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 #include "amdgpu_reset.h"
39
40 struct mm_struct;
41
42 #include "kfd_priv.h"
43 #include "kfd_device_queue_manager.h"
44 #include "kfd_svm.h"
45 #include "kfd_smi_events.h"
46 #include "kfd_debug.h"
47
48 /*
49 * List of struct kfd_process (field kfd_process).
50 * Unique/indexed by mm_struct*
51 */
52 DEFINE_HASHTABLE(kfd_processes_table, KFD_PROCESS_TABLE_SIZE);
53 DEFINE_MUTEX(kfd_processes_mutex);
54
55 DEFINE_SRCU(kfd_processes_srcu);
56
57 /* For process termination handling */
58 static struct workqueue_struct *kfd_process_wq;
59
60 /* Ordered, single-threaded workqueue for restoring evicted
61 * processes. Restoring multiple processes concurrently under memory
62 * pressure can lead to processes blocking each other from validating
63 * their BOs and result in a live-lock situation where processes
64 * remain evicted indefinitely.
65 */
66 static struct workqueue_struct *kfd_restore_wq;
67
68 static struct kfd_process *find_process(const struct task_struct *thread,
69 bool ref);
70 static void kfd_process_ref_release(struct kref *ref);
71
72 static void evict_process_worker(struct work_struct *work);
73 static void restore_process_worker(struct work_struct *work);
74
75 static void kfd_process_device_destroy_cwsr_dgpu(struct kfd_process_device *pdd);
76
77 struct kfd_procfs_tree {
78 struct kobject *kobj;
79 };
80
81 static struct kfd_procfs_tree procfs;
82
83 /*
84 * Structure for SDMA activity tracking
85 */
86 struct kfd_sdma_activity_handler_workarea {
87 struct work_struct sdma_activity_work;
88 struct kfd_process_device *pdd;
89 uint64_t sdma_activity_counter;
90 };
91
92 struct temp_sdma_queue_list {
93 uint64_t __user *rptr;
94 uint64_t sdma_val;
95 unsigned int queue_id;
96 struct list_head list;
97 };
98
kfd_sdma_activity_worker(struct work_struct * work)99 static void kfd_sdma_activity_worker(struct work_struct *work)
100 {
101 struct kfd_sdma_activity_handler_workarea *workarea;
102 struct kfd_process_device *pdd;
103 uint64_t val;
104 struct mm_struct *mm;
105 struct queue *q;
106 struct qcm_process_device *qpd;
107 struct device_queue_manager *dqm;
108 int ret = 0;
109 struct temp_sdma_queue_list sdma_q_list;
110 struct temp_sdma_queue_list *sdma_q, *next;
111
112 workarea = container_of(work, struct kfd_sdma_activity_handler_workarea,
113 sdma_activity_work);
114
115 pdd = workarea->pdd;
116 if (!pdd)
117 return;
118 dqm = pdd->dev->dqm;
119 qpd = &pdd->qpd;
120 if (!dqm || !qpd)
121 return;
122 /*
123 * Total SDMA activity is current SDMA activity + past SDMA activity
124 * Past SDMA count is stored in pdd.
125 * To get the current activity counters for all active SDMA queues,
126 * we loop over all SDMA queues and get their counts from user-space.
127 *
128 * We cannot call get_user() with dqm_lock held as it can cause
129 * a circular lock dependency situation. To read the SDMA stats,
130 * we need to do the following:
131 *
132 * 1. Create a temporary list of SDMA queue nodes from the qpd->queues_list,
133 * with dqm_lock/dqm_unlock().
134 * 2. Call get_user() for each node in temporary list without dqm_lock.
135 * Save the SDMA count for each node and also add the count to the total
136 * SDMA count counter.
137 * Its possible, during this step, a few SDMA queue nodes got deleted
138 * from the qpd->queues_list.
139 * 3. Do a second pass over qpd->queues_list to check if any nodes got deleted.
140 * If any node got deleted, its SDMA count would be captured in the sdma
141 * past activity counter. So subtract the SDMA counter stored in step 2
142 * for this node from the total SDMA count.
143 */
144 INIT_LIST_HEAD(&sdma_q_list.list);
145
146 /*
147 * Create the temp list of all SDMA queues
148 */
149 dqm_lock(dqm);
150
151 list_for_each_entry(q, &qpd->queues_list, list) {
152 if ((q->properties.type != KFD_QUEUE_TYPE_SDMA) &&
153 (q->properties.type != KFD_QUEUE_TYPE_SDMA_XGMI))
154 continue;
155
156 sdma_q = kzalloc_obj(struct temp_sdma_queue_list, GFP_KERNEL);
157 if (!sdma_q) {
158 dqm_unlock(dqm);
159 goto cleanup;
160 }
161
162 INIT_LIST_HEAD(&sdma_q->list);
163 sdma_q->rptr = (uint64_t __user *)q->properties.read_ptr;
164 sdma_q->queue_id = q->properties.queue_id;
165 list_add_tail(&sdma_q->list, &sdma_q_list.list);
166 }
167
168 /*
169 * If the temp list is empty, then no SDMA queues nodes were found in
170 * qpd->queues_list. Return the past activity count as the total sdma
171 * count
172 */
173 if (list_empty(&sdma_q_list.list)) {
174 workarea->sdma_activity_counter = pdd->sdma_past_activity_counter;
175 dqm_unlock(dqm);
176 return;
177 }
178
179 dqm_unlock(dqm);
180
181 /*
182 * Get the usage count for each SDMA queue in temp_list.
183 */
184 mm = get_task_mm(pdd->process->lead_thread);
185 if (!mm)
186 goto cleanup;
187
188 kthread_use_mm(mm);
189
190 list_for_each_entry(sdma_q, &sdma_q_list.list, list) {
191 val = 0;
192 ret = read_sdma_queue_counter(sdma_q->rptr, &val);
193 if (ret) {
194 pr_debug("Failed to read SDMA queue active counter for queue id: %d",
195 sdma_q->queue_id);
196 } else {
197 sdma_q->sdma_val = val;
198 workarea->sdma_activity_counter += val;
199 }
200 }
201
202 kthread_unuse_mm(mm);
203 mmput(mm);
204
205 /*
206 * Do a second iteration over qpd_queues_list to check if any SDMA
207 * nodes got deleted while fetching SDMA counter.
208 */
209 dqm_lock(dqm);
210
211 workarea->sdma_activity_counter += pdd->sdma_past_activity_counter;
212
213 list_for_each_entry(q, &qpd->queues_list, list) {
214 if (list_empty(&sdma_q_list.list))
215 break;
216
217 if ((q->properties.type != KFD_QUEUE_TYPE_SDMA) &&
218 (q->properties.type != KFD_QUEUE_TYPE_SDMA_XGMI))
219 continue;
220
221 list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
222 if (((uint64_t __user *)q->properties.read_ptr == sdma_q->rptr) &&
223 (sdma_q->queue_id == q->properties.queue_id)) {
224 list_del(&sdma_q->list);
225 kfree(sdma_q);
226 break;
227 }
228 }
229 }
230
231 dqm_unlock(dqm);
232
233 /*
234 * If temp list is not empty, it implies some queues got deleted
235 * from qpd->queues_list during SDMA usage read. Subtract the SDMA
236 * count for each node from the total SDMA count.
237 */
238 list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
239 workarea->sdma_activity_counter -= sdma_q->sdma_val;
240 list_del(&sdma_q->list);
241 kfree(sdma_q);
242 }
243
244 return;
245
246 cleanup:
247 list_for_each_entry_safe(sdma_q, next, &sdma_q_list.list, list) {
248 list_del(&sdma_q->list);
249 kfree(sdma_q);
250 }
251 }
252
253 /**
254 * kfd_get_cu_occupancy - Collect number of waves in-flight on this device
255 * by current process. Translates acquired wave count into number of compute units
256 * that are occupied.
257 *
258 * @attr: Handle of attribute that allows reporting of wave count. The attribute
259 * handle encapsulates GPU device it is associated with, thereby allowing collection
260 * of waves in flight, etc
261 * @buffer: Handle of user provided buffer updated with wave count
262 *
263 * Return: Number of bytes written to user buffer or an error value
264 */
kfd_get_cu_occupancy(struct attribute * attr,char * buffer)265 static int kfd_get_cu_occupancy(struct attribute *attr, char *buffer)
266 {
267 int cu_cnt;
268 int wave_cnt;
269 int max_waves_per_cu;
270 struct kfd_node *dev = NULL;
271 struct kfd_process *proc = NULL;
272 struct kfd_process_device *pdd = NULL;
273 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 pid %d\n",
286 dev->id, (int)proc->lead_thread->pid);
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 = kzalloc_objs(*cu_occupancy, AMDGPU_MAX_QUEUES,
295 GFP_KERNEL);
296 if (!cu_occupancy)
297 return -ENOMEM;
298
299 /*
300 * For GFX 9.4.3, fetch the CU occupancy from the first XCC in the partition.
301 * For AQL queues, because of cooperative dispatch we multiply the wave count
302 * by number of XCCs in the partition to get the total wave counts across all
303 * XCCs in the partition.
304 * For PM4 queues, there is no cooperative dispatch so wave_cnt stay as it is.
305 */
306 dev->kfd2kgd->get_cu_occupancy(dev->adev, cu_occupancy,
307 &max_waves_per_cu, ffs(dev->xcc_mask) - 1);
308
309 for (i = 0; i < AMDGPU_MAX_QUEUES; i++) {
310 if (cu_occupancy[i].wave_cnt != 0 &&
311 kfd_dqm_is_queue_in_process(dev->dqm, &pdd->qpd,
312 cu_occupancy[i].doorbell_off,
313 &queue_format)) {
314 if (unlikely(queue_format == KFD_QUEUE_FORMAT_PM4))
315 wave_cnt += cu_occupancy[i].wave_cnt;
316 else
317 wave_cnt += (NUM_XCC(dev->xcc_mask) *
318 cu_occupancy[i].wave_cnt);
319 }
320 }
321
322 /* Translate wave count to number of compute units */
323 cu_cnt = (wave_cnt + (max_waves_per_cu - 1)) / max_waves_per_cu;
324 kfree(cu_occupancy);
325 return snprintf(buffer, PAGE_SIZE, "%d\n", cu_cnt);
326 }
327
kfd_procfs_show(struct kobject * kobj,struct attribute * attr,char * buffer)328 static ssize_t kfd_procfs_show(struct kobject *kobj, struct attribute *attr,
329 char *buffer)
330 {
331 if (strcmp(attr->name, "pasid") == 0)
332 return snprintf(buffer, PAGE_SIZE, "%d\n", 0);
333 else if (strncmp(attr->name, "vram_", 5) == 0) {
334 struct kfd_process_device *pdd = container_of(attr, struct kfd_process_device,
335 attr_vram);
336 return snprintf(buffer, PAGE_SIZE, "%llu\n", atomic64_read(&pdd->vram_usage));
337 } else if (strncmp(attr->name, "sdma_", 5) == 0) {
338 struct kfd_process_device *pdd = container_of(attr, struct kfd_process_device,
339 attr_sdma);
340 struct kfd_sdma_activity_handler_workarea sdma_activity_work_handler;
341
342 INIT_WORK_ONSTACK(&sdma_activity_work_handler.sdma_activity_work,
343 kfd_sdma_activity_worker);
344
345 sdma_activity_work_handler.pdd = pdd;
346 sdma_activity_work_handler.sdma_activity_counter = 0;
347
348 schedule_work(&sdma_activity_work_handler.sdma_activity_work);
349
350 flush_work(&sdma_activity_work_handler.sdma_activity_work);
351 destroy_work_on_stack(&sdma_activity_work_handler.sdma_activity_work);
352
353 return snprintf(buffer, PAGE_SIZE, "%llu\n",
354 (sdma_activity_work_handler.sdma_activity_counter)/
355 SDMA_ACTIVITY_DIVISOR);
356 } else {
357 pr_err("Invalid attribute");
358 return -EINVAL;
359 }
360
361 return 0;
362 }
363
kfd_procfs_kobj_release(struct kobject * kobj)364 static void kfd_procfs_kobj_release(struct kobject *kobj)
365 {
366 kfree(kobj);
367 }
368
369 static const struct sysfs_ops kfd_procfs_ops = {
370 .show = kfd_procfs_show,
371 };
372
373 static const struct kobj_type procfs_type = {
374 .release = kfd_procfs_kobj_release,
375 .sysfs_ops = &kfd_procfs_ops,
376 };
377
kfd_procfs_init(void)378 void kfd_procfs_init(void)
379 {
380 int ret = 0;
381
382 procfs.kobj = kfd_alloc_struct(procfs.kobj);
383 if (!procfs.kobj)
384 return;
385
386 ret = kobject_init_and_add(procfs.kobj, &procfs_type,
387 &kfd_device->kobj, "proc");
388 if (ret) {
389 pr_warn("Could not create procfs proc folder");
390 /* If we fail to create the procfs, clean up */
391 kfd_procfs_shutdown();
392 }
393 }
394
kfd_procfs_shutdown(void)395 void kfd_procfs_shutdown(void)
396 {
397 if (procfs.kobj) {
398 kobject_del(procfs.kobj);
399 kobject_put(procfs.kobj);
400 procfs.kobj = NULL;
401 }
402 }
403
kfd_procfs_queue_show(struct kobject * kobj,struct attribute * attr,char * buffer)404 static ssize_t kfd_procfs_queue_show(struct kobject *kobj,
405 struct attribute *attr, char *buffer)
406 {
407 struct queue *q = container_of(kobj, struct queue, kobj);
408
409 if (!strcmp(attr->name, "size"))
410 return snprintf(buffer, PAGE_SIZE, "%llu",
411 q->properties.queue_size);
412 else if (!strcmp(attr->name, "type"))
413 return snprintf(buffer, PAGE_SIZE, "%d", q->properties.type);
414 else if (!strcmp(attr->name, "gpuid"))
415 return snprintf(buffer, PAGE_SIZE, "%u", q->device->id);
416 else
417 pr_err("Invalid attribute");
418
419 return 0;
420 }
421
kfd_procfs_stats_show(struct kobject * kobj,struct attribute * attr,char * buffer)422 static ssize_t kfd_procfs_stats_show(struct kobject *kobj,
423 struct attribute *attr, char *buffer)
424 {
425 if (strcmp(attr->name, "evicted_ms") == 0) {
426 struct kfd_process_device *pdd = container_of(attr,
427 struct kfd_process_device,
428 attr_evict);
429 uint64_t evict_jiffies;
430
431 evict_jiffies = atomic64_read(&pdd->evict_duration_counter);
432
433 return snprintf(buffer,
434 PAGE_SIZE,
435 "%llu\n",
436 jiffies64_to_msecs(evict_jiffies));
437
438 /* Sysfs handle that gets CU occupancy is per device */
439 } else if (strcmp(attr->name, "cu_occupancy") == 0) {
440 return kfd_get_cu_occupancy(attr, buffer);
441 } else {
442 pr_err("Invalid attribute");
443 }
444
445 return 0;
446 }
447
kfd_sysfs_counters_show(struct kobject * kobj,struct attribute * attr,char * buf)448 static ssize_t kfd_sysfs_counters_show(struct kobject *kobj,
449 struct attribute *attr, char *buf)
450 {
451 struct kfd_process_device *pdd;
452
453 if (!strcmp(attr->name, "faults")) {
454 pdd = container_of(attr, struct kfd_process_device,
455 attr_faults);
456 return sysfs_emit(buf, "%llu\n", READ_ONCE(pdd->faults));
457 }
458 if (!strcmp(attr->name, "page_in")) {
459 pdd = container_of(attr, struct kfd_process_device,
460 attr_page_in);
461 return sysfs_emit(buf, "%llu\n", READ_ONCE(pdd->page_in));
462 }
463 if (!strcmp(attr->name, "page_out")) {
464 pdd = container_of(attr, struct kfd_process_device,
465 attr_page_out);
466 return sysfs_emit(buf, "%llu\n", READ_ONCE(pdd->page_out));
467 }
468 return 0;
469 }
470
471 static struct attribute attr_queue_size = {
472 .name = "size",
473 .mode = KFD_SYSFS_FILE_MODE
474 };
475
476 static struct attribute attr_queue_type = {
477 .name = "type",
478 .mode = KFD_SYSFS_FILE_MODE
479 };
480
481 static struct attribute attr_queue_gpuid = {
482 .name = "gpuid",
483 .mode = KFD_SYSFS_FILE_MODE
484 };
485
486 static struct attribute *procfs_queue_attrs[] = {
487 &attr_queue_size,
488 &attr_queue_type,
489 &attr_queue_gpuid,
490 NULL
491 };
492 ATTRIBUTE_GROUPS(procfs_queue);
493
494 static const struct sysfs_ops procfs_queue_ops = {
495 .show = kfd_procfs_queue_show,
496 };
497
498 static const struct kobj_type procfs_queue_type = {
499 .sysfs_ops = &procfs_queue_ops,
500 .default_groups = procfs_queue_groups,
501 };
502
503 static const struct sysfs_ops procfs_stats_ops = {
504 .show = kfd_procfs_stats_show,
505 };
506
507 static const struct kobj_type procfs_stats_type = {
508 .sysfs_ops = &procfs_stats_ops,
509 .release = kfd_procfs_kobj_release,
510 };
511
512 static const struct sysfs_ops sysfs_counters_ops = {
513 .show = kfd_sysfs_counters_show,
514 };
515
516 static const struct kobj_type sysfs_counters_type = {
517 .sysfs_ops = &sysfs_counters_ops,
518 .release = kfd_procfs_kobj_release,
519 };
520
kfd_procfs_add_queue(struct queue * q)521 int kfd_procfs_add_queue(struct queue *q)
522 {
523 struct kfd_process *proc;
524 int ret;
525
526 if (!q || !q->process)
527 return -EINVAL;
528 proc = q->process;
529
530 /* Create proc/<pid>/queues/<queue id> folder */
531 if (!proc->kobj_queues)
532 return -EFAULT;
533 ret = kobject_init_and_add(&q->kobj, &procfs_queue_type,
534 proc->kobj_queues, "%u", q->properties.queue_id);
535 if (ret < 0) {
536 pr_warn("Creating proc/<pid>/queues/%u failed",
537 q->properties.queue_id);
538 kobject_put(&q->kobj);
539 return ret;
540 }
541
542 return 0;
543 }
544
kfd_sysfs_create_file(struct kobject * kobj,struct attribute * attr,char * name)545 static void kfd_sysfs_create_file(struct kobject *kobj, struct attribute *attr,
546 char *name)
547 {
548 int ret;
549
550 if (!kobj || !attr || !name)
551 return;
552
553 attr->name = name;
554 attr->mode = KFD_SYSFS_FILE_MODE;
555 sysfs_attr_init(attr);
556
557 ret = sysfs_create_file(kobj, attr);
558 if (ret)
559 pr_warn("Create sysfs %s/%s failed %d", kobj->name, name, ret);
560 }
561
kfd_procfs_add_sysfs_stats(struct kfd_process * p)562 static void kfd_procfs_add_sysfs_stats(struct kfd_process *p)
563 {
564 int ret;
565 int i;
566 char stats_dir_filename[MAX_SYSFS_FILENAME_LEN];
567
568 if (!p || !p->kobj)
569 return;
570
571 /*
572 * Create sysfs files for each GPU:
573 * - proc/<pid>/stats_<gpuid>/
574 * - proc/<pid>/stats_<gpuid>/evicted_ms
575 * - proc/<pid>/stats_<gpuid>/cu_occupancy
576 */
577 for (i = 0; i < p->n_pdds; i++) {
578 struct kfd_process_device *pdd = p->pdds[i];
579
580 snprintf(stats_dir_filename, MAX_SYSFS_FILENAME_LEN,
581 "stats_%u", pdd->dev->id);
582 pdd->kobj_stats = kfd_alloc_struct(pdd->kobj_stats);
583 if (!pdd->kobj_stats)
584 return;
585
586 ret = kobject_init_and_add(pdd->kobj_stats,
587 &procfs_stats_type,
588 p->kobj,
589 stats_dir_filename);
590
591 if (ret) {
592 pr_warn("Creating KFD proc/stats_%s folder failed",
593 stats_dir_filename);
594 kobject_put(pdd->kobj_stats);
595 pdd->kobj_stats = NULL;
596 return;
597 }
598
599 kfd_sysfs_create_file(pdd->kobj_stats, &pdd->attr_evict,
600 "evicted_ms");
601 /* Add sysfs file to report compute unit occupancy */
602 if (pdd->dev->kfd2kgd->get_cu_occupancy)
603 kfd_sysfs_create_file(pdd->kobj_stats,
604 &pdd->attr_cu_occupancy,
605 "cu_occupancy");
606 }
607 }
608
kfd_procfs_add_sysfs_counters(struct kfd_process * p)609 static void kfd_procfs_add_sysfs_counters(struct kfd_process *p)
610 {
611 int ret = 0;
612 int i;
613 char counters_dir_filename[MAX_SYSFS_FILENAME_LEN];
614
615 if (!p || !p->kobj)
616 return;
617
618 /*
619 * Create sysfs files for each GPU which supports SVM
620 * - proc/<pid>/counters_<gpuid>/
621 * - proc/<pid>/counters_<gpuid>/faults
622 * - proc/<pid>/counters_<gpuid>/page_in
623 * - proc/<pid>/counters_<gpuid>/page_out
624 */
625 for_each_set_bit(i, p->svms.bitmap_supported, p->n_pdds) {
626 struct kfd_process_device *pdd = p->pdds[i];
627 struct kobject *kobj_counters;
628
629 snprintf(counters_dir_filename, MAX_SYSFS_FILENAME_LEN,
630 "counters_%u", pdd->dev->id);
631 kobj_counters = kfd_alloc_struct(kobj_counters);
632 if (!kobj_counters)
633 return;
634
635 ret = kobject_init_and_add(kobj_counters, &sysfs_counters_type,
636 p->kobj, counters_dir_filename);
637 if (ret) {
638 pr_warn("Creating KFD proc/%s folder failed",
639 counters_dir_filename);
640 kobject_put(kobj_counters);
641 return;
642 }
643
644 pdd->kobj_counters = kobj_counters;
645 kfd_sysfs_create_file(kobj_counters, &pdd->attr_faults,
646 "faults");
647 kfd_sysfs_create_file(kobj_counters, &pdd->attr_page_in,
648 "page_in");
649 kfd_sysfs_create_file(kobj_counters, &pdd->attr_page_out,
650 "page_out");
651 }
652 }
653
kfd_procfs_add_sysfs_files(struct kfd_process * p)654 static void kfd_procfs_add_sysfs_files(struct kfd_process *p)
655 {
656 int i;
657
658 if (!p || !p->kobj)
659 return;
660
661 /*
662 * Create sysfs files for each GPU:
663 * - proc/<pid>/vram_<gpuid>
664 * - proc/<pid>/sdma_<gpuid>
665 */
666 for (i = 0; i < p->n_pdds; i++) {
667 struct kfd_process_device *pdd = p->pdds[i];
668
669 snprintf(pdd->vram_filename, MAX_SYSFS_FILENAME_LEN, "vram_%u",
670 pdd->dev->id);
671 kfd_sysfs_create_file(p->kobj, &pdd->attr_vram,
672 pdd->vram_filename);
673
674 snprintf(pdd->sdma_filename, MAX_SYSFS_FILENAME_LEN, "sdma_%u",
675 pdd->dev->id);
676 kfd_sysfs_create_file(p->kobj, &pdd->attr_sdma,
677 pdd->sdma_filename);
678 }
679 }
680
kfd_procfs_del_queue(struct queue * q)681 void kfd_procfs_del_queue(struct queue *q)
682 {
683 if (!q)
684 return;
685
686 kobject_del(&q->kobj);
687 kobject_put(&q->kobj);
688 }
689
kfd_process_create_wq(void)690 int kfd_process_create_wq(void)
691 {
692 if (!kfd_process_wq)
693 kfd_process_wq = alloc_workqueue("kfd_process_wq", 0, 0);
694 if (!kfd_restore_wq)
695 kfd_restore_wq = alloc_ordered_workqueue("kfd_restore_wq",
696 WQ_FREEZABLE);
697
698 if (!kfd_process_wq || !kfd_restore_wq) {
699 kfd_process_destroy_wq();
700 return -ENOMEM;
701 }
702
703 return 0;
704 }
705
kfd_process_destroy_wq(void)706 void kfd_process_destroy_wq(void)
707 {
708 if (kfd_process_wq) {
709 destroy_workqueue(kfd_process_wq);
710 kfd_process_wq = NULL;
711 }
712 if (kfd_restore_wq) {
713 destroy_workqueue(kfd_restore_wq);
714 kfd_restore_wq = NULL;
715 }
716 }
717
kfd_process_free_gpuvm(struct kgd_mem * mem,struct kfd_process_device * pdd,void ** kptr)718 static void kfd_process_free_gpuvm(struct kgd_mem *mem,
719 struct kfd_process_device *pdd, void **kptr)
720 {
721 struct kfd_node *dev = pdd->dev;
722
723 if (kptr && *kptr) {
724 amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(mem);
725 *kptr = NULL;
726 }
727
728 amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(dev->adev, mem, pdd->drm_priv);
729 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(dev->adev, mem, pdd->drm_priv,
730 NULL);
731 }
732
733 /* kfd_process_alloc_gpuvm - Allocate GPU VM for the KFD process
734 * This function should be only called right after the process
735 * is created and when kfd_processes_mutex is still being held
736 * to avoid concurrency. Because of that exclusiveness, we do
737 * not need to take p->mutex.
738 */
kfd_process_alloc_gpuvm(struct kfd_process_device * pdd,uint64_t gpu_va,uint32_t size,uint32_t flags,struct kgd_mem ** mem,void ** kptr)739 static int kfd_process_alloc_gpuvm(struct kfd_process_device *pdd,
740 uint64_t gpu_va, uint32_t size,
741 uint32_t flags, struct kgd_mem **mem, void **kptr)
742 {
743 struct kfd_node *kdev = pdd->dev;
744 int err;
745
746 err = amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu(kdev->adev, gpu_va, size,
747 pdd->drm_priv, mem, NULL,
748 flags, false);
749 if (err)
750 goto err_alloc_mem;
751
752 err = amdgpu_amdkfd_gpuvm_map_memory_to_gpu(kdev->adev, *mem,
753 pdd->drm_priv);
754 if (err)
755 goto err_map_mem;
756
757 err = amdgpu_amdkfd_gpuvm_sync_memory(kdev->adev, *mem, true);
758 if (err) {
759 pr_debug("Sync memory failed, wait interrupted by user signal\n");
760 goto sync_memory_failed;
761 }
762
763 if (kptr) {
764 err = amdgpu_amdkfd_gpuvm_map_gtt_bo_to_kernel(
765 (struct kgd_mem *)*mem, kptr, NULL);
766 if (err) {
767 pr_debug("Map GTT BO to kernel failed\n");
768 goto sync_memory_failed;
769 }
770 }
771
772 return err;
773
774 sync_memory_failed:
775 amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(kdev->adev, *mem, pdd->drm_priv);
776
777 err_map_mem:
778 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(kdev->adev, *mem, pdd->drm_priv,
779 NULL);
780 err_alloc_mem:
781 *mem = NULL;
782 *kptr = NULL;
783 return err;
784 }
785
786 /* kfd_process_device_reserve_ib_mem - Reserve memory inside the
787 * process for IB usage The memory reserved is for KFD to submit
788 * IB to AMDGPU from kernel. If the memory is reserved
789 * successfully, ib_kaddr will have the CPU/kernel
790 * address. Check ib_kaddr before accessing the memory.
791 */
kfd_process_device_reserve_ib_mem(struct kfd_process_device * pdd)792 static int kfd_process_device_reserve_ib_mem(struct kfd_process_device *pdd)
793 {
794 struct qcm_process_device *qpd = &pdd->qpd;
795 uint32_t flags = KFD_IOC_ALLOC_MEM_FLAGS_GTT |
796 KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE |
797 KFD_IOC_ALLOC_MEM_FLAGS_WRITABLE |
798 KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
799 struct kgd_mem *mem;
800 void *kaddr;
801 int ret;
802
803 if (qpd->ib_kaddr || !qpd->ib_base)
804 return 0;
805
806 /* ib_base is only set for dGPU */
807 ret = kfd_process_alloc_gpuvm(pdd, qpd->ib_base, PAGE_SIZE, flags,
808 &mem, &kaddr);
809 if (ret)
810 return ret;
811
812 qpd->ib_mem = mem;
813 qpd->ib_kaddr = kaddr;
814
815 return 0;
816 }
817
kfd_process_device_destroy_ib_mem(struct kfd_process_device * pdd)818 static void kfd_process_device_destroy_ib_mem(struct kfd_process_device *pdd)
819 {
820 struct qcm_process_device *qpd = &pdd->qpd;
821
822 if (!qpd->ib_kaddr || !qpd->ib_base)
823 return;
824
825 kfd_process_free_gpuvm(qpd->ib_mem, pdd, &qpd->ib_kaddr);
826 }
827
kfd_create_process_sysfs(struct kfd_process * process)828 int kfd_create_process_sysfs(struct kfd_process *process)
829 {
830 struct kfd_process *primary_process;
831 int ret;
832
833 if (process->kobj) {
834 pr_warn("kobject already exists for the kfd_process\n");
835 return -EINVAL;
836 }
837
838 process->kobj = kfd_alloc_struct(process->kobj);
839 if (!process->kobj) {
840 pr_warn("Creating procfs kobject failed");
841 return -ENOMEM;
842 }
843
844 if (process->context_id == KFD_CONTEXT_ID_PRIMARY)
845 ret = kobject_init_and_add(process->kobj, &procfs_type,
846 procfs.kobj, "%d",
847 (int)process->lead_thread->pid);
848 else {
849 primary_process = kfd_lookup_process_by_mm(process->lead_thread->mm);
850 if (!primary_process)
851 return -ESRCH;
852
853 ret = kobject_init_and_add(process->kobj, &procfs_type,
854 primary_process->kobj, "context_%u",
855 process->context_id);
856 kfd_unref_process(primary_process);
857 }
858
859 if (ret) {
860 pr_warn("Creating procfs pid directory failed");
861 kobject_put(process->kobj);
862 return ret;
863 }
864
865 kfd_sysfs_create_file(process->kobj, &process->attr_pasid,
866 "pasid");
867
868 process->kobj_queues = kobject_create_and_add("queues",
869 process->kobj);
870 if (!process->kobj_queues)
871 pr_warn("Creating KFD proc/queues folder failed");
872
873 kfd_procfs_add_sysfs_stats(process);
874 kfd_procfs_add_sysfs_files(process);
875 kfd_procfs_add_sysfs_counters(process);
876
877 return 0;
878 }
879
kfd_process_alloc_id(struct kfd_process * process)880 static int kfd_process_alloc_id(struct kfd_process *process)
881 {
882 int ret;
883 struct kfd_process *primary_process;
884
885 /* already assign 0xFFFF when create */
886 if (process->context_id == KFD_CONTEXT_ID_PRIMARY)
887 return 0;
888
889 primary_process = kfd_lookup_process_by_mm(process->lead_thread->mm);
890 if (!primary_process)
891 return -ESRCH;
892
893 /* id range: KFD_CONTEXT_ID_MIN to 0xFFFE */
894 ret = ida_alloc_range(&primary_process->id_table, KFD_CONTEXT_ID_MIN,
895 KFD_CONTEXT_ID_PRIMARY - 1, GFP_KERNEL);
896 if (ret < 0)
897 goto out;
898
899 process->context_id = ret;
900 ret = 0;
901
902 out:
903 kfd_unref_process(primary_process);
904
905 return ret;
906 }
907
kfd_process_free_id(struct kfd_process * process)908 static void kfd_process_free_id(struct kfd_process *process)
909 {
910 struct kfd_process *primary_process;
911
912 if (process->context_id != KFD_CONTEXT_ID_PRIMARY)
913 return;
914
915 primary_process = kfd_lookup_process_by_mm(process->lead_thread->mm);
916 if (!primary_process)
917 return;
918
919 ida_free(&primary_process->id_table, process->context_id);
920
921 kfd_unref_process(primary_process);
922 }
923
kfd_create_process(struct task_struct * thread)924 struct kfd_process *kfd_create_process(struct task_struct *thread)
925 {
926 struct kfd_process *process;
927 int ret;
928
929 if (!(thread->mm && mmget_not_zero(thread->mm)))
930 return ERR_PTR(-EINVAL);
931
932 /* If the process just called exec(3), it is possible that the
933 * cleanup of the kfd_process (following the release of the mm
934 * of the old process image) is still in the cleanup work queue.
935 * Make sure to drain any job before trying to recreate any
936 * resource for this process.
937 */
938 flush_workqueue(kfd_process_wq);
939
940 /*
941 * take kfd processes mutex before starting of process creation
942 * so there won't be a case where two threads of the same process
943 * create two kfd_process structures
944 */
945 mutex_lock(&kfd_processes_mutex);
946
947 if (kfd_gpu_node_num() <= 0) {
948 pr_warn("no gpu node! Cannot create KFD process");
949 process = ERR_PTR(-EINVAL);
950 goto out;
951 }
952
953 if (kfd_is_locked(NULL)) {
954 pr_debug("KFD is locked! Cannot create process");
955 process = ERR_PTR(-EINVAL);
956 goto out;
957 }
958
959 /* A prior open of /dev/kfd could have already created the process.
960 * find_process will increase process kref in this case
961 */
962 process = find_process(thread, true);
963 if (process) {
964 pr_debug("Process already found\n");
965 } else {
966 process = create_process(thread, true);
967 if (IS_ERR(process))
968 goto out;
969
970 if (!procfs.kobj)
971 goto out;
972
973 ret = kfd_create_process_sysfs(process);
974 if (ret)
975 pr_warn("Failed to create sysfs entry for the kfd_process");
976
977 kfd_debugfs_add_process(process);
978
979 init_waitqueue_head(&process->wait_irq_drain);
980 }
981 out:
982 mutex_unlock(&kfd_processes_mutex);
983 mmput(thread->mm);
984
985 return process;
986 }
987
find_process_by_mm(const struct mm_struct * mm)988 static struct kfd_process *find_process_by_mm(const struct mm_struct *mm)
989 {
990 struct kfd_process *process;
991
992 hash_for_each_possible_rcu(kfd_processes_table, process,
993 kfd_processes, (uintptr_t)mm)
994 if (process->mm == mm && process->context_id == KFD_CONTEXT_ID_PRIMARY)
995 return process;
996
997 return NULL;
998 }
999
find_process(const struct task_struct * thread,bool ref)1000 static struct kfd_process *find_process(const struct task_struct *thread,
1001 bool ref)
1002 {
1003 struct kfd_process *p;
1004 int idx;
1005
1006 idx = srcu_read_lock(&kfd_processes_srcu);
1007 p = find_process_by_mm(thread->mm);
1008 if (p && ref)
1009 kref_get(&p->ref);
1010 srcu_read_unlock(&kfd_processes_srcu, idx);
1011
1012 return p;
1013 }
1014
kfd_unref_process(struct kfd_process * p)1015 void kfd_unref_process(struct kfd_process *p)
1016 {
1017 kref_put(&p->ref, kfd_process_ref_release);
1018 }
1019
1020 /* This increments the process->ref counter. */
kfd_lookup_process_by_pid(struct pid * pid)1021 struct kfd_process *kfd_lookup_process_by_pid(struct pid *pid)
1022 {
1023 struct task_struct *task = NULL;
1024 struct kfd_process *p = NULL;
1025
1026 if (!pid) {
1027 task = current;
1028 get_task_struct(task);
1029 } else {
1030 task = get_pid_task(pid, PIDTYPE_PID);
1031 }
1032
1033 if (task) {
1034 p = find_process(task, true);
1035 put_task_struct(task);
1036 }
1037
1038 return p;
1039 }
1040
kfd_process_device_free_bos(struct kfd_process_device * pdd)1041 static void kfd_process_device_free_bos(struct kfd_process_device *pdd)
1042 {
1043 struct kfd_process *p = pdd->process;
1044 void *mem;
1045 int id;
1046 int i;
1047
1048 /*
1049 * Remove all handles from idr and release appropriate
1050 * local memory object
1051 */
1052 idr_for_each_entry(&pdd->alloc_idr, mem, id) {
1053
1054 for (i = 0; i < p->n_pdds; i++) {
1055 struct kfd_process_device *peer_pdd = p->pdds[i];
1056
1057 if (!peer_pdd->drm_priv)
1058 continue;
1059 amdgpu_amdkfd_gpuvm_unmap_memory_from_gpu(
1060 peer_pdd->dev->adev, mem, peer_pdd->drm_priv);
1061 }
1062
1063 amdgpu_amdkfd_gpuvm_free_memory_of_gpu(pdd->dev->adev, mem,
1064 pdd->drm_priv, NULL);
1065 kfd_process_device_remove_obj_handle(pdd, id);
1066 }
1067 }
1068
1069 /*
1070 * Just kunmap and unpin signal BO here. It will be freed in
1071 * kfd_process_free_outstanding_kfd_bos()
1072 */
kfd_process_kunmap_signal_bo(struct kfd_process * p)1073 static void kfd_process_kunmap_signal_bo(struct kfd_process *p)
1074 {
1075 struct kfd_process_device *pdd;
1076 struct kfd_node *kdev;
1077 void *mem;
1078
1079 kdev = kfd_device_by_id(GET_GPU_ID(p->signal_handle));
1080 if (!kdev)
1081 return;
1082
1083 mutex_lock(&p->mutex);
1084
1085 pdd = kfd_get_process_device_data(kdev, p);
1086 if (!pdd)
1087 goto out;
1088
1089 mem = kfd_process_device_translate_handle(
1090 pdd, GET_IDR_HANDLE(p->signal_handle));
1091 if (!mem)
1092 goto out;
1093
1094 amdgpu_amdkfd_gpuvm_unmap_gtt_bo_from_kernel(mem);
1095
1096 out:
1097 mutex_unlock(&p->mutex);
1098 }
1099
kfd_process_free_outstanding_kfd_bos(struct kfd_process * p)1100 static void kfd_process_free_outstanding_kfd_bos(struct kfd_process *p)
1101 {
1102 int i;
1103
1104 for (i = 0; i < p->n_pdds; i++)
1105 kfd_process_device_free_bos(p->pdds[i]);
1106 }
1107
kfd_process_destroy_pdds(struct kfd_process * p)1108 static void kfd_process_destroy_pdds(struct kfd_process *p)
1109 {
1110 int i;
1111
1112 for (i = 0; i < p->n_pdds; i++) {
1113 struct kfd_process_device *pdd = p->pdds[i];
1114
1115 kfd_smi_event_process(pdd, false);
1116
1117 pr_debug("Releasing pdd (topology id %d, for pid %d)\n",
1118 pdd->dev->id, p->lead_thread->pid);
1119 kfd_process_device_destroy_cwsr_dgpu(pdd);
1120 kfd_process_device_destroy_ib_mem(pdd);
1121
1122 if (pdd->drm_file)
1123 fput(pdd->drm_file);
1124
1125 if (pdd->qpd.cwsr_kaddr && !pdd->qpd.cwsr_base)
1126 free_pages((unsigned long)pdd->qpd.cwsr_kaddr,
1127 get_order(KFD_CWSR_TBA_TMA_SIZE));
1128
1129 idr_destroy(&pdd->alloc_idr);
1130
1131 kfd_free_process_doorbells(pdd->dev->kfd, pdd);
1132
1133 if (pdd->dev->kfd->shared_resources.enable_mes &&
1134 pdd->proc_ctx_cpu_ptr)
1135 amdgpu_amdkfd_free_kernel_mem(pdd->dev->adev,
1136 &pdd->proc_ctx_bo);
1137 /*
1138 * before destroying pdd, make sure to report availability
1139 * for auto suspend
1140 */
1141 if (pdd->runtime_inuse) {
1142 pm_runtime_put_autosuspend(adev_to_drm(pdd->dev->adev)->dev);
1143 pdd->runtime_inuse = false;
1144 }
1145
1146 atomic_dec(&pdd->dev->kfd->kfd_processes_count);
1147
1148 kfree(pdd);
1149 p->pdds[i] = NULL;
1150 }
1151 p->n_pdds = 0;
1152 }
1153
kfd_process_remove_sysfs(struct kfd_process * p)1154 static void kfd_process_remove_sysfs(struct kfd_process *p)
1155 {
1156 struct kfd_process_device *pdd;
1157 int i;
1158
1159 if (!p->kobj)
1160 return;
1161
1162 sysfs_remove_file(p->kobj, &p->attr_pasid);
1163 kobject_del(p->kobj_queues);
1164 kobject_put(p->kobj_queues);
1165 p->kobj_queues = NULL;
1166
1167 for (i = 0; i < p->n_pdds; i++) {
1168 pdd = p->pdds[i];
1169
1170 sysfs_remove_file(p->kobj, &pdd->attr_vram);
1171 sysfs_remove_file(p->kobj, &pdd->attr_sdma);
1172
1173 sysfs_remove_file(pdd->kobj_stats, &pdd->attr_evict);
1174 if (pdd->dev->kfd2kgd->get_cu_occupancy)
1175 sysfs_remove_file(pdd->kobj_stats,
1176 &pdd->attr_cu_occupancy);
1177 kobject_del(pdd->kobj_stats);
1178 kobject_put(pdd->kobj_stats);
1179 pdd->kobj_stats = NULL;
1180 }
1181
1182 for_each_set_bit(i, p->svms.bitmap_supported, p->n_pdds) {
1183 pdd = p->pdds[i];
1184
1185 sysfs_remove_file(pdd->kobj_counters, &pdd->attr_faults);
1186 sysfs_remove_file(pdd->kobj_counters, &pdd->attr_page_in);
1187 sysfs_remove_file(pdd->kobj_counters, &pdd->attr_page_out);
1188 kobject_del(pdd->kobj_counters);
1189 kobject_put(pdd->kobj_counters);
1190 pdd->kobj_counters = NULL;
1191 }
1192
1193 kobject_del(p->kobj);
1194 kobject_put(p->kobj);
1195 p->kobj = NULL;
1196 }
1197
1198 /*
1199 * If any GPU is ongoing reset, wait for reset complete.
1200 */
kfd_process_wait_gpu_reset_complete(struct kfd_process * p)1201 static void kfd_process_wait_gpu_reset_complete(struct kfd_process *p)
1202 {
1203 int i;
1204
1205 for (i = 0; i < p->n_pdds; i++)
1206 flush_workqueue(p->pdds[i]->dev->adev->reset_domain->wq);
1207 }
1208
1209 /* No process locking is needed in this function, because the process
1210 * is not findable any more. We must assume that no other thread is
1211 * using it any more, otherwise we couldn't safely free the process
1212 * structure in the end.
1213 */
kfd_process_wq_release(struct work_struct * work)1214 static void kfd_process_wq_release(struct work_struct *work)
1215 {
1216 struct kfd_process *p = container_of(work, struct kfd_process,
1217 release_work);
1218 struct dma_fence *ef;
1219
1220 /*
1221 * If GPU in reset, user queues may still running, wait for reset complete.
1222 */
1223 kfd_process_wait_gpu_reset_complete(p);
1224
1225 /* Signal the eviction fence after user mode queues are
1226 * destroyed. This allows any BOs to be freed without
1227 * triggering pointless evictions or waiting for fences.
1228 */
1229 synchronize_rcu();
1230 ef = rcu_access_pointer(p->ef);
1231 if (ef)
1232 dma_fence_signal(ef);
1233
1234 if (p->context_id != KFD_CONTEXT_ID_PRIMARY)
1235 kfd_process_free_id(p);
1236 else
1237 ida_destroy(&p->id_table);
1238
1239 kfd_debugfs_remove_process(p);
1240
1241 kfd_process_kunmap_signal_bo(p);
1242 kfd_process_free_outstanding_kfd_bos(p);
1243 svm_range_list_fini(p);
1244
1245 kfd_process_destroy_pdds(p);
1246 dma_fence_put(ef);
1247
1248 kfd_event_free_process(p);
1249
1250 mutex_destroy(&p->mutex);
1251
1252 put_task_struct(p->lead_thread);
1253
1254 /* the last step is removing process entries under /sys
1255 * to indicate the process has been terminated.
1256 */
1257 kfd_process_remove_sysfs(p);
1258
1259 kfree(p);
1260 }
1261
kfd_process_ref_release(struct kref * ref)1262 static void kfd_process_ref_release(struct kref *ref)
1263 {
1264 struct kfd_process *p = container_of(ref, struct kfd_process, ref);
1265
1266 INIT_WORK(&p->release_work, kfd_process_wq_release);
1267 queue_work(kfd_process_wq, &p->release_work);
1268 }
1269
kfd_process_alloc_notifier(struct mm_struct * mm)1270 static struct mmu_notifier *kfd_process_alloc_notifier(struct mm_struct *mm)
1271 {
1272 /* This increments p->ref counter if kfd process p exists */
1273 struct kfd_process *p = kfd_lookup_process_by_mm(mm);
1274
1275 return p ? &p->mmu_notifier : ERR_PTR(-ESRCH);
1276 }
1277
kfd_process_free_notifier(struct mmu_notifier * mn)1278 static void kfd_process_free_notifier(struct mmu_notifier *mn)
1279 {
1280 kfd_unref_process(container_of(mn, struct kfd_process, mmu_notifier));
1281 }
1282
kfd_process_table_remove(struct kfd_process * p)1283 static void kfd_process_table_remove(struct kfd_process *p)
1284 {
1285 mutex_lock(&kfd_processes_mutex);
1286 /*
1287 * Do early return if table is empty.
1288 *
1289 * This could potentially happen if this function is called concurrently
1290 * by mmu_notifier and by kfd_cleanup_pocesses.
1291 *
1292 */
1293 if (hash_empty(kfd_processes_table)) {
1294 mutex_unlock(&kfd_processes_mutex);
1295 return;
1296 }
1297 hash_del_rcu(&p->kfd_processes);
1298 mutex_unlock(&kfd_processes_mutex);
1299 synchronize_srcu(&kfd_processes_srcu);
1300 }
1301
kfd_process_notifier_release_internal(struct kfd_process * p)1302 void kfd_process_notifier_release_internal(struct kfd_process *p)
1303 {
1304 int i;
1305
1306 kfd_process_table_remove(p);
1307 cancel_delayed_work_sync(&p->eviction_work);
1308 cancel_delayed_work_sync(&p->restore_work);
1309
1310 /*
1311 * Dequeue and destroy user queues, it is not safe for GPU to access
1312 * system memory after mmu release notifier callback returns because
1313 * exit_mmap free process memory afterwards.
1314 */
1315 kfd_process_dequeue_from_all_devices(p);
1316 pqm_uninit(&p->pqm);
1317
1318 for (i = 0; i < p->n_pdds; i++) {
1319 struct kfd_process_device *pdd = p->pdds[i];
1320
1321 /* re-enable GFX OFF since runtime enable with ttmp setup disabled it. */
1322 if (!kfd_dbg_is_rlc_restore_supported(pdd->dev) && p->runtime_info.ttmp_setup)
1323 amdgpu_gfx_off_ctrl(pdd->dev->adev, true);
1324 }
1325
1326 /* Indicate to other users that MM is no longer valid */
1327 p->mm = NULL;
1328 kfd_dbg_trap_disable(p);
1329
1330 if (atomic_read(&p->debugged_process_count) > 0) {
1331 struct kfd_process *target;
1332 unsigned int temp;
1333 int idx = srcu_read_lock(&kfd_processes_srcu);
1334
1335 hash_for_each_rcu(kfd_processes_table, temp, target, kfd_processes) {
1336 if (target->debugger_process && target->debugger_process == p) {
1337 mutex_lock_nested(&target->mutex, 1);
1338 kfd_dbg_trap_disable(target);
1339 mutex_unlock(&target->mutex);
1340 if (atomic_read(&p->debugged_process_count) == 0)
1341 break;
1342 }
1343 }
1344
1345 srcu_read_unlock(&kfd_processes_srcu, idx);
1346 }
1347
1348 if (p->context_id == KFD_CONTEXT_ID_PRIMARY)
1349 mmu_notifier_put(&p->mmu_notifier);
1350 }
1351
kfd_process_notifier_release(struct mmu_notifier * mn,struct mm_struct * mm)1352 static void kfd_process_notifier_release(struct mmu_notifier *mn,
1353 struct mm_struct *mm)
1354 {
1355 struct kfd_process *p;
1356
1357 /*
1358 * The kfd_process structure can not be free because the
1359 * mmu_notifier srcu is read locked
1360 */
1361 p = container_of(mn, struct kfd_process, mmu_notifier);
1362 if (WARN_ON(p->mm != mm))
1363 return;
1364
1365 kfd_process_notifier_release_internal(p);
1366 }
1367
1368 static const struct mmu_notifier_ops kfd_process_mmu_notifier_ops = {
1369 .release = kfd_process_notifier_release,
1370 .alloc_notifier = kfd_process_alloc_notifier,
1371 .free_notifier = kfd_process_free_notifier,
1372 };
1373
1374 /*
1375 * This code handles the case when driver is being unloaded before all
1376 * mm_struct are released. We need to safely free the kfd_process and
1377 * avoid race conditions with mmu_notifier that might try to free them.
1378 *
1379 */
kfd_cleanup_processes(void)1380 void kfd_cleanup_processes(void)
1381 {
1382 struct kfd_process *p;
1383 struct hlist_node *p_temp;
1384 unsigned int temp;
1385 HLIST_HEAD(cleanup_list);
1386
1387 /*
1388 * Move all remaining kfd_process from the process table to a
1389 * temp list for processing. Once done, callback from mmu_notifier
1390 * release will not see the kfd_process in the table and do early return,
1391 * avoiding double free issues.
1392 */
1393 mutex_lock(&kfd_processes_mutex);
1394 hash_for_each_safe(kfd_processes_table, temp, p_temp, p, kfd_processes) {
1395 hash_del_rcu(&p->kfd_processes);
1396 synchronize_srcu(&kfd_processes_srcu);
1397 hlist_add_head(&p->kfd_processes, &cleanup_list);
1398 }
1399 mutex_unlock(&kfd_processes_mutex);
1400
1401 hlist_for_each_entry_safe(p, p_temp, &cleanup_list, kfd_processes)
1402 kfd_process_notifier_release_internal(p);
1403
1404 /*
1405 * Ensures that all outstanding free_notifier get called, triggering
1406 * the release of the kfd_process struct.
1407 */
1408 mmu_notifier_synchronize();
1409 }
1410
kfd_process_init_cwsr_apu(struct kfd_process * p,struct file * filep)1411 int kfd_process_init_cwsr_apu(struct kfd_process *p, struct file *filep)
1412 {
1413 unsigned long offset;
1414 int i;
1415
1416 if (p->has_cwsr)
1417 return 0;
1418
1419 for (i = 0; i < p->n_pdds; i++) {
1420 struct kfd_node *dev = p->pdds[i]->dev;
1421 struct qcm_process_device *qpd = &p->pdds[i]->qpd;
1422
1423 if (!dev->kfd->cwsr_enabled || qpd->cwsr_kaddr || qpd->cwsr_base)
1424 continue;
1425
1426 offset = KFD_MMAP_TYPE_RESERVED_MEM | KFD_MMAP_GPU_ID(dev->id);
1427 qpd->tba_addr = (int64_t)vm_mmap(filep, 0,
1428 KFD_CWSR_TBA_TMA_SIZE, PROT_READ | PROT_EXEC,
1429 MAP_SHARED, offset);
1430
1431 if (IS_ERR_VALUE(qpd->tba_addr)) {
1432 int err = qpd->tba_addr;
1433
1434 dev_err(dev->adev->dev,
1435 "Failure to set tba address. error %d.\n", err);
1436 qpd->tba_addr = 0;
1437 qpd->cwsr_kaddr = NULL;
1438 return err;
1439 }
1440
1441 memcpy(qpd->cwsr_kaddr, dev->kfd->cwsr_isa, dev->kfd->cwsr_isa_size);
1442
1443 kfd_process_set_trap_debug_flag(qpd, p->debug_trap_enabled);
1444
1445 qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET;
1446 pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n",
1447 qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr);
1448 }
1449
1450 p->has_cwsr = true;
1451
1452 return 0;
1453 }
1454
kfd_process_device_init_cwsr_dgpu(struct kfd_process_device * pdd)1455 static int kfd_process_device_init_cwsr_dgpu(struct kfd_process_device *pdd)
1456 {
1457 struct kfd_node *dev = pdd->dev;
1458 struct qcm_process_device *qpd = &pdd->qpd;
1459 uint32_t flags = KFD_IOC_ALLOC_MEM_FLAGS_GTT
1460 | KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE
1461 | KFD_IOC_ALLOC_MEM_FLAGS_EXECUTABLE;
1462 struct kgd_mem *mem;
1463 void *kaddr;
1464 int ret;
1465
1466 if (!dev->kfd->cwsr_enabled || qpd->cwsr_kaddr || !qpd->cwsr_base)
1467 return 0;
1468
1469 /* cwsr_base is only set for dGPU */
1470 ret = kfd_process_alloc_gpuvm(pdd, qpd->cwsr_base,
1471 KFD_CWSR_TBA_TMA_SIZE, flags, &mem, &kaddr);
1472 if (ret)
1473 return ret;
1474
1475 qpd->cwsr_mem = mem;
1476 qpd->cwsr_kaddr = kaddr;
1477 qpd->tba_addr = qpd->cwsr_base;
1478
1479 memcpy(qpd->cwsr_kaddr, dev->kfd->cwsr_isa, dev->kfd->cwsr_isa_size);
1480
1481 kfd_process_set_trap_debug_flag(&pdd->qpd,
1482 pdd->process->debug_trap_enabled);
1483
1484 qpd->tma_addr = qpd->tba_addr + KFD_CWSR_TMA_OFFSET;
1485 pr_debug("set tba :0x%llx, tma:0x%llx, cwsr_kaddr:%p for pqm.\n",
1486 qpd->tba_addr, qpd->tma_addr, qpd->cwsr_kaddr);
1487
1488 return 0;
1489 }
1490
kfd_process_device_destroy_cwsr_dgpu(struct kfd_process_device * pdd)1491 static void kfd_process_device_destroy_cwsr_dgpu(struct kfd_process_device *pdd)
1492 {
1493 struct kfd_node *dev = pdd->dev;
1494 struct qcm_process_device *qpd = &pdd->qpd;
1495
1496 if (!dev->kfd->cwsr_enabled || !qpd->cwsr_kaddr || !qpd->cwsr_base)
1497 return;
1498
1499 kfd_process_free_gpuvm(qpd->cwsr_mem, pdd, &qpd->cwsr_kaddr);
1500 }
1501
kfd_process_set_trap_handler(struct qcm_process_device * qpd,uint64_t tba_addr,uint64_t tma_addr)1502 void kfd_process_set_trap_handler(struct qcm_process_device *qpd,
1503 uint64_t tba_addr,
1504 uint64_t tma_addr)
1505 {
1506 if (qpd->cwsr_kaddr) {
1507 /* KFD trap handler is bound, record as second-level TBA/TMA
1508 * in first-level TMA. First-level trap will jump to second.
1509 */
1510 uint64_t *tma =
1511 (uint64_t *)(qpd->cwsr_kaddr + KFD_CWSR_TMA_OFFSET);
1512 tma[0] = tba_addr;
1513 tma[1] = tma_addr;
1514 } else {
1515 /* No trap handler bound, bind as first-level TBA/TMA. */
1516 qpd->tba_addr = tba_addr;
1517 qpd->tma_addr = tma_addr;
1518 }
1519 }
1520
kfd_process_xnack_mode(struct kfd_process * p,bool supported)1521 bool kfd_process_xnack_mode(struct kfd_process *p, bool supported)
1522 {
1523 int i;
1524
1525 /* On most GFXv9 GPUs, the retry mode in the SQ must match the
1526 * boot time retry setting. Mixing processes with different
1527 * XNACK/retry settings can hang the GPU.
1528 *
1529 * Different GPUs can have different noretry settings depending
1530 * on HW bugs or limitations. We need to find at least one
1531 * XNACK mode for this process that's compatible with all GPUs.
1532 * Fortunately GPUs with retry enabled (noretry=0) can run code
1533 * built for XNACK-off. On GFXv9 it may perform slower.
1534 *
1535 * Therefore applications built for XNACK-off can always be
1536 * supported and will be our fallback if any GPU does not
1537 * support retry.
1538 */
1539 for (i = 0; i < p->n_pdds; i++) {
1540 struct kfd_node *dev = p->pdds[i]->dev;
1541
1542 /* Only consider GFXv9 and higher GPUs. Older GPUs don't
1543 * support the SVM APIs and don't need to be considered
1544 * for the XNACK mode selection.
1545 */
1546 if (!KFD_IS_SOC15(dev))
1547 continue;
1548 /* Aldebaran can always support XNACK because it can support
1549 * per-process XNACK mode selection. But let the dev->noretry
1550 * setting still influence the default XNACK mode.
1551 */
1552 if (supported && KFD_SUPPORT_XNACK_PER_PROCESS(dev)) {
1553 if (!amdgpu_sriov_xnack_support(dev->kfd->adev)) {
1554 pr_debug("SRIOV platform xnack not supported\n");
1555 return false;
1556 }
1557 continue;
1558 }
1559
1560 /* GFXv10 and later GPUs do not support shader preemption
1561 * during page faults. This can lead to poor QoS for queue
1562 * management and memory-manager-related preemptions or
1563 * even deadlocks.
1564 */
1565 if (KFD_GC_VERSION(dev) >= IP_VERSION(10, 1, 1) &&
1566 KFD_GC_VERSION(dev) < IP_VERSION(12, 1, 0))
1567 return false;
1568
1569 if (dev->kfd->noretry)
1570 return false;
1571 }
1572
1573 return true;
1574 }
1575
kfd_process_set_trap_debug_flag(struct qcm_process_device * qpd,bool enabled)1576 void kfd_process_set_trap_debug_flag(struct qcm_process_device *qpd,
1577 bool enabled)
1578 {
1579 if (qpd->cwsr_kaddr) {
1580 uint64_t *tma =
1581 (uint64_t *)(qpd->cwsr_kaddr + KFD_CWSR_TMA_OFFSET);
1582 tma[2] = enabled;
1583 }
1584 }
1585
1586 /*
1587 * On return the kfd_process is fully operational and will be freed when the
1588 * mm is released
1589 */
create_process(const struct task_struct * thread,bool primary)1590 struct kfd_process *create_process(const struct task_struct *thread, bool primary)
1591 {
1592 struct kfd_process *process;
1593 struct mmu_notifier *mn;
1594 int err = -ENOMEM;
1595
1596 process = kzalloc_obj(*process, GFP_KERNEL);
1597 if (!process)
1598 goto err_alloc_process;
1599
1600 kref_init(&process->ref);
1601 mutex_init(&process->mutex);
1602 process->mm = thread->mm;
1603 process->lead_thread = thread->group_leader;
1604 process->n_pdds = 0;
1605 process->queues_paused = false;
1606
1607 INIT_DELAYED_WORK(&process->eviction_work, evict_process_worker);
1608 INIT_DELAYED_WORK(&process->restore_work, restore_process_worker);
1609 process->last_restore_timestamp = get_jiffies_64();
1610 err = kfd_event_init_process(process);
1611 if (err)
1612 goto err_event_init;
1613 process->is_32bit_user_mode = in_compat_syscall();
1614 process->debug_trap_enabled = false;
1615 process->debugger_process = NULL;
1616 process->exception_enable_mask = 0;
1617 atomic_set(&process->debugged_process_count, 0);
1618 sema_init(&process->runtime_enable_sema, 0);
1619
1620 err = pqm_init(&process->pqm, process);
1621 if (err != 0)
1622 goto err_process_pqm_init;
1623
1624 /* init process apertures*/
1625 err = kfd_init_apertures(process);
1626 if (err != 0)
1627 goto err_init_apertures;
1628
1629 /* Check XNACK support after PDDs are created in kfd_init_apertures */
1630 process->xnack_enabled = kfd_process_xnack_mode(process, false);
1631
1632 err = svm_range_list_init(process);
1633 if (err)
1634 goto err_init_svm_range_list;
1635
1636 /* alloc_notifier needs to find the process in the hash table */
1637 hash_add_rcu(kfd_processes_table, &process->kfd_processes,
1638 (uintptr_t)process->mm);
1639
1640 /* Avoid free_notifier to start kfd_process_wq_release if
1641 * mmu_notifier_get failed because of pending signal.
1642 */
1643 kref_get(&process->ref);
1644
1645 /* MMU notifier registration must be the last call that can fail
1646 * because after this point we cannot unwind the process creation.
1647 * After this point, mmu_notifier_put will trigger the cleanup by
1648 * dropping the last process reference in the free_notifier.
1649 */
1650 if (primary) {
1651 process->context_id = KFD_CONTEXT_ID_PRIMARY;
1652 mn = mmu_notifier_get(&kfd_process_mmu_notifier_ops, process->mm);
1653 if (IS_ERR(mn)) {
1654 err = PTR_ERR(mn);
1655 goto err_register_notifier;
1656 }
1657 BUG_ON(mn != &process->mmu_notifier);
1658 ida_init(&process->id_table);
1659 }
1660
1661 err = kfd_process_alloc_id(process);
1662 if (err) {
1663 pr_err("Creating kfd process: failed to alloc an id\n");
1664 goto err_alloc_id;
1665 }
1666
1667 kfd_unref_process(process);
1668 get_task_struct(process->lead_thread);
1669
1670 INIT_WORK(&process->debug_event_workarea, debug_event_write_work_handler);
1671
1672 return process;
1673
1674 err_alloc_id:
1675 kfd_process_free_id(process);
1676 err_register_notifier:
1677 hash_del_rcu(&process->kfd_processes);
1678 svm_range_list_fini(process);
1679 err_init_svm_range_list:
1680 kfd_process_free_outstanding_kfd_bos(process);
1681 kfd_process_destroy_pdds(process);
1682 err_init_apertures:
1683 pqm_uninit(&process->pqm);
1684 err_process_pqm_init:
1685 kfd_event_free_process(process);
1686 err_event_init:
1687 mutex_destroy(&process->mutex);
1688 kfree(process);
1689 err_alloc_process:
1690 return ERR_PTR(err);
1691 }
1692
kfd_get_process_device_data(struct kfd_node * dev,struct kfd_process * p)1693 struct kfd_process_device *kfd_get_process_device_data(struct kfd_node *dev,
1694 struct kfd_process *p)
1695 {
1696 int i;
1697
1698 for (i = 0; i < p->n_pdds; i++)
1699 if (p->pdds[i]->dev == dev)
1700 return p->pdds[i];
1701
1702 return NULL;
1703 }
1704
kfd_create_process_device_data(struct kfd_node * dev,struct kfd_process * p)1705 struct kfd_process_device *kfd_create_process_device_data(struct kfd_node *dev,
1706 struct kfd_process *p)
1707 {
1708 struct kfd_process_device *pdd = NULL;
1709
1710 if (WARN_ON_ONCE(p->n_pdds >= MAX_GPU_INSTANCE))
1711 return NULL;
1712 pdd = kzalloc_obj(*pdd, GFP_KERNEL);
1713 if (!pdd)
1714 return NULL;
1715
1716 pdd->dev = dev;
1717 INIT_LIST_HEAD(&pdd->qpd.queues_list);
1718 INIT_LIST_HEAD(&pdd->qpd.priv_queue_list);
1719 pdd->qpd.dqm = dev->dqm;
1720 pdd->qpd.pqm = &p->pqm;
1721 pdd->qpd.evicted = 0;
1722 pdd->qpd.mapped_gws_queue = false;
1723 pdd->process = p;
1724 pdd->bound = PDD_UNBOUND;
1725 pdd->already_dequeued = false;
1726 pdd->runtime_inuse = false;
1727 atomic64_set(&pdd->vram_usage, 0);
1728 pdd->sdma_past_activity_counter = 0;
1729 pdd->user_gpu_id = dev->id;
1730 atomic64_set(&pdd->evict_duration_counter, 0);
1731
1732 p->pdds[p->n_pdds++] = pdd;
1733 if (kfd_dbg_is_per_vmid_supported(pdd->dev))
1734 pdd->spi_dbg_override = pdd->dev->kfd2kgd->disable_debug_trap(
1735 pdd->dev->adev,
1736 false,
1737 0);
1738
1739 /* Init idr used for memory handle translation */
1740 idr_init(&pdd->alloc_idr);
1741
1742 atomic_inc(&dev->kfd->kfd_processes_count);
1743
1744 return pdd;
1745 }
1746
1747 /**
1748 * kfd_process_device_init_vm - Initialize a VM for a process-device
1749 *
1750 * @pdd: The process-device
1751 * @drm_file: Optional pointer to a DRM file descriptor
1752 *
1753 * If @drm_file is specified, it will be used to acquire the VM from
1754 * that file descriptor. If successful, the @pdd takes ownership of
1755 * the file descriptor.
1756 *
1757 * If @drm_file is NULL, a new VM is created.
1758 *
1759 * Returns 0 on success, -errno on failure.
1760 */
kfd_process_device_init_vm(struct kfd_process_device * pdd,struct file * drm_file)1761 int kfd_process_device_init_vm(struct kfd_process_device *pdd,
1762 struct file *drm_file)
1763 {
1764 struct amdgpu_fpriv *drv_priv;
1765 struct amdgpu_vm *avm;
1766 struct kfd_process *p;
1767 struct dma_fence *ef;
1768 struct kfd_node *dev;
1769 int ret;
1770
1771 if (pdd->drm_priv)
1772 return -EBUSY;
1773
1774 ret = amdgpu_file_to_fpriv(drm_file, &drv_priv);
1775 if (ret)
1776 return ret;
1777 avm = &drv_priv->vm;
1778
1779 p = pdd->process;
1780 dev = pdd->dev;
1781
1782 ret = amdgpu_amdkfd_gpuvm_acquire_process_vm(dev->adev, avm,
1783 &p->kgd_process_info,
1784 p->ef ? NULL : &ef);
1785 if (ret) {
1786 dev_err(dev->adev->dev, "Failed to create process VM object\n");
1787 return ret;
1788 }
1789
1790 if (!p->ef)
1791 RCU_INIT_POINTER(p->ef, ef);
1792
1793 pdd->drm_priv = drm_file->private_data;
1794
1795 ret = kfd_process_device_reserve_ib_mem(pdd);
1796 if (ret)
1797 goto err_reserve_ib_mem;
1798 ret = kfd_process_device_init_cwsr_dgpu(pdd);
1799 if (ret)
1800 goto err_init_cwsr;
1801
1802 if (unlikely(!avm->pasid)) {
1803 dev_warn(pdd->dev->adev->dev, "WARN: vm %p has no pasid associated",
1804 avm);
1805 ret = -EINVAL;
1806 goto err_get_pasid;
1807 }
1808
1809 pdd->pasid = avm->pasid;
1810 pdd->drm_file = drm_file;
1811
1812 kfd_smi_event_process(pdd, true);
1813
1814 return 0;
1815
1816 err_get_pasid:
1817 kfd_process_device_destroy_cwsr_dgpu(pdd);
1818 err_init_cwsr:
1819 kfd_process_device_destroy_ib_mem(pdd);
1820 err_reserve_ib_mem:
1821 pdd->drm_priv = NULL;
1822 amdgpu_amdkfd_gpuvm_destroy_cb(dev->adev, avm);
1823
1824 return ret;
1825 }
1826
1827 /*
1828 * Direct the IOMMU to bind the process (specifically the pasid->mm)
1829 * to the device.
1830 * Unbinding occurs when the process dies or the device is removed.
1831 *
1832 * Assumes that the process lock is held.
1833 */
kfd_bind_process_to_device(struct kfd_node * dev,struct kfd_process * p)1834 struct kfd_process_device *kfd_bind_process_to_device(struct kfd_node *dev,
1835 struct kfd_process *p)
1836 {
1837 struct kfd_process_device *pdd;
1838 int err;
1839
1840 pdd = kfd_get_process_device_data(dev, p);
1841 if (!pdd) {
1842 dev_err(dev->adev->dev, "Process device data doesn't exist\n");
1843 return ERR_PTR(-ENOMEM);
1844 }
1845
1846 if (!pdd->drm_priv)
1847 return ERR_PTR(-ENODEV);
1848
1849 /*
1850 * signal runtime-pm system to auto resume and prevent
1851 * further runtime suspend once device pdd is created until
1852 * pdd is destroyed.
1853 */
1854 if (!pdd->runtime_inuse) {
1855 err = pm_runtime_get_sync(adev_to_drm(dev->adev)->dev);
1856 if (err < 0) {
1857 pm_runtime_put_autosuspend(adev_to_drm(dev->adev)->dev);
1858 return ERR_PTR(err);
1859 }
1860 }
1861
1862 /*
1863 * make sure that runtime_usage counter is incremented just once
1864 * per pdd
1865 */
1866 pdd->runtime_inuse = true;
1867
1868 return pdd;
1869 }
1870
1871 /* Create specific handle mapped to mem from process local memory idr
1872 * Assumes that the process lock is held.
1873 */
kfd_process_device_create_obj_handle(struct kfd_process_device * pdd,void * mem)1874 int kfd_process_device_create_obj_handle(struct kfd_process_device *pdd,
1875 void *mem)
1876 {
1877 return idr_alloc(&pdd->alloc_idr, mem, 0, 0, GFP_KERNEL);
1878 }
1879
1880 /* Translate specific handle from process local memory idr
1881 * Assumes that the process lock is held.
1882 */
kfd_process_device_translate_handle(struct kfd_process_device * pdd,int handle)1883 void *kfd_process_device_translate_handle(struct kfd_process_device *pdd,
1884 int handle)
1885 {
1886 if (handle < 0)
1887 return NULL;
1888
1889 return idr_find(&pdd->alloc_idr, handle);
1890 }
1891
1892 /* Remove specific handle from process local memory idr
1893 * Assumes that the process lock is held.
1894 */
kfd_process_device_remove_obj_handle(struct kfd_process_device * pdd,int handle)1895 void kfd_process_device_remove_obj_handle(struct kfd_process_device *pdd,
1896 int handle)
1897 {
1898 if (handle >= 0)
1899 idr_remove(&pdd->alloc_idr, handle);
1900 }
1901
kfd_lookup_process_device_by_pasid(u32 pasid)1902 static struct kfd_process_device *kfd_lookup_process_device_by_pasid(u32 pasid)
1903 {
1904 struct kfd_process_device *ret_p = NULL;
1905 struct kfd_process *p;
1906 unsigned int temp;
1907 int i;
1908
1909 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1910 for (i = 0; i < p->n_pdds; i++) {
1911 if (p->pdds[i]->pasid == pasid) {
1912 ret_p = p->pdds[i];
1913 break;
1914 }
1915 }
1916 if (ret_p)
1917 break;
1918 }
1919 return ret_p;
1920 }
1921
1922 /* This increments the process->ref counter. */
kfd_lookup_process_by_pasid(u32 pasid,struct kfd_process_device ** pdd)1923 struct kfd_process *kfd_lookup_process_by_pasid(u32 pasid,
1924 struct kfd_process_device **pdd)
1925 {
1926 struct kfd_process_device *ret_p;
1927
1928 int idx = srcu_read_lock(&kfd_processes_srcu);
1929
1930 ret_p = kfd_lookup_process_device_by_pasid(pasid);
1931 if (ret_p) {
1932 if (pdd)
1933 *pdd = ret_p;
1934 kref_get(&ret_p->process->ref);
1935
1936 srcu_read_unlock(&kfd_processes_srcu, idx);
1937 return ret_p->process;
1938 }
1939
1940 srcu_read_unlock(&kfd_processes_srcu, idx);
1941
1942 if (pdd)
1943 *pdd = NULL;
1944
1945 return NULL;
1946 }
1947
1948 /* This increments the process->ref counter. */
kfd_lookup_process_by_mm(const struct mm_struct * mm)1949 struct kfd_process *kfd_lookup_process_by_mm(const struct mm_struct *mm)
1950 {
1951 struct kfd_process *p;
1952
1953 int idx = srcu_read_lock(&kfd_processes_srcu);
1954
1955 p = find_process_by_mm(mm);
1956 if (p)
1957 kref_get(&p->ref);
1958
1959 srcu_read_unlock(&kfd_processes_srcu, idx);
1960
1961 return p;
1962 }
1963
1964 /* This increments the process->ref counter. */
kfd_lookup_process_by_id(const struct mm_struct * mm,u16 id)1965 struct kfd_process *kfd_lookup_process_by_id(const struct mm_struct *mm, u16 id)
1966 {
1967 struct kfd_process *p, *ret_p = NULL;
1968 unsigned int temp;
1969
1970 int idx = srcu_read_lock(&kfd_processes_srcu);
1971
1972 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1973 if (p->mm == mm && p->context_id == id) {
1974 kref_get(&p->ref);
1975 ret_p = p;
1976 break;
1977 }
1978 }
1979
1980 srcu_read_unlock(&kfd_processes_srcu, idx);
1981
1982 return ret_p;
1983 }
1984
1985 /* kfd_process_evict_queues - Evict all user queues of a process
1986 *
1987 * Eviction is reference-counted per process-device. This means multiple
1988 * evictions from different sources can be nested safely.
1989 */
kfd_process_evict_queues(struct kfd_process * p,uint32_t trigger)1990 int kfd_process_evict_queues(struct kfd_process *p, uint32_t trigger)
1991 {
1992 int r = 0;
1993 int i;
1994 unsigned int n_evicted = 0;
1995
1996 for (i = 0; i < p->n_pdds; i++) {
1997 struct kfd_process_device *pdd = p->pdds[i];
1998 struct device *dev = pdd->dev->adev->dev;
1999
2000 kfd_smi_event_queue_eviction(pdd->dev, p->lead_thread->pid,
2001 trigger);
2002
2003 r = pdd->dev->dqm->ops.evict_process_queues(pdd->dev->dqm,
2004 &pdd->qpd);
2005 /* evict return -EIO if HWS is hang or asic is resetting, in this case
2006 * we would like to set all the queues to be in evicted state to prevent
2007 * them been add back since they actually not be saved right now.
2008 */
2009 if (r && r != -EIO) {
2010 dev_err(dev, "Failed to evict process queues\n");
2011 goto fail;
2012 }
2013 n_evicted++;
2014
2015 pdd->dev->dqm->is_hws_hang = false;
2016 }
2017
2018 return r;
2019
2020 fail:
2021 /* To keep state consistent, roll back partial eviction by
2022 * restoring queues
2023 */
2024 for (i = 0; i < p->n_pdds; i++) {
2025 struct kfd_process_device *pdd = p->pdds[i];
2026
2027 if (n_evicted == 0)
2028 break;
2029
2030 kfd_smi_event_queue_restore(pdd->dev, p->lead_thread->pid);
2031
2032 if (pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm,
2033 &pdd->qpd))
2034 dev_err(pdd->dev->adev->dev,
2035 "Failed to restore queues\n");
2036
2037 n_evicted--;
2038 }
2039
2040 return r;
2041 }
2042
2043 /* kfd_process_restore_queues - Restore all user queues of a process */
kfd_process_restore_queues(struct kfd_process * p)2044 int kfd_process_restore_queues(struct kfd_process *p)
2045 {
2046 int r, ret = 0;
2047 int i;
2048
2049 for (i = 0; i < p->n_pdds; i++) {
2050 struct kfd_process_device *pdd = p->pdds[i];
2051 struct device *dev = pdd->dev->adev->dev;
2052
2053 kfd_smi_event_queue_restore(pdd->dev, p->lead_thread->pid);
2054
2055 r = pdd->dev->dqm->ops.restore_process_queues(pdd->dev->dqm,
2056 &pdd->qpd);
2057 if (r) {
2058 dev_err(dev, "Failed to restore process queues\n");
2059 if (!ret)
2060 ret = r;
2061 }
2062 }
2063
2064 return ret;
2065 }
2066
kfd_process_gpuidx_from_gpuid(struct kfd_process * p,uint32_t gpu_id)2067 int kfd_process_gpuidx_from_gpuid(struct kfd_process *p, uint32_t gpu_id)
2068 {
2069 int i;
2070
2071 for (i = 0; i < p->n_pdds; i++)
2072 if (p->pdds[i] && gpu_id == p->pdds[i]->user_gpu_id)
2073 return i;
2074 return -EINVAL;
2075 }
2076
2077 int
kfd_process_gpuid_from_node(struct kfd_process * p,struct kfd_node * node,uint32_t * gpuid,uint32_t * gpuidx)2078 kfd_process_gpuid_from_node(struct kfd_process *p, struct kfd_node *node,
2079 uint32_t *gpuid, uint32_t *gpuidx)
2080 {
2081 int i;
2082
2083 for (i = 0; i < p->n_pdds; i++)
2084 if (p->pdds[i] && p->pdds[i]->dev == node) {
2085 *gpuid = p->pdds[i]->user_gpu_id;
2086 *gpuidx = i;
2087 return 0;
2088 }
2089 return -EINVAL;
2090 }
2091
signal_eviction_fence(struct kfd_process * p)2092 static bool signal_eviction_fence(struct kfd_process *p)
2093 {
2094 struct dma_fence *ef;
2095 bool ret;
2096
2097 rcu_read_lock();
2098 ef = dma_fence_get_rcu_safe(&p->ef);
2099 rcu_read_unlock();
2100 if (!ef)
2101 return true;
2102
2103 ret = dma_fence_check_and_signal(ef);
2104 dma_fence_put(ef);
2105
2106 return ret;
2107 }
2108
evict_process_worker(struct work_struct * work)2109 static void evict_process_worker(struct work_struct *work)
2110 {
2111 int ret;
2112 struct kfd_process *p;
2113 struct delayed_work *dwork;
2114
2115 dwork = to_delayed_work(work);
2116
2117 /* Process termination destroys this worker thread. So during the
2118 * lifetime of this thread, kfd_process p will be valid
2119 */
2120 p = container_of(dwork, struct kfd_process, eviction_work);
2121
2122 pr_debug("Started evicting process pid %d\n", p->lead_thread->pid);
2123 ret = kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_TRIGGER_TTM);
2124 if (!ret) {
2125 /* If another thread already signaled the eviction fence,
2126 * they are responsible stopping the queues and scheduling
2127 * the restore work.
2128 */
2129 if (signal_eviction_fence(p) ||
2130 mod_delayed_work(kfd_restore_wq, &p->restore_work,
2131 msecs_to_jiffies(PROCESS_RESTORE_TIME_MS)))
2132 kfd_process_restore_queues(p);
2133
2134 pr_debug("Finished evicting process pid %d\n", p->lead_thread->pid);
2135 } else
2136 pr_err("Failed to evict queues of process pid %d\n", p->lead_thread->pid);
2137 }
2138
restore_process_helper(struct kfd_process * p)2139 static int restore_process_helper(struct kfd_process *p)
2140 {
2141 int ret = 0;
2142
2143 /* VMs may not have been acquired yet during debugging. */
2144 if (p->kgd_process_info) {
2145 ret = amdgpu_amdkfd_gpuvm_restore_process_bos(
2146 p->kgd_process_info, &p->ef);
2147 if (ret)
2148 return ret;
2149 }
2150
2151 ret = kfd_process_restore_queues(p);
2152 if (!ret)
2153 pr_debug("Finished restoring process pid %d\n",
2154 p->lead_thread->pid);
2155 else
2156 pr_err("Failed to restore queues of process pid %d\n",
2157 p->lead_thread->pid);
2158
2159 return ret;
2160 }
2161
restore_process_worker(struct work_struct * work)2162 static void restore_process_worker(struct work_struct *work)
2163 {
2164 struct delayed_work *dwork;
2165 struct kfd_process *p;
2166 int ret = 0;
2167
2168 dwork = to_delayed_work(work);
2169
2170 /* Process termination destroys this worker thread. So during the
2171 * lifetime of this thread, kfd_process p will be valid
2172 */
2173 p = container_of(dwork, struct kfd_process, restore_work);
2174 pr_debug("Started restoring process pasid %d\n", (int)p->lead_thread->pid);
2175
2176 /* Setting last_restore_timestamp before successful restoration.
2177 * Otherwise this would have to be set by KGD (restore_process_bos)
2178 * before KFD BOs are unreserved. If not, the process can be evicted
2179 * again before the timestamp is set.
2180 * If restore fails, the timestamp will be set again in the next
2181 * attempt. This would mean that the minimum GPU quanta would be
2182 * PROCESS_ACTIVE_TIME_MS - (time to execute the following two
2183 * functions)
2184 */
2185
2186 p->last_restore_timestamp = get_jiffies_64();
2187
2188 ret = restore_process_helper(p);
2189 if (ret) {
2190 pr_debug("Failed to restore BOs of process pid %d, retry after %d ms\n",
2191 p->lead_thread->pid, PROCESS_BACK_OFF_TIME_MS);
2192 if (mod_delayed_work(kfd_restore_wq, &p->restore_work,
2193 msecs_to_jiffies(PROCESS_RESTORE_TIME_MS)))
2194 kfd_process_restore_queues(p);
2195 }
2196 }
2197
kfd_suspend_all_processes(void)2198 void kfd_suspend_all_processes(void)
2199 {
2200 struct kfd_process *p;
2201 unsigned int temp;
2202 int idx = srcu_read_lock(&kfd_processes_srcu);
2203
2204 WARN(debug_evictions, "Evicting all processes");
2205 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
2206 if (kfd_process_evict_queues(p, KFD_QUEUE_EVICTION_TRIGGER_SUSPEND))
2207 pr_err("Failed to suspend process pid %d\n", p->lead_thread->pid);
2208 signal_eviction_fence(p);
2209 }
2210 srcu_read_unlock(&kfd_processes_srcu, idx);
2211 }
2212
kfd_resume_all_processes(void)2213 int kfd_resume_all_processes(void)
2214 {
2215 struct kfd_process *p;
2216 unsigned int temp;
2217 int ret = 0, idx = srcu_read_lock(&kfd_processes_srcu);
2218
2219 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
2220 if (restore_process_helper(p)) {
2221 pr_err("Restore process pid %d failed during resume\n",
2222 p->lead_thread->pid);
2223 ret = -EFAULT;
2224 }
2225 }
2226 srcu_read_unlock(&kfd_processes_srcu, idx);
2227 return ret;
2228 }
2229
kfd_reserved_mem_mmap(struct kfd_node * dev,struct kfd_process * process,struct vm_area_struct * vma)2230 int kfd_reserved_mem_mmap(struct kfd_node *dev, struct kfd_process *process,
2231 struct vm_area_struct *vma)
2232 {
2233 struct kfd_process_device *pdd;
2234 struct qcm_process_device *qpd;
2235
2236 if ((vma->vm_end - vma->vm_start) != KFD_CWSR_TBA_TMA_SIZE) {
2237 dev_err(dev->adev->dev, "Incorrect CWSR mapping size.\n");
2238 return -EINVAL;
2239 }
2240
2241 pdd = kfd_get_process_device_data(dev, process);
2242 if (!pdd)
2243 return -EINVAL;
2244 qpd = &pdd->qpd;
2245
2246 qpd->cwsr_kaddr = (void *)__get_free_pages(GFP_KERNEL | __GFP_ZERO,
2247 get_order(KFD_CWSR_TBA_TMA_SIZE));
2248 if (!qpd->cwsr_kaddr) {
2249 dev_err(dev->adev->dev,
2250 "Error allocating per process CWSR buffer.\n");
2251 return -ENOMEM;
2252 }
2253
2254 vm_flags_set(vma, VM_IO | VM_DONTCOPY | VM_DONTEXPAND
2255 | VM_NORESERVE | VM_DONTDUMP | VM_PFNMAP);
2256 /* Mapping pages to user process */
2257 return remap_pfn_range(vma, vma->vm_start,
2258 PFN_DOWN(__pa(qpd->cwsr_kaddr)),
2259 KFD_CWSR_TBA_TMA_SIZE, vma->vm_page_prot);
2260 }
2261
2262 /* assumes caller holds process lock. */
kfd_process_drain_interrupts(struct kfd_process_device * pdd)2263 int kfd_process_drain_interrupts(struct kfd_process_device *pdd)
2264 {
2265 uint32_t irq_drain_fence[8];
2266 uint8_t node_id = 0;
2267 int r = 0;
2268
2269 if (!KFD_IS_SOC15(pdd->dev))
2270 return 0;
2271
2272 pdd->process->irq_drain_is_open = true;
2273
2274 memset(irq_drain_fence, 0, sizeof(irq_drain_fence));
2275 irq_drain_fence[0] = (KFD_IRQ_FENCE_SOURCEID << 8) |
2276 KFD_IRQ_FENCE_CLIENTID;
2277 irq_drain_fence[3] = pdd->pasid;
2278
2279 /*
2280 * For GFX 9.4.3/9.5.0, send the NodeId also in IH cookie DW[3]
2281 */
2282 if (KFD_GC_VERSION(pdd->dev->kfd) == IP_VERSION(9, 4, 3) ||
2283 KFD_GC_VERSION(pdd->dev->kfd) == IP_VERSION(9, 4, 4) ||
2284 KFD_GC_VERSION(pdd->dev->kfd) == IP_VERSION(9, 5, 0) ||
2285 KFD_GC_VERSION(pdd->dev->kfd) == IP_VERSION(12, 1, 0)) {
2286 node_id = ffs(pdd->dev->interrupt_bitmap) - 1;
2287 irq_drain_fence[3] |= node_id << 16;
2288 }
2289
2290 /* ensure stale irqs scheduled KFD interrupts and send drain fence. */
2291 if (amdgpu_amdkfd_send_close_event_drain_irq(pdd->dev->adev,
2292 irq_drain_fence)) {
2293 pdd->process->irq_drain_is_open = false;
2294 return 0;
2295 }
2296
2297 r = wait_event_interruptible(pdd->process->wait_irq_drain,
2298 !READ_ONCE(pdd->process->irq_drain_is_open));
2299 if (r)
2300 pdd->process->irq_drain_is_open = false;
2301
2302 return r;
2303 }
2304
kfd_process_close_interrupt_drain(unsigned int pasid)2305 void kfd_process_close_interrupt_drain(unsigned int pasid)
2306 {
2307 struct kfd_process *p;
2308
2309 p = kfd_lookup_process_by_pasid(pasid, NULL);
2310
2311 if (!p)
2312 return;
2313
2314 WRITE_ONCE(p->irq_drain_is_open, false);
2315 wake_up_all(&p->wait_irq_drain);
2316 kfd_unref_process(p);
2317 }
2318
2319 struct send_exception_work_handler_workarea {
2320 struct work_struct work;
2321 struct kfd_process *p;
2322 unsigned int queue_id;
2323 uint64_t error_reason;
2324 };
2325
send_exception_work_handler(struct work_struct * work)2326 static void send_exception_work_handler(struct work_struct *work)
2327 {
2328 struct send_exception_work_handler_workarea *workarea;
2329 struct kfd_process *p;
2330 struct queue *q;
2331 struct mm_struct *mm;
2332 struct kfd_context_save_area_header __user *csa_header;
2333 uint64_t __user *err_payload_ptr;
2334 uint64_t cur_err;
2335 uint32_t ev_id;
2336
2337 workarea = container_of(work,
2338 struct send_exception_work_handler_workarea,
2339 work);
2340 p = workarea->p;
2341
2342 mm = get_task_mm(p->lead_thread);
2343
2344 if (!mm)
2345 return;
2346
2347 kthread_use_mm(mm);
2348
2349 q = pqm_get_user_queue(&p->pqm, workarea->queue_id);
2350
2351 if (!q)
2352 goto out;
2353
2354 csa_header = (void __user *)q->properties.ctx_save_restore_area_address;
2355
2356 get_user(err_payload_ptr, (uint64_t __user **)&csa_header->err_payload_addr);
2357 get_user(cur_err, err_payload_ptr);
2358 cur_err |= workarea->error_reason;
2359 put_user(cur_err, err_payload_ptr);
2360 get_user(ev_id, &csa_header->err_event_id);
2361
2362 kfd_set_event(p, ev_id);
2363
2364 out:
2365 kthread_unuse_mm(mm);
2366 mmput(mm);
2367 }
2368
kfd_send_exception_to_runtime(struct kfd_process * p,unsigned int queue_id,uint64_t error_reason)2369 int kfd_send_exception_to_runtime(struct kfd_process *p,
2370 unsigned int queue_id,
2371 uint64_t error_reason)
2372 {
2373 struct send_exception_work_handler_workarea worker;
2374
2375 INIT_WORK_ONSTACK(&worker.work, send_exception_work_handler);
2376
2377 worker.p = p;
2378 worker.queue_id = queue_id;
2379 worker.error_reason = error_reason;
2380
2381 schedule_work(&worker.work);
2382 flush_work(&worker.work);
2383 destroy_work_on_stack(&worker.work);
2384
2385 return 0;
2386 }
2387
kfd_process_device_data_by_id(struct kfd_process * p,uint32_t gpu_id)2388 struct kfd_process_device *kfd_process_device_data_by_id(struct kfd_process *p, uint32_t gpu_id)
2389 {
2390 int i;
2391
2392 if (gpu_id) {
2393 for (i = 0; i < p->n_pdds; i++) {
2394 struct kfd_process_device *pdd = p->pdds[i];
2395
2396 if (pdd->user_gpu_id == gpu_id)
2397 return pdd;
2398 }
2399 }
2400 return NULL;
2401 }
2402
kfd_process_get_user_gpu_id(struct kfd_process * p,uint32_t actual_gpu_id)2403 int kfd_process_get_user_gpu_id(struct kfd_process *p, uint32_t actual_gpu_id)
2404 {
2405 int i;
2406
2407 if (!actual_gpu_id)
2408 return 0;
2409
2410 for (i = 0; i < p->n_pdds; i++) {
2411 struct kfd_process_device *pdd = p->pdds[i];
2412
2413 if (pdd->dev->id == actual_gpu_id)
2414 return pdd->user_gpu_id;
2415 }
2416 return -EINVAL;
2417 }
2418
2419 #if defined(CONFIG_DEBUG_FS)
2420
kfd_debugfs_mqds_by_process(struct seq_file * m,void * data)2421 int kfd_debugfs_mqds_by_process(struct seq_file *m, void *data)
2422 {
2423 struct kfd_process *p;
2424 unsigned int temp;
2425 int r = 0;
2426
2427 int idx = srcu_read_lock(&kfd_processes_srcu);
2428
2429 hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
2430 seq_printf(m, "Process %d PASID %d:\n",
2431 p->lead_thread->tgid, p->lead_thread->pid);
2432
2433 mutex_lock(&p->mutex);
2434 r = pqm_debugfs_mqds(m, &p->pqm);
2435 mutex_unlock(&p->mutex);
2436
2437 if (r)
2438 break;
2439 }
2440
2441 srcu_read_unlock(&kfd_processes_srcu, idx);
2442
2443 return r;
2444 }
2445
2446 #endif
2447