xref: /linux/drivers/gpu/drm/xe/xe_exec_queue.c (revision c7546e2c3cb739a3c1a2f5acaf9bb629d401afe5)
1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2021 Intel Corporation
4  */
5 
6 #include "xe_exec_queue.h"
7 
8 #include <linux/nospec.h>
9 
10 #include <drm/drm_device.h>
11 #include <drm/drm_file.h>
12 #include <uapi/drm/xe_drm.h>
13 
14 #include "xe_device.h"
15 #include "xe_gt.h"
16 #include "xe_hw_engine_class_sysfs.h"
17 #include "xe_hw_engine_group.h"
18 #include "xe_hw_fence.h"
19 #include "xe_lrc.h"
20 #include "xe_macros.h"
21 #include "xe_migrate.h"
22 #include "xe_pm.h"
23 #include "xe_ring_ops_types.h"
24 #include "xe_trace.h"
25 #include "xe_vm.h"
26 
27 enum xe_exec_queue_sched_prop {
28 	XE_EXEC_QUEUE_JOB_TIMEOUT = 0,
29 	XE_EXEC_QUEUE_TIMESLICE = 1,
30 	XE_EXEC_QUEUE_PREEMPT_TIMEOUT = 2,
31 	XE_EXEC_QUEUE_SCHED_PROP_MAX = 3,
32 };
33 
34 static int exec_queue_user_extensions(struct xe_device *xe, struct xe_exec_queue *q,
35 				      u64 extensions, int ext_number);
36 
37 static void __xe_exec_queue_free(struct xe_exec_queue *q)
38 {
39 	if (q->vm)
40 		xe_vm_put(q->vm);
41 
42 	if (q->xef)
43 		xe_file_put(q->xef);
44 
45 	kfree(q);
46 }
47 
48 static struct xe_exec_queue *__xe_exec_queue_alloc(struct xe_device *xe,
49 						   struct xe_vm *vm,
50 						   u32 logical_mask,
51 						   u16 width, struct xe_hw_engine *hwe,
52 						   u32 flags, u64 extensions)
53 {
54 	struct xe_exec_queue *q;
55 	struct xe_gt *gt = hwe->gt;
56 	int err;
57 
58 	/* only kernel queues can be permanent */
59 	XE_WARN_ON((flags & EXEC_QUEUE_FLAG_PERMANENT) && !(flags & EXEC_QUEUE_FLAG_KERNEL));
60 
61 	q = kzalloc(struct_size(q, lrc, width), GFP_KERNEL);
62 	if (!q)
63 		return ERR_PTR(-ENOMEM);
64 
65 	kref_init(&q->refcount);
66 	q->flags = flags;
67 	q->hwe = hwe;
68 	q->gt = gt;
69 	q->class = hwe->class;
70 	q->width = width;
71 	q->logical_mask = logical_mask;
72 	q->fence_irq = &gt->fence_irq[hwe->class];
73 	q->ring_ops = gt->ring_ops[hwe->class];
74 	q->ops = gt->exec_queue_ops;
75 	INIT_LIST_HEAD(&q->lr.link);
76 	INIT_LIST_HEAD(&q->multi_gt_link);
77 	INIT_LIST_HEAD(&q->hw_engine_group_link);
78 
79 	q->sched_props.timeslice_us = hwe->eclass->sched_props.timeslice_us;
80 	q->sched_props.preempt_timeout_us =
81 				hwe->eclass->sched_props.preempt_timeout_us;
82 	q->sched_props.job_timeout_ms =
83 				hwe->eclass->sched_props.job_timeout_ms;
84 	if (q->flags & EXEC_QUEUE_FLAG_KERNEL &&
85 	    q->flags & EXEC_QUEUE_FLAG_HIGH_PRIORITY)
86 		q->sched_props.priority = XE_EXEC_QUEUE_PRIORITY_KERNEL;
87 	else
88 		q->sched_props.priority = XE_EXEC_QUEUE_PRIORITY_NORMAL;
89 
90 	if (vm)
91 		q->vm = xe_vm_get(vm);
92 
93 	if (extensions) {
94 		/*
95 		 * may set q->usm, must come before xe_lrc_create(),
96 		 * may overwrite q->sched_props, must come before q->ops->init()
97 		 */
98 		err = exec_queue_user_extensions(xe, q, extensions, 0);
99 		if (err) {
100 			__xe_exec_queue_free(q);
101 			return ERR_PTR(err);
102 		}
103 	}
104 
105 	return q;
106 }
107 
108 static int __xe_exec_queue_init(struct xe_exec_queue *q)
109 {
110 	struct xe_vm *vm = q->vm;
111 	int i, err;
112 
113 	if (vm) {
114 		err = xe_vm_lock(vm, true);
115 		if (err)
116 			return err;
117 	}
118 
119 	for (i = 0; i < q->width; ++i) {
120 		q->lrc[i] = xe_lrc_create(q->hwe, q->vm, SZ_16K);
121 		if (IS_ERR(q->lrc[i])) {
122 			err = PTR_ERR(q->lrc[i]);
123 			goto err_unlock;
124 		}
125 	}
126 
127 	if (vm)
128 		xe_vm_unlock(vm);
129 
130 	err = q->ops->init(q);
131 	if (err)
132 		goto err_lrc;
133 
134 	return 0;
135 
136 err_unlock:
137 	if (vm)
138 		xe_vm_unlock(vm);
139 err_lrc:
140 	for (i = i - 1; i >= 0; --i)
141 		xe_lrc_put(q->lrc[i]);
142 	return err;
143 }
144 
145 struct xe_exec_queue *xe_exec_queue_create(struct xe_device *xe, struct xe_vm *vm,
146 					   u32 logical_mask, u16 width,
147 					   struct xe_hw_engine *hwe, u32 flags,
148 					   u64 extensions)
149 {
150 	struct xe_exec_queue *q;
151 	int err;
152 
153 	q = __xe_exec_queue_alloc(xe, vm, logical_mask, width, hwe, flags,
154 				  extensions);
155 	if (IS_ERR(q))
156 		return q;
157 
158 	err = __xe_exec_queue_init(q);
159 	if (err)
160 		goto err_post_alloc;
161 
162 	return q;
163 
164 err_post_alloc:
165 	__xe_exec_queue_free(q);
166 	return ERR_PTR(err);
167 }
168 
169 struct xe_exec_queue *xe_exec_queue_create_class(struct xe_device *xe, struct xe_gt *gt,
170 						 struct xe_vm *vm,
171 						 enum xe_engine_class class,
172 						 u32 flags, u64 extensions)
173 {
174 	struct xe_hw_engine *hwe, *hwe0 = NULL;
175 	enum xe_hw_engine_id id;
176 	u32 logical_mask = 0;
177 
178 	for_each_hw_engine(hwe, gt, id) {
179 		if (xe_hw_engine_is_reserved(hwe))
180 			continue;
181 
182 		if (hwe->class == class) {
183 			logical_mask |= BIT(hwe->logical_instance);
184 			if (!hwe0)
185 				hwe0 = hwe;
186 		}
187 	}
188 
189 	if (!logical_mask)
190 		return ERR_PTR(-ENODEV);
191 
192 	return xe_exec_queue_create(xe, vm, logical_mask, 1, hwe0, flags, extensions);
193 }
194 
195 /**
196  * xe_exec_queue_create_bind() - Create bind exec queue.
197  * @xe: Xe device.
198  * @tile: tile which bind exec queue belongs to.
199  * @flags: exec queue creation flags
200  * @extensions: exec queue creation extensions
201  *
202  * Normalize bind exec queue creation. Bind exec queue is tied to migration VM
203  * for access to physical memory required for page table programming. On a
204  * faulting devices the reserved copy engine instance must be used to avoid
205  * deadlocking (user binds cannot get stuck behind faults as kernel binds which
206  * resolve faults depend on user binds). On non-faulting devices any copy engine
207  * can be used.
208  *
209  * Returns exec queue on success, ERR_PTR on failure
210  */
211 struct xe_exec_queue *xe_exec_queue_create_bind(struct xe_device *xe,
212 						struct xe_tile *tile,
213 						u32 flags, u64 extensions)
214 {
215 	struct xe_gt *gt = tile->primary_gt;
216 	struct xe_exec_queue *q;
217 	struct xe_vm *migrate_vm;
218 
219 	migrate_vm = xe_migrate_get_vm(tile->migrate);
220 	if (xe->info.has_usm) {
221 		struct xe_hw_engine *hwe = xe_gt_hw_engine(gt,
222 							   XE_ENGINE_CLASS_COPY,
223 							   gt->usm.reserved_bcs_instance,
224 							   false);
225 
226 		if (!hwe) {
227 			xe_vm_put(migrate_vm);
228 			return ERR_PTR(-EINVAL);
229 		}
230 
231 		q = xe_exec_queue_create(xe, migrate_vm,
232 					 BIT(hwe->logical_instance), 1, hwe,
233 					 flags, extensions);
234 	} else {
235 		q = xe_exec_queue_create_class(xe, gt, migrate_vm,
236 					       XE_ENGINE_CLASS_COPY, flags,
237 					       extensions);
238 	}
239 	xe_vm_put(migrate_vm);
240 
241 	return q;
242 }
243 
244 void xe_exec_queue_destroy(struct kref *ref)
245 {
246 	struct xe_exec_queue *q = container_of(ref, struct xe_exec_queue, refcount);
247 	struct xe_exec_queue *eq, *next;
248 
249 	xe_exec_queue_last_fence_put_unlocked(q);
250 	if (!(q->flags & EXEC_QUEUE_FLAG_BIND_ENGINE_CHILD)) {
251 		list_for_each_entry_safe(eq, next, &q->multi_gt_list,
252 					 multi_gt_link)
253 			xe_exec_queue_put(eq);
254 	}
255 
256 	q->ops->fini(q);
257 }
258 
259 void xe_exec_queue_fini(struct xe_exec_queue *q)
260 {
261 	int i;
262 
263 	for (i = 0; i < q->width; ++i)
264 		xe_lrc_put(q->lrc[i]);
265 	__xe_exec_queue_free(q);
266 }
267 
268 void xe_exec_queue_assign_name(struct xe_exec_queue *q, u32 instance)
269 {
270 	switch (q->class) {
271 	case XE_ENGINE_CLASS_RENDER:
272 		snprintf(q->name, sizeof(q->name), "rcs%d", instance);
273 		break;
274 	case XE_ENGINE_CLASS_VIDEO_DECODE:
275 		snprintf(q->name, sizeof(q->name), "vcs%d", instance);
276 		break;
277 	case XE_ENGINE_CLASS_VIDEO_ENHANCE:
278 		snprintf(q->name, sizeof(q->name), "vecs%d", instance);
279 		break;
280 	case XE_ENGINE_CLASS_COPY:
281 		snprintf(q->name, sizeof(q->name), "bcs%d", instance);
282 		break;
283 	case XE_ENGINE_CLASS_COMPUTE:
284 		snprintf(q->name, sizeof(q->name), "ccs%d", instance);
285 		break;
286 	case XE_ENGINE_CLASS_OTHER:
287 		snprintf(q->name, sizeof(q->name), "gsccs%d", instance);
288 		break;
289 	default:
290 		XE_WARN_ON(q->class);
291 	}
292 }
293 
294 struct xe_exec_queue *xe_exec_queue_lookup(struct xe_file *xef, u32 id)
295 {
296 	struct xe_exec_queue *q;
297 
298 	mutex_lock(&xef->exec_queue.lock);
299 	q = xa_load(&xef->exec_queue.xa, id);
300 	if (q)
301 		xe_exec_queue_get(q);
302 	mutex_unlock(&xef->exec_queue.lock);
303 
304 	return q;
305 }
306 
307 enum xe_exec_queue_priority
308 xe_exec_queue_device_get_max_priority(struct xe_device *xe)
309 {
310 	return capable(CAP_SYS_NICE) ? XE_EXEC_QUEUE_PRIORITY_HIGH :
311 				       XE_EXEC_QUEUE_PRIORITY_NORMAL;
312 }
313 
314 static int exec_queue_set_priority(struct xe_device *xe, struct xe_exec_queue *q,
315 				   u64 value)
316 {
317 	if (XE_IOCTL_DBG(xe, value > XE_EXEC_QUEUE_PRIORITY_HIGH))
318 		return -EINVAL;
319 
320 	if (XE_IOCTL_DBG(xe, value > xe_exec_queue_device_get_max_priority(xe)))
321 		return -EPERM;
322 
323 	q->sched_props.priority = value;
324 	return 0;
325 }
326 
327 static bool xe_exec_queue_enforce_schedule_limit(void)
328 {
329 #if IS_ENABLED(CONFIG_DRM_XE_ENABLE_SCHEDTIMEOUT_LIMIT)
330 	return true;
331 #else
332 	return !capable(CAP_SYS_NICE);
333 #endif
334 }
335 
336 static void
337 xe_exec_queue_get_prop_minmax(struct xe_hw_engine_class_intf *eclass,
338 			      enum xe_exec_queue_sched_prop prop,
339 			      u32 *min, u32 *max)
340 {
341 	switch (prop) {
342 	case XE_EXEC_QUEUE_JOB_TIMEOUT:
343 		*min = eclass->sched_props.job_timeout_min;
344 		*max = eclass->sched_props.job_timeout_max;
345 		break;
346 	case XE_EXEC_QUEUE_TIMESLICE:
347 		*min = eclass->sched_props.timeslice_min;
348 		*max = eclass->sched_props.timeslice_max;
349 		break;
350 	case XE_EXEC_QUEUE_PREEMPT_TIMEOUT:
351 		*min = eclass->sched_props.preempt_timeout_min;
352 		*max = eclass->sched_props.preempt_timeout_max;
353 		break;
354 	default:
355 		break;
356 	}
357 #if IS_ENABLED(CONFIG_DRM_XE_ENABLE_SCHEDTIMEOUT_LIMIT)
358 	if (capable(CAP_SYS_NICE)) {
359 		switch (prop) {
360 		case XE_EXEC_QUEUE_JOB_TIMEOUT:
361 			*min = XE_HW_ENGINE_JOB_TIMEOUT_MIN;
362 			*max = XE_HW_ENGINE_JOB_TIMEOUT_MAX;
363 			break;
364 		case XE_EXEC_QUEUE_TIMESLICE:
365 			*min = XE_HW_ENGINE_TIMESLICE_MIN;
366 			*max = XE_HW_ENGINE_TIMESLICE_MAX;
367 			break;
368 		case XE_EXEC_QUEUE_PREEMPT_TIMEOUT:
369 			*min = XE_HW_ENGINE_PREEMPT_TIMEOUT_MIN;
370 			*max = XE_HW_ENGINE_PREEMPT_TIMEOUT_MAX;
371 			break;
372 		default:
373 			break;
374 		}
375 	}
376 #endif
377 }
378 
379 static int exec_queue_set_timeslice(struct xe_device *xe, struct xe_exec_queue *q,
380 				    u64 value)
381 {
382 	u32 min = 0, max = 0;
383 
384 	xe_exec_queue_get_prop_minmax(q->hwe->eclass,
385 				      XE_EXEC_QUEUE_TIMESLICE, &min, &max);
386 
387 	if (xe_exec_queue_enforce_schedule_limit() &&
388 	    !xe_hw_engine_timeout_in_range(value, min, max))
389 		return -EINVAL;
390 
391 	q->sched_props.timeslice_us = value;
392 	return 0;
393 }
394 
395 typedef int (*xe_exec_queue_set_property_fn)(struct xe_device *xe,
396 					     struct xe_exec_queue *q,
397 					     u64 value);
398 
399 static const xe_exec_queue_set_property_fn exec_queue_set_property_funcs[] = {
400 	[DRM_XE_EXEC_QUEUE_SET_PROPERTY_PRIORITY] = exec_queue_set_priority,
401 	[DRM_XE_EXEC_QUEUE_SET_PROPERTY_TIMESLICE] = exec_queue_set_timeslice,
402 };
403 
404 static int exec_queue_user_ext_set_property(struct xe_device *xe,
405 					    struct xe_exec_queue *q,
406 					    u64 extension)
407 {
408 	u64 __user *address = u64_to_user_ptr(extension);
409 	struct drm_xe_ext_set_property ext;
410 	int err;
411 	u32 idx;
412 
413 	err = __copy_from_user(&ext, address, sizeof(ext));
414 	if (XE_IOCTL_DBG(xe, err))
415 		return -EFAULT;
416 
417 	if (XE_IOCTL_DBG(xe, ext.property >=
418 			 ARRAY_SIZE(exec_queue_set_property_funcs)) ||
419 	    XE_IOCTL_DBG(xe, ext.pad) ||
420 	    XE_IOCTL_DBG(xe, ext.property != DRM_XE_EXEC_QUEUE_SET_PROPERTY_PRIORITY &&
421 			 ext.property != DRM_XE_EXEC_QUEUE_SET_PROPERTY_TIMESLICE))
422 		return -EINVAL;
423 
424 	idx = array_index_nospec(ext.property, ARRAY_SIZE(exec_queue_set_property_funcs));
425 	if (!exec_queue_set_property_funcs[idx])
426 		return -EINVAL;
427 
428 	return exec_queue_set_property_funcs[idx](xe, q, ext.value);
429 }
430 
431 typedef int (*xe_exec_queue_user_extension_fn)(struct xe_device *xe,
432 					       struct xe_exec_queue *q,
433 					       u64 extension);
434 
435 static const xe_exec_queue_user_extension_fn exec_queue_user_extension_funcs[] = {
436 	[DRM_XE_EXEC_QUEUE_EXTENSION_SET_PROPERTY] = exec_queue_user_ext_set_property,
437 };
438 
439 #define MAX_USER_EXTENSIONS	16
440 static int exec_queue_user_extensions(struct xe_device *xe, struct xe_exec_queue *q,
441 				      u64 extensions, int ext_number)
442 {
443 	u64 __user *address = u64_to_user_ptr(extensions);
444 	struct drm_xe_user_extension ext;
445 	int err;
446 	u32 idx;
447 
448 	if (XE_IOCTL_DBG(xe, ext_number >= MAX_USER_EXTENSIONS))
449 		return -E2BIG;
450 
451 	err = __copy_from_user(&ext, address, sizeof(ext));
452 	if (XE_IOCTL_DBG(xe, err))
453 		return -EFAULT;
454 
455 	if (XE_IOCTL_DBG(xe, ext.pad) ||
456 	    XE_IOCTL_DBG(xe, ext.name >=
457 			 ARRAY_SIZE(exec_queue_user_extension_funcs)))
458 		return -EINVAL;
459 
460 	idx = array_index_nospec(ext.name,
461 				 ARRAY_SIZE(exec_queue_user_extension_funcs));
462 	err = exec_queue_user_extension_funcs[idx](xe, q, extensions);
463 	if (XE_IOCTL_DBG(xe, err))
464 		return err;
465 
466 	if (ext.next_extension)
467 		return exec_queue_user_extensions(xe, q, ext.next_extension,
468 						  ++ext_number);
469 
470 	return 0;
471 }
472 
473 static u32 calc_validate_logical_mask(struct xe_device *xe, struct xe_gt *gt,
474 				      struct drm_xe_engine_class_instance *eci,
475 				      u16 width, u16 num_placements)
476 {
477 	int len = width * num_placements;
478 	int i, j, n;
479 	u16 class;
480 	u16 gt_id;
481 	u32 return_mask = 0, prev_mask;
482 
483 	if (XE_IOCTL_DBG(xe, !xe_device_uc_enabled(xe) &&
484 			 len > 1))
485 		return 0;
486 
487 	for (i = 0; i < width; ++i) {
488 		u32 current_mask = 0;
489 
490 		for (j = 0; j < num_placements; ++j) {
491 			struct xe_hw_engine *hwe;
492 
493 			n = j * width + i;
494 
495 			hwe = xe_hw_engine_lookup(xe, eci[n]);
496 			if (XE_IOCTL_DBG(xe, !hwe))
497 				return 0;
498 
499 			if (XE_IOCTL_DBG(xe, xe_hw_engine_is_reserved(hwe)))
500 				return 0;
501 
502 			if (XE_IOCTL_DBG(xe, n && eci[n].gt_id != gt_id) ||
503 			    XE_IOCTL_DBG(xe, n && eci[n].engine_class != class))
504 				return 0;
505 
506 			class = eci[n].engine_class;
507 			gt_id = eci[n].gt_id;
508 
509 			if (width == 1 || !i)
510 				return_mask |= BIT(eci[n].engine_instance);
511 			current_mask |= BIT(eci[n].engine_instance);
512 		}
513 
514 		/* Parallel submissions must be logically contiguous */
515 		if (i && XE_IOCTL_DBG(xe, current_mask != prev_mask << 1))
516 			return 0;
517 
518 		prev_mask = current_mask;
519 	}
520 
521 	return return_mask;
522 }
523 
524 int xe_exec_queue_create_ioctl(struct drm_device *dev, void *data,
525 			       struct drm_file *file)
526 {
527 	struct xe_device *xe = to_xe_device(dev);
528 	struct xe_file *xef = to_xe_file(file);
529 	struct drm_xe_exec_queue_create *args = data;
530 	struct drm_xe_engine_class_instance eci[XE_HW_ENGINE_MAX_INSTANCE];
531 	struct drm_xe_engine_class_instance __user *user_eci =
532 		u64_to_user_ptr(args->instances);
533 	struct xe_hw_engine *hwe;
534 	struct xe_vm *vm;
535 	struct xe_gt *gt;
536 	struct xe_tile *tile;
537 	struct xe_exec_queue *q = NULL;
538 	u32 logical_mask;
539 	u32 id;
540 	u32 len;
541 	int err;
542 
543 	if (XE_IOCTL_DBG(xe, args->flags) ||
544 	    XE_IOCTL_DBG(xe, args->reserved[0] || args->reserved[1]))
545 		return -EINVAL;
546 
547 	len = args->width * args->num_placements;
548 	if (XE_IOCTL_DBG(xe, !len || len > XE_HW_ENGINE_MAX_INSTANCE))
549 		return -EINVAL;
550 
551 	err = __copy_from_user(eci, user_eci,
552 			       sizeof(struct drm_xe_engine_class_instance) *
553 			       len);
554 	if (XE_IOCTL_DBG(xe, err))
555 		return -EFAULT;
556 
557 	if (XE_IOCTL_DBG(xe, eci[0].gt_id >= xe->info.gt_count))
558 		return -EINVAL;
559 
560 	if (eci[0].engine_class == DRM_XE_ENGINE_CLASS_VM_BIND) {
561 		if (XE_IOCTL_DBG(xe, args->width != 1) ||
562 		    XE_IOCTL_DBG(xe, args->num_placements != 1) ||
563 		    XE_IOCTL_DBG(xe, eci[0].engine_instance != 0))
564 			return -EINVAL;
565 
566 		for_each_tile(tile, xe, id) {
567 			struct xe_exec_queue *new;
568 			u32 flags = EXEC_QUEUE_FLAG_VM;
569 
570 			if (id)
571 				flags |= EXEC_QUEUE_FLAG_BIND_ENGINE_CHILD;
572 
573 			new = xe_exec_queue_create_bind(xe, tile, flags,
574 							args->extensions);
575 			if (IS_ERR(new)) {
576 				err = PTR_ERR(new);
577 				if (q)
578 					goto put_exec_queue;
579 				return err;
580 			}
581 			if (id == 0)
582 				q = new;
583 			else
584 				list_add_tail(&new->multi_gt_list,
585 					      &q->multi_gt_link);
586 		}
587 	} else {
588 		gt = xe_device_get_gt(xe, eci[0].gt_id);
589 		logical_mask = calc_validate_logical_mask(xe, gt, eci,
590 							  args->width,
591 							  args->num_placements);
592 		if (XE_IOCTL_DBG(xe, !logical_mask))
593 			return -EINVAL;
594 
595 		hwe = xe_hw_engine_lookup(xe, eci[0]);
596 		if (XE_IOCTL_DBG(xe, !hwe))
597 			return -EINVAL;
598 
599 		vm = xe_vm_lookup(xef, args->vm_id);
600 		if (XE_IOCTL_DBG(xe, !vm))
601 			return -ENOENT;
602 
603 		err = down_read_interruptible(&vm->lock);
604 		if (err) {
605 			xe_vm_put(vm);
606 			return err;
607 		}
608 
609 		if (XE_IOCTL_DBG(xe, xe_vm_is_closed_or_banned(vm))) {
610 			up_read(&vm->lock);
611 			xe_vm_put(vm);
612 			return -ENOENT;
613 		}
614 
615 		q = xe_exec_queue_create(xe, vm, logical_mask,
616 					 args->width, hwe, 0,
617 					 args->extensions);
618 		up_read(&vm->lock);
619 		xe_vm_put(vm);
620 		if (IS_ERR(q))
621 			return PTR_ERR(q);
622 
623 		if (xe_vm_in_preempt_fence_mode(vm)) {
624 			q->lr.context = dma_fence_context_alloc(1);
625 
626 			err = xe_vm_add_compute_exec_queue(vm, q);
627 			if (XE_IOCTL_DBG(xe, err))
628 				goto put_exec_queue;
629 		}
630 
631 		if (q->vm && q->hwe->hw_engine_group) {
632 			err = xe_hw_engine_group_add_exec_queue(q->hwe->hw_engine_group, q);
633 			if (err)
634 				goto put_exec_queue;
635 		}
636 	}
637 
638 	mutex_lock(&xef->exec_queue.lock);
639 	err = xa_alloc(&xef->exec_queue.xa, &id, q, xa_limit_32b, GFP_KERNEL);
640 	mutex_unlock(&xef->exec_queue.lock);
641 	if (err)
642 		goto kill_exec_queue;
643 
644 	args->exec_queue_id = id;
645 	q->xef = xe_file_get(xef);
646 
647 	return 0;
648 
649 kill_exec_queue:
650 	xe_exec_queue_kill(q);
651 put_exec_queue:
652 	xe_exec_queue_put(q);
653 	return err;
654 }
655 
656 int xe_exec_queue_get_property_ioctl(struct drm_device *dev, void *data,
657 				     struct drm_file *file)
658 {
659 	struct xe_device *xe = to_xe_device(dev);
660 	struct xe_file *xef = to_xe_file(file);
661 	struct drm_xe_exec_queue_get_property *args = data;
662 	struct xe_exec_queue *q;
663 	int ret;
664 
665 	if (XE_IOCTL_DBG(xe, args->reserved[0] || args->reserved[1]))
666 		return -EINVAL;
667 
668 	q = xe_exec_queue_lookup(xef, args->exec_queue_id);
669 	if (XE_IOCTL_DBG(xe, !q))
670 		return -ENOENT;
671 
672 	switch (args->property) {
673 	case DRM_XE_EXEC_QUEUE_GET_PROPERTY_BAN:
674 		args->value = q->ops->reset_status(q);
675 		ret = 0;
676 		break;
677 	default:
678 		ret = -EINVAL;
679 	}
680 
681 	xe_exec_queue_put(q);
682 
683 	return ret;
684 }
685 
686 /**
687  * xe_exec_queue_is_lr() - Whether an exec_queue is long-running
688  * @q: The exec_queue
689  *
690  * Return: True if the exec_queue is long-running, false otherwise.
691  */
692 bool xe_exec_queue_is_lr(struct xe_exec_queue *q)
693 {
694 	return q->vm && xe_vm_in_lr_mode(q->vm) &&
695 		!(q->flags & EXEC_QUEUE_FLAG_VM);
696 }
697 
698 static s32 xe_exec_queue_num_job_inflight(struct xe_exec_queue *q)
699 {
700 	return q->lrc[0]->fence_ctx.next_seqno - xe_lrc_seqno(q->lrc[0]) - 1;
701 }
702 
703 /**
704  * xe_exec_queue_ring_full() - Whether an exec_queue's ring is full
705  * @q: The exec_queue
706  *
707  * Return: True if the exec_queue's ring is full, false otherwise.
708  */
709 bool xe_exec_queue_ring_full(struct xe_exec_queue *q)
710 {
711 	struct xe_lrc *lrc = q->lrc[0];
712 	s32 max_job = lrc->ring.size / MAX_JOB_SIZE_BYTES;
713 
714 	return xe_exec_queue_num_job_inflight(q) >= max_job;
715 }
716 
717 /**
718  * xe_exec_queue_is_idle() - Whether an exec_queue is idle.
719  * @q: The exec_queue
720  *
721  * FIXME: Need to determine what to use as the short-lived
722  * timeline lock for the exec_queues, so that the return value
723  * of this function becomes more than just an advisory
724  * snapshot in time. The timeline lock must protect the
725  * seqno from racing submissions on the same exec_queue.
726  * Typically vm->resv, but user-created timeline locks use the migrate vm
727  * and never grabs the migrate vm->resv so we have a race there.
728  *
729  * Return: True if the exec_queue is idle, false otherwise.
730  */
731 bool xe_exec_queue_is_idle(struct xe_exec_queue *q)
732 {
733 	if (xe_exec_queue_is_parallel(q)) {
734 		int i;
735 
736 		for (i = 0; i < q->width; ++i) {
737 			if (xe_lrc_seqno(q->lrc[i]) !=
738 			    q->lrc[i]->fence_ctx.next_seqno - 1)
739 				return false;
740 		}
741 
742 		return true;
743 	}
744 
745 	return xe_lrc_seqno(q->lrc[0]) ==
746 		q->lrc[0]->fence_ctx.next_seqno - 1;
747 }
748 
749 /**
750  * xe_exec_queue_update_run_ticks() - Update run time in ticks for this exec queue
751  * from hw
752  * @q: The exec queue
753  *
754  * Update the timestamp saved by HW for this exec queue and save run ticks
755  * calculated by using the delta from last update.
756  */
757 void xe_exec_queue_update_run_ticks(struct xe_exec_queue *q)
758 {
759 	struct xe_file *xef;
760 	struct xe_lrc *lrc;
761 	u32 old_ts, new_ts;
762 
763 	/*
764 	 * Jobs that are run during driver load may use an exec_queue, but are
765 	 * not associated with a user xe file, so avoid accumulating busyness
766 	 * for kernel specific work.
767 	 */
768 	if (!q->vm || !q->vm->xef)
769 		return;
770 
771 	xef = q->vm->xef;
772 
773 	/*
774 	 * Only sample the first LRC. For parallel submission, all of them are
775 	 * scheduled together and we compensate that below by multiplying by
776 	 * width - this may introduce errors if that premise is not true and
777 	 * they don't exit 100% aligned. On the other hand, looping through
778 	 * the LRCs and reading them in different time could also introduce
779 	 * errors.
780 	 */
781 	lrc = q->lrc[0];
782 	new_ts = xe_lrc_update_timestamp(lrc, &old_ts);
783 	xef->run_ticks[q->class] += (new_ts - old_ts) * q->width;
784 }
785 
786 /**
787  * xe_exec_queue_kill - permanently stop all execution from an exec queue
788  * @q: The exec queue
789  *
790  * This function permanently stops all activity on an exec queue. If the queue
791  * is actively executing on the HW, it will be kicked off the engine; any
792  * pending jobs are discarded and all future submissions are rejected.
793  * This function is safe to call multiple times.
794  */
795 void xe_exec_queue_kill(struct xe_exec_queue *q)
796 {
797 	struct xe_exec_queue *eq = q, *next;
798 
799 	list_for_each_entry_safe(eq, next, &eq->multi_gt_list,
800 				 multi_gt_link) {
801 		q->ops->kill(eq);
802 		xe_vm_remove_compute_exec_queue(q->vm, eq);
803 	}
804 
805 	q->ops->kill(q);
806 	xe_vm_remove_compute_exec_queue(q->vm, q);
807 }
808 
809 int xe_exec_queue_destroy_ioctl(struct drm_device *dev, void *data,
810 				struct drm_file *file)
811 {
812 	struct xe_device *xe = to_xe_device(dev);
813 	struct xe_file *xef = to_xe_file(file);
814 	struct drm_xe_exec_queue_destroy *args = data;
815 	struct xe_exec_queue *q;
816 
817 	if (XE_IOCTL_DBG(xe, args->pad) ||
818 	    XE_IOCTL_DBG(xe, args->reserved[0] || args->reserved[1]))
819 		return -EINVAL;
820 
821 	mutex_lock(&xef->exec_queue.lock);
822 	q = xa_erase(&xef->exec_queue.xa, args->exec_queue_id);
823 	mutex_unlock(&xef->exec_queue.lock);
824 	if (XE_IOCTL_DBG(xe, !q))
825 		return -ENOENT;
826 
827 	if (q->vm && q->hwe->hw_engine_group)
828 		xe_hw_engine_group_del_exec_queue(q->hwe->hw_engine_group, q);
829 
830 	xe_exec_queue_kill(q);
831 
832 	trace_xe_exec_queue_close(q);
833 	xe_exec_queue_put(q);
834 
835 	return 0;
836 }
837 
838 static void xe_exec_queue_last_fence_lockdep_assert(struct xe_exec_queue *q,
839 						    struct xe_vm *vm)
840 {
841 	if (q->flags & EXEC_QUEUE_FLAG_VM) {
842 		lockdep_assert_held(&vm->lock);
843 	} else {
844 		xe_vm_assert_held(vm);
845 		lockdep_assert_held(&q->hwe->hw_engine_group->mode_sem);
846 	}
847 }
848 
849 /**
850  * xe_exec_queue_last_fence_put() - Drop ref to last fence
851  * @q: The exec queue
852  * @vm: The VM the engine does a bind or exec for
853  */
854 void xe_exec_queue_last_fence_put(struct xe_exec_queue *q, struct xe_vm *vm)
855 {
856 	xe_exec_queue_last_fence_lockdep_assert(q, vm);
857 
858 	xe_exec_queue_last_fence_put_unlocked(q);
859 }
860 
861 /**
862  * xe_exec_queue_last_fence_put_unlocked() - Drop ref to last fence unlocked
863  * @q: The exec queue
864  *
865  * Only safe to be called from xe_exec_queue_destroy().
866  */
867 void xe_exec_queue_last_fence_put_unlocked(struct xe_exec_queue *q)
868 {
869 	if (q->last_fence) {
870 		dma_fence_put(q->last_fence);
871 		q->last_fence = NULL;
872 	}
873 }
874 
875 /**
876  * xe_exec_queue_last_fence_get() - Get last fence
877  * @q: The exec queue
878  * @vm: The VM the engine does a bind or exec for
879  *
880  * Get last fence, takes a ref
881  *
882  * Returns: last fence if not signaled, dma fence stub if signaled
883  */
884 struct dma_fence *xe_exec_queue_last_fence_get(struct xe_exec_queue *q,
885 					       struct xe_vm *vm)
886 {
887 	struct dma_fence *fence;
888 
889 	xe_exec_queue_last_fence_lockdep_assert(q, vm);
890 
891 	if (q->last_fence &&
892 	    test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &q->last_fence->flags))
893 		xe_exec_queue_last_fence_put(q, vm);
894 
895 	fence = q->last_fence ? q->last_fence : dma_fence_get_stub();
896 	dma_fence_get(fence);
897 	return fence;
898 }
899 
900 /**
901  * xe_exec_queue_last_fence_get_for_resume() - Get last fence
902  * @q: The exec queue
903  * @vm: The VM the engine does a bind or exec for
904  *
905  * Get last fence, takes a ref. Only safe to be called in the context of
906  * resuming the hw engine group's long-running exec queue, when the group
907  * semaphore is held.
908  *
909  * Returns: last fence if not signaled, dma fence stub if signaled
910  */
911 struct dma_fence *xe_exec_queue_last_fence_get_for_resume(struct xe_exec_queue *q,
912 							  struct xe_vm *vm)
913 {
914 	struct dma_fence *fence;
915 
916 	lockdep_assert_held_write(&q->hwe->hw_engine_group->mode_sem);
917 
918 	if (q->last_fence &&
919 	    test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &q->last_fence->flags))
920 		xe_exec_queue_last_fence_put_unlocked(q);
921 
922 	fence = q->last_fence ? q->last_fence : dma_fence_get_stub();
923 	dma_fence_get(fence);
924 	return fence;
925 }
926 
927 /**
928  * xe_exec_queue_last_fence_set() - Set last fence
929  * @q: The exec queue
930  * @vm: The VM the engine does a bind or exec for
931  * @fence: The fence
932  *
933  * Set the last fence for the engine. Increases reference count for fence, when
934  * closing engine xe_exec_queue_last_fence_put should be called.
935  */
936 void xe_exec_queue_last_fence_set(struct xe_exec_queue *q, struct xe_vm *vm,
937 				  struct dma_fence *fence)
938 {
939 	xe_exec_queue_last_fence_lockdep_assert(q, vm);
940 
941 	xe_exec_queue_last_fence_put(q, vm);
942 	q->last_fence = dma_fence_get(fence);
943 }
944 
945 /**
946  * xe_exec_queue_last_fence_test_dep - Test last fence dependency of queue
947  * @q: The exec queue
948  * @vm: The VM the engine does a bind or exec for
949  *
950  * Returns:
951  * -ETIME if there exists an unsignalled last fence dependency, zero otherwise.
952  */
953 int xe_exec_queue_last_fence_test_dep(struct xe_exec_queue *q, struct xe_vm *vm)
954 {
955 	struct dma_fence *fence;
956 	int err = 0;
957 
958 	fence = xe_exec_queue_last_fence_get(q, vm);
959 	if (fence) {
960 		err = test_bit(DMA_FENCE_FLAG_SIGNALED_BIT, &fence->flags) ?
961 			0 : -ETIME;
962 		dma_fence_put(fence);
963 	}
964 
965 	return err;
966 }
967