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// SPDX-License-Identifier: GPL-2.0 OR MIT
2/*
3 * Copyright 2014-2022 Advanced Micro Devices, Inc.
4 *
5 * Permission is hereby granted, free of charge, to any person obtaining a
6 * copy of this software and associated documentation files (the "Software"),
7 * to deal in the Software without restriction, including without limitation
8 * the rights to use, copy, modify, merge, publish, distribute, sublicense,
9 * and/or sell copies of the Software, and to permit persons to whom the
10 * Software is furnished to do so, subject to the following conditions:
11 *
12 * The above copyright notice and this permission notice shall be included in
13 * all copies or substantial portions of the Software.
14 *
15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
18 * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
19 * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
20 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
21 * OTHER DEALINGS IN THE SOFTWARE.
22 */
23
24#include <linux/mutex.h>
25#include <linux/log2.h>
26#include <linux/sched.h>
27#include <linux/sched/mm.h>
28#include <linux/sched/task.h>
29#include <linux/mmu_context.h>
30#include <linux/slab.h>
31#include <linux/notifier.h>
32#include <linux/compat.h>
33#include <linux/mman.h>
34#include <linux/file.h>
35#include <linux/pm_runtime.h>
36#include "amdgpu_amdkfd.h"
37#include "amdgpu.h"
38
39struct mm_struct;
40
41#include "kfd_priv.h"
42#include "kfd_device_queue_manager.h"
43#include "kfd_svm.h"
44#include "kfd_smi_events.h"
45#include "kfd_debug.h"
46
47/*
48 * List of struct kfd_process (field kfd_process).
49 * Unique/indexed by mm_struct*
50 */
51DEFINE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
52DEFINE_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,
68 bool ref);
69static void kfd_process_ref_release(struct kref *ref);
70static struct kfd_process *create_process(const struct task_struct *thread);
71
72static void evict_process_worker(struct work_struct *work);
73static void restore_process_worker(struct work_struct *work);
74
75static void kfd_process_device_destroy_cwsr_dgpu(struct kfd_process_device *pdd);
76
77struct kfd_procfs_tree {
78 struct kobject *kobj;
79};
80
81static struct kfd_procfs_tree procfs;
82
83/*
84 * Structure for SDMA activity tracking
85 */
86struct kfd_sdma_activity_handler_workarea {
87 struct work_struct sdma_activity_work;
88 struct kfd_process_device *pdd;
89 uint64_t sdma_activity_counter;
90};
91
92struct temp_sdma_queue_list {
93 uint64_t __user *rptr;
94 uint64_t sdma_val;
95 unsigned int queue_id;
96 struct list_head list;
97};
98
99static void kfd_sdma_activity_worker(struct work_struct *work)
100{
101 struct kfd_sdma_activity_handler_workarea *workarea;
102 struct kfd_process_device *pdd;
103 uint64_t val;
104 struct mm_struct *mm;
105 struct queue *q;
106 struct qcm_process_device *qpd;
107 struct device_queue_manager *dqm;
108 int ret = 0;
109 struct temp_sdma_queue_list sdma_q_list;
110 struct temp_sdma_queue_list *sdma_q, *next;
111
112 workarea = container_of(work, struct kfd_sdma_activity_handler_workarea,
113 sdma_activity_work);
114
115 pdd = workarea->pdd;
116 if (!pdd)
117 return;
118 dqm = pdd->dev->dqm;
119 qpd = &pdd->qpd;
120 if (!dqm || !qpd)
121 return;
122 /*
123 * Total SDMA activity is current SDMA activity + past SDMA activity
124 * Past SDMA count is stored in pdd.
125 * To get the current activity counters for all active SDMA queues,
126 * we loop over all SDMA queues and get their counts from user-space.
127 *
128 * We cannot call get_user() with dqm_lock held as it can cause
129 * a circular lock dependency situation. To read the SDMA stats,
130 * we need to do the following:
131 *
132 * 1. Create a temporary list of SDMA queue nodes from the qpd->queues_list,
133 * with dqm_lock/dqm_unlock().
134 * 2. Call get_user() for each node in temporary list without dqm_lock.
135 * Save the SDMA count for each node and also add the count to the total
136 * SDMA count counter.
137 * Its possible, during this step, a few SDMA queue nodes got deleted
138 * from the qpd->queues_list.
139 * 3. Do a second pass over qpd->queues_list to check if any nodes got deleted.
140 * If any node got deleted, its SDMA count would be captured in the sdma
141 * past activity counter. So subtract the SDMA counter stored in step 2
142 * for this node from the total SDMA count.
143 */
144 INIT_LIST_HEAD(&sdma_q_list.list);
145
146 /*
147 * Create the temp list of all SDMA queues
148 */
149 dqm_lock(dqm);
150
151 list_for_each_entry(q, &qpd->queues_list, list) {
152 if ((q->properties.type != KFD_QUEUE_TYPE_SDMA) &&
153 (q->properties.type != KFD_QUEUE_TYPE_SDMA_XGMI))
154 continue;
155
156 sdma_q = kzalloc(sizeof(struct temp_sdma_queue_list), GFP_KERNEL);
157 if (!sdma_q) {
158 dqm_unlock(dqm);
159 goto cleanup;
160 }
161
162 INIT_LIST_HEAD(&sdma_q->list);
163 sdma_q->rptr = (uint64_t __user *)q->properties.read_ptr;
164 sdma_q->queue_id = q->properties.queue_id;
165 list_add_tail(&sdma_q->list, &sdma_q_list.list);
166 }
167
168 /*
169 * If the temp list is empty, then no SDMA queues nodes were found in
170 * qpd->queues_list. Return the past activity count as the total sdma
171 * count
172 */
173 if (list_empty(&sdma_q_list.list)) {
174 workarea->sdma_activity_counter = pdd->sdma_past_activity_counter;
175 dqm_unlock(dqm);
176 return;
177 }
178
179 dqm_unlock(dqm);
180
181 /*
182 * Get the usage count for each SDMA queue in temp_list.
183 */
184 mm = get_task_mm(pdd->process->lead_thread);
185 if (!mm)
186 goto cleanup;
187
188 kthread_use_mm(mm);
189
190 list_for_each_entry(sdma_q, &sdma_q_list.list, list) {
191 val = 0;
192 ret = read_sdma_queue_counter(sdma_q->rptr, &val);
193 if (ret) {
194 pr_debug("Failed to read SDMA queue active counter for queue id: %d",
195 sdma_q->queue_id);
196 } else {
197 sdma_q->sdma_val = val;
198 workarea->sdma_activity_counter += val;
199 }
200 }
201
202 kthread_unuse_mm(mm);
203 mmput(mm);
204
205 /*
206 * Do a second iteration over qpd_queues_list to check if any SDMA
207 * nodes got deleted while fetching SDMA counter.
208 */
209 dqm_lock(dqm);
210
211 workarea->sdma_activity_counter += pdd->sdma_past_activity_counter;
212
213 list_for_each_entry(q, &qpd->queues_list, list) {
214 if (list_empty(&sdma_q_list.list))
215 break;
216
217 if ((q->properties.type != KFD_QUEUE_TYPE_SDMA) &&
218 (q->properties.type != KFD_QUEUE_TYPE_SDMA_XGMI))
219 continue;
220
221 list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
222 if (((uint64_t __user *)q->properties.read_ptr == sdma_q->rptr) &&
223 (sdma_q->queue_id == q->properties.queue_id)) {
224 list_del(&sdma_q->list);
225 kfree(sdma_q);
226 break;
227 }
228 }
229 }
230
231 dqm_unlock(dqm);
232
233 /*
234 * If temp list is not empty, it implies some queues got deleted
235 * from qpd->queues_list during SDMA usage read. Subtract the SDMA
236 * count for each node from the total SDMA count.
237 */
238 list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
239 workarea->sdma_activity_counter -= sdma_q->sdma_val;
240 list_del(&sdma_q->list);
241 kfree(sdma_q);
242 }
243
244 return;
245
246cleanup:
247 list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
248 list_del(&sdma_q->list);
249 kfree(sdma_q);
250 }
251}
252
253/**
254 * kfd_get_cu_occupancy - Collect number of waves in-flight on this device
255 * by current process. Translates acquired wave count into number of compute units
256 * that are occupied.
257 *
258 * @attr: Handle of attribute that allows reporting of wave count. The attribute
259 * handle encapsulates GPU device it is associated with, thereby allowing collection
260 * of waves in flight, etc
261 * @buffer: Handle of user provided buffer updated with wave count
262 *
263 * Return: Number of bytes written to user buffer or an error value
264 */
265static int kfd_get_cu_occupancy(struct attribute *attr, char *buffer)
266{
267 int cu_cnt;
268 int wave_cnt;
269 int max_waves_per_cu;
270 struct kfd_node *dev = NULL;
271 struct kfd_process *proc = NULL;
272 struct kfd_process_device *pdd = NULL;
273 int i;
274 struct kfd_cu_occupancy *cu_occupancy;
275 u32 queue_format;
276
277 pdd = container_of(attr, struct kfd_process_device, attr_cu_occupancy);
278 dev = pdd->dev;
279 if (dev->kfd2kgd->get_cu_occupancy == NULL)
280 return -EINVAL;
281
282 cu_cnt = 0;
283 proc = pdd->process;
284 if (pdd->qpd.queue_count == 0) {
285 pr_debug("Gpu-Id: %d has no active queues for process %d\n",
286 dev->id, proc->pasid);
287 return snprintf(buffer, PAGE_SIZE, "%d\n", cu_cnt);
288 }
289
290 /* Collect wave count from device if it supports */
291 wave_cnt = 0;
292 max_waves_per_cu = 0;
293
294 cu_occupancy = kcalloc(AMDGPU_MAX_QUEUES, sizeof(*cu_occupancy), GFP_KERNEL);
295 if (!cu_occupancy)
296 return -ENOMEM;
297
298 /*
299 * For GFX 9.4.3, fetch the CU occupancy from the first XCC in the partition.
300 * For AQL queues, because of cooperative dispatch we multiply the wave count
301 * by number of XCCs in the partition to get the total wave counts across all
302 * XCCs in the partition.
303 * For PM4 queues, there is no cooperative dispatch so wave_cnt stay as it is.
304 */
305 dev->kfd2kgd->get_cu_occupancy(dev->adev, cu_occupancy,
306 &max_waves_per_cu, ffs(dev->xcc_mask) - 1);
307
308 for (i = 0; i < AMDGPU_MAX_QUEUES; i++) {
309 if (cu_occupancy[i].wave_cnt != 0 &&
310 kfd_dqm_is_queue_in_process(dev->dqm, &pdd->qpd,
311 cu_occupancy[i].doorbell_off,
312 &queue_format)) {
313 if (unlikely(queue_format == KFD_QUEUE_FORMAT_PM4))
314 wave_cnt += cu_occupancy[i].wave_cnt;
315 else
316 wave_cnt += (NUM_XCC(dev->xcc_mask) *
317 cu_occupancy[i].wave_cnt);
318 }
319 }
320
321 /* Translate wave count to number of compute units */
322 cu_cnt = (wave_cnt + (max_waves_per_cu - 1)) / max_waves_per_cu;
323 kfree(cu_occupancy);
324 return snprintf(buffer, PAGE_SIZE, "%d\n", cu_cnt);
325}
326
327static ssize_t kfd_procfs_show(struct kobject *kobj, struct attribute *attr,
328 char *buffer)
329{
330 if (strcmp(attr->name, "pasid") == 0) {
331 struct kfd_process *p = container_of(attr, struct kfd_process,
332 attr_pasid);
333
334 return snprintf(buffer, PAGE_SIZE, "%d\n", p->pasid);
335 } else if (strncmp(attr->name, "vram_", 5) == 0) {
336 struct kfd_process_device *pdd = container_of(attr, struct kfd_process_device,
337 attr_vram);
338 return snprintf(buffer, PAGE_SIZE, "%llu\n", atomic64_read(&pdd->vram_usage));
339 } else if (strncmp(attr->name, "sdma_", 5) == 0) {
340 struct kfd_process_device *pdd = container_of(attr, struct kfd_process_device,
341 attr_sdma);
342 struct kfd_sdma_activity_handler_workarea sdma_activity_work_handler;
343
344 INIT_WORK_ONSTACK(&sdma_activity_work_handler.sdma_activity_work,
345 kfd_sdma_activity_worker);
346
347 sdma_activity_work_handler.pdd = pdd;
348 sdma_activity_work_handler.sdma_activity_counter = 0;
349
350 schedule_work(&sdma_activity_work_handler.sdma_activity_work);
351
352 flush_work(&sdma_activity_work_handler.sdma_activity_work);
353 destroy_work_on_stack(&sdma_activity_work_handler.sdma_activity_work);
354
355 return snprintf(buffer, PAGE_SIZE, "%llu\n",
356 (sdma_activity_work_handler.sdma_activity_counter)/
357 SDMA_ACTIVITY_DIVISOR);
358 } else {
359 pr_err("Invalid attribute");
360 return -EINVAL;
361 }
362
363 return 0;
364}
365
366static void kfd_procfs_kobj_release(struct kobject *kobj)
367{
368 kfree(kobj);
369}
370
371static const struct sysfs_ops kfd_procfs_ops = {
372 .show = kfd_procfs_show,
373};
374
375static const struct kobj_type procfs_type = {
376 .release = kfd_procfs_kobj_release,
377 .sysfs_ops = &kfd_procfs_ops,
378};
379
380void kfd_procfs_init(void)
381{
382 int ret = 0;
383
384 procfs.kobj = kfd_alloc_struct(procfs.kobj);
385 if (!procfs.kobj)
386 return;
387
388 ret = kobject_init_and_add(procfs.kobj, &procfs_type,
389 &kfd_device->kobj, "proc");
390 if (ret) {
391 pr_warn("Could not create procfs proc folder");
392 /* If we fail to create the procfs, clean up */
393 kfd_procfs_shutdown();
394 }
395}
396
397void kfd_procfs_shutdown(void)
398{
399 if (procfs.kobj) {
400 kobject_del(procfs.kobj);
401 kobject_put(procfs.kobj);
402 procfs.kobj = NULL;
403 }
404}
405
406static ssize_t kfd_procfs_queue_show(struct kobject *kobj,
407 struct attribute *attr, char *buffer)
408{
409 struct queue *q = container_of(kobj, struct queue, kobj);
410
411 if (!strcmp(attr->name, "size"))
412 return snprintf(buffer, PAGE_SIZE, "%llu",
413 q->properties.queue_size);
414 else if (!strcmp(attr->name, "type"))
415 return snprintf(buffer, PAGE_SIZE, "%d", q->properties.type);
416 else if (!strcmp(attr->name, "gpuid"))
417 return snprintf(buffer, PAGE_SIZE, "%u", q->device->id);
418 else
419 pr_err("Invalid attribute");
420
421 return 0;
422}
423
424static ssize_t kfd_procfs_stats_show(struct kobject *kobj,
425 struct attribute *attr, char *buffer)
426{
427 if (strcmp(attr->name, "evicted_ms") == 0) {
428 struct kfd_process_device *pdd = container_of(attr,
429 struct kfd_process_device,
430 attr_evict);
431 uint64_t evict_jiffies;
432
433 evict_jiffies = atomic64_read(&pdd->evict_duration_counter);
434
435 return snprintf(buffer,
436 PAGE_SIZE,
437 "%llu\n",
438 jiffies64_to_msecs(evict_jiffies));
439
440 /* Sysfs handle that gets CU occupancy is per device */
441 } else if (strcmp(attr->name, "cu_occupancy") == 0) {
442 return kfd_get_cu_occupancy(attr, buffer);
443 } else {
444 pr_err("Invalid attribute");
445 }
446
447 return 0;
448}
449
450static ssize_t kfd_sysfs_counters_show(struct kobject *kobj,
451 struct attribute *attr, char *buf)
452{
453 struct kfd_process_device *pdd;
454
455 if (!strcmp(attr->name, "faults")) {
456 pdd = container_of(attr, struct kfd_process_device,
457 attr_faults);
458 return sysfs_emit(buf, "%llu\n", READ_ONCE(pdd->faults));
459 }
460 if (!strcmp(attr->name, "page_in")) {
461 pdd = container_of(attr, struct kfd_process_device,
462 attr_page_in);
463 return sysfs_emit(buf, "%llu\n", READ_ONCE(pdd->page_in));
464 }
465 if (!strcmp(attr->name, "page_out")) {
466 pdd = container_of(attr, struct kfd_process_device,
467 attr_page_out);
468 return sysfs_emit(buf, "%llu\n", READ_ONCE(pdd->page_out));
469 }
470 return 0;
471}
472
473static struct attribute attr_queue_size = {
474 .name = "size",
475 .mode = KFD_SYSFS_FILE_MODE
476};
477
478static struct attribute attr_queue_type = {
479 .name = "type",
480 .mode = KFD_SYSFS_FILE_MODE
481};
482
483static struct attribute attr_queue_gpuid = {
484 .name = "gpuid",
485 .mode = KFD_SYSFS_FILE_MODE
486};
487
488static struct attribute *procfs_queue_attrs[] = {
489 &attr_queue_size,
490 &attr_queue_type,
491 &attr_queue_gpuid,
492 NULL
493};
494ATTRIBUTE_GROUPS(procfs_queue);
495
496static const struct sysfs_ops procfs_queue_ops = {
497 .show = kfd_procfs_queue_show,
498};
499
500static const struct kobj_type procfs_queue_type = {
501 .sysfs_ops = &procfs_queue_ops,
502 .default_groups = procfs_queue_groups,
503};
504
505static const struct sysfs_ops procfs_stats_ops = {
506 .show = kfd_procfs_stats_show,
507};
508
509static const struct kobj_type procfs_stats_type = {
510 .sysfs_ops = &procfs_stats_ops,
511 .release = kfd_procfs_kobj_release,
512};
513
514static const struct sysfs_ops sysfs_counters_ops = {
515 .show = kfd_sysfs_counters_show,
516};
517
518static const struct kobj_type sysfs_counters_type = {
519 .sysfs_ops = &sysfs_counters_ops,
520 .release = kfd_procfs_kobj_release,
521};
522
523int kfd_procfs_add_queue(struct queue *q)
524{
525 struct kfd_process *proc;
526 int ret;
527
528 if (!q || !q->process)
529 return -EINVAL;
530 proc = q->process;
531
532 /* Create proc/<pid>/queues/<queue id> folder */
533 if (!proc->kobj_queues)
534 return -EFAULT;
535 ret = kobject_init_and_add(&q->kobj, &procfs_queue_type,
536 proc->kobj_queues, "%u", q->properties.queue_id);
537 if (ret < 0) {
538 pr_warn("Creating proc/<pid>/queues/%u failed",
539 q->properties.queue_id);
540 kobject_put(&q->kobj);
541 return ret;
542 }
543
544 return 0;
545}
546
547static void kfd_sysfs_create_file(struct kobject *kobj, struct attribute *attr,
548 char *name)
549{
550 int ret;
551
552 if (!kobj || !attr || !name)
553 return;
554
555 attr->name = name;
556 attr->mode = KFD_SYSFS_FILE_MODE;
557 sysfs_attr_init(attr);
558
559 ret = sysfs_create_file(kobj, attr);
560 if (ret)
561 pr_warn("Create sysfs %s/%s failed %d", kobj->name, name, ret);
562}
563
564static void kfd_procfs_add_sysfs_stats(struct kfd_process *p)
565{
566 int ret;
567 int i;
568 char stats_dir_filename[MAX_SYSFS_FILENAME_LEN];
569
570 if (!p || !p->kobj)
571 return;
572
573 /*
574 * Create sysfs files for each GPU:
575 * - proc/<pid>/stats_<gpuid>/
576 * - proc/<pid>/stats_<gpuid>/evicted_ms
577 * - proc/<pid>/stats_<gpuid>/cu_occupancy
578 */
579 for (i = 0; i < p->n_pdds; i++) {
580 struct kfd_process_device *pdd = p->pdds[i];
581
582 snprintf(stats_dir_filename, MAX_SYSFS_FILENAME_LEN,
583 "stats_%u", pdd->dev->id);
584 pdd->kobj_stats = kfd_alloc_struct(pdd->kobj_stats);
585 if (!pdd->kobj_stats)
586 return;
587
588 ret = kobject_init_and_add(pdd->kobj_stats,
589 &procfs_stats_type,
590 p->kobj,
591 stats_dir_filename);
592
593 if (ret) {
594 pr_warn("Creating KFD proc/stats_%s folder failed",
595 stats_dir_filename);
596 kobject_put(pdd->kobj_stats);
597 pdd->kobj_stats = NULL;
598 return;
599 }
600
601 kfd_sysfs_create_file(pdd->kobj_stats, &pdd->attr_evict,
602 "evicted_ms");
603 /* Add sysfs file to report compute unit occupancy */
604 if (pdd->dev->kfd2kgd->get_cu_occupancy)
605 kfd_sysfs_create_file(pdd->kobj_stats,
606 &pdd->attr_cu_occupancy,
607 "cu_occupancy");
608 }
609}
610
611static void kfd_procfs_add_sysfs_counters(struct kfd_process *p)
612{
613 int ret = 0;
614 int i;
615 char counters_dir_filename[MAX_SYSFS_FILENAME_LEN];
616
617 if (!p || !p->kobj)
618 return;
619
620 /*
621 * Create sysfs files for each GPU which supports SVM
622 * - proc/<pid>/counters_<gpuid>/
623 * - proc/<pid>/counters_<gpuid>/faults
624 * - proc/<pid>/counters_<gpuid>/page_in
625 * - proc/<pid>/counters_<gpuid>/page_out
626 */
627 for_each_set_bit(i, p->svms.bitmap_supported, p->n_pdds) {
628 struct kfd_process_device *pdd = p->pdds[i];
629 struct kobject *kobj_counters;
630
631 snprintf(counters_dir_filename, MAX_SYSFS_FILENAME_LEN,
632 "counters_%u", pdd->dev->id);
633 kobj_counters = kfd_alloc_struct(kobj_counters);
634 if (!kobj_counters)
635 return;
636
637 ret = kobject_init_and_add(kobj_counters, &sysfs_counters_type,
638 p->kobj, counters_dir_filename);
639 if (ret) {
640 pr_warn("Creating KFD proc/%s folder failed",
641 counters_dir_filename);
642 kobject_put(kobj_counters);
643 return;
644 }
645
646 pdd->kobj_counters = kobj_counters;
647 kfd_sysfs_create_file(kobj_counters, &pdd->attr_faults,
648 "faults");
649 kfd_sysfs_create_file(kobj_counters, &pdd->attr_page_in,
650 "page_in");
651 kfd_sysfs_create_file(kobj_counters, &pdd->attr_page_out,
652 "page_out");
653 }
654}
655
656static void kfd_procfs_add_sysfs_files(struct kfd_process *p)
657{
658 int i;
659
660 if (!p || !p->kobj)
661 return;
662
663 /*
664 * Create sysfs files for each GPU:
665 * - proc/<pid>/vram_<gpuid>
666 * - proc/<pid>/sdma_<gpuid>
667 */
668 for (i = 0; i < p->n_pdds; i++) {
669 struct kfd_process_device *pdd = p->pdds[i];
670
671 snprintf(pdd->vram_filename, MAX_SYSFS_FILENAME_LEN, "vram_%u",
672 pdd->dev->id);
673 kfd_sysfs_create_file(p->kobj, &pdd->attr_vram,
674 pdd->vram_filename);
675
676 snprintf(pdd->sdma_filename, MAX_SYSFS_FILENAME_LEN, "sdma_%u",
677 pdd->dev->id);
678 kfd_sysfs_create_file(p->kobj, &pdd->attr_sdma,
679 pdd->sdma_filename);
680 }
681}
682
683void kfd_procfs_del_queue(struct queue *q)
684{
685 if (!q)
686 return;
687
688 kobject_del(&q->kobj);
689 kobject_put(&q->kobj);
690}
691
692int kfd_process_create_wq(void)
693{
694 if (!kfd_process_wq)
695 kfd_process_wq = alloc_workqueue("kfd_process_wq", 0, 0);
696 if (!kfd_restore_wq)
697 kfd_restore_wq = alloc_ordered_workqueue("kfd_restore_wq",
698 WQ_FREEZABLE);
699
700 if (!kfd_process_wq || !kfd_restore_wq) {
701 kfd_process_destroy_wq();
702 return -ENOMEM;
703 }
704
705 return 0;
706}
707
708void kfd_process_destroy_wq(void)
709{
710 if (kfd_process_wq) {
711 destroy_workqueue(kfd_process_wq);
712 kfd_process_wq = NULL;
713 }
714 if (kfd_restore_wq) {
715 destroy_workqueue(kfd_restore_wq);
716 kfd_restore_wq = NULL;
717 }
718}
719
720static void kfd_process_free_gpuvm(struct kgd_mem *mem,
721 struct kfd_process_device *pdd, void **kptr)
722{
723 struct kfd_node *dev = pdd->dev;
724
725 if (kptr && *kptr) {
726 amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(mem);
727 *kptr = NULL;
728 }
729
730 amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(dev->adev, mem, pdd->drm_priv);
731 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->adev, mem, pdd->drm_priv,
732 NULL);
733}
734
735/* kfd_process_alloc_gpuvm - Allocate GPU VM for the KFD process
736 * This function should be only called right after the process
737 * is created and when kfd_processes_mutex is still being held
738 * to avoid concurrency. Because of that exclusiveness, we do
739 * not need to take p->mutex.
740 */
741static int kfd_process_alloc_gpuvm(struct kfd_process_device *pdd,
742 uint64_t gpu_va, uint32_t size,
743 uint32_t flags, struct kgd_mem **mem, void **kptr)
744{
745 struct kfd_node *kdev = pdd->dev;
746 int err;
747
748 err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(kdev->adev, gpu_va, size,
749 pdd->drm_priv, mem, NULL,
750 flags, false);
751 if (err)
752 goto err_alloc_mem;
753
754 err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(kdev->adev, *mem,
755 pdd->drm_priv);
756 if (err)
757 goto err_map_mem;
758
759 err = amdgpu_amdkfd_gpuvm_sync_memory(kdev->adev, *mem, true);
760 if (err) {
761 pr_debug("Sync memory failed, wait interrupted by user signal\n");
762 goto sync_memory_failed;
763 }
764
765 if (kptr) {
766 err = amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(
767 (struct kgd_mem *)*mem, kptr, NULL);
768 if (err) {
769 pr_debug("Map GTT BO to kernel failed\n");
770 goto sync_memory_failed;
771 }
772 }
773
774 return err;
775
776sync_memory_failed:
777 amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(kdev->adev, *mem, pdd->drm_priv);
778
779err_map_mem:
780 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(kdev->adev, *mem, pdd->drm_priv,
781 NULL);
782err_alloc_mem:
783 *mem = NULL;
784 *kptr = NULL;
785 return err;
786}
787
788/* kfd_process_device_reserve_ib_mem - Reserve memory inside the
789 * process for IB usage The memory reserved is for KFD to submit
790 * IB to AMDGPU from kernel. If the memory is reserved
791 * successfully, ib_kaddr will have the CPU/kernel
792 * address. Check ib_kaddr before accessing the memory.
793 */
794static int kfd_process_device_reserve_ib_mem(struct kfd_process_device *pdd)
795{
796 struct qcm_process_device *qpd = &pdd->qpd;
797 uint32_t flags = KFD_IOC_ALLOC_MEM_FLAGS_GTT |
798 KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE |
799 KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE |
800 KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
801 struct kgd_mem *mem;
802 void *kaddr;
803 int ret;
804
805 if (qpd->ib_kaddr || !qpd->ib_base)
806 return 0;
807
808 /* ib_base is only set for dGPU */
809 ret = kfd_process_alloc_gpuvm(pdd, qpd->ib_base, PAGE_SIZE, flags,
810 &mem, &kaddr);
811 if (ret)
812 return ret;
813
814 qpd->ib_mem = mem;
815 qpd->ib_kaddr = kaddr;
816
817 return 0;
818}
819
820static void kfd_process_device_destroy_ib_mem(struct kfd_process_device *pdd)
821{
822 struct qcm_process_device *qpd = &pdd->qpd;
823
824 if (!qpd->ib_kaddr || !qpd->ib_base)
825 return;
826
827 kfd_process_free_gpuvm(qpd->ib_mem, pdd, &qpd->ib_kaddr);
828}
829
830struct kfd_process *kfd_create_process(struct task_struct *thread)
831{
832 struct kfd_process *process;
833 int ret;
834
835 if (!(thread->mm && mmget_not_zero(thread->mm)))
836 return ERR_PTR(-EINVAL);
837
838 /* Only the pthreads threading model is supported. */
839 if (thread->group_leader->mm != thread->mm) {
840 mmput(thread->mm);
841 return ERR_PTR(-EINVAL);
842 }
843
844 /*
845 * take kfd processes mutex before starting of process creation
846 * so there won't be a case where two threads of the same process
847 * create two kfd_process structures
848 */
849 mutex_lock(&kfd_processes_mutex);
850
851 if (kfd_is_locked()) {
852 pr_debug("KFD is locked! Cannot create process");
853 process = ERR_PTR(-EINVAL);
854 goto out;
855 }
856
857 /* A prior open of /dev/kfd could have already created the process.
858 * find_process will increase process kref in this case
859 */
860 process = find_process(thread, true);
861 if (process) {
862 pr_debug("Process already found\n");
863 } else {
864 /* If the process just called exec(3), it is possible that the
865 * cleanup of the kfd_process (following the release of the mm
866 * of the old process image) is still in the cleanup work queue.
867 * Make sure to drain any job before trying to recreate any
868 * resource for this process.
869 */
870 flush_workqueue(kfd_process_wq);
871
872 process = create_process(thread);
873 if (IS_ERR(process))
874 goto out;
875
876 if (!procfs.kobj)
877 goto out;
878
879 process->kobj = kfd_alloc_struct(process->kobj);
880 if (!process->kobj) {
881 pr_warn("Creating procfs kobject failed");
882 goto out;
883 }
884 ret = kobject_init_and_add(process->kobj, &procfs_type,
885 procfs.kobj, "%d",
886 (int)process->lead_thread->pid);
887 if (ret) {
888 pr_warn("Creating procfs pid directory failed");
889 kobject_put(process->kobj);
890 goto out;
891 }
892
893 kfd_sysfs_create_file(process->kobj, &process->attr_pasid,
894 "pasid");
895
896 process->kobj_queues = kobject_create_and_add("queues",
897 process->kobj);
898 if (!process->kobj_queues)
899 pr_warn("Creating KFD proc/queues folder failed");
900
901 kfd_procfs_add_sysfs_stats(process);
902 kfd_procfs_add_sysfs_files(process);
903 kfd_procfs_add_sysfs_counters(process);
904
905 init_waitqueue_head(&process->wait_irq_drain);
906 }
907out:
908 mutex_unlock(&kfd_processes_mutex);
909 mmput(thread->mm);
910
911 return process;
912}
913
914struct kfd_process *kfd_get_process(const struct task_struct *thread)
915{
916 struct kfd_process *process;
917
918 if (!thread->mm)
919 return ERR_PTR(-EINVAL);
920
921 /* Only the pthreads threading model is supported. */
922 if (thread->group_leader->mm != thread->mm)
923 return ERR_PTR(-EINVAL);
924
925 process = find_process(thread, false);
926 if (!process)
927 return ERR_PTR(-EINVAL);
928
929 return process;
930}
931
932static struct kfd_process *find_process_by_mm(const struct mm_struct *mm)
933{
934 struct kfd_process *process;
935
936 hash_for_each_possible_rcu(kfd_processes_table, process,
937 kfd_processes, (uintptr_t)mm)
938 if (process->mm == mm)
939 return process;
940
941 return NULL;
942}
943
944static struct kfd_process *find_process(const struct task_struct *thread,
945 bool ref)
946{
947 struct kfd_process *p;
948 int idx;
949
950 idx = srcu_read_lock(&kfd_processes_srcu);
951 p = find_process_by_mm(thread->mm);
952 if (p && ref)
953 kref_get(&p->ref);
954 srcu_read_unlock(&kfd_processes_srcu, idx);
955
956 return p;
957}
958
959void kfd_unref_process(struct kfd_process *p)
960{
961 kref_put(&p->ref, kfd_process_ref_release);
962}
963
964/* This increments the process->ref counter. */
965struct kfd_process *kfd_lookup_process_by_pid(struct pid *pid)
966{
967 struct task_struct *task = NULL;
968 struct kfd_process *p = NULL;
969
970 if (!pid) {
971 task = current;
972 get_task_struct(task);
973 } else {
974 task = get_pid_task(pid, PIDTYPE_PID);
975 }
976
977 if (task) {
978 p = find_process(task, true);
979 put_task_struct(task);
980 }
981
982 return p;
983}
984
985static void kfd_process_device_free_bos(struct kfd_process_device *pdd)
986{
987 struct kfd_process *p = pdd->process;
988 void *mem;
989 int id;
990 int i;
991
992 /*
993 * Remove all handles from idr and release appropriate
994 * local memory object
995 */
996 idr_for_each_entry(&pdd->alloc_idr, mem, id) {
997
998 for (i = 0; i < p->n_pdds; i++) {
999 struct kfd_process_device *peer_pdd = p->pdds[i];
1000
1001 if (!peer_pdd->drm_priv)
1002 continue;
1003 amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
1004 peer_pdd->dev->adev, mem, peer_pdd->drm_priv);
1005 }
1006
1007 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, mem,
1008 pdd->drm_priv, NULL);
1009 kfd_process_device_remove_obj_handle(pdd, id);
1010 }
1011}
1012
1013/*
1014 * Just kunmap and unpin signal BO here. It will be freed in
1015 * kfd_process_free_outstanding_kfd_bos()
1016 */
1017static void kfd_process_kunmap_signal_bo(struct kfd_process *p)
1018{
1019 struct kfd_process_device *pdd;
1020 struct kfd_node *kdev;
1021 void *mem;
1022
1023 kdev = kfd_device_by_id(GET_GPU_ID(p->signal_handle));
1024 if (!kdev)
1025 return;
1026
1027 mutex_lock(&p->mutex);
1028
1029 pdd = kfd_get_process_device_data(kdev, p);
1030 if (!pdd)
1031 goto out;
1032
1033 mem = kfd_process_device_translate_handle(
1034 pdd, GET_IDR_HANDLE(p->signal_handle));
1035 if (!mem)
1036 goto out;
1037
1038 amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(mem);
1039
1040out:
1041 mutex_unlock(&p->mutex);
1042}
1043
1044static void kfd_process_free_outstanding_kfd_bos(struct kfd_process *p)
1045{
1046 int i;
1047
1048 for (i = 0; i < p->n_pdds; i++)
1049 kfd_process_device_free_bos(p->pdds[i]);
1050}
1051
1052static void kfd_process_destroy_pdds(struct kfd_process *p)
1053{
1054 int i;
1055
1056 for (i = 0; i < p->n_pdds; i++) {
1057 struct kfd_process_device *pdd = p->pdds[i];
1058
1059 pr_debug("Releasing pdd (topology id %d) for process (pasid 0x%x)\n",
1060 pdd->dev->id, p->pasid);
1061
1062 kfd_process_device_destroy_cwsr_dgpu(pdd);
1063 kfd_process_device_destroy_ib_mem(pdd);
1064
1065 if (pdd->drm_file) {
1066 amdgpu_amdkfd_gpuvm_release_process_vm(
1067 pdd->dev->adev, pdd->drm_priv);
1068 fput(pdd->drm_file);
1069 }
1070
1071 if (pdd->qpd.cwsr_kaddr && !pdd->qpd.cwsr_base)
1072 free_pages((unsigned long)pdd->qpd.cwsr_kaddr,
1073 get_order(KFD_CWSR_TBA_TMA_SIZE));
1074
1075 idr_destroy(&pdd->alloc_idr);
1076
1077 kfd_free_process_doorbells(pdd->dev->kfd, pdd);
1078
1079 if (pdd->dev->kfd->shared_resources.enable_mes &&
1080 pdd->proc_ctx_cpu_ptr)
1081 amdgpu_amdkfd_free_gtt_mem(pdd->dev->adev,
1082 &pdd->proc_ctx_bo);
1083 /*
1084 * before destroying pdd, make sure to report availability
1085 * for auto suspend
1086 */
1087 if (pdd->runtime_inuse) {
1088 pm_runtime_mark_last_busy(adev_to_drm(pdd->dev->adev)->dev);
1089 pm_runtime_put_autosuspend(adev_to_drm(pdd->dev->adev)->dev);
1090 pdd->runtime_inuse = false;
1091 }
1092
1093 kfree(pdd);
1094 p->pdds[i] = NULL;
1095 }
1096 p->n_pdds = 0;
1097}
1098
1099static void kfd_process_remove_sysfs(struct kfd_process *p)
1100{
1101 struct kfd_process_device *pdd;
1102 int i;
1103
1104 if (!p->kobj)
1105 return;
1106
1107 sysfs_remove_file(p->kobj, &p->attr_pasid);
1108 kobject_del(p->kobj_queues);
1109 kobject_put(p->kobj_queues);
1110 p->kobj_queues = NULL;
1111
1112 for (i = 0; i < p->n_pdds; i++) {
1113 pdd = p->pdds[i];
1114
1115 sysfs_remove_file(p->kobj, &pdd->attr_vram);
1116 sysfs_remove_file(p->kobj, &pdd->attr_sdma);
1117
1118 sysfs_remove_file(pdd->kobj_stats, &pdd->attr_evict);
1119 if (pdd->dev->kfd2kgd->get_cu_occupancy)
1120 sysfs_remove_file(pdd->kobj_stats,
1121 &pdd->attr_cu_occupancy);
1122 kobject_del(pdd->kobj_stats);
1123 kobject_put(pdd->kobj_stats);
1124 pdd->kobj_stats = NULL;
1125 }
1126
1127 for_each_set_bit(i, p->svms.bitmap_supported, p->n_pdds) {
1128 pdd = p->pdds[i];
1129
1130 sysfs_remove_file(pdd->kobj_counters, &pdd->attr_faults);
1131 sysfs_remove_file(pdd->kobj_counters, &pdd->attr_page_in);
1132 sysfs_remove_file(pdd->kobj_counters, &pdd->attr_page_out);
1133 kobject_del(pdd->kobj_counters);
1134 kobject_put(pdd->kobj_counters);
1135 pdd->kobj_counters = NULL;
1136 }
1137
1138 kobject_del(p->kobj);
1139 kobject_put(p->kobj);
1140 p->kobj = NULL;
1141}
1142
1143/* No process locking is needed in this function, because the process
1144 * is not findable any more. We must assume that no other thread is
1145 * using it any more, otherwise we couldn't safely free the process
1146 * structure in the end.
1147 */
1148static void kfd_process_wq_release(struct work_struct *work)
1149{
1150 struct kfd_process *p = container_of(work, struct kfd_process,
1151 release_work);
1152 struct dma_fence *ef;
1153
1154 kfd_process_dequeue_from_all_devices(p);
1155 pqm_uninit(&p->pqm);
1156
1157 /* Signal the eviction fence after user mode queues are
1158 * destroyed. This allows any BOs to be freed without
1159 * triggering pointless evictions or waiting for fences.
1160 */
1161 synchronize_rcu();
1162 ef = rcu_access_pointer(p->ef);
1163 if (ef)
1164 dma_fence_signal(ef);
1165
1166 kfd_process_remove_sysfs(p);
1167
1168 kfd_process_kunmap_signal_bo(p);
1169 kfd_process_free_outstanding_kfd_bos(p);
1170 svm_range_list_fini(p);
1171
1172 kfd_process_destroy_pdds(p);
1173 dma_fence_put(ef);
1174
1175 kfd_event_free_process(p);
1176
1177 kfd_pasid_free(p->pasid);
1178 mutex_destroy(&p->mutex);
1179
1180 put_task_struct(p->lead_thread);
1181
1182 kfree(p);
1183}
1184
1185static void kfd_process_ref_release(struct kref *ref)
1186{
1187 struct kfd_process *p = container_of(ref, struct kfd_process, ref);
1188
1189 INIT_WORK(&p->release_work, kfd_process_wq_release);
1190 queue_work(kfd_process_wq, &p->release_work);
1191}
1192
1193static struct mmu_notifier *kfd_process_alloc_notifier(struct mm_struct *mm)
1194{
1195 /* This increments p->ref counter if kfd process p exists */
1196 struct kfd_process *p = kfd_lookup_process_by_mm(mm);
1197
1198 return p ? &p->mmu_notifier : ERR_PTR(-ESRCH);
1199}
1200
1201static void kfd_process_free_notifier(struct mmu_notifier *mn)
1202{
1203 kfd_unref_process(container_of(mn, struct kfd_process, mmu_notifier));
1204}
1205
1206static void kfd_process_notifier_release_internal(struct kfd_process *p)
1207{
1208 int i;
1209
1210 cancel_delayed_work_sync(&p->eviction_work);
1211 cancel_delayed_work_sync(&p->restore_work);
1212
1213 for (i = 0; i < p->n_pdds; i++) {
1214 struct kfd_process_device *pdd = p->pdds[i];
1215
1216 /* re-enable GFX OFF since runtime enable with ttmp setup disabled it. */
1217 if (!kfd_dbg_is_rlc_restore_supported(pdd->dev) && p->runtime_info.ttmp_setup)
1218 amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
1219 }
1220
1221 /* Indicate to other users that MM is no longer valid */
1222 p->mm = NULL;
1223 kfd_dbg_trap_disable(p);
1224
1225 if (atomic_read(&p->debugged_process_count) > 0) {
1226 struct kfd_process *target;
1227 unsigned int temp;
1228 int idx = srcu_read_lock(&kfd_processes_srcu);
1229
1230 hash_for_each_rcu(kfd_processes_table, temp, target, kfd_processes) {
1231 if (target->debugger_process && target->debugger_process == p) {
1232 mutex_lock_nested(&target->mutex, 1);
1233 kfd_dbg_trap_disable(target);
1234 mutex_unlock(&target->mutex);
1235 if (atomic_read(&p->debugged_process_count) == 0)
1236 break;
1237 }
1238 }
1239
1240 srcu_read_unlock(&kfd_processes_srcu, idx);
1241 }
1242
1243 mmu_notifier_put(&p->mmu_notifier);
1244}
1245
1246static void kfd_process_notifier_release(struct mmu_notifier *mn,
1247 struct mm_struct *mm)
1248{
1249 struct kfd_process *p;
1250
1251 /*
1252 * The kfd_process structure can not be free because the
1253 * mmu_notifier srcu is read locked
1254 */
1255 p = container_of(mn, struct kfd_process, mmu_notifier);
1256 if (WARN_ON(p->mm != mm))
1257 return;
1258
1259 mutex_lock(&kfd_processes_mutex);
1260 /*
1261 * Do early return if table is empty.
1262 *
1263 * This could potentially happen if this function is called concurrently
1264 * by mmu_notifier and by kfd_cleanup_pocesses.
1265 *
1266 */
1267 if (hash_empty(kfd_processes_table)) {
1268 mutex_unlock(&kfd_processes_mutex);
1269 return;
1270 }
1271 hash_del_rcu(&p->kfd_processes);
1272 mutex_unlock(&kfd_processes_mutex);
1273 synchronize_srcu(&kfd_processes_srcu);
1274
1275 kfd_process_notifier_release_internal(p);
1276}
1277
1278static const struct mmu_notifier_ops kfd_process_mmu_notifier_ops = {
1279 .release = kfd_process_notifier_release,
1280 .alloc_notifier = kfd_process_alloc_notifier,
1281 .free_notifier = kfd_process_free_notifier,
1282};
1283
1284/*
1285 * This code handles the case when driver is being unloaded before all
1286 * mm_struct are released. We need to safely free the kfd_process and
1287 * avoid race conditions with mmu_notifier that might try to free them.
1288 *
1289 */
1290void kfd_cleanup_processes(void)
1291{
1292 struct kfd_process *p;
1293 struct hlist_node *p_temp;
1294 unsigned int temp;
1295 HLIST_HEAD(cleanup_list);
1296
1297 /*
1298 * Move all remaining kfd_process from the process table to a
1299 * temp list for processing. Once done, callback from mmu_notifier
1300 * release will not see the kfd_process in the table and do early return,
1301 * avoiding double free issues.
1302 */
1303 mutex_lock(&kfd_processes_mutex);
1304 hash_for_each_safe(kfd_processes_table, temp, p_temp, p, kfd_processes) {
1305 hash_del_rcu(&p->kfd_processes);
1306 synchronize_srcu(&kfd_processes_srcu);
1307 hlist_add_head(&p->kfd_processes, &cleanup_list);
1308 }
1309 mutex_unlock(&kfd_processes_mutex);
1310
1311 hlist_for_each_entry_safe(p, p_temp, &cleanup_list, kfd_processes)
1312 kfd_process_notifier_release_internal(p);
1313
1314 /*
1315 * Ensures that all outstanding free_notifier get called, triggering
1316 * the release of the kfd_process struct.
1317 */
1318 mmu_notifier_synchronize();
1319}
1320
1321int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep)
1322{
1323 unsigned long offset;
1324 int i;
1325
1326 if (p->has_cwsr)
1327 return 0;
1328
1329 for (i = 0; i < p->n_pdds; i++) {
1330 struct kfd_node *dev = p->pdds[i]->dev;
1331 struct qcm_process_device *qpd = &p->pdds[i]->qpd;
1332
1333 if (!dev->kfd->cwsr_enabled || qpd->cwsr_kaddr || qpd->cwsr_base)
1334 continue;
1335
1336 offset = KFD_MMAP_TYPE_RESERVED_MEM | KFD_MMAP_GPU_ID(dev->id);
1337 qpd->tba_addr = (int64_t)vm_mmap(filep, 0,
1338 KFD_CWSR_TBA_TMA_SIZE, PROT_READ | PROT_EXEC,
1339 MAP_SHARED, offset);
1340
1341 if (IS_ERR_VALUE(qpd->tba_addr)) {
1342 int err = qpd->tba_addr;
1343
1344 dev_err(dev->adev->dev,
1345 "Failure to set tba address. error %d.\n", err);
1346 qpd->tba_addr = 0;
1347 qpd->cwsr_kaddr = NULL;
1348 return err;
1349 }
1350
1351 memcpy(qpd->cwsr_kaddr, dev->kfd->cwsr_isa, dev->kfd->cwsr_isa_size);
1352
1353 kfd_process_set_trap_debug_flag(qpd, p->debug_trap_enabled);
1354
1355 qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET;
1356 pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n",
1357 qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr);
1358 }
1359
1360 p->has_cwsr = true;
1361
1362 return 0;
1363}
1364
1365static int kfd_process_device_init_cwsr_dgpu(struct kfd_process_device *pdd)
1366{
1367 struct kfd_node *dev = pdd->dev;
1368 struct qcm_process_device *qpd = &pdd->qpd;
1369 uint32_t flags = KFD_IOC_ALLOC_MEM_FLAGS_GTT
1370 | KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE
1371 | KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
1372 struct kgd_mem *mem;
1373 void *kaddr;
1374 int ret;
1375
1376 if (!dev->kfd->cwsr_enabled || qpd->cwsr_kaddr || !qpd->cwsr_base)
1377 return 0;
1378
1379 /* cwsr_base is only set for dGPU */
1380 ret = kfd_process_alloc_gpuvm(pdd, qpd->cwsr_base,
1381 KFD_CWSR_TBA_TMA_SIZE, flags, &mem, &kaddr);
1382 if (ret)
1383 return ret;
1384
1385 qpd->cwsr_mem = mem;
1386 qpd->cwsr_kaddr = kaddr;
1387 qpd->tba_addr = qpd->cwsr_base;
1388
1389 memcpy(qpd->cwsr_kaddr, dev->kfd->cwsr_isa, dev->kfd->cwsr_isa_size);
1390
1391 kfd_process_set_trap_debug_flag(&pdd->qpd,
1392 pdd->process->debug_trap_enabled);
1393
1394 qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET;
1395 pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n",
1396 qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr);
1397
1398 return 0;
1399}
1400
1401static void kfd_process_device_destroy_cwsr_dgpu(struct kfd_process_device *pdd)
1402{
1403 struct kfd_node *dev = pdd->dev;
1404 struct qcm_process_device *qpd = &pdd->qpd;
1405
1406 if (!dev->kfd->cwsr_enabled || !qpd->cwsr_kaddr || !qpd->cwsr_base)
1407 return;
1408
1409 kfd_process_free_gpuvm(qpd->cwsr_mem, pdd, &qpd->cwsr_kaddr);
1410}
1411
1412void kfd_process_set_trap_handler(struct qcm_process_device *qpd,
1413 uint64_t tba_addr,
1414 uint64_t tma_addr)
1415{
1416 if (qpd->cwsr_kaddr) {
1417 /* KFD trap handler is bound, record as second-level TBA/TMA
1418 * in first-level TMA. First-level trap will jump to second.
1419 */
1420 uint64_t *tma =
1421 (uint64_t *)(qpd->cwsr_kaddr + KFD_CWSR_TMA_OFFSET);
1422 tma[0] = tba_addr;
1423 tma[1] = tma_addr;
1424 } else {
1425 /* No trap handler bound, bind as first-level TBA/TMA. */
1426 qpd->tba_addr = tba_addr;
1427 qpd->tma_addr = tma_addr;
1428 }
1429}
1430
1431bool kfd_process_xnack_mode(struct kfd_process *p, bool supported)
1432{
1433 int i;
1434
1435 /* On most GFXv9 GPUs, the retry mode in the SQ must match the
1436 * boot time retry setting. Mixing processes with different
1437 * XNACK/retry settings can hang the GPU.
1438 *
1439 * Different GPUs can have different noretry settings depending
1440 * on HW bugs or limitations. We need to find at least one
1441 * XNACK mode for this process that's compatible with all GPUs.
1442 * Fortunately GPUs with retry enabled (noretry=0) can run code
1443 * built for XNACK-off. On GFXv9 it may perform slower.
1444 *
1445 * Therefore applications built for XNACK-off can always be
1446 * supported and will be our fallback if any GPU does not
1447 * support retry.
1448 */
1449 for (i = 0; i < p->n_pdds; i++) {
1450 struct kfd_node *dev = p->pdds[i]->dev;
1451
1452 /* Only consider GFXv9 and higher GPUs. Older GPUs don't
1453 * support the SVM APIs and don't need to be considered
1454 * for the XNACK mode selection.
1455 */
1456 if (!KFD_IS_SOC15(dev))
1457 continue;
1458 /* Aldebaran can always support XNACK because it can support
1459 * per-process XNACK mode selection. But let the dev->noretry
1460 * setting still influence the default XNACK mode.
1461 */
1462 if (supported && KFD_SUPPORT_XNACK_PER_PROCESS(dev)) {
1463 if (!amdgpu_sriov_xnack_support(dev->kfd->adev)) {
1464 pr_debug("SRIOV platform xnack not supported\n");
1465 return false;
1466 }
1467 continue;
1468 }
1469
1470 /* GFXv10 and later GPUs do not support shader preemption
1471 * during page faults. This can lead to poor QoS for queue
1472 * management and memory-manager-related preemptions or
1473 * even deadlocks.
1474 */
1475 if (KFD_GC_VERSION(dev) >= IP_VERSION(10, 1, 1))
1476 return false;
1477
1478 if (dev->kfd->noretry)
1479 return false;
1480 }
1481
1482 return true;
1483}
1484
1485void kfd_process_set_trap_debug_flag(struct qcm_process_device *qpd,
1486 bool enabled)
1487{
1488 if (qpd->cwsr_kaddr) {
1489 uint64_t *tma =
1490 (uint64_t *)(qpd->cwsr_kaddr + KFD_CWSR_TMA_OFFSET);
1491 tma[2] = enabled;
1492 }
1493}
1494
1495/*
1496 * On return the kfd_process is fully operational and will be freed when the
1497 * mm is released
1498 */
1499static struct kfd_process *create_process(const struct task_struct *thread)
1500{
1501 struct kfd_process *process;
1502 struct mmu_notifier *mn;
1503 int err = -ENOMEM;
1504
1505 process = kzalloc(sizeof(*process), GFP_KERNEL);
1506 if (!process)
1507 goto err_alloc_process;
1508
1509 kref_init(&process->ref);
1510 mutex_init(&process->mutex);
1511 process->mm = thread->mm;
1512 process->lead_thread = thread->group_leader;
1513 process->n_pdds = 0;
1514 process->queues_paused = false;
1515 INIT_DELAYED_WORK(&process->eviction_work, evict_process_worker);
1516 INIT_DELAYED_WORK(&process->restore_work, restore_process_worker);
1517 process->last_restore_timestamp = get_jiffies_64();
1518 err = kfd_event_init_process(process);
1519 if (err)
1520 goto err_event_init;
1521 process->is_32bit_user_mode = in_compat_syscall();
1522 process->debug_trap_enabled = false;
1523 process->debugger_process = NULL;
1524 process->exception_enable_mask = 0;
1525 atomic_set(&process->debugged_process_count, 0);
1526 sema_init(&process->runtime_enable_sema, 0);
1527
1528 process->pasid = kfd_pasid_alloc();
1529 if (process->pasid == 0) {
1530 err = -ENOSPC;
1531 goto err_alloc_pasid;
1532 }
1533
1534 err = pqm_init(&process->pqm, process);
1535 if (err != 0)
1536 goto err_process_pqm_init;
1537
1538 /* init process apertures*/
1539 err = kfd_init_apertures(process);
1540 if (err != 0)
1541 goto err_init_apertures;
1542
1543 /* Check XNACK support after PDDs are created in kfd_init_apertures */
1544 process->xnack_enabled = kfd_process_xnack_mode(process, false);
1545
1546 err = svm_range_list_init(process);
1547 if (err)
1548 goto err_init_svm_range_list;
1549
1550 /* alloc_notifier needs to find the process in the hash table */
1551 hash_add_rcu(kfd_processes_table, &process->kfd_processes,
1552 (uintptr_t)process->mm);
1553
1554 /* Avoid free_notifier to start kfd_process_wq_release if
1555 * mmu_notifier_get failed because of pending signal.
1556 */
1557 kref_get(&process->ref);
1558
1559 /* MMU notifier registration must be the last call that can fail
1560 * because after this point we cannot unwind the process creation.
1561 * After this point, mmu_notifier_put will trigger the cleanup by
1562 * dropping the last process reference in the free_notifier.
1563 */
1564 mn = mmu_notifier_get(&kfd_process_mmu_notifier_ops, process->mm);
1565 if (IS_ERR(mn)) {
1566 err = PTR_ERR(mn);
1567 goto err_register_notifier;
1568 }
1569 BUG_ON(mn != &process->mmu_notifier);
1570
1571 kfd_unref_process(process);
1572 get_task_struct(process->lead_thread);
1573
1574 INIT_WORK(&process->debug_event_workarea, debug_event_write_work_handler);
1575
1576 return process;
1577
1578err_register_notifier:
1579 hash_del_rcu(&process->kfd_processes);
1580 svm_range_list_fini(process);
1581err_init_svm_range_list:
1582 kfd_process_free_outstanding_kfd_bos(process);
1583 kfd_process_destroy_pdds(process);
1584err_init_apertures:
1585 pqm_uninit(&process->pqm);
1586err_process_pqm_init:
1587 kfd_pasid_free(process->pasid);
1588err_alloc_pasid:
1589 kfd_event_free_process(process);
1590err_event_init:
1591 mutex_destroy(&process->mutex);
1592 kfree(process);
1593err_alloc_process:
1594 return ERR_PTR(err);
1595}
1596
1597struct kfd_process_device *kfd_get_process_device_data(struct kfd_node *dev,
1598 struct kfd_process *p)
1599{
1600 int i;
1601
1602 for (i = 0; i < p->n_pdds; i++)
1603 if (p->pdds[i]->dev == dev)
1604 return p->pdds[i];
1605
1606 return NULL;
1607}
1608
1609struct kfd_process_device *kfd_create_process_device_data(struct kfd_node *dev,
1610 struct kfd_process *p)
1611{
1612 struct kfd_process_device *pdd = NULL;
1613
1614 if (WARN_ON_ONCE(p->n_pdds >= MAX_GPU_INSTANCE))
1615 return NULL;
1616 pdd = kzalloc(sizeof(*pdd), GFP_KERNEL);
1617 if (!pdd)
1618 return NULL;
1619
1620 pdd->dev = dev;
1621 INIT_LIST_HEAD(&pdd->qpd.queues_list);
1622 INIT_LIST_HEAD(&pdd->qpd.priv_queue_list);
1623 pdd->qpd.dqm = dev->dqm;
1624 pdd->qpd.pqm = &p->pqm;
1625 pdd->qpd.evicted = 0;
1626 pdd->qpd.mapped_gws_queue = false;
1627 pdd->process = p;
1628 pdd->bound = PDD_UNBOUND;
1629 pdd->already_dequeued = false;
1630 pdd->runtime_inuse = false;
1631 atomic64_set(&pdd->vram_usage, 0);
1632 pdd->sdma_past_activity_counter = 0;
1633 pdd->user_gpu_id = dev->id;
1634 atomic64_set(&pdd->evict_duration_counter, 0);
1635
1636 p->pdds[p->n_pdds++] = pdd;
1637 if (kfd_dbg_is_per_vmid_supported(pdd->dev))
1638 pdd->spi_dbg_override = pdd->dev->kfd2kgd->disable_debug_trap(
1639 pdd->dev->adev,
1640 false,
1641 0);
1642
1643 /* Init idr used for memory handle translation */
1644 idr_init(&pdd->alloc_idr);
1645
1646 return pdd;
1647}
1648
1649/**
1650 * kfd_process_device_init_vm - Initialize a VM for a process-device
1651 *
1652 * @pdd: The process-device
1653 * @drm_file: Optional pointer to a DRM file descriptor
1654 *
1655 * If @drm_file is specified, it will be used to acquire the VM from
1656 * that file descriptor. If successful, the @pdd takes ownership of
1657 * the file descriptor.
1658 *
1659 * If @drm_file is NULL, a new VM is created.
1660 *
1661 * Returns 0 on success, -errno on failure.
1662 */
1663int kfd_process_device_init_vm(struct kfd_process_device *pdd,
1664 struct file *drm_file)
1665{
1666 struct amdgpu_fpriv *drv_priv;
1667 struct amdgpu_vm *avm;
1668 struct kfd_process *p;
1669 struct dma_fence *ef;
1670 struct kfd_node *dev;
1671 int ret;
1672
1673 if (!drm_file)
1674 return -EINVAL;
1675
1676 if (pdd->drm_priv)
1677 return -EBUSY;
1678
1679 ret = amdgpu_file_to_fpriv(drm_file, &drv_priv);
1680 if (ret)
1681 return ret;
1682 avm = &drv_priv->vm;
1683
1684 p = pdd->process;
1685 dev = pdd->dev;
1686
1687 ret = amdgpu_amdkfd_gpuvm_acquire_process_vm(dev->adev, avm,
1688 &p->kgd_process_info,
1689 p->ef ? NULL : &ef);
1690 if (ret) {
1691 dev_err(dev->adev->dev, "Failed to create process VM object\n");
1692 return ret;
1693 }
1694
1695 if (!p->ef)
1696 RCU_INIT_POINTER(p->ef, ef);
1697
1698 pdd->drm_priv = drm_file->private_data;
1699
1700 ret = kfd_process_device_reserve_ib_mem(pdd);
1701 if (ret)
1702 goto err_reserve_ib_mem;
1703 ret = kfd_process_device_init_cwsr_dgpu(pdd);
1704 if (ret)
1705 goto err_init_cwsr;
1706
1707 ret = amdgpu_amdkfd_gpuvm_set_vm_pasid(dev->adev, avm, p->pasid);
1708 if (ret)
1709 goto err_set_pasid;
1710
1711 pdd->drm_file = drm_file;
1712
1713 return 0;
1714
1715err_set_pasid:
1716 kfd_process_device_destroy_cwsr_dgpu(pdd);
1717err_init_cwsr:
1718 kfd_process_device_destroy_ib_mem(pdd);
1719err_reserve_ib_mem:
1720 pdd->drm_priv = NULL;
1721 amdgpu_amdkfd_gpuvm_destroy_cb(dev->adev, avm);
1722
1723 return ret;
1724}
1725
1726/*
1727 * Direct the IOMMU to bind the process (specifically the pasid->mm)
1728 * to the device.
1729 * Unbinding occurs when the process dies or the device is removed.
1730 *
1731 * Assumes that the process lock is held.
1732 */
1733struct kfd_process_device *kfd_bind_process_to_device(struct kfd_node *dev,
1734 struct kfd_process *p)
1735{
1736 struct kfd_process_device *pdd;
1737 int err;
1738
1739 pdd = kfd_get_process_device_data(dev, p);
1740 if (!pdd) {
1741 dev_err(dev->adev->dev, "Process device data doesn't exist\n");
1742 return ERR_PTR(-ENOMEM);
1743 }
1744
1745 if (!pdd->drm_priv)
1746 return ERR_PTR(-ENODEV);
1747
1748 /*
1749 * signal runtime-pm system to auto resume and prevent
1750 * further runtime suspend once device pdd is created until
1751 * pdd is destroyed.
1752 */
1753 if (!pdd->runtime_inuse) {
1754 err = pm_runtime_get_sync(adev_to_drm(dev->adev)->dev);
1755 if (err < 0) {
1756 pm_runtime_put_autosuspend(adev_to_drm(dev->adev)->dev);
1757 return ERR_PTR(err);
1758 }
1759 }
1760
1761 /*
1762 * make sure that runtime_usage counter is incremented just once
1763 * per pdd
1764 */
1765 pdd->runtime_inuse = true;
1766
1767 return pdd;
1768}
1769
1770/* Create specific handle mapped to mem from process local memory idr
1771 * Assumes that the process lock is held.
1772 */
1773int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
1774 void *mem)
1775{
1776 return idr_alloc(&pdd->alloc_idr, mem, 0, 0, GFP_KERNEL);
1777}
1778
1779/* Translate specific handle from process local memory idr
1780 * Assumes that the process lock is held.
1781 */
1782void *kfd_process_device_translate_handle(struct kfd_process_device *pdd,
1783 int handle)
1784{
1785 if (handle < 0)
1786 return NULL;
1787
1788 return idr_find(&pdd->alloc_idr, handle);
1789}
1790
1791/* Remove specific handle from process local memory idr
1792 * Assumes that the process lock is held.
1793 */
1794void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,
1795 int handle)
1796{
1797 if (handle >= 0)
1798 idr_remove(&pdd->alloc_idr, handle);
1799}
1800
1801/* This increments the process->ref counter. */
1802struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid)
1803{
1804 struct kfd_process *p, *ret_p = NULL;
1805 unsigned int temp;
1806
1807 int idx = srcu_read_lock(&kfd_processes_srcu);
1808
1809 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1810 if (p->pasid == pasid) {
1811 kref_get(&p->ref);
1812 ret_p = p;
1813 break;
1814 }
1815 }
1816
1817 srcu_read_unlock(&kfd_processes_srcu, idx);
1818
1819 return ret_p;
1820}
1821
1822/* This increments the process->ref counter. */
1823struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm)
1824{
1825 struct kfd_process *p;
1826
1827 int idx = srcu_read_lock(&kfd_processes_srcu);
1828
1829 p = find_process_by_mm(mm);
1830 if (p)
1831 kref_get(&p->ref);
1832
1833 srcu_read_unlock(&kfd_processes_srcu, idx);
1834
1835 return p;
1836}
1837
1838/* kfd_process_evict_queues - Evict all user queues of a process
1839 *
1840 * Eviction is reference-counted per process-device. This means multiple
1841 * evictions from different sources can be nested safely.
1842 */
1843int kfd_process_evict_queues(struct kfd_process *p, uint32_t trigger)
1844{
1845 int r = 0;
1846 int i;
1847 unsigned int n_evicted = 0;
1848
1849 for (i = 0; i < p->n_pdds; i++) {
1850 struct kfd_process_device *pdd = p->pdds[i];
1851 struct device *dev = pdd->dev->adev->dev;
1852
1853 kfd_smi_event_queue_eviction(pdd->dev, p->lead_thread->pid,
1854 trigger);
1855
1856 r = pdd->dev->dqm->ops.evict_process_queues(pdd->dev->dqm,
1857 &pdd->qpd);
1858 /* evict return -EIO if HWS is hang or asic is resetting, in this case
1859 * we would like to set all the queues to be in evicted state to prevent
1860 * them been add back since they actually not be saved right now.
1861 */
1862 if (r && r != -EIO) {
1863 dev_err(dev, "Failed to evict process queues\n");
1864 goto fail;
1865 }
1866 n_evicted++;
1867
1868 pdd->dev->dqm->is_hws_hang = false;
1869 }
1870
1871 return r;
1872
1873fail:
1874 /* To keep state consistent, roll back partial eviction by
1875 * restoring queues
1876 */
1877 for (i = 0; i < p->n_pdds; i++) {
1878 struct kfd_process_device *pdd = p->pdds[i];
1879
1880 if (n_evicted == 0)
1881 break;
1882
1883 kfd_smi_event_queue_restore(pdd->dev, p->lead_thread->pid);
1884
1885 if (pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm,
1886 &pdd->qpd))
1887 dev_err(pdd->dev->adev->dev,
1888 "Failed to restore queues\n");
1889
1890 n_evicted--;
1891 }
1892
1893 return r;
1894}
1895
1896/* kfd_process_restore_queues - Restore all user queues of a process */
1897int kfd_process_restore_queues(struct kfd_process *p)
1898{
1899 int r, ret = 0;
1900 int i;
1901
1902 for (i = 0; i < p->n_pdds; i++) {
1903 struct kfd_process_device *pdd = p->pdds[i];
1904 struct device *dev = pdd->dev->adev->dev;
1905
1906 kfd_smi_event_queue_restore(pdd->dev, p->lead_thread->pid);
1907
1908 r = pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm,
1909 &pdd->qpd);
1910 if (r) {
1911 dev_err(dev, "Failed to restore process queues\n");
1912 if (!ret)
1913 ret = r;
1914 }
1915 }
1916
1917 return ret;
1918}
1919
1920int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id)
1921{
1922 int i;
1923
1924 for (i = 0; i < p->n_pdds; i++)
1925 if (p->pdds[i] && gpu_id == p->pdds[i]->user_gpu_id)
1926 return i;
1927 return -EINVAL;
1928}
1929
1930int
1931kfd_process_gpuid_from_node(struct kfd_process *p, struct kfd_node *node,
1932 uint32_t *gpuid, uint32_t *gpuidx)
1933{
1934 int i;
1935
1936 for (i = 0; i < p->n_pdds; i++)
1937 if (p->pdds[i] && p->pdds[i]->dev == node) {
1938 *gpuid = p->pdds[i]->user_gpu_id;
1939 *gpuidx = i;
1940 return 0;
1941 }
1942 return -EINVAL;
1943}
1944
1945static int signal_eviction_fence(struct kfd_process *p)
1946{
1947 struct dma_fence *ef;
1948 int ret;
1949
1950 rcu_read_lock();
1951 ef = dma_fence_get_rcu_safe(&p->ef);
1952 rcu_read_unlock();
1953 if (!ef)
1954 return -EINVAL;
1955
1956 ret = dma_fence_signal(ef);
1957 dma_fence_put(ef);
1958
1959 return ret;
1960}
1961
1962static void evict_process_worker(struct work_struct *work)
1963{
1964 int ret;
1965 struct kfd_process *p;
1966 struct delayed_work *dwork;
1967
1968 dwork = to_delayed_work(work);
1969
1970 /* Process termination destroys this worker thread. So during the
1971 * lifetime of this thread, kfd_process p will be valid
1972 */
1973 p = container_of(dwork, struct kfd_process, eviction_work);
1974
1975 pr_debug("Started evicting pasid 0x%x\n", p->pasid);
1976 ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_TRIGGER_TTM);
1977 if (!ret) {
1978 /* If another thread already signaled the eviction fence,
1979 * they are responsible stopping the queues and scheduling
1980 * the restore work.
1981 */
1982 if (signal_eviction_fence(p) ||
1983 mod_delayed_work(kfd_restore_wq, &p->restore_work,
1984 msecs_to_jiffies(PROCESS_RESTORE_TIME_MS)))
1985 kfd_process_restore_queues(p);
1986
1987 pr_debug("Finished evicting pasid 0x%x\n", p->pasid);
1988 } else
1989 pr_err("Failed to evict queues of pasid 0x%x\n", p->pasid);
1990}
1991
1992static int restore_process_helper(struct kfd_process *p)
1993{
1994 int ret = 0;
1995
1996 /* VMs may not have been acquired yet during debugging. */
1997 if (p->kgd_process_info) {
1998 ret = amdgpu_amdkfd_gpuvm_restore_process_bos(
1999 p->kgd_process_info, &p->ef);
2000 if (ret)
2001 return ret;
2002 }
2003
2004 ret = kfd_process_restore_queues(p);
2005 if (!ret)
2006 pr_debug("Finished restoring pasid 0x%x\n", p->pasid);
2007 else
2008 pr_err("Failed to restore queues of pasid 0x%x\n", p->pasid);
2009
2010 return ret;
2011}
2012
2013static void restore_process_worker(struct work_struct *work)
2014{
2015 struct delayed_work *dwork;
2016 struct kfd_process *p;
2017 int ret = 0;
2018
2019 dwork = to_delayed_work(work);
2020
2021 /* Process termination destroys this worker thread. So during the
2022 * lifetime of this thread, kfd_process p will be valid
2023 */
2024 p = container_of(dwork, struct kfd_process, restore_work);
2025 pr_debug("Started restoring pasid 0x%x\n", p->pasid);
2026
2027 /* Setting last_restore_timestamp before successful restoration.
2028 * Otherwise this would have to be set by KGD (restore_process_bos)
2029 * before KFD BOs are unreserved. If not, the process can be evicted
2030 * again before the timestamp is set.
2031 * If restore fails, the timestamp will be set again in the next
2032 * attempt. This would mean that the minimum GPU quanta would be
2033 * PROCESS_ACTIVE_TIME_MS - (time to execute the following two
2034 * functions)
2035 */
2036
2037 p->last_restore_timestamp = get_jiffies_64();
2038
2039 ret = restore_process_helper(p);
2040 if (ret) {
2041 pr_debug("Failed to restore BOs of pasid 0x%x, retry after %d ms\n",
2042 p->pasid, PROCESS_BACK_OFF_TIME_MS);
2043 if (mod_delayed_work(kfd_restore_wq, &p->restore_work,
2044 msecs_to_jiffies(PROCESS_RESTORE_TIME_MS)))
2045 kfd_process_restore_queues(p);
2046 }
2047}
2048
2049void kfd_suspend_all_processes(void)
2050{
2051 struct kfd_process *p;
2052 unsigned int temp;
2053 int idx = srcu_read_lock(&kfd_processes_srcu);
2054
2055 WARN(debug_evictions, "Evicting all processes");
2056 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
2057 if (kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_TRIGGER_SUSPEND))
2058 pr_err("Failed to suspend process 0x%x\n", p->pasid);
2059 signal_eviction_fence(p);
2060 }
2061 srcu_read_unlock(&kfd_processes_srcu, idx);
2062}
2063
2064int kfd_resume_all_processes(void)
2065{
2066 struct kfd_process *p;
2067 unsigned int temp;
2068 int ret = 0, idx = srcu_read_lock(&kfd_processes_srcu);
2069
2070 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
2071 if (restore_process_helper(p)) {
2072 pr_err("Restore process %d failed during resume\n",
2073 p->pasid);
2074 ret = -EFAULT;
2075 }
2076 }
2077 srcu_read_unlock(&kfd_processes_srcu, idx);
2078 return ret;
2079}
2080
2081int kfd_reserved_mem_mmap(struct kfd_node *dev, struct kfd_process *process,
2082 struct vm_area_struct *vma)
2083{
2084 struct kfd_process_device *pdd;
2085 struct qcm_process_device *qpd;
2086
2087 if ((vma->vm_end - vma->vm_start) != KFD_CWSR_TBA_TMA_SIZE) {
2088 dev_err(dev->adev->dev, "Incorrect CWSR mapping size.\n");
2089 return -EINVAL;
2090 }
2091
2092 pdd = kfd_get_process_device_data(dev, process);
2093 if (!pdd)
2094 return -EINVAL;
2095 qpd = &pdd->qpd;
2096
2097 qpd->cwsr_kaddr = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
2098 get_order(KFD_CWSR_TBA_TMA_SIZE));
2099 if (!qpd->cwsr_kaddr) {
2100 dev_err(dev->adev->dev,
2101 "Error allocating per process CWSR buffer.\n");
2102 return -ENOMEM;
2103 }
2104
2105 vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND
2106 | VM_NORESERVE | VM_DONTDUMP | VM_PFNMAP);
2107 /* Mapping pages to user process */
2108 return remap_pfn_range(vma, vma->vm_start,
2109 PFN_DOWN(__pa(qpd->cwsr_kaddr)),
2110 KFD_CWSR_TBA_TMA_SIZE, vma->vm_page_prot);
2111}
2112
2113/* assumes caller holds process lock. */
2114int kfd_process_drain_interrupts(struct kfd_process_device *pdd)
2115{
2116 uint32_t irq_drain_fence[8];
2117 uint8_t node_id = 0;
2118 int r = 0;
2119
2120 if (!KFD_IS_SOC15(pdd->dev))
2121 return 0;
2122
2123 pdd->process->irq_drain_is_open = true;
2124
2125 memset(irq_drain_fence, 0, sizeof(irq_drain_fence));
2126 irq_drain_fence[0] = (KFD_IRQ_FENCE_SOURCEID << 8) |
2127 KFD_IRQ_FENCE_CLIENTID;
2128 irq_drain_fence[3] = pdd->process->pasid;
2129
2130 /*
2131 * For GFX 9.4.3, send the NodeId also in IH cookie DW[3]
2132 */
2133 if (KFD_GC_VERSION(pdd->dev->kfd) == IP_VERSION(9, 4, 3) ||
2134 KFD_GC_VERSION(pdd->dev->kfd) == IP_VERSION(9, 4, 4)) {
2135 node_id = ffs(pdd->dev->interrupt_bitmap) - 1;
2136 irq_drain_fence[3] |= node_id << 16;
2137 }
2138
2139 /* ensure stale irqs scheduled KFD interrupts and send drain fence. */
2140 if (amdgpu_amdkfd_send_close_event_drain_irq(pdd->dev->adev,
2141 irq_drain_fence)) {
2142 pdd->process->irq_drain_is_open = false;
2143 return 0;
2144 }
2145
2146 r = wait_event_interruptible(pdd->process->wait_irq_drain,
2147 !READ_ONCE(pdd->process->irq_drain_is_open));
2148 if (r)
2149 pdd->process->irq_drain_is_open = false;
2150
2151 return r;
2152}
2153
2154void kfd_process_close_interrupt_drain(unsigned int pasid)
2155{
2156 struct kfd_process *p;
2157
2158 p = kfd_lookup_process_by_pasid(pasid);
2159
2160 if (!p)
2161 return;
2162
2163 WRITE_ONCE(p->irq_drain_is_open, false);
2164 wake_up_all(&p->wait_irq_drain);
2165 kfd_unref_process(p);
2166}
2167
2168struct send_exception_work_handler_workarea {
2169 struct work_struct work;
2170 struct kfd_process *p;
2171 unsigned int queue_id;
2172 uint64_t error_reason;
2173};
2174
2175static void send_exception_work_handler(struct work_struct *work)
2176{
2177 struct send_exception_work_handler_workarea *workarea;
2178 struct kfd_process *p;
2179 struct queue *q;
2180 struct mm_struct *mm;
2181 struct kfd_context_save_area_header __user *csa_header;
2182 uint64_t __user *err_payload_ptr;
2183 uint64_t cur_err;
2184 uint32_t ev_id;
2185
2186 workarea = container_of(work,
2187 struct send_exception_work_handler_workarea,
2188 work);
2189 p = workarea->p;
2190
2191 mm = get_task_mm(p->lead_thread);
2192
2193 if (!mm)
2194 return;
2195
2196 kthread_use_mm(mm);
2197
2198 q = pqm_get_user_queue(&p->pqm, workarea->queue_id);
2199
2200 if (!q)
2201 goto out;
2202
2203 csa_header = (void __user *)q->properties.ctx_save_restore_area_address;
2204
2205 get_user(err_payload_ptr, (uint64_t __user **)&csa_header->err_payload_addr);
2206 get_user(cur_err, err_payload_ptr);
2207 cur_err |= workarea->error_reason;
2208 put_user(cur_err, err_payload_ptr);
2209 get_user(ev_id, &csa_header->err_event_id);
2210
2211 kfd_set_event(p, ev_id);
2212
2213out:
2214 kthread_unuse_mm(mm);
2215 mmput(mm);
2216}
2217
2218int kfd_send_exception_to_runtime(struct kfd_process *p,
2219 unsigned int queue_id,
2220 uint64_t error_reason)
2221{
2222 struct send_exception_work_handler_workarea worker;
2223
2224 INIT_WORK_ONSTACK(&worker.work, send_exception_work_handler);
2225
2226 worker.p = p;
2227 worker.queue_id = queue_id;
2228 worker.error_reason = error_reason;
2229
2230 schedule_work(&worker.work);
2231 flush_work(&worker.work);
2232 destroy_work_on_stack(&worker.work);
2233
2234 return 0;
2235}
2236
2237struct kfd_process_device *kfd_process_device_data_by_id(struct kfd_process *p, uint32_t gpu_id)
2238{
2239 int i;
2240
2241 if (gpu_id) {
2242 for (i = 0; i < p->n_pdds; i++) {
2243 struct kfd_process_device *pdd = p->pdds[i];
2244
2245 if (pdd->user_gpu_id == gpu_id)
2246 return pdd;
2247 }
2248 }
2249 return NULL;
2250}
2251
2252int kfd_process_get_user_gpu_id(struct kfd_process *p, uint32_t actual_gpu_id)
2253{
2254 int i;
2255
2256 if (!actual_gpu_id)
2257 return 0;
2258
2259 for (i = 0; i < p->n_pdds; i++) {
2260 struct kfd_process_device *pdd = p->pdds[i];
2261
2262 if (pdd->dev->id == actual_gpu_id)
2263 return pdd->user_gpu_id;
2264 }
2265 return -EINVAL;
2266}
2267
2268#if defined(CONFIG_DEBUG_FS)
2269
2270int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data)
2271{
2272 struct kfd_process *p;
2273 unsigned int temp;
2274 int r = 0;
2275
2276 int idx = srcu_read_lock(&kfd_processes_srcu);
2277
2278 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
2279 seq_printf(m, "Process %d PASID 0x%x:\n",
2280 p->lead_thread->tgid, p->pasid);
2281
2282 mutex_lock(&p->mutex);
2283 r = pqm_debugfs_mqds(m, &p->pqm);
2284 mutex_unlock(&p->mutex);
2285
2286 if (r)
2287 break;
2288 }
2289
2290 srcu_read_unlock(&kfd_processes_srcu, idx);
2291
2292 return r;
2293}
2294
2295#endif