1 /*
2  * Copyright © 2016 Intel Corporation
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the next
12  * paragraph) shall be included in all copies or substantial portions of the
13  * 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 AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21  * IN THE SOFTWARE.
22  *
23  */
24 
25 #include <drm/drm_print.h>
26 
27 #include "gem/i915_gem_context.h"
28 
29 #include "i915_drv.h"
30 
31 #include "intel_context.h"
32 #include "intel_engine.h"
33 #include "intel_engine_pm.h"
34 #include "intel_engine_pool.h"
35 #include "intel_engine_user.h"
36 #include "intel_gt.h"
37 #include "intel_gt_requests.h"
38 #include "intel_lrc.h"
39 #include "intel_reset.h"
40 #include "intel_ring.h"
41 
42 /* Haswell does have the CXT_SIZE register however it does not appear to be
43  * valid. Now, docs explain in dwords what is in the context object. The full
44  * size is 70720 bytes, however, the power context and execlist context will
45  * never be saved (power context is stored elsewhere, and execlists don't work
46  * on HSW) - so the final size, including the extra state required for the
47  * Resource Streamer, is 66944 bytes, which rounds to 17 pages.
48  */
49 #define HSW_CXT_TOTAL_SIZE		(17 * PAGE_SIZE)
50 
51 #define DEFAULT_LR_CONTEXT_RENDER_SIZE	(22 * PAGE_SIZE)
52 #define GEN8_LR_CONTEXT_RENDER_SIZE	(20 * PAGE_SIZE)
53 #define GEN9_LR_CONTEXT_RENDER_SIZE	(22 * PAGE_SIZE)
54 #define GEN10_LR_CONTEXT_RENDER_SIZE	(18 * PAGE_SIZE)
55 #define GEN11_LR_CONTEXT_RENDER_SIZE	(14 * PAGE_SIZE)
56 
57 #define GEN8_LR_CONTEXT_OTHER_SIZE	( 2 * PAGE_SIZE)
58 
59 #define MAX_MMIO_BASES 3
60 struct engine_info {
61 	unsigned int hw_id;
62 	u8 class;
63 	u8 instance;
64 	/* mmio bases table *must* be sorted in reverse gen order */
65 	struct engine_mmio_base {
66 		u32 gen : 8;
67 		u32 base : 24;
68 	} mmio_bases[MAX_MMIO_BASES];
69 };
70 
71 static const struct engine_info intel_engines[] = {
72 	[RCS0] = {
73 		.hw_id = RCS0_HW,
74 		.class = RENDER_CLASS,
75 		.instance = 0,
76 		.mmio_bases = {
77 			{ .gen = 1, .base = RENDER_RING_BASE }
78 		},
79 	},
80 	[BCS0] = {
81 		.hw_id = BCS0_HW,
82 		.class = COPY_ENGINE_CLASS,
83 		.instance = 0,
84 		.mmio_bases = {
85 			{ .gen = 6, .base = BLT_RING_BASE }
86 		},
87 	},
88 	[VCS0] = {
89 		.hw_id = VCS0_HW,
90 		.class = VIDEO_DECODE_CLASS,
91 		.instance = 0,
92 		.mmio_bases = {
93 			{ .gen = 11, .base = GEN11_BSD_RING_BASE },
94 			{ .gen = 6, .base = GEN6_BSD_RING_BASE },
95 			{ .gen = 4, .base = BSD_RING_BASE }
96 		},
97 	},
98 	[VCS1] = {
99 		.hw_id = VCS1_HW,
100 		.class = VIDEO_DECODE_CLASS,
101 		.instance = 1,
102 		.mmio_bases = {
103 			{ .gen = 11, .base = GEN11_BSD2_RING_BASE },
104 			{ .gen = 8, .base = GEN8_BSD2_RING_BASE }
105 		},
106 	},
107 	[VCS2] = {
108 		.hw_id = VCS2_HW,
109 		.class = VIDEO_DECODE_CLASS,
110 		.instance = 2,
111 		.mmio_bases = {
112 			{ .gen = 11, .base = GEN11_BSD3_RING_BASE }
113 		},
114 	},
115 	[VCS3] = {
116 		.hw_id = VCS3_HW,
117 		.class = VIDEO_DECODE_CLASS,
118 		.instance = 3,
119 		.mmio_bases = {
120 			{ .gen = 11, .base = GEN11_BSD4_RING_BASE }
121 		},
122 	},
123 	[VECS0] = {
124 		.hw_id = VECS0_HW,
125 		.class = VIDEO_ENHANCEMENT_CLASS,
126 		.instance = 0,
127 		.mmio_bases = {
128 			{ .gen = 11, .base = GEN11_VEBOX_RING_BASE },
129 			{ .gen = 7, .base = VEBOX_RING_BASE }
130 		},
131 	},
132 	[VECS1] = {
133 		.hw_id = VECS1_HW,
134 		.class = VIDEO_ENHANCEMENT_CLASS,
135 		.instance = 1,
136 		.mmio_bases = {
137 			{ .gen = 11, .base = GEN11_VEBOX2_RING_BASE }
138 		},
139 	},
140 };
141 
142 /**
143  * intel_engine_context_size() - return the size of the context for an engine
144  * @gt: the gt
145  * @class: engine class
146  *
147  * Each engine class may require a different amount of space for a context
148  * image.
149  *
150  * Return: size (in bytes) of an engine class specific context image
151  *
152  * Note: this size includes the HWSP, which is part of the context image
153  * in LRC mode, but does not include the "shared data page" used with
154  * GuC submission. The caller should account for this if using the GuC.
155  */
156 u32 intel_engine_context_size(struct intel_gt *gt, u8 class)
157 {
158 	struct intel_uncore *uncore = gt->uncore;
159 	u32 cxt_size;
160 
161 	BUILD_BUG_ON(I915_GTT_PAGE_SIZE != PAGE_SIZE);
162 
163 	switch (class) {
164 	case RENDER_CLASS:
165 		switch (INTEL_GEN(gt->i915)) {
166 		default:
167 			MISSING_CASE(INTEL_GEN(gt->i915));
168 			return DEFAULT_LR_CONTEXT_RENDER_SIZE;
169 		case 12:
170 		case 11:
171 			return GEN11_LR_CONTEXT_RENDER_SIZE;
172 		case 10:
173 			return GEN10_LR_CONTEXT_RENDER_SIZE;
174 		case 9:
175 			return GEN9_LR_CONTEXT_RENDER_SIZE;
176 		case 8:
177 			return GEN8_LR_CONTEXT_RENDER_SIZE;
178 		case 7:
179 			if (IS_HASWELL(gt->i915))
180 				return HSW_CXT_TOTAL_SIZE;
181 
182 			cxt_size = intel_uncore_read(uncore, GEN7_CXT_SIZE);
183 			return round_up(GEN7_CXT_TOTAL_SIZE(cxt_size) * 64,
184 					PAGE_SIZE);
185 		case 6:
186 			cxt_size = intel_uncore_read(uncore, CXT_SIZE);
187 			return round_up(GEN6_CXT_TOTAL_SIZE(cxt_size) * 64,
188 					PAGE_SIZE);
189 		case 5:
190 		case 4:
191 			/*
192 			 * There is a discrepancy here between the size reported
193 			 * by the register and the size of the context layout
194 			 * in the docs. Both are described as authorative!
195 			 *
196 			 * The discrepancy is on the order of a few cachelines,
197 			 * but the total is under one page (4k), which is our
198 			 * minimum allocation anyway so it should all come
199 			 * out in the wash.
200 			 */
201 			cxt_size = intel_uncore_read(uncore, CXT_SIZE) + 1;
202 			DRM_DEBUG_DRIVER("gen%d CXT_SIZE = %d bytes [0x%08x]\n",
203 					 INTEL_GEN(gt->i915),
204 					 cxt_size * 64,
205 					 cxt_size - 1);
206 			return round_up(cxt_size * 64, PAGE_SIZE);
207 		case 3:
208 		case 2:
209 		/* For the special day when i810 gets merged. */
210 		case 1:
211 			return 0;
212 		}
213 		break;
214 	default:
215 		MISSING_CASE(class);
216 		/* fall through */
217 	case VIDEO_DECODE_CLASS:
218 	case VIDEO_ENHANCEMENT_CLASS:
219 	case COPY_ENGINE_CLASS:
220 		if (INTEL_GEN(gt->i915) < 8)
221 			return 0;
222 		return GEN8_LR_CONTEXT_OTHER_SIZE;
223 	}
224 }
225 
226 static u32 __engine_mmio_base(struct drm_i915_private *i915,
227 			      const struct engine_mmio_base *bases)
228 {
229 	int i;
230 
231 	for (i = 0; i < MAX_MMIO_BASES; i++)
232 		if (INTEL_GEN(i915) >= bases[i].gen)
233 			break;
234 
235 	GEM_BUG_ON(i == MAX_MMIO_BASES);
236 	GEM_BUG_ON(!bases[i].base);
237 
238 	return bases[i].base;
239 }
240 
241 static void __sprint_engine_name(struct intel_engine_cs *engine)
242 {
243 	/*
244 	 * Before we know what the uABI name for this engine will be,
245 	 * we still would like to keep track of this engine in the debug logs.
246 	 * We throw in a ' here as a reminder that this isn't its final name.
247 	 */
248 	GEM_WARN_ON(snprintf(engine->name, sizeof(engine->name), "%s'%u",
249 			     intel_engine_class_repr(engine->class),
250 			     engine->instance) >= sizeof(engine->name));
251 }
252 
253 void intel_engine_set_hwsp_writemask(struct intel_engine_cs *engine, u32 mask)
254 {
255 	/*
256 	 * Though they added more rings on g4x/ilk, they did not add
257 	 * per-engine HWSTAM until gen6.
258 	 */
259 	if (INTEL_GEN(engine->i915) < 6 && engine->class != RENDER_CLASS)
260 		return;
261 
262 	if (INTEL_GEN(engine->i915) >= 3)
263 		ENGINE_WRITE(engine, RING_HWSTAM, mask);
264 	else
265 		ENGINE_WRITE16(engine, RING_HWSTAM, mask);
266 }
267 
268 static void intel_engine_sanitize_mmio(struct intel_engine_cs *engine)
269 {
270 	/* Mask off all writes into the unknown HWSP */
271 	intel_engine_set_hwsp_writemask(engine, ~0u);
272 }
273 
274 static int intel_engine_setup(struct intel_gt *gt, enum intel_engine_id id)
275 {
276 	const struct engine_info *info = &intel_engines[id];
277 	struct intel_engine_cs *engine;
278 
279 	BUILD_BUG_ON(MAX_ENGINE_CLASS >= BIT(GEN11_ENGINE_CLASS_WIDTH));
280 	BUILD_BUG_ON(MAX_ENGINE_INSTANCE >= BIT(GEN11_ENGINE_INSTANCE_WIDTH));
281 
282 	if (GEM_DEBUG_WARN_ON(id >= ARRAY_SIZE(gt->engine)))
283 		return -EINVAL;
284 
285 	if (GEM_DEBUG_WARN_ON(info->class > MAX_ENGINE_CLASS))
286 		return -EINVAL;
287 
288 	if (GEM_DEBUG_WARN_ON(info->instance > MAX_ENGINE_INSTANCE))
289 		return -EINVAL;
290 
291 	if (GEM_DEBUG_WARN_ON(gt->engine_class[info->class][info->instance]))
292 		return -EINVAL;
293 
294 	engine = kzalloc(sizeof(*engine), GFP_KERNEL);
295 	if (!engine)
296 		return -ENOMEM;
297 
298 	BUILD_BUG_ON(BITS_PER_TYPE(engine->mask) < I915_NUM_ENGINES);
299 
300 	engine->id = id;
301 	engine->legacy_idx = INVALID_ENGINE;
302 	engine->mask = BIT(id);
303 	engine->i915 = gt->i915;
304 	engine->gt = gt;
305 	engine->uncore = gt->uncore;
306 	engine->hw_id = engine->guc_id = info->hw_id;
307 	engine->mmio_base = __engine_mmio_base(gt->i915, info->mmio_bases);
308 
309 	engine->class = info->class;
310 	engine->instance = info->instance;
311 	__sprint_engine_name(engine);
312 
313 	engine->props.heartbeat_interval_ms =
314 		CONFIG_DRM_I915_HEARTBEAT_INTERVAL;
315 	engine->props.preempt_timeout_ms =
316 		CONFIG_DRM_I915_PREEMPT_TIMEOUT;
317 	engine->props.stop_timeout_ms =
318 		CONFIG_DRM_I915_STOP_TIMEOUT;
319 	engine->props.timeslice_duration_ms =
320 		CONFIG_DRM_I915_TIMESLICE_DURATION;
321 
322 	engine->context_size = intel_engine_context_size(gt, engine->class);
323 	if (WARN_ON(engine->context_size > BIT(20)))
324 		engine->context_size = 0;
325 	if (engine->context_size)
326 		DRIVER_CAPS(gt->i915)->has_logical_contexts = true;
327 
328 	/* Nothing to do here, execute in order of dependencies */
329 	engine->schedule = NULL;
330 
331 	ewma__engine_latency_init(&engine->latency);
332 	seqlock_init(&engine->stats.lock);
333 
334 	ATOMIC_INIT_NOTIFIER_HEAD(&engine->context_status_notifier);
335 
336 	/* Scrub mmio state on takeover */
337 	intel_engine_sanitize_mmio(engine);
338 
339 	gt->engine_class[info->class][info->instance] = engine;
340 	gt->engine[id] = engine;
341 
342 	gt->i915->engine[id] = engine;
343 
344 	return 0;
345 }
346 
347 static void __setup_engine_capabilities(struct intel_engine_cs *engine)
348 {
349 	struct drm_i915_private *i915 = engine->i915;
350 
351 	if (engine->class == VIDEO_DECODE_CLASS) {
352 		/*
353 		 * HEVC support is present on first engine instance
354 		 * before Gen11 and on all instances afterwards.
355 		 */
356 		if (INTEL_GEN(i915) >= 11 ||
357 		    (INTEL_GEN(i915) >= 9 && engine->instance == 0))
358 			engine->uabi_capabilities |=
359 				I915_VIDEO_CLASS_CAPABILITY_HEVC;
360 
361 		/*
362 		 * SFC block is present only on even logical engine
363 		 * instances.
364 		 */
365 		if ((INTEL_GEN(i915) >= 11 &&
366 		     RUNTIME_INFO(i915)->vdbox_sfc_access & engine->mask) ||
367 		    (INTEL_GEN(i915) >= 9 && engine->instance == 0))
368 			engine->uabi_capabilities |=
369 				I915_VIDEO_AND_ENHANCE_CLASS_CAPABILITY_SFC;
370 	} else if (engine->class == VIDEO_ENHANCEMENT_CLASS) {
371 		if (INTEL_GEN(i915) >= 9)
372 			engine->uabi_capabilities |=
373 				I915_VIDEO_AND_ENHANCE_CLASS_CAPABILITY_SFC;
374 	}
375 }
376 
377 static void intel_setup_engine_capabilities(struct intel_gt *gt)
378 {
379 	struct intel_engine_cs *engine;
380 	enum intel_engine_id id;
381 
382 	for_each_engine(engine, gt, id)
383 		__setup_engine_capabilities(engine);
384 }
385 
386 /**
387  * intel_engines_release() - free the resources allocated for Command Streamers
388  * @gt: pointer to struct intel_gt
389  */
390 void intel_engines_release(struct intel_gt *gt)
391 {
392 	struct intel_engine_cs *engine;
393 	enum intel_engine_id id;
394 
395 	/* Decouple the backend; but keep the layout for late GPU resets */
396 	for_each_engine(engine, gt, id) {
397 		if (!engine->release)
398 			continue;
399 
400 		engine->release(engine);
401 		engine->release = NULL;
402 
403 		memset(&engine->reset, 0, sizeof(engine->reset));
404 
405 		gt->i915->engine[id] = NULL;
406 	}
407 }
408 
409 void intel_engines_free(struct intel_gt *gt)
410 {
411 	struct intel_engine_cs *engine;
412 	enum intel_engine_id id;
413 
414 	for_each_engine(engine, gt, id) {
415 		kfree(engine);
416 		gt->engine[id] = NULL;
417 	}
418 }
419 
420 /**
421  * intel_engines_init_mmio() - allocate and prepare the Engine Command Streamers
422  * @gt: pointer to struct intel_gt
423  *
424  * Return: non-zero if the initialization failed.
425  */
426 int intel_engines_init_mmio(struct intel_gt *gt)
427 {
428 	struct drm_i915_private *i915 = gt->i915;
429 	struct intel_device_info *device_info = mkwrite_device_info(i915);
430 	const unsigned int engine_mask = INTEL_INFO(i915)->engine_mask;
431 	unsigned int mask = 0;
432 	unsigned int i;
433 	int err;
434 
435 	WARN_ON(engine_mask == 0);
436 	WARN_ON(engine_mask &
437 		GENMASK(BITS_PER_TYPE(mask) - 1, I915_NUM_ENGINES));
438 
439 	if (i915_inject_probe_failure(i915))
440 		return -ENODEV;
441 
442 	for (i = 0; i < ARRAY_SIZE(intel_engines); i++) {
443 		if (!HAS_ENGINE(i915, i))
444 			continue;
445 
446 		err = intel_engine_setup(gt, i);
447 		if (err)
448 			goto cleanup;
449 
450 		mask |= BIT(i);
451 	}
452 
453 	/*
454 	 * Catch failures to update intel_engines table when the new engines
455 	 * are added to the driver by a warning and disabling the forgotten
456 	 * engines.
457 	 */
458 	if (WARN_ON(mask != engine_mask))
459 		device_info->engine_mask = mask;
460 
461 	RUNTIME_INFO(i915)->num_engines = hweight32(mask);
462 
463 	intel_gt_check_and_clear_faults(gt);
464 
465 	intel_setup_engine_capabilities(gt);
466 
467 	return 0;
468 
469 cleanup:
470 	intel_engines_free(gt);
471 	return err;
472 }
473 
474 void intel_engine_init_execlists(struct intel_engine_cs *engine)
475 {
476 	struct intel_engine_execlists * const execlists = &engine->execlists;
477 
478 	execlists->port_mask = 1;
479 	GEM_BUG_ON(!is_power_of_2(execlists_num_ports(execlists)));
480 	GEM_BUG_ON(execlists_num_ports(execlists) > EXECLIST_MAX_PORTS);
481 
482 	memset(execlists->pending, 0, sizeof(execlists->pending));
483 	execlists->active =
484 		memset(execlists->inflight, 0, sizeof(execlists->inflight));
485 
486 	execlists->queue_priority_hint = INT_MIN;
487 	execlists->queue = RB_ROOT_CACHED;
488 }
489 
490 static void cleanup_status_page(struct intel_engine_cs *engine)
491 {
492 	struct i915_vma *vma;
493 
494 	/* Prevent writes into HWSP after returning the page to the system */
495 	intel_engine_set_hwsp_writemask(engine, ~0u);
496 
497 	vma = fetch_and_zero(&engine->status_page.vma);
498 	if (!vma)
499 		return;
500 
501 	if (!HWS_NEEDS_PHYSICAL(engine->i915))
502 		i915_vma_unpin(vma);
503 
504 	i915_gem_object_unpin_map(vma->obj);
505 	i915_gem_object_put(vma->obj);
506 }
507 
508 static int pin_ggtt_status_page(struct intel_engine_cs *engine,
509 				struct i915_vma *vma)
510 {
511 	unsigned int flags;
512 
513 	flags = PIN_GLOBAL;
514 	if (!HAS_LLC(engine->i915) && i915_ggtt_has_aperture(engine->gt->ggtt))
515 		/*
516 		 * On g33, we cannot place HWS above 256MiB, so
517 		 * restrict its pinning to the low mappable arena.
518 		 * Though this restriction is not documented for
519 		 * gen4, gen5, or byt, they also behave similarly
520 		 * and hang if the HWS is placed at the top of the
521 		 * GTT. To generalise, it appears that all !llc
522 		 * platforms have issues with us placing the HWS
523 		 * above the mappable region (even though we never
524 		 * actually map it).
525 		 */
526 		flags |= PIN_MAPPABLE;
527 	else
528 		flags |= PIN_HIGH;
529 
530 	return i915_vma_pin(vma, 0, 0, flags);
531 }
532 
533 static int init_status_page(struct intel_engine_cs *engine)
534 {
535 	struct drm_i915_gem_object *obj;
536 	struct i915_vma *vma;
537 	void *vaddr;
538 	int ret;
539 
540 	/*
541 	 * Though the HWS register does support 36bit addresses, historically
542 	 * we have had hangs and corruption reported due to wild writes if
543 	 * the HWS is placed above 4G. We only allow objects to be allocated
544 	 * in GFP_DMA32 for i965, and no earlier physical address users had
545 	 * access to more than 4G.
546 	 */
547 	obj = i915_gem_object_create_internal(engine->i915, PAGE_SIZE);
548 	if (IS_ERR(obj)) {
549 		DRM_ERROR("Failed to allocate status page\n");
550 		return PTR_ERR(obj);
551 	}
552 
553 	i915_gem_object_set_cache_coherency(obj, I915_CACHE_LLC);
554 
555 	vma = i915_vma_instance(obj, &engine->gt->ggtt->vm, NULL);
556 	if (IS_ERR(vma)) {
557 		ret = PTR_ERR(vma);
558 		goto err;
559 	}
560 
561 	vaddr = i915_gem_object_pin_map(obj, I915_MAP_WB);
562 	if (IS_ERR(vaddr)) {
563 		ret = PTR_ERR(vaddr);
564 		goto err;
565 	}
566 
567 	engine->status_page.addr = memset(vaddr, 0, PAGE_SIZE);
568 	engine->status_page.vma = vma;
569 
570 	if (!HWS_NEEDS_PHYSICAL(engine->i915)) {
571 		ret = pin_ggtt_status_page(engine, vma);
572 		if (ret)
573 			goto err_unpin;
574 	}
575 
576 	return 0;
577 
578 err_unpin:
579 	i915_gem_object_unpin_map(obj);
580 err:
581 	i915_gem_object_put(obj);
582 	return ret;
583 }
584 
585 static int engine_setup_common(struct intel_engine_cs *engine)
586 {
587 	int err;
588 
589 	init_llist_head(&engine->barrier_tasks);
590 
591 	err = init_status_page(engine);
592 	if (err)
593 		return err;
594 
595 	intel_engine_init_active(engine, ENGINE_PHYSICAL);
596 	intel_engine_init_breadcrumbs(engine);
597 	intel_engine_init_execlists(engine);
598 	intel_engine_init_cmd_parser(engine);
599 	intel_engine_init__pm(engine);
600 	intel_engine_init_retire(engine);
601 
602 	intel_engine_pool_init(&engine->pool);
603 
604 	/* Use the whole device by default */
605 	engine->sseu =
606 		intel_sseu_from_device_info(&RUNTIME_INFO(engine->i915)->sseu);
607 
608 	intel_engine_init_workarounds(engine);
609 	intel_engine_init_whitelist(engine);
610 	intel_engine_init_ctx_wa(engine);
611 
612 	return 0;
613 }
614 
615 struct measure_breadcrumb {
616 	struct i915_request rq;
617 	struct intel_timeline timeline;
618 	struct intel_ring ring;
619 	u32 cs[1024];
620 };
621 
622 static int measure_breadcrumb_dw(struct intel_engine_cs *engine)
623 {
624 	struct measure_breadcrumb *frame;
625 	int dw = -ENOMEM;
626 
627 	GEM_BUG_ON(!engine->gt->scratch);
628 
629 	frame = kzalloc(sizeof(*frame), GFP_KERNEL);
630 	if (!frame)
631 		return -ENOMEM;
632 
633 	if (intel_timeline_init(&frame->timeline,
634 				engine->gt,
635 				engine->status_page.vma))
636 		goto out_frame;
637 
638 	mutex_lock(&frame->timeline.mutex);
639 
640 	frame->ring.vaddr = frame->cs;
641 	frame->ring.size = sizeof(frame->cs);
642 	frame->ring.effective_size = frame->ring.size;
643 	intel_ring_update_space(&frame->ring);
644 
645 	frame->rq.i915 = engine->i915;
646 	frame->rq.engine = engine;
647 	frame->rq.ring = &frame->ring;
648 	rcu_assign_pointer(frame->rq.timeline, &frame->timeline);
649 
650 	dw = intel_timeline_pin(&frame->timeline);
651 	if (dw < 0)
652 		goto out_timeline;
653 
654 	spin_lock_irq(&engine->active.lock);
655 	dw = engine->emit_fini_breadcrumb(&frame->rq, frame->cs) - frame->cs;
656 	spin_unlock_irq(&engine->active.lock);
657 
658 	GEM_BUG_ON(dw & 1); /* RING_TAIL must be qword aligned */
659 
660 	intel_timeline_unpin(&frame->timeline);
661 
662 out_timeline:
663 	mutex_unlock(&frame->timeline.mutex);
664 	intel_timeline_fini(&frame->timeline);
665 out_frame:
666 	kfree(frame);
667 	return dw;
668 }
669 
670 void
671 intel_engine_init_active(struct intel_engine_cs *engine, unsigned int subclass)
672 {
673 	INIT_LIST_HEAD(&engine->active.requests);
674 	INIT_LIST_HEAD(&engine->active.hold);
675 
676 	spin_lock_init(&engine->active.lock);
677 	lockdep_set_subclass(&engine->active.lock, subclass);
678 
679 	/*
680 	 * Due to an interesting quirk in lockdep's internal debug tracking,
681 	 * after setting a subclass we must ensure the lock is used. Otherwise,
682 	 * nr_unused_locks is incremented once too often.
683 	 */
684 #ifdef CONFIG_DEBUG_LOCK_ALLOC
685 	local_irq_disable();
686 	lock_map_acquire(&engine->active.lock.dep_map);
687 	lock_map_release(&engine->active.lock.dep_map);
688 	local_irq_enable();
689 #endif
690 }
691 
692 static struct intel_context *
693 create_kernel_context(struct intel_engine_cs *engine)
694 {
695 	static struct lock_class_key kernel;
696 	struct intel_context *ce;
697 	int err;
698 
699 	ce = intel_context_create(engine);
700 	if (IS_ERR(ce))
701 		return ce;
702 
703 	__set_bit(CONTEXT_BARRIER_BIT, &ce->flags);
704 
705 	err = intel_context_pin(ce); /* perma-pin so it is always available */
706 	if (err) {
707 		intel_context_put(ce);
708 		return ERR_PTR(err);
709 	}
710 
711 	/*
712 	 * Give our perma-pinned kernel timelines a separate lockdep class,
713 	 * so that we can use them from within the normal user timelines
714 	 * should we need to inject GPU operations during their request
715 	 * construction.
716 	 */
717 	lockdep_set_class(&ce->timeline->mutex, &kernel);
718 
719 	return ce;
720 }
721 
722 /**
723  * intel_engines_init_common - initialize cengine state which might require hw access
724  * @engine: Engine to initialize.
725  *
726  * Initializes @engine@ structure members shared between legacy and execlists
727  * submission modes which do require hardware access.
728  *
729  * Typcally done at later stages of submission mode specific engine setup.
730  *
731  * Returns zero on success or an error code on failure.
732  */
733 static int engine_init_common(struct intel_engine_cs *engine)
734 {
735 	struct intel_context *ce;
736 	int ret;
737 
738 	engine->set_default_submission(engine);
739 
740 	ret = measure_breadcrumb_dw(engine);
741 	if (ret < 0)
742 		return ret;
743 
744 	engine->emit_fini_breadcrumb_dw = ret;
745 
746 	/*
747 	 * We may need to do things with the shrinker which
748 	 * require us to immediately switch back to the default
749 	 * context. This can cause a problem as pinning the
750 	 * default context also requires GTT space which may not
751 	 * be available. To avoid this we always pin the default
752 	 * context.
753 	 */
754 	ce = create_kernel_context(engine);
755 	if (IS_ERR(ce))
756 		return PTR_ERR(ce);
757 
758 	engine->kernel_context = ce;
759 
760 	return 0;
761 }
762 
763 int intel_engines_init(struct intel_gt *gt)
764 {
765 	int (*setup)(struct intel_engine_cs *engine);
766 	struct intel_engine_cs *engine;
767 	enum intel_engine_id id;
768 	int err;
769 
770 	if (HAS_EXECLISTS(gt->i915))
771 		setup = intel_execlists_submission_setup;
772 	else
773 		setup = intel_ring_submission_setup;
774 
775 	for_each_engine(engine, gt, id) {
776 		err = engine_setup_common(engine);
777 		if (err)
778 			return err;
779 
780 		err = setup(engine);
781 		if (err)
782 			return err;
783 
784 		err = engine_init_common(engine);
785 		if (err)
786 			return err;
787 
788 		intel_engine_add_user(engine);
789 	}
790 
791 	return 0;
792 }
793 
794 /**
795  * intel_engines_cleanup_common - cleans up the engine state created by
796  *                                the common initiailizers.
797  * @engine: Engine to cleanup.
798  *
799  * This cleans up everything created by the common helpers.
800  */
801 void intel_engine_cleanup_common(struct intel_engine_cs *engine)
802 {
803 	GEM_BUG_ON(!list_empty(&engine->active.requests));
804 	tasklet_kill(&engine->execlists.tasklet); /* flush the callback */
805 
806 	cleanup_status_page(engine);
807 
808 	intel_engine_fini_retire(engine);
809 	intel_engine_pool_fini(&engine->pool);
810 	intel_engine_fini_breadcrumbs(engine);
811 	intel_engine_cleanup_cmd_parser(engine);
812 
813 	if (engine->default_state)
814 		i915_gem_object_put(engine->default_state);
815 
816 	if (engine->kernel_context) {
817 		intel_context_unpin(engine->kernel_context);
818 		intel_context_put(engine->kernel_context);
819 	}
820 	GEM_BUG_ON(!llist_empty(&engine->barrier_tasks));
821 
822 	intel_wa_list_free(&engine->ctx_wa_list);
823 	intel_wa_list_free(&engine->wa_list);
824 	intel_wa_list_free(&engine->whitelist);
825 }
826 
827 u64 intel_engine_get_active_head(const struct intel_engine_cs *engine)
828 {
829 	struct drm_i915_private *i915 = engine->i915;
830 
831 	u64 acthd;
832 
833 	if (INTEL_GEN(i915) >= 8)
834 		acthd = ENGINE_READ64(engine, RING_ACTHD, RING_ACTHD_UDW);
835 	else if (INTEL_GEN(i915) >= 4)
836 		acthd = ENGINE_READ(engine, RING_ACTHD);
837 	else
838 		acthd = ENGINE_READ(engine, ACTHD);
839 
840 	return acthd;
841 }
842 
843 u64 intel_engine_get_last_batch_head(const struct intel_engine_cs *engine)
844 {
845 	u64 bbaddr;
846 
847 	if (INTEL_GEN(engine->i915) >= 8)
848 		bbaddr = ENGINE_READ64(engine, RING_BBADDR, RING_BBADDR_UDW);
849 	else
850 		bbaddr = ENGINE_READ(engine, RING_BBADDR);
851 
852 	return bbaddr;
853 }
854 
855 static unsigned long stop_timeout(const struct intel_engine_cs *engine)
856 {
857 	if (in_atomic() || irqs_disabled()) /* inside atomic preempt-reset? */
858 		return 0;
859 
860 	/*
861 	 * If we are doing a normal GPU reset, we can take our time and allow
862 	 * the engine to quiesce. We've stopped submission to the engine, and
863 	 * if we wait long enough an innocent context should complete and
864 	 * leave the engine idle. So they should not be caught unaware by
865 	 * the forthcoming GPU reset (which usually follows the stop_cs)!
866 	 */
867 	return READ_ONCE(engine->props.stop_timeout_ms);
868 }
869 
870 int intel_engine_stop_cs(struct intel_engine_cs *engine)
871 {
872 	struct intel_uncore *uncore = engine->uncore;
873 	const u32 base = engine->mmio_base;
874 	const i915_reg_t mode = RING_MI_MODE(base);
875 	int err;
876 
877 	if (INTEL_GEN(engine->i915) < 3)
878 		return -ENODEV;
879 
880 	ENGINE_TRACE(engine, "\n");
881 
882 	intel_uncore_write_fw(uncore, mode, _MASKED_BIT_ENABLE(STOP_RING));
883 
884 	err = 0;
885 	if (__intel_wait_for_register_fw(uncore,
886 					 mode, MODE_IDLE, MODE_IDLE,
887 					 1000, stop_timeout(engine),
888 					 NULL)) {
889 		ENGINE_TRACE(engine, "timed out on STOP_RING -> IDLE\n");
890 		err = -ETIMEDOUT;
891 	}
892 
893 	/* A final mmio read to let GPU writes be hopefully flushed to memory */
894 	intel_uncore_posting_read_fw(uncore, mode);
895 
896 	return err;
897 }
898 
899 void intel_engine_cancel_stop_cs(struct intel_engine_cs *engine)
900 {
901 	ENGINE_TRACE(engine, "\n");
902 
903 	ENGINE_WRITE_FW(engine, RING_MI_MODE, _MASKED_BIT_DISABLE(STOP_RING));
904 }
905 
906 const char *i915_cache_level_str(struct drm_i915_private *i915, int type)
907 {
908 	switch (type) {
909 	case I915_CACHE_NONE: return " uncached";
910 	case I915_CACHE_LLC: return HAS_LLC(i915) ? " LLC" : " snooped";
911 	case I915_CACHE_L3_LLC: return " L3+LLC";
912 	case I915_CACHE_WT: return " WT";
913 	default: return "";
914 	}
915 }
916 
917 static u32
918 read_subslice_reg(const struct intel_engine_cs *engine,
919 		  int slice, int subslice, i915_reg_t reg)
920 {
921 	struct drm_i915_private *i915 = engine->i915;
922 	struct intel_uncore *uncore = engine->uncore;
923 	u32 mcr_mask, mcr_ss, mcr, old_mcr, val;
924 	enum forcewake_domains fw_domains;
925 
926 	if (INTEL_GEN(i915) >= 11) {
927 		mcr_mask = GEN11_MCR_SLICE_MASK | GEN11_MCR_SUBSLICE_MASK;
928 		mcr_ss = GEN11_MCR_SLICE(slice) | GEN11_MCR_SUBSLICE(subslice);
929 	} else {
930 		mcr_mask = GEN8_MCR_SLICE_MASK | GEN8_MCR_SUBSLICE_MASK;
931 		mcr_ss = GEN8_MCR_SLICE(slice) | GEN8_MCR_SUBSLICE(subslice);
932 	}
933 
934 	fw_domains = intel_uncore_forcewake_for_reg(uncore, reg,
935 						    FW_REG_READ);
936 	fw_domains |= intel_uncore_forcewake_for_reg(uncore,
937 						     GEN8_MCR_SELECTOR,
938 						     FW_REG_READ | FW_REG_WRITE);
939 
940 	spin_lock_irq(&uncore->lock);
941 	intel_uncore_forcewake_get__locked(uncore, fw_domains);
942 
943 	old_mcr = mcr = intel_uncore_read_fw(uncore, GEN8_MCR_SELECTOR);
944 
945 	mcr &= ~mcr_mask;
946 	mcr |= mcr_ss;
947 	intel_uncore_write_fw(uncore, GEN8_MCR_SELECTOR, mcr);
948 
949 	val = intel_uncore_read_fw(uncore, reg);
950 
951 	mcr &= ~mcr_mask;
952 	mcr |= old_mcr & mcr_mask;
953 
954 	intel_uncore_write_fw(uncore, GEN8_MCR_SELECTOR, mcr);
955 
956 	intel_uncore_forcewake_put__locked(uncore, fw_domains);
957 	spin_unlock_irq(&uncore->lock);
958 
959 	return val;
960 }
961 
962 /* NB: please notice the memset */
963 void intel_engine_get_instdone(const struct intel_engine_cs *engine,
964 			       struct intel_instdone *instdone)
965 {
966 	struct drm_i915_private *i915 = engine->i915;
967 	const struct sseu_dev_info *sseu = &RUNTIME_INFO(i915)->sseu;
968 	struct intel_uncore *uncore = engine->uncore;
969 	u32 mmio_base = engine->mmio_base;
970 	int slice;
971 	int subslice;
972 
973 	memset(instdone, 0, sizeof(*instdone));
974 
975 	switch (INTEL_GEN(i915)) {
976 	default:
977 		instdone->instdone =
978 			intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
979 
980 		if (engine->id != RCS0)
981 			break;
982 
983 		instdone->slice_common =
984 			intel_uncore_read(uncore, GEN7_SC_INSTDONE);
985 		for_each_instdone_slice_subslice(i915, sseu, slice, subslice) {
986 			instdone->sampler[slice][subslice] =
987 				read_subslice_reg(engine, slice, subslice,
988 						  GEN7_SAMPLER_INSTDONE);
989 			instdone->row[slice][subslice] =
990 				read_subslice_reg(engine, slice, subslice,
991 						  GEN7_ROW_INSTDONE);
992 		}
993 		break;
994 	case 7:
995 		instdone->instdone =
996 			intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
997 
998 		if (engine->id != RCS0)
999 			break;
1000 
1001 		instdone->slice_common =
1002 			intel_uncore_read(uncore, GEN7_SC_INSTDONE);
1003 		instdone->sampler[0][0] =
1004 			intel_uncore_read(uncore, GEN7_SAMPLER_INSTDONE);
1005 		instdone->row[0][0] =
1006 			intel_uncore_read(uncore, GEN7_ROW_INSTDONE);
1007 
1008 		break;
1009 	case 6:
1010 	case 5:
1011 	case 4:
1012 		instdone->instdone =
1013 			intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
1014 		if (engine->id == RCS0)
1015 			/* HACK: Using the wrong struct member */
1016 			instdone->slice_common =
1017 				intel_uncore_read(uncore, GEN4_INSTDONE1);
1018 		break;
1019 	case 3:
1020 	case 2:
1021 		instdone->instdone = intel_uncore_read(uncore, GEN2_INSTDONE);
1022 		break;
1023 	}
1024 }
1025 
1026 static bool ring_is_idle(struct intel_engine_cs *engine)
1027 {
1028 	bool idle = true;
1029 
1030 	if (I915_SELFTEST_ONLY(!engine->mmio_base))
1031 		return true;
1032 
1033 	if (!intel_engine_pm_get_if_awake(engine))
1034 		return true;
1035 
1036 	/* First check that no commands are left in the ring */
1037 	if ((ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR) !=
1038 	    (ENGINE_READ(engine, RING_TAIL) & TAIL_ADDR))
1039 		idle = false;
1040 
1041 	/* No bit for gen2, so assume the CS parser is idle */
1042 	if (INTEL_GEN(engine->i915) > 2 &&
1043 	    !(ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE))
1044 		idle = false;
1045 
1046 	intel_engine_pm_put(engine);
1047 
1048 	return idle;
1049 }
1050 
1051 void intel_engine_flush_submission(struct intel_engine_cs *engine)
1052 {
1053 	struct tasklet_struct *t = &engine->execlists.tasklet;
1054 
1055 	if (__tasklet_is_scheduled(t)) {
1056 		local_bh_disable();
1057 		if (tasklet_trylock(t)) {
1058 			/* Must wait for any GPU reset in progress. */
1059 			if (__tasklet_is_enabled(t))
1060 				t->func(t->data);
1061 			tasklet_unlock(t);
1062 		}
1063 		local_bh_enable();
1064 	}
1065 
1066 	/* Otherwise flush the tasklet if it was running on another cpu */
1067 	tasklet_unlock_wait(t);
1068 }
1069 
1070 /**
1071  * intel_engine_is_idle() - Report if the engine has finished process all work
1072  * @engine: the intel_engine_cs
1073  *
1074  * Return true if there are no requests pending, nothing left to be submitted
1075  * to hardware, and that the engine is idle.
1076  */
1077 bool intel_engine_is_idle(struct intel_engine_cs *engine)
1078 {
1079 	/* More white lies, if wedged, hw state is inconsistent */
1080 	if (intel_gt_is_wedged(engine->gt))
1081 		return true;
1082 
1083 	if (!intel_engine_pm_is_awake(engine))
1084 		return true;
1085 
1086 	/* Waiting to drain ELSP? */
1087 	if (execlists_active(&engine->execlists)) {
1088 		synchronize_hardirq(engine->i915->drm.pdev->irq);
1089 
1090 		intel_engine_flush_submission(engine);
1091 
1092 		if (execlists_active(&engine->execlists))
1093 			return false;
1094 	}
1095 
1096 	/* ELSP is empty, but there are ready requests? E.g. after reset */
1097 	if (!RB_EMPTY_ROOT(&engine->execlists.queue.rb_root))
1098 		return false;
1099 
1100 	/* Ring stopped? */
1101 	return ring_is_idle(engine);
1102 }
1103 
1104 bool intel_engines_are_idle(struct intel_gt *gt)
1105 {
1106 	struct intel_engine_cs *engine;
1107 	enum intel_engine_id id;
1108 
1109 	/*
1110 	 * If the driver is wedged, HW state may be very inconsistent and
1111 	 * report that it is still busy, even though we have stopped using it.
1112 	 */
1113 	if (intel_gt_is_wedged(gt))
1114 		return true;
1115 
1116 	/* Already parked (and passed an idleness test); must still be idle */
1117 	if (!READ_ONCE(gt->awake))
1118 		return true;
1119 
1120 	for_each_engine(engine, gt, id) {
1121 		if (!intel_engine_is_idle(engine))
1122 			return false;
1123 	}
1124 
1125 	return true;
1126 }
1127 
1128 void intel_engines_reset_default_submission(struct intel_gt *gt)
1129 {
1130 	struct intel_engine_cs *engine;
1131 	enum intel_engine_id id;
1132 
1133 	for_each_engine(engine, gt, id)
1134 		engine->set_default_submission(engine);
1135 }
1136 
1137 bool intel_engine_can_store_dword(struct intel_engine_cs *engine)
1138 {
1139 	switch (INTEL_GEN(engine->i915)) {
1140 	case 2:
1141 		return false; /* uses physical not virtual addresses */
1142 	case 3:
1143 		/* maybe only uses physical not virtual addresses */
1144 		return !(IS_I915G(engine->i915) || IS_I915GM(engine->i915));
1145 	case 4:
1146 		return !IS_I965G(engine->i915); /* who knows! */
1147 	case 6:
1148 		return engine->class != VIDEO_DECODE_CLASS; /* b0rked */
1149 	default:
1150 		return true;
1151 	}
1152 }
1153 
1154 static int print_sched_attr(struct drm_i915_private *i915,
1155 			    const struct i915_sched_attr *attr,
1156 			    char *buf, int x, int len)
1157 {
1158 	if (attr->priority == I915_PRIORITY_INVALID)
1159 		return x;
1160 
1161 	x += snprintf(buf + x, len - x,
1162 		      " prio=%d", attr->priority);
1163 
1164 	return x;
1165 }
1166 
1167 static void print_request(struct drm_printer *m,
1168 			  struct i915_request *rq,
1169 			  const char *prefix)
1170 {
1171 	const char *name = rq->fence.ops->get_timeline_name(&rq->fence);
1172 	char buf[80] = "";
1173 	int x = 0;
1174 
1175 	x = print_sched_attr(rq->i915, &rq->sched.attr, buf, x, sizeof(buf));
1176 
1177 	drm_printf(m, "%s %llx:%llx%s%s %s @ %dms: %s\n",
1178 		   prefix,
1179 		   rq->fence.context, rq->fence.seqno,
1180 		   i915_request_completed(rq) ? "!" :
1181 		   i915_request_started(rq) ? "*" :
1182 		   "",
1183 		   test_bit(DMA_FENCE_FLAG_SIGNALED_BIT,
1184 			    &rq->fence.flags) ? "+" :
1185 		   test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT,
1186 			    &rq->fence.flags) ? "-" :
1187 		   "",
1188 		   buf,
1189 		   jiffies_to_msecs(jiffies - rq->emitted_jiffies),
1190 		   name);
1191 }
1192 
1193 static void hexdump(struct drm_printer *m, const void *buf, size_t len)
1194 {
1195 	const size_t rowsize = 8 * sizeof(u32);
1196 	const void *prev = NULL;
1197 	bool skip = false;
1198 	size_t pos;
1199 
1200 	for (pos = 0; pos < len; pos += rowsize) {
1201 		char line[128];
1202 
1203 		if (prev && !memcmp(prev, buf + pos, rowsize)) {
1204 			if (!skip) {
1205 				drm_printf(m, "*\n");
1206 				skip = true;
1207 			}
1208 			continue;
1209 		}
1210 
1211 		WARN_ON_ONCE(hex_dump_to_buffer(buf + pos, len - pos,
1212 						rowsize, sizeof(u32),
1213 						line, sizeof(line),
1214 						false) >= sizeof(line));
1215 		drm_printf(m, "[%04zx] %s\n", pos, line);
1216 
1217 		prev = buf + pos;
1218 		skip = false;
1219 	}
1220 }
1221 
1222 static struct intel_timeline *get_timeline(struct i915_request *rq)
1223 {
1224 	struct intel_timeline *tl;
1225 
1226 	/*
1227 	 * Even though we are holding the engine->active.lock here, there
1228 	 * is no control over the submission queue per-se and we are
1229 	 * inspecting the active state at a random point in time, with an
1230 	 * unknown queue. Play safe and make sure the timeline remains valid.
1231 	 * (Only being used for pretty printing, one extra kref shouldn't
1232 	 * cause a camel stampede!)
1233 	 */
1234 	rcu_read_lock();
1235 	tl = rcu_dereference(rq->timeline);
1236 	if (!kref_get_unless_zero(&tl->kref))
1237 		tl = NULL;
1238 	rcu_read_unlock();
1239 
1240 	return tl;
1241 }
1242 
1243 static const char *repr_timer(const struct timer_list *t)
1244 {
1245 	if (!READ_ONCE(t->expires))
1246 		return "inactive";
1247 
1248 	if (timer_pending(t))
1249 		return "active";
1250 
1251 	return "expired";
1252 }
1253 
1254 static void intel_engine_print_registers(struct intel_engine_cs *engine,
1255 					 struct drm_printer *m)
1256 {
1257 	struct drm_i915_private *dev_priv = engine->i915;
1258 	struct intel_engine_execlists * const execlists = &engine->execlists;
1259 	u64 addr;
1260 
1261 	if (engine->id == RENDER_CLASS && IS_GEN_RANGE(dev_priv, 4, 7))
1262 		drm_printf(m, "\tCCID: 0x%08x\n", ENGINE_READ(engine, CCID));
1263 	drm_printf(m, "\tRING_START: 0x%08x\n",
1264 		   ENGINE_READ(engine, RING_START));
1265 	drm_printf(m, "\tRING_HEAD:  0x%08x\n",
1266 		   ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR);
1267 	drm_printf(m, "\tRING_TAIL:  0x%08x\n",
1268 		   ENGINE_READ(engine, RING_TAIL) & TAIL_ADDR);
1269 	drm_printf(m, "\tRING_CTL:   0x%08x%s\n",
1270 		   ENGINE_READ(engine, RING_CTL),
1271 		   ENGINE_READ(engine, RING_CTL) & (RING_WAIT | RING_WAIT_SEMAPHORE) ? " [waiting]" : "");
1272 	if (INTEL_GEN(engine->i915) > 2) {
1273 		drm_printf(m, "\tRING_MODE:  0x%08x%s\n",
1274 			   ENGINE_READ(engine, RING_MI_MODE),
1275 			   ENGINE_READ(engine, RING_MI_MODE) & (MODE_IDLE) ? " [idle]" : "");
1276 	}
1277 
1278 	if (INTEL_GEN(dev_priv) >= 6) {
1279 		drm_printf(m, "\tRING_IMR: %08x\n",
1280 			   ENGINE_READ(engine, RING_IMR));
1281 	}
1282 
1283 	addr = intel_engine_get_active_head(engine);
1284 	drm_printf(m, "\tACTHD:  0x%08x_%08x\n",
1285 		   upper_32_bits(addr), lower_32_bits(addr));
1286 	addr = intel_engine_get_last_batch_head(engine);
1287 	drm_printf(m, "\tBBADDR: 0x%08x_%08x\n",
1288 		   upper_32_bits(addr), lower_32_bits(addr));
1289 	if (INTEL_GEN(dev_priv) >= 8)
1290 		addr = ENGINE_READ64(engine, RING_DMA_FADD, RING_DMA_FADD_UDW);
1291 	else if (INTEL_GEN(dev_priv) >= 4)
1292 		addr = ENGINE_READ(engine, RING_DMA_FADD);
1293 	else
1294 		addr = ENGINE_READ(engine, DMA_FADD_I8XX);
1295 	drm_printf(m, "\tDMA_FADDR: 0x%08x_%08x\n",
1296 		   upper_32_bits(addr), lower_32_bits(addr));
1297 	if (INTEL_GEN(dev_priv) >= 4) {
1298 		drm_printf(m, "\tIPEIR: 0x%08x\n",
1299 			   ENGINE_READ(engine, RING_IPEIR));
1300 		drm_printf(m, "\tIPEHR: 0x%08x\n",
1301 			   ENGINE_READ(engine, RING_IPEHR));
1302 	} else {
1303 		drm_printf(m, "\tIPEIR: 0x%08x\n", ENGINE_READ(engine, IPEIR));
1304 		drm_printf(m, "\tIPEHR: 0x%08x\n", ENGINE_READ(engine, IPEHR));
1305 	}
1306 
1307 	if (HAS_EXECLISTS(dev_priv)) {
1308 		struct i915_request * const *port, *rq;
1309 		const u32 *hws =
1310 			&engine->status_page.addr[I915_HWS_CSB_BUF0_INDEX];
1311 		const u8 num_entries = execlists->csb_size;
1312 		unsigned int idx;
1313 		u8 read, write;
1314 
1315 		drm_printf(m, "\tExeclist tasklet queued? %s (%s), preempt? %s, timeslice? %s\n",
1316 			   yesno(test_bit(TASKLET_STATE_SCHED,
1317 					  &engine->execlists.tasklet.state)),
1318 			   enableddisabled(!atomic_read(&engine->execlists.tasklet.count)),
1319 			   repr_timer(&engine->execlists.preempt),
1320 			   repr_timer(&engine->execlists.timer));
1321 
1322 		read = execlists->csb_head;
1323 		write = READ_ONCE(*execlists->csb_write);
1324 
1325 		drm_printf(m, "\tExeclist status: 0x%08x %08x; CSB read:%d, write:%d, entries:%d\n",
1326 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_LO),
1327 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_HI),
1328 			   read, write, num_entries);
1329 
1330 		if (read >= num_entries)
1331 			read = 0;
1332 		if (write >= num_entries)
1333 			write = 0;
1334 		if (read > write)
1335 			write += num_entries;
1336 		while (read < write) {
1337 			idx = ++read % num_entries;
1338 			drm_printf(m, "\tExeclist CSB[%d]: 0x%08x, context: %d\n",
1339 				   idx, hws[idx * 2], hws[idx * 2 + 1]);
1340 		}
1341 
1342 		execlists_active_lock_bh(execlists);
1343 		rcu_read_lock();
1344 		for (port = execlists->active; (rq = *port); port++) {
1345 			char hdr[80];
1346 			int len;
1347 
1348 			len = snprintf(hdr, sizeof(hdr),
1349 				       "\t\tActive[%d]: ",
1350 				       (int)(port - execlists->active));
1351 			if (!i915_request_signaled(rq)) {
1352 				struct intel_timeline *tl = get_timeline(rq);
1353 
1354 				len += snprintf(hdr + len, sizeof(hdr) - len,
1355 						"ring:{start:%08x, hwsp:%08x, seqno:%08x}, ",
1356 						i915_ggtt_offset(rq->ring->vma),
1357 						tl ? tl->hwsp_offset : 0,
1358 						hwsp_seqno(rq));
1359 
1360 				if (tl)
1361 					intel_timeline_put(tl);
1362 			}
1363 			snprintf(hdr + len, sizeof(hdr) - len, "rq: ");
1364 			print_request(m, rq, hdr);
1365 		}
1366 		for (port = execlists->pending; (rq = *port); port++) {
1367 			struct intel_timeline *tl = get_timeline(rq);
1368 			char hdr[80];
1369 
1370 			snprintf(hdr, sizeof(hdr),
1371 				 "\t\tPending[%d] ring:{start:%08x, hwsp:%08x, seqno:%08x}, rq: ",
1372 				 (int)(port - execlists->pending),
1373 				 i915_ggtt_offset(rq->ring->vma),
1374 				 tl ? tl->hwsp_offset : 0,
1375 				 hwsp_seqno(rq));
1376 			print_request(m, rq, hdr);
1377 
1378 			if (tl)
1379 				intel_timeline_put(tl);
1380 		}
1381 		rcu_read_unlock();
1382 		execlists_active_unlock_bh(execlists);
1383 	} else if (INTEL_GEN(dev_priv) > 6) {
1384 		drm_printf(m, "\tPP_DIR_BASE: 0x%08x\n",
1385 			   ENGINE_READ(engine, RING_PP_DIR_BASE));
1386 		drm_printf(m, "\tPP_DIR_BASE_READ: 0x%08x\n",
1387 			   ENGINE_READ(engine, RING_PP_DIR_BASE_READ));
1388 		drm_printf(m, "\tPP_DIR_DCLV: 0x%08x\n",
1389 			   ENGINE_READ(engine, RING_PP_DIR_DCLV));
1390 	}
1391 }
1392 
1393 static void print_request_ring(struct drm_printer *m, struct i915_request *rq)
1394 {
1395 	void *ring;
1396 	int size;
1397 
1398 	drm_printf(m,
1399 		   "[head %04x, postfix %04x, tail %04x, batch 0x%08x_%08x]:\n",
1400 		   rq->head, rq->postfix, rq->tail,
1401 		   rq->batch ? upper_32_bits(rq->batch->node.start) : ~0u,
1402 		   rq->batch ? lower_32_bits(rq->batch->node.start) : ~0u);
1403 
1404 	size = rq->tail - rq->head;
1405 	if (rq->tail < rq->head)
1406 		size += rq->ring->size;
1407 
1408 	ring = kmalloc(size, GFP_ATOMIC);
1409 	if (ring) {
1410 		const void *vaddr = rq->ring->vaddr;
1411 		unsigned int head = rq->head;
1412 		unsigned int len = 0;
1413 
1414 		if (rq->tail < head) {
1415 			len = rq->ring->size - head;
1416 			memcpy(ring, vaddr + head, len);
1417 			head = 0;
1418 		}
1419 		memcpy(ring + len, vaddr + head, size - len);
1420 
1421 		hexdump(m, ring, size);
1422 		kfree(ring);
1423 	}
1424 }
1425 
1426 static unsigned long list_count(struct list_head *list)
1427 {
1428 	struct list_head *pos;
1429 	unsigned long count = 0;
1430 
1431 	list_for_each(pos, list)
1432 		count++;
1433 
1434 	return count;
1435 }
1436 
1437 void intel_engine_dump(struct intel_engine_cs *engine,
1438 		       struct drm_printer *m,
1439 		       const char *header, ...)
1440 {
1441 	struct i915_gpu_error * const error = &engine->i915->gpu_error;
1442 	struct i915_request *rq;
1443 	intel_wakeref_t wakeref;
1444 	unsigned long flags;
1445 
1446 	if (header) {
1447 		va_list ap;
1448 
1449 		va_start(ap, header);
1450 		drm_vprintf(m, header, &ap);
1451 		va_end(ap);
1452 	}
1453 
1454 	if (intel_gt_is_wedged(engine->gt))
1455 		drm_printf(m, "*** WEDGED ***\n");
1456 
1457 	drm_printf(m, "\tAwake? %d\n", atomic_read(&engine->wakeref.count));
1458 	drm_printf(m, "\tBarriers?: %s\n",
1459 		   yesno(!llist_empty(&engine->barrier_tasks)));
1460 	drm_printf(m, "\tLatency: %luus\n",
1461 		   ewma__engine_latency_read(&engine->latency));
1462 
1463 	rcu_read_lock();
1464 	rq = READ_ONCE(engine->heartbeat.systole);
1465 	if (rq)
1466 		drm_printf(m, "\tHeartbeat: %d ms ago\n",
1467 			   jiffies_to_msecs(jiffies - rq->emitted_jiffies));
1468 	rcu_read_unlock();
1469 	drm_printf(m, "\tReset count: %d (global %d)\n",
1470 		   i915_reset_engine_count(error, engine),
1471 		   i915_reset_count(error));
1472 
1473 	drm_printf(m, "\tRequests:\n");
1474 
1475 	spin_lock_irqsave(&engine->active.lock, flags);
1476 	rq = intel_engine_find_active_request(engine);
1477 	if (rq) {
1478 		struct intel_timeline *tl = get_timeline(rq);
1479 
1480 		print_request(m, rq, "\t\tactive ");
1481 
1482 		drm_printf(m, "\t\tring->start:  0x%08x\n",
1483 			   i915_ggtt_offset(rq->ring->vma));
1484 		drm_printf(m, "\t\tring->head:   0x%08x\n",
1485 			   rq->ring->head);
1486 		drm_printf(m, "\t\tring->tail:   0x%08x\n",
1487 			   rq->ring->tail);
1488 		drm_printf(m, "\t\tring->emit:   0x%08x\n",
1489 			   rq->ring->emit);
1490 		drm_printf(m, "\t\tring->space:  0x%08x\n",
1491 			   rq->ring->space);
1492 
1493 		if (tl) {
1494 			drm_printf(m, "\t\tring->hwsp:   0x%08x\n",
1495 				   tl->hwsp_offset);
1496 			intel_timeline_put(tl);
1497 		}
1498 
1499 		print_request_ring(m, rq);
1500 
1501 		if (rq->context->lrc_reg_state) {
1502 			drm_printf(m, "Logical Ring Context:\n");
1503 			hexdump(m, rq->context->lrc_reg_state, PAGE_SIZE);
1504 		}
1505 	}
1506 	drm_printf(m, "\tOn hold?: %lu\n", list_count(&engine->active.hold));
1507 	spin_unlock_irqrestore(&engine->active.lock, flags);
1508 
1509 	drm_printf(m, "\tMMIO base:  0x%08x\n", engine->mmio_base);
1510 	wakeref = intel_runtime_pm_get_if_in_use(engine->uncore->rpm);
1511 	if (wakeref) {
1512 		intel_engine_print_registers(engine, m);
1513 		intel_runtime_pm_put(engine->uncore->rpm, wakeref);
1514 	} else {
1515 		drm_printf(m, "\tDevice is asleep; skipping register dump\n");
1516 	}
1517 
1518 	intel_execlists_show_requests(engine, m, print_request, 8);
1519 
1520 	drm_printf(m, "HWSP:\n");
1521 	hexdump(m, engine->status_page.addr, PAGE_SIZE);
1522 
1523 	drm_printf(m, "Idle? %s\n", yesno(intel_engine_is_idle(engine)));
1524 
1525 	intel_engine_print_breadcrumbs(engine, m);
1526 }
1527 
1528 /**
1529  * intel_enable_engine_stats() - Enable engine busy tracking on engine
1530  * @engine: engine to enable stats collection
1531  *
1532  * Start collecting the engine busyness data for @engine.
1533  *
1534  * Returns 0 on success or a negative error code.
1535  */
1536 int intel_enable_engine_stats(struct intel_engine_cs *engine)
1537 {
1538 	struct intel_engine_execlists *execlists = &engine->execlists;
1539 	unsigned long flags;
1540 	int err = 0;
1541 
1542 	if (!intel_engine_supports_stats(engine))
1543 		return -ENODEV;
1544 
1545 	execlists_active_lock_bh(execlists);
1546 	write_seqlock_irqsave(&engine->stats.lock, flags);
1547 
1548 	if (unlikely(engine->stats.enabled == ~0)) {
1549 		err = -EBUSY;
1550 		goto unlock;
1551 	}
1552 
1553 	if (engine->stats.enabled++ == 0) {
1554 		struct i915_request * const *port;
1555 		struct i915_request *rq;
1556 
1557 		engine->stats.enabled_at = ktime_get();
1558 
1559 		/* XXX submission method oblivious? */
1560 		for (port = execlists->active; (rq = *port); port++)
1561 			engine->stats.active++;
1562 
1563 		for (port = execlists->pending; (rq = *port); port++) {
1564 			/* Exclude any contexts already counted in active */
1565 			if (!intel_context_inflight_count(rq->context))
1566 				engine->stats.active++;
1567 		}
1568 
1569 		if (engine->stats.active)
1570 			engine->stats.start = engine->stats.enabled_at;
1571 	}
1572 
1573 unlock:
1574 	write_sequnlock_irqrestore(&engine->stats.lock, flags);
1575 	execlists_active_unlock_bh(execlists);
1576 
1577 	return err;
1578 }
1579 
1580 static ktime_t __intel_engine_get_busy_time(struct intel_engine_cs *engine)
1581 {
1582 	ktime_t total = engine->stats.total;
1583 
1584 	/*
1585 	 * If the engine is executing something at the moment
1586 	 * add it to the total.
1587 	 */
1588 	if (engine->stats.active)
1589 		total = ktime_add(total,
1590 				  ktime_sub(ktime_get(), engine->stats.start));
1591 
1592 	return total;
1593 }
1594 
1595 /**
1596  * intel_engine_get_busy_time() - Return current accumulated engine busyness
1597  * @engine: engine to report on
1598  *
1599  * Returns accumulated time @engine was busy since engine stats were enabled.
1600  */
1601 ktime_t intel_engine_get_busy_time(struct intel_engine_cs *engine)
1602 {
1603 	unsigned int seq;
1604 	ktime_t total;
1605 
1606 	do {
1607 		seq = read_seqbegin(&engine->stats.lock);
1608 		total = __intel_engine_get_busy_time(engine);
1609 	} while (read_seqretry(&engine->stats.lock, seq));
1610 
1611 	return total;
1612 }
1613 
1614 /**
1615  * intel_disable_engine_stats() - Disable engine busy tracking on engine
1616  * @engine: engine to disable stats collection
1617  *
1618  * Stops collecting the engine busyness data for @engine.
1619  */
1620 void intel_disable_engine_stats(struct intel_engine_cs *engine)
1621 {
1622 	unsigned long flags;
1623 
1624 	if (!intel_engine_supports_stats(engine))
1625 		return;
1626 
1627 	write_seqlock_irqsave(&engine->stats.lock, flags);
1628 	WARN_ON_ONCE(engine->stats.enabled == 0);
1629 	if (--engine->stats.enabled == 0) {
1630 		engine->stats.total = __intel_engine_get_busy_time(engine);
1631 		engine->stats.active = 0;
1632 	}
1633 	write_sequnlock_irqrestore(&engine->stats.lock, flags);
1634 }
1635 
1636 static bool match_ring(struct i915_request *rq)
1637 {
1638 	u32 ring = ENGINE_READ(rq->engine, RING_START);
1639 
1640 	return ring == i915_ggtt_offset(rq->ring->vma);
1641 }
1642 
1643 struct i915_request *
1644 intel_engine_find_active_request(struct intel_engine_cs *engine)
1645 {
1646 	struct i915_request *request, *active = NULL;
1647 
1648 	/*
1649 	 * We are called by the error capture, reset and to dump engine
1650 	 * state at random points in time. In particular, note that neither is
1651 	 * crucially ordered with an interrupt. After a hang, the GPU is dead
1652 	 * and we assume that no more writes can happen (we waited long enough
1653 	 * for all writes that were in transaction to be flushed) - adding an
1654 	 * extra delay for a recent interrupt is pointless. Hence, we do
1655 	 * not need an engine->irq_seqno_barrier() before the seqno reads.
1656 	 * At all other times, we must assume the GPU is still running, but
1657 	 * we only care about the snapshot of this moment.
1658 	 */
1659 	lockdep_assert_held(&engine->active.lock);
1660 	list_for_each_entry(request, &engine->active.requests, sched.link) {
1661 		if (i915_request_completed(request))
1662 			continue;
1663 
1664 		if (!i915_request_started(request))
1665 			continue;
1666 
1667 		/* More than one preemptible request may match! */
1668 		if (!match_ring(request))
1669 			continue;
1670 
1671 		active = request;
1672 		break;
1673 	}
1674 
1675 	return active;
1676 }
1677 
1678 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1679 #include "mock_engine.c"
1680 #include "selftest_engine.c"
1681 #include "selftest_engine_cs.c"
1682 #endif
1683