878 return -ENOMEM;
1879 }
1880
1881 vma->vm_flags |= VM_IO | VM_DONTCOPY | VM_DONTEXPAND
1882 | VM_NORESERVE | VM_DONTDUMP | VM_PFNMAP;
1883 /* Mapping pages to user process */
1884 return remap_pfn_range(vma, vma->vm_start,
1885 PFN_DOWN(__pa(qpd->cwsr_kaddr)),
1886 KFD_CWSR_TBA_TMA_SIZE, vma->vm_page_prot);
1887}
1888
1889void kfd_flush_tlb(struct kfd_process_device *pdd, enum TLB_FLUSH_TYPE type)
1890{
1891 struct kfd_dev *dev = pdd->dev;
1892
1893 if (dev->dqm->sched_policy == KFD_SCHED_POLICY_NO_HWS) {
1894 /* Nothing to flush until a VMID is assigned, which
1895 * only happens when the first queue is created.
1896 */
1897 if (pdd->qpd.vmid)
1898 amdgpu_amdkfd_flush_gpu_tlb_vmid(dev->kgd,
1899 pdd->qpd.vmid);
1900 } else {
1901 amdgpu_amdkfd_flush_gpu_tlb_pasid(dev->kgd,
1902 pdd->process->pasid, type);
1903 }
1904}
1905
1906#if defined(CONFIG_DEBUG_FS)
1907
1908int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data)
1909{
1910 struct kfd_process *p;
1911 unsigned int temp;
1912 int r = 0;
1913
1914 int idx = srcu_read_lock(&kfd_processes_srcu);
1915
1916 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1917 seq_printf(m, "Process %d PASID 0x%x:\n",
1918 p->lead_thread->tgid, p->pasid);
1919
1920 mutex_lock(&p->mutex);
1921 r = pqm_debugfs_mqds(m, &p->pqm);
1922 mutex_unlock(&p->mutex);
1923
1924 if (r)
1925 break;
1926 }
1927
1928 srcu_read_unlock(&kfd_processes_srcu, idx);
1929
1930 return r;
1931}
1932
1933#endif
1934