1 // SPDX-License-Identifier: MIT
2 /*
3  * Copyright © 2016 Intel Corporation
4  */
5 
6 #include <drm/drm_print.h>
7 
8 #include "gem/i915_gem_context.h"
9 
10 #include "i915_drv.h"
11 
12 #include "intel_breadcrumbs.h"
13 #include "intel_context.h"
14 #include "intel_engine.h"
15 #include "intel_engine_pm.h"
16 #include "intel_engine_user.h"
17 #include "intel_execlists_submission.h"
18 #include "intel_gt.h"
19 #include "intel_gt_requests.h"
20 #include "intel_gt_pm.h"
21 #include "intel_lrc_reg.h"
22 #include "intel_reset.h"
23 #include "intel_ring.h"
24 #include "uc/intel_guc_submission.h"
25 
26 /* Haswell does have the CXT_SIZE register however it does not appear to be
27  * valid. Now, docs explain in dwords what is in the context object. The full
28  * size is 70720 bytes, however, the power context and execlist context will
29  * never be saved (power context is stored elsewhere, and execlists don't work
30  * on HSW) - so the final size, including the extra state required for the
31  * Resource Streamer, is 66944 bytes, which rounds to 17 pages.
32  */
33 #define HSW_CXT_TOTAL_SIZE		(17 * PAGE_SIZE)
34 
35 #define DEFAULT_LR_CONTEXT_RENDER_SIZE	(22 * PAGE_SIZE)
36 #define GEN8_LR_CONTEXT_RENDER_SIZE	(20 * PAGE_SIZE)
37 #define GEN9_LR_CONTEXT_RENDER_SIZE	(22 * PAGE_SIZE)
38 #define GEN11_LR_CONTEXT_RENDER_SIZE	(14 * PAGE_SIZE)
39 
40 #define GEN8_LR_CONTEXT_OTHER_SIZE	( 2 * PAGE_SIZE)
41 
42 #define MAX_MMIO_BASES 3
43 struct engine_info {
44 	u8 class;
45 	u8 instance;
46 	/* mmio bases table *must* be sorted in reverse graphics_ver order */
47 	struct engine_mmio_base {
48 		u32 graphics_ver : 8;
49 		u32 base : 24;
50 	} mmio_bases[MAX_MMIO_BASES];
51 };
52 
53 static const struct engine_info intel_engines[] = {
54 	[RCS0] = {
55 		.class = RENDER_CLASS,
56 		.instance = 0,
57 		.mmio_bases = {
58 			{ .graphics_ver = 1, .base = RENDER_RING_BASE }
59 		},
60 	},
61 	[BCS0] = {
62 		.class = COPY_ENGINE_CLASS,
63 		.instance = 0,
64 		.mmio_bases = {
65 			{ .graphics_ver = 6, .base = BLT_RING_BASE }
66 		},
67 	},
68 	[VCS0] = {
69 		.class = VIDEO_DECODE_CLASS,
70 		.instance = 0,
71 		.mmio_bases = {
72 			{ .graphics_ver = 11, .base = GEN11_BSD_RING_BASE },
73 			{ .graphics_ver = 6, .base = GEN6_BSD_RING_BASE },
74 			{ .graphics_ver = 4, .base = BSD_RING_BASE }
75 		},
76 	},
77 	[VCS1] = {
78 		.class = VIDEO_DECODE_CLASS,
79 		.instance = 1,
80 		.mmio_bases = {
81 			{ .graphics_ver = 11, .base = GEN11_BSD2_RING_BASE },
82 			{ .graphics_ver = 8, .base = GEN8_BSD2_RING_BASE }
83 		},
84 	},
85 	[VCS2] = {
86 		.class = VIDEO_DECODE_CLASS,
87 		.instance = 2,
88 		.mmio_bases = {
89 			{ .graphics_ver = 11, .base = GEN11_BSD3_RING_BASE }
90 		},
91 	},
92 	[VCS3] = {
93 		.class = VIDEO_DECODE_CLASS,
94 		.instance = 3,
95 		.mmio_bases = {
96 			{ .graphics_ver = 11, .base = GEN11_BSD4_RING_BASE }
97 		},
98 	},
99 	[VCS4] = {
100 		.class = VIDEO_DECODE_CLASS,
101 		.instance = 4,
102 		.mmio_bases = {
103 			{ .graphics_ver = 12, .base = XEHP_BSD5_RING_BASE }
104 		},
105 	},
106 	[VCS5] = {
107 		.class = VIDEO_DECODE_CLASS,
108 		.instance = 5,
109 		.mmio_bases = {
110 			{ .graphics_ver = 12, .base = XEHP_BSD6_RING_BASE }
111 		},
112 	},
113 	[VCS6] = {
114 		.class = VIDEO_DECODE_CLASS,
115 		.instance = 6,
116 		.mmio_bases = {
117 			{ .graphics_ver = 12, .base = XEHP_BSD7_RING_BASE }
118 		},
119 	},
120 	[VCS7] = {
121 		.class = VIDEO_DECODE_CLASS,
122 		.instance = 7,
123 		.mmio_bases = {
124 			{ .graphics_ver = 12, .base = XEHP_BSD8_RING_BASE }
125 		},
126 	},
127 	[VECS0] = {
128 		.class = VIDEO_ENHANCEMENT_CLASS,
129 		.instance = 0,
130 		.mmio_bases = {
131 			{ .graphics_ver = 11, .base = GEN11_VEBOX_RING_BASE },
132 			{ .graphics_ver = 7, .base = VEBOX_RING_BASE }
133 		},
134 	},
135 	[VECS1] = {
136 		.class = VIDEO_ENHANCEMENT_CLASS,
137 		.instance = 1,
138 		.mmio_bases = {
139 			{ .graphics_ver = 11, .base = GEN11_VEBOX2_RING_BASE }
140 		},
141 	},
142 	[VECS2] = {
143 		.class = VIDEO_ENHANCEMENT_CLASS,
144 		.instance = 2,
145 		.mmio_bases = {
146 			{ .graphics_ver = 12, .base = XEHP_VEBOX3_RING_BASE }
147 		},
148 	},
149 	[VECS3] = {
150 		.class = VIDEO_ENHANCEMENT_CLASS,
151 		.instance = 3,
152 		.mmio_bases = {
153 			{ .graphics_ver = 12, .base = XEHP_VEBOX4_RING_BASE }
154 		},
155 	},
156 };
157 
158 /**
159  * intel_engine_context_size() - return the size of the context for an engine
160  * @gt: the gt
161  * @class: engine class
162  *
163  * Each engine class may require a different amount of space for a context
164  * image.
165  *
166  * Return: size (in bytes) of an engine class specific context image
167  *
168  * Note: this size includes the HWSP, which is part of the context image
169  * in LRC mode, but does not include the "shared data page" used with
170  * GuC submission. The caller should account for this if using the GuC.
171  */
172 u32 intel_engine_context_size(struct intel_gt *gt, u8 class)
173 {
174 	struct intel_uncore *uncore = gt->uncore;
175 	u32 cxt_size;
176 
177 	BUILD_BUG_ON(I915_GTT_PAGE_SIZE != PAGE_SIZE);
178 
179 	switch (class) {
180 	case RENDER_CLASS:
181 		switch (GRAPHICS_VER(gt->i915)) {
182 		default:
183 			MISSING_CASE(GRAPHICS_VER(gt->i915));
184 			return DEFAULT_LR_CONTEXT_RENDER_SIZE;
185 		case 12:
186 		case 11:
187 			return GEN11_LR_CONTEXT_RENDER_SIZE;
188 		case 9:
189 			return GEN9_LR_CONTEXT_RENDER_SIZE;
190 		case 8:
191 			return GEN8_LR_CONTEXT_RENDER_SIZE;
192 		case 7:
193 			if (IS_HASWELL(gt->i915))
194 				return HSW_CXT_TOTAL_SIZE;
195 
196 			cxt_size = intel_uncore_read(uncore, GEN7_CXT_SIZE);
197 			return round_up(GEN7_CXT_TOTAL_SIZE(cxt_size) * 64,
198 					PAGE_SIZE);
199 		case 6:
200 			cxt_size = intel_uncore_read(uncore, CXT_SIZE);
201 			return round_up(GEN6_CXT_TOTAL_SIZE(cxt_size) * 64,
202 					PAGE_SIZE);
203 		case 5:
204 		case 4:
205 			/*
206 			 * There is a discrepancy here between the size reported
207 			 * by the register and the size of the context layout
208 			 * in the docs. Both are described as authorative!
209 			 *
210 			 * The discrepancy is on the order of a few cachelines,
211 			 * but the total is under one page (4k), which is our
212 			 * minimum allocation anyway so it should all come
213 			 * out in the wash.
214 			 */
215 			cxt_size = intel_uncore_read(uncore, CXT_SIZE) + 1;
216 			drm_dbg(&gt->i915->drm,
217 				"graphics_ver = %d CXT_SIZE = %d bytes [0x%08x]\n",
218 				GRAPHICS_VER(gt->i915), cxt_size * 64,
219 				cxt_size - 1);
220 			return round_up(cxt_size * 64, PAGE_SIZE);
221 		case 3:
222 		case 2:
223 		/* For the special day when i810 gets merged. */
224 		case 1:
225 			return 0;
226 		}
227 		break;
228 	default:
229 		MISSING_CASE(class);
230 		fallthrough;
231 	case VIDEO_DECODE_CLASS:
232 	case VIDEO_ENHANCEMENT_CLASS:
233 	case COPY_ENGINE_CLASS:
234 		if (GRAPHICS_VER(gt->i915) < 8)
235 			return 0;
236 		return GEN8_LR_CONTEXT_OTHER_SIZE;
237 	}
238 }
239 
240 static u32 __engine_mmio_base(struct drm_i915_private *i915,
241 			      const struct engine_mmio_base *bases)
242 {
243 	int i;
244 
245 	for (i = 0; i < MAX_MMIO_BASES; i++)
246 		if (GRAPHICS_VER(i915) >= bases[i].graphics_ver)
247 			break;
248 
249 	GEM_BUG_ON(i == MAX_MMIO_BASES);
250 	GEM_BUG_ON(!bases[i].base);
251 
252 	return bases[i].base;
253 }
254 
255 static void __sprint_engine_name(struct intel_engine_cs *engine)
256 {
257 	/*
258 	 * Before we know what the uABI name for this engine will be,
259 	 * we still would like to keep track of this engine in the debug logs.
260 	 * We throw in a ' here as a reminder that this isn't its final name.
261 	 */
262 	GEM_WARN_ON(snprintf(engine->name, sizeof(engine->name), "%s'%u",
263 			     intel_engine_class_repr(engine->class),
264 			     engine->instance) >= sizeof(engine->name));
265 }
266 
267 void intel_engine_set_hwsp_writemask(struct intel_engine_cs *engine, u32 mask)
268 {
269 	/*
270 	 * Though they added more rings on g4x/ilk, they did not add
271 	 * per-engine HWSTAM until gen6.
272 	 */
273 	if (GRAPHICS_VER(engine->i915) < 6 && engine->class != RENDER_CLASS)
274 		return;
275 
276 	if (GRAPHICS_VER(engine->i915) >= 3)
277 		ENGINE_WRITE(engine, RING_HWSTAM, mask);
278 	else
279 		ENGINE_WRITE16(engine, RING_HWSTAM, mask);
280 }
281 
282 static void intel_engine_sanitize_mmio(struct intel_engine_cs *engine)
283 {
284 	/* Mask off all writes into the unknown HWSP */
285 	intel_engine_set_hwsp_writemask(engine, ~0u);
286 }
287 
288 static void nop_irq_handler(struct intel_engine_cs *engine, u16 iir)
289 {
290 	GEM_DEBUG_WARN_ON(iir);
291 }
292 
293 static int intel_engine_setup(struct intel_gt *gt, enum intel_engine_id id,
294 			      u8 logical_instance)
295 {
296 	const struct engine_info *info = &intel_engines[id];
297 	struct drm_i915_private *i915 = gt->i915;
298 	struct intel_engine_cs *engine;
299 	u8 guc_class;
300 
301 	BUILD_BUG_ON(MAX_ENGINE_CLASS >= BIT(GEN11_ENGINE_CLASS_WIDTH));
302 	BUILD_BUG_ON(MAX_ENGINE_INSTANCE >= BIT(GEN11_ENGINE_INSTANCE_WIDTH));
303 	BUILD_BUG_ON(I915_MAX_VCS > (MAX_ENGINE_INSTANCE + 1));
304 	BUILD_BUG_ON(I915_MAX_VECS > (MAX_ENGINE_INSTANCE + 1));
305 
306 	if (GEM_DEBUG_WARN_ON(id >= ARRAY_SIZE(gt->engine)))
307 		return -EINVAL;
308 
309 	if (GEM_DEBUG_WARN_ON(info->class > MAX_ENGINE_CLASS))
310 		return -EINVAL;
311 
312 	if (GEM_DEBUG_WARN_ON(info->instance > MAX_ENGINE_INSTANCE))
313 		return -EINVAL;
314 
315 	if (GEM_DEBUG_WARN_ON(gt->engine_class[info->class][info->instance]))
316 		return -EINVAL;
317 
318 	engine = kzalloc(sizeof(*engine), GFP_KERNEL);
319 	if (!engine)
320 		return -ENOMEM;
321 
322 	BUILD_BUG_ON(BITS_PER_TYPE(engine->mask) < I915_NUM_ENGINES);
323 
324 	INIT_LIST_HEAD(&engine->pinned_contexts_list);
325 	engine->id = id;
326 	engine->legacy_idx = INVALID_ENGINE;
327 	engine->mask = BIT(id);
328 	if (GRAPHICS_VER(gt->i915) >= 11) {
329 		static const u32 engine_reset_domains[] = {
330 			[RCS0]  = GEN11_GRDOM_RENDER,
331 			[BCS0]  = GEN11_GRDOM_BLT,
332 			[VCS0]  = GEN11_GRDOM_MEDIA,
333 			[VCS1]  = GEN11_GRDOM_MEDIA2,
334 			[VCS2]  = GEN11_GRDOM_MEDIA3,
335 			[VCS3]  = GEN11_GRDOM_MEDIA4,
336 			[VCS4]  = GEN11_GRDOM_MEDIA5,
337 			[VCS5]  = GEN11_GRDOM_MEDIA6,
338 			[VCS6]  = GEN11_GRDOM_MEDIA7,
339 			[VCS7]  = GEN11_GRDOM_MEDIA8,
340 			[VECS0] = GEN11_GRDOM_VECS,
341 			[VECS1] = GEN11_GRDOM_VECS2,
342 			[VECS2] = GEN11_GRDOM_VECS3,
343 			[VECS3] = GEN11_GRDOM_VECS4,
344 		};
345 		GEM_BUG_ON(id >= ARRAY_SIZE(engine_reset_domains) ||
346 			   !engine_reset_domains[id]);
347 		engine->reset_domain = engine_reset_domains[id];
348 	} else {
349 		static const u32 engine_reset_domains[] = {
350 			[RCS0]  = GEN6_GRDOM_RENDER,
351 			[BCS0]  = GEN6_GRDOM_BLT,
352 			[VCS0]  = GEN6_GRDOM_MEDIA,
353 			[VCS1]  = GEN8_GRDOM_MEDIA2,
354 			[VECS0] = GEN6_GRDOM_VECS,
355 		};
356 		GEM_BUG_ON(id >= ARRAY_SIZE(engine_reset_domains) ||
357 			   !engine_reset_domains[id]);
358 		engine->reset_domain = engine_reset_domains[id];
359 	}
360 	engine->i915 = i915;
361 	engine->gt = gt;
362 	engine->uncore = gt->uncore;
363 	guc_class = engine_class_to_guc_class(info->class);
364 	engine->guc_id = MAKE_GUC_ID(guc_class, info->instance);
365 	engine->mmio_base = __engine_mmio_base(i915, info->mmio_bases);
366 
367 	engine->irq_handler = nop_irq_handler;
368 
369 	engine->class = info->class;
370 	engine->instance = info->instance;
371 	engine->logical_mask = BIT(logical_instance);
372 	__sprint_engine_name(engine);
373 
374 	engine->props.heartbeat_interval_ms =
375 		CONFIG_DRM_I915_HEARTBEAT_INTERVAL;
376 	engine->props.max_busywait_duration_ns =
377 		CONFIG_DRM_I915_MAX_REQUEST_BUSYWAIT;
378 	engine->props.preempt_timeout_ms =
379 		CONFIG_DRM_I915_PREEMPT_TIMEOUT;
380 	engine->props.stop_timeout_ms =
381 		CONFIG_DRM_I915_STOP_TIMEOUT;
382 	engine->props.timeslice_duration_ms =
383 		CONFIG_DRM_I915_TIMESLICE_DURATION;
384 
385 	/* Override to uninterruptible for OpenCL workloads. */
386 	if (GRAPHICS_VER(i915) == 12 && engine->class == RENDER_CLASS)
387 		engine->props.preempt_timeout_ms = 0;
388 
389 	engine->defaults = engine->props; /* never to change again */
390 
391 	engine->context_size = intel_engine_context_size(gt, engine->class);
392 	if (WARN_ON(engine->context_size > BIT(20)))
393 		engine->context_size = 0;
394 	if (engine->context_size)
395 		DRIVER_CAPS(i915)->has_logical_contexts = true;
396 
397 	ewma__engine_latency_init(&engine->latency);
398 	seqcount_init(&engine->stats.execlists.lock);
399 
400 	ATOMIC_INIT_NOTIFIER_HEAD(&engine->context_status_notifier);
401 
402 	/* Scrub mmio state on takeover */
403 	intel_engine_sanitize_mmio(engine);
404 
405 	gt->engine_class[info->class][info->instance] = engine;
406 	gt->engine[id] = engine;
407 
408 	return 0;
409 }
410 
411 static void __setup_engine_capabilities(struct intel_engine_cs *engine)
412 {
413 	struct drm_i915_private *i915 = engine->i915;
414 
415 	if (engine->class == VIDEO_DECODE_CLASS) {
416 		/*
417 		 * HEVC support is present on first engine instance
418 		 * before Gen11 and on all instances afterwards.
419 		 */
420 		if (GRAPHICS_VER(i915) >= 11 ||
421 		    (GRAPHICS_VER(i915) >= 9 && engine->instance == 0))
422 			engine->uabi_capabilities |=
423 				I915_VIDEO_CLASS_CAPABILITY_HEVC;
424 
425 		/*
426 		 * SFC block is present only on even logical engine
427 		 * instances.
428 		 */
429 		if ((GRAPHICS_VER(i915) >= 11 &&
430 		     (engine->gt->info.vdbox_sfc_access &
431 		      BIT(engine->instance))) ||
432 		    (GRAPHICS_VER(i915) >= 9 && engine->instance == 0))
433 			engine->uabi_capabilities |=
434 				I915_VIDEO_AND_ENHANCE_CLASS_CAPABILITY_SFC;
435 	} else if (engine->class == VIDEO_ENHANCEMENT_CLASS) {
436 		if (GRAPHICS_VER(i915) >= 9 &&
437 		    engine->gt->info.sfc_mask & BIT(engine->instance))
438 			engine->uabi_capabilities |=
439 				I915_VIDEO_AND_ENHANCE_CLASS_CAPABILITY_SFC;
440 	}
441 }
442 
443 static void intel_setup_engine_capabilities(struct intel_gt *gt)
444 {
445 	struct intel_engine_cs *engine;
446 	enum intel_engine_id id;
447 
448 	for_each_engine(engine, gt, id)
449 		__setup_engine_capabilities(engine);
450 }
451 
452 /**
453  * intel_engines_release() - free the resources allocated for Command Streamers
454  * @gt: pointer to struct intel_gt
455  */
456 void intel_engines_release(struct intel_gt *gt)
457 {
458 	struct intel_engine_cs *engine;
459 	enum intel_engine_id id;
460 
461 	/*
462 	 * Before we release the resources held by engine, we must be certain
463 	 * that the HW is no longer accessing them -- having the GPU scribble
464 	 * to or read from a page being used for something else causes no end
465 	 * of fun.
466 	 *
467 	 * The GPU should be reset by this point, but assume the worst just
468 	 * in case we aborted before completely initialising the engines.
469 	 */
470 	GEM_BUG_ON(intel_gt_pm_is_awake(gt));
471 	if (!INTEL_INFO(gt->i915)->gpu_reset_clobbers_display)
472 		__intel_gt_reset(gt, ALL_ENGINES);
473 
474 	/* Decouple the backend; but keep the layout for late GPU resets */
475 	for_each_engine(engine, gt, id) {
476 		if (!engine->release)
477 			continue;
478 
479 		intel_wakeref_wait_for_idle(&engine->wakeref);
480 		GEM_BUG_ON(intel_engine_pm_is_awake(engine));
481 
482 		engine->release(engine);
483 		engine->release = NULL;
484 
485 		memset(&engine->reset, 0, sizeof(engine->reset));
486 	}
487 }
488 
489 void intel_engine_free_request_pool(struct intel_engine_cs *engine)
490 {
491 	if (!engine->request_pool)
492 		return;
493 
494 	kmem_cache_free(i915_request_slab_cache(), engine->request_pool);
495 }
496 
497 void intel_engines_free(struct intel_gt *gt)
498 {
499 	struct intel_engine_cs *engine;
500 	enum intel_engine_id id;
501 
502 	/* Free the requests! dma-resv keeps fences around for an eternity */
503 	rcu_barrier();
504 
505 	for_each_engine(engine, gt, id) {
506 		intel_engine_free_request_pool(engine);
507 		kfree(engine);
508 		gt->engine[id] = NULL;
509 	}
510 }
511 
512 static
513 bool gen11_vdbox_has_sfc(struct intel_gt *gt,
514 			 unsigned int physical_vdbox,
515 			 unsigned int logical_vdbox, u16 vdbox_mask)
516 {
517 	struct drm_i915_private *i915 = gt->i915;
518 
519 	/*
520 	 * In Gen11, only even numbered logical VDBOXes are hooked
521 	 * up to an SFC (Scaler & Format Converter) unit.
522 	 * In Gen12, Even numbered physical instance always are connected
523 	 * to an SFC. Odd numbered physical instances have SFC only if
524 	 * previous even instance is fused off.
525 	 *
526 	 * Starting with Xe_HP, there's also a dedicated SFC_ENABLE field
527 	 * in the fuse register that tells us whether a specific SFC is present.
528 	 */
529 	if ((gt->info.sfc_mask & BIT(physical_vdbox / 2)) == 0)
530 		return false;
531 	else if (GRAPHICS_VER(i915) == 12)
532 		return (physical_vdbox % 2 == 0) ||
533 			!(BIT(physical_vdbox - 1) & vdbox_mask);
534 	else if (GRAPHICS_VER(i915) == 11)
535 		return logical_vdbox % 2 == 0;
536 
537 	MISSING_CASE(GRAPHICS_VER(i915));
538 	return false;
539 }
540 
541 /*
542  * Determine which engines are fused off in our particular hardware.
543  * Note that we have a catch-22 situation where we need to be able to access
544  * the blitter forcewake domain to read the engine fuses, but at the same time
545  * we need to know which engines are available on the system to know which
546  * forcewake domains are present. We solve this by intializing the forcewake
547  * domains based on the full engine mask in the platform capabilities before
548  * calling this function and pruning the domains for fused-off engines
549  * afterwards.
550  */
551 static intel_engine_mask_t init_engine_mask(struct intel_gt *gt)
552 {
553 	struct drm_i915_private *i915 = gt->i915;
554 	struct intel_gt_info *info = &gt->info;
555 	struct intel_uncore *uncore = gt->uncore;
556 	unsigned int logical_vdbox = 0;
557 	unsigned int i;
558 	u32 media_fuse, fuse1;
559 	u16 vdbox_mask;
560 	u16 vebox_mask;
561 
562 	info->engine_mask = INTEL_INFO(i915)->platform_engine_mask;
563 
564 	if (GRAPHICS_VER(i915) < 11)
565 		return info->engine_mask;
566 
567 	/*
568 	 * On newer platforms the fusing register is called 'enable' and has
569 	 * enable semantics, while on older platforms it is called 'disable'
570 	 * and bits have disable semantices.
571 	 */
572 	media_fuse = intel_uncore_read(uncore, GEN11_GT_VEBOX_VDBOX_DISABLE);
573 	if (GRAPHICS_VER_FULL(i915) < IP_VER(12, 50))
574 		media_fuse = ~media_fuse;
575 
576 	vdbox_mask = media_fuse & GEN11_GT_VDBOX_DISABLE_MASK;
577 	vebox_mask = (media_fuse & GEN11_GT_VEBOX_DISABLE_MASK) >>
578 		      GEN11_GT_VEBOX_DISABLE_SHIFT;
579 
580 	if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 50)) {
581 		fuse1 = intel_uncore_read(uncore, HSW_PAVP_FUSE1);
582 		gt->info.sfc_mask = REG_FIELD_GET(XEHP_SFC_ENABLE_MASK, fuse1);
583 	} else {
584 		gt->info.sfc_mask = ~0;
585 	}
586 
587 	for (i = 0; i < I915_MAX_VCS; i++) {
588 		if (!HAS_ENGINE(gt, _VCS(i))) {
589 			vdbox_mask &= ~BIT(i);
590 			continue;
591 		}
592 
593 		if (!(BIT(i) & vdbox_mask)) {
594 			info->engine_mask &= ~BIT(_VCS(i));
595 			drm_dbg(&i915->drm, "vcs%u fused off\n", i);
596 			continue;
597 		}
598 
599 		if (gen11_vdbox_has_sfc(gt, i, logical_vdbox, vdbox_mask))
600 			gt->info.vdbox_sfc_access |= BIT(i);
601 		logical_vdbox++;
602 	}
603 	drm_dbg(&i915->drm, "vdbox enable: %04x, instances: %04lx\n",
604 		vdbox_mask, VDBOX_MASK(gt));
605 	GEM_BUG_ON(vdbox_mask != VDBOX_MASK(gt));
606 
607 	for (i = 0; i < I915_MAX_VECS; i++) {
608 		if (!HAS_ENGINE(gt, _VECS(i))) {
609 			vebox_mask &= ~BIT(i);
610 			continue;
611 		}
612 
613 		if (!(BIT(i) & vebox_mask)) {
614 			info->engine_mask &= ~BIT(_VECS(i));
615 			drm_dbg(&i915->drm, "vecs%u fused off\n", i);
616 		}
617 	}
618 	drm_dbg(&i915->drm, "vebox enable: %04x, instances: %04lx\n",
619 		vebox_mask, VEBOX_MASK(gt));
620 	GEM_BUG_ON(vebox_mask != VEBOX_MASK(gt));
621 
622 	return info->engine_mask;
623 }
624 
625 static void populate_logical_ids(struct intel_gt *gt, u8 *logical_ids,
626 				 u8 class, const u8 *map, u8 num_instances)
627 {
628 	int i, j;
629 	u8 current_logical_id = 0;
630 
631 	for (j = 0; j < num_instances; ++j) {
632 		for (i = 0; i < ARRAY_SIZE(intel_engines); ++i) {
633 			if (!HAS_ENGINE(gt, i) ||
634 			    intel_engines[i].class != class)
635 				continue;
636 
637 			if (intel_engines[i].instance == map[j]) {
638 				logical_ids[intel_engines[i].instance] =
639 					current_logical_id++;
640 				break;
641 			}
642 		}
643 	}
644 }
645 
646 static void setup_logical_ids(struct intel_gt *gt, u8 *logical_ids, u8 class)
647 {
648 	int i;
649 	u8 map[MAX_ENGINE_INSTANCE + 1];
650 
651 	for (i = 0; i < MAX_ENGINE_INSTANCE + 1; ++i)
652 		map[i] = i;
653 	populate_logical_ids(gt, logical_ids, class, map, ARRAY_SIZE(map));
654 }
655 
656 /**
657  * intel_engines_init_mmio() - allocate and prepare the Engine Command Streamers
658  * @gt: pointer to struct intel_gt
659  *
660  * Return: non-zero if the initialization failed.
661  */
662 int intel_engines_init_mmio(struct intel_gt *gt)
663 {
664 	struct drm_i915_private *i915 = gt->i915;
665 	const unsigned int engine_mask = init_engine_mask(gt);
666 	unsigned int mask = 0;
667 	unsigned int i, class;
668 	u8 logical_ids[MAX_ENGINE_INSTANCE + 1];
669 	int err;
670 
671 	drm_WARN_ON(&i915->drm, engine_mask == 0);
672 	drm_WARN_ON(&i915->drm, engine_mask &
673 		    GENMASK(BITS_PER_TYPE(mask) - 1, I915_NUM_ENGINES));
674 
675 	if (i915_inject_probe_failure(i915))
676 		return -ENODEV;
677 
678 	for (class = 0; class < MAX_ENGINE_CLASS + 1; ++class) {
679 		setup_logical_ids(gt, logical_ids, class);
680 
681 		for (i = 0; i < ARRAY_SIZE(intel_engines); ++i) {
682 			u8 instance = intel_engines[i].instance;
683 
684 			if (intel_engines[i].class != class ||
685 			    !HAS_ENGINE(gt, i))
686 				continue;
687 
688 			err = intel_engine_setup(gt, i,
689 						 logical_ids[instance]);
690 			if (err)
691 				goto cleanup;
692 
693 			mask |= BIT(i);
694 		}
695 	}
696 
697 	/*
698 	 * Catch failures to update intel_engines table when the new engines
699 	 * are added to the driver by a warning and disabling the forgotten
700 	 * engines.
701 	 */
702 	if (drm_WARN_ON(&i915->drm, mask != engine_mask))
703 		gt->info.engine_mask = mask;
704 
705 	gt->info.num_engines = hweight32(mask);
706 
707 	intel_gt_check_and_clear_faults(gt);
708 
709 	intel_setup_engine_capabilities(gt);
710 
711 	intel_uncore_prune_engine_fw_domains(gt->uncore, gt);
712 
713 	return 0;
714 
715 cleanup:
716 	intel_engines_free(gt);
717 	return err;
718 }
719 
720 void intel_engine_init_execlists(struct intel_engine_cs *engine)
721 {
722 	struct intel_engine_execlists * const execlists = &engine->execlists;
723 
724 	execlists->port_mask = 1;
725 	GEM_BUG_ON(!is_power_of_2(execlists_num_ports(execlists)));
726 	GEM_BUG_ON(execlists_num_ports(execlists) > EXECLIST_MAX_PORTS);
727 
728 	memset(execlists->pending, 0, sizeof(execlists->pending));
729 	execlists->active =
730 		memset(execlists->inflight, 0, sizeof(execlists->inflight));
731 }
732 
733 static void cleanup_status_page(struct intel_engine_cs *engine)
734 {
735 	struct i915_vma *vma;
736 
737 	/* Prevent writes into HWSP after returning the page to the system */
738 	intel_engine_set_hwsp_writemask(engine, ~0u);
739 
740 	vma = fetch_and_zero(&engine->status_page.vma);
741 	if (!vma)
742 		return;
743 
744 	if (!HWS_NEEDS_PHYSICAL(engine->i915))
745 		i915_vma_unpin(vma);
746 
747 	i915_gem_object_unpin_map(vma->obj);
748 	i915_gem_object_put(vma->obj);
749 }
750 
751 static int pin_ggtt_status_page(struct intel_engine_cs *engine,
752 				struct i915_gem_ww_ctx *ww,
753 				struct i915_vma *vma)
754 {
755 	unsigned int flags;
756 
757 	if (!HAS_LLC(engine->i915) && i915_ggtt_has_aperture(engine->gt->ggtt))
758 		/*
759 		 * On g33, we cannot place HWS above 256MiB, so
760 		 * restrict its pinning to the low mappable arena.
761 		 * Though this restriction is not documented for
762 		 * gen4, gen5, or byt, they also behave similarly
763 		 * and hang if the HWS is placed at the top of the
764 		 * GTT. To generalise, it appears that all !llc
765 		 * platforms have issues with us placing the HWS
766 		 * above the mappable region (even though we never
767 		 * actually map it).
768 		 */
769 		flags = PIN_MAPPABLE;
770 	else
771 		flags = PIN_HIGH;
772 
773 	return i915_ggtt_pin(vma, ww, 0, flags);
774 }
775 
776 static int init_status_page(struct intel_engine_cs *engine)
777 {
778 	struct drm_i915_gem_object *obj;
779 	struct i915_gem_ww_ctx ww;
780 	struct i915_vma *vma;
781 	void *vaddr;
782 	int ret;
783 
784 	INIT_LIST_HEAD(&engine->status_page.timelines);
785 
786 	/*
787 	 * Though the HWS register does support 36bit addresses, historically
788 	 * we have had hangs and corruption reported due to wild writes if
789 	 * the HWS is placed above 4G. We only allow objects to be allocated
790 	 * in GFP_DMA32 for i965, and no earlier physical address users had
791 	 * access to more than 4G.
792 	 */
793 	obj = i915_gem_object_create_internal(engine->i915, PAGE_SIZE);
794 	if (IS_ERR(obj)) {
795 		drm_err(&engine->i915->drm,
796 			"Failed to allocate status page\n");
797 		return PTR_ERR(obj);
798 	}
799 
800 	i915_gem_object_set_cache_coherency(obj, I915_CACHE_LLC);
801 
802 	vma = i915_vma_instance(obj, &engine->gt->ggtt->vm, NULL);
803 	if (IS_ERR(vma)) {
804 		ret = PTR_ERR(vma);
805 		goto err_put;
806 	}
807 
808 	i915_gem_ww_ctx_init(&ww, true);
809 retry:
810 	ret = i915_gem_object_lock(obj, &ww);
811 	if (!ret && !HWS_NEEDS_PHYSICAL(engine->i915))
812 		ret = pin_ggtt_status_page(engine, &ww, vma);
813 	if (ret)
814 		goto err;
815 
816 	vaddr = i915_gem_object_pin_map(obj, I915_MAP_WB);
817 	if (IS_ERR(vaddr)) {
818 		ret = PTR_ERR(vaddr);
819 		goto err_unpin;
820 	}
821 
822 	engine->status_page.addr = memset(vaddr, 0, PAGE_SIZE);
823 	engine->status_page.vma = vma;
824 
825 err_unpin:
826 	if (ret)
827 		i915_vma_unpin(vma);
828 err:
829 	if (ret == -EDEADLK) {
830 		ret = i915_gem_ww_ctx_backoff(&ww);
831 		if (!ret)
832 			goto retry;
833 	}
834 	i915_gem_ww_ctx_fini(&ww);
835 err_put:
836 	if (ret)
837 		i915_gem_object_put(obj);
838 	return ret;
839 }
840 
841 static int engine_setup_common(struct intel_engine_cs *engine)
842 {
843 	int err;
844 
845 	init_llist_head(&engine->barrier_tasks);
846 
847 	err = init_status_page(engine);
848 	if (err)
849 		return err;
850 
851 	engine->breadcrumbs = intel_breadcrumbs_create(engine);
852 	if (!engine->breadcrumbs) {
853 		err = -ENOMEM;
854 		goto err_status;
855 	}
856 
857 	engine->sched_engine = i915_sched_engine_create(ENGINE_PHYSICAL);
858 	if (!engine->sched_engine) {
859 		err = -ENOMEM;
860 		goto err_sched_engine;
861 	}
862 	engine->sched_engine->private_data = engine;
863 
864 	err = intel_engine_init_cmd_parser(engine);
865 	if (err)
866 		goto err_cmd_parser;
867 
868 	intel_engine_init_execlists(engine);
869 	intel_engine_init__pm(engine);
870 	intel_engine_init_retire(engine);
871 
872 	/* Use the whole device by default */
873 	engine->sseu =
874 		intel_sseu_from_device_info(&engine->gt->info.sseu);
875 
876 	intel_engine_init_workarounds(engine);
877 	intel_engine_init_whitelist(engine);
878 	intel_engine_init_ctx_wa(engine);
879 
880 	if (GRAPHICS_VER(engine->i915) >= 12)
881 		engine->flags |= I915_ENGINE_HAS_RELATIVE_MMIO;
882 
883 	return 0;
884 
885 err_cmd_parser:
886 	i915_sched_engine_put(engine->sched_engine);
887 err_sched_engine:
888 	intel_breadcrumbs_put(engine->breadcrumbs);
889 err_status:
890 	cleanup_status_page(engine);
891 	return err;
892 }
893 
894 struct measure_breadcrumb {
895 	struct i915_request rq;
896 	struct intel_ring ring;
897 	u32 cs[2048];
898 };
899 
900 static int measure_breadcrumb_dw(struct intel_context *ce)
901 {
902 	struct intel_engine_cs *engine = ce->engine;
903 	struct measure_breadcrumb *frame;
904 	int dw;
905 
906 	GEM_BUG_ON(!engine->gt->scratch);
907 
908 	frame = kzalloc(sizeof(*frame), GFP_KERNEL);
909 	if (!frame)
910 		return -ENOMEM;
911 
912 	frame->rq.engine = engine;
913 	frame->rq.context = ce;
914 	rcu_assign_pointer(frame->rq.timeline, ce->timeline);
915 	frame->rq.hwsp_seqno = ce->timeline->hwsp_seqno;
916 
917 	frame->ring.vaddr = frame->cs;
918 	frame->ring.size = sizeof(frame->cs);
919 	frame->ring.wrap =
920 		BITS_PER_TYPE(frame->ring.size) - ilog2(frame->ring.size);
921 	frame->ring.effective_size = frame->ring.size;
922 	intel_ring_update_space(&frame->ring);
923 	frame->rq.ring = &frame->ring;
924 
925 	mutex_lock(&ce->timeline->mutex);
926 	spin_lock_irq(&engine->sched_engine->lock);
927 
928 	dw = engine->emit_fini_breadcrumb(&frame->rq, frame->cs) - frame->cs;
929 
930 	spin_unlock_irq(&engine->sched_engine->lock);
931 	mutex_unlock(&ce->timeline->mutex);
932 
933 	GEM_BUG_ON(dw & 1); /* RING_TAIL must be qword aligned */
934 
935 	kfree(frame);
936 	return dw;
937 }
938 
939 struct intel_context *
940 intel_engine_create_pinned_context(struct intel_engine_cs *engine,
941 				   struct i915_address_space *vm,
942 				   unsigned int ring_size,
943 				   unsigned int hwsp,
944 				   struct lock_class_key *key,
945 				   const char *name)
946 {
947 	struct intel_context *ce;
948 	int err;
949 
950 	ce = intel_context_create(engine);
951 	if (IS_ERR(ce))
952 		return ce;
953 
954 	__set_bit(CONTEXT_BARRIER_BIT, &ce->flags);
955 	ce->timeline = page_pack_bits(NULL, hwsp);
956 	ce->ring = NULL;
957 	ce->ring_size = ring_size;
958 
959 	i915_vm_put(ce->vm);
960 	ce->vm = i915_vm_get(vm);
961 
962 	err = intel_context_pin(ce); /* perma-pin so it is always available */
963 	if (err) {
964 		intel_context_put(ce);
965 		return ERR_PTR(err);
966 	}
967 
968 	list_add_tail(&ce->pinned_contexts_link, &engine->pinned_contexts_list);
969 
970 	/*
971 	 * Give our perma-pinned kernel timelines a separate lockdep class,
972 	 * so that we can use them from within the normal user timelines
973 	 * should we need to inject GPU operations during their request
974 	 * construction.
975 	 */
976 	lockdep_set_class_and_name(&ce->timeline->mutex, key, name);
977 
978 	return ce;
979 }
980 
981 void intel_engine_destroy_pinned_context(struct intel_context *ce)
982 {
983 	struct intel_engine_cs *engine = ce->engine;
984 	struct i915_vma *hwsp = engine->status_page.vma;
985 
986 	GEM_BUG_ON(ce->timeline->hwsp_ggtt != hwsp);
987 
988 	mutex_lock(&hwsp->vm->mutex);
989 	list_del(&ce->timeline->engine_link);
990 	mutex_unlock(&hwsp->vm->mutex);
991 
992 	list_del(&ce->pinned_contexts_link);
993 	intel_context_unpin(ce);
994 	intel_context_put(ce);
995 }
996 
997 static struct intel_context *
998 create_kernel_context(struct intel_engine_cs *engine)
999 {
1000 	static struct lock_class_key kernel;
1001 
1002 	return intel_engine_create_pinned_context(engine, engine->gt->vm, SZ_4K,
1003 						  I915_GEM_HWS_SEQNO_ADDR,
1004 						  &kernel, "kernel_context");
1005 }
1006 
1007 /**
1008  * intel_engines_init_common - initialize cengine state which might require hw access
1009  * @engine: Engine to initialize.
1010  *
1011  * Initializes @engine@ structure members shared between legacy and execlists
1012  * submission modes which do require hardware access.
1013  *
1014  * Typcally done at later stages of submission mode specific engine setup.
1015  *
1016  * Returns zero on success or an error code on failure.
1017  */
1018 static int engine_init_common(struct intel_engine_cs *engine)
1019 {
1020 	struct intel_context *ce;
1021 	int ret;
1022 
1023 	engine->set_default_submission(engine);
1024 
1025 	/*
1026 	 * We may need to do things with the shrinker which
1027 	 * require us to immediately switch back to the default
1028 	 * context. This can cause a problem as pinning the
1029 	 * default context also requires GTT space which may not
1030 	 * be available. To avoid this we always pin the default
1031 	 * context.
1032 	 */
1033 	ce = create_kernel_context(engine);
1034 	if (IS_ERR(ce))
1035 		return PTR_ERR(ce);
1036 
1037 	ret = measure_breadcrumb_dw(ce);
1038 	if (ret < 0)
1039 		goto err_context;
1040 
1041 	engine->emit_fini_breadcrumb_dw = ret;
1042 	engine->kernel_context = ce;
1043 
1044 	return 0;
1045 
1046 err_context:
1047 	intel_engine_destroy_pinned_context(ce);
1048 	return ret;
1049 }
1050 
1051 int intel_engines_init(struct intel_gt *gt)
1052 {
1053 	int (*setup)(struct intel_engine_cs *engine);
1054 	struct intel_engine_cs *engine;
1055 	enum intel_engine_id id;
1056 	int err;
1057 
1058 	if (intel_uc_uses_guc_submission(&gt->uc)) {
1059 		gt->submission_method = INTEL_SUBMISSION_GUC;
1060 		setup = intel_guc_submission_setup;
1061 	} else if (HAS_EXECLISTS(gt->i915)) {
1062 		gt->submission_method = INTEL_SUBMISSION_ELSP;
1063 		setup = intel_execlists_submission_setup;
1064 	} else {
1065 		gt->submission_method = INTEL_SUBMISSION_RING;
1066 		setup = intel_ring_submission_setup;
1067 	}
1068 
1069 	for_each_engine(engine, gt, id) {
1070 		err = engine_setup_common(engine);
1071 		if (err)
1072 			return err;
1073 
1074 		err = setup(engine);
1075 		if (err)
1076 			return err;
1077 
1078 		err = engine_init_common(engine);
1079 		if (err)
1080 			return err;
1081 
1082 		intel_engine_add_user(engine);
1083 	}
1084 
1085 	return 0;
1086 }
1087 
1088 /**
1089  * intel_engines_cleanup_common - cleans up the engine state created by
1090  *                                the common initiailizers.
1091  * @engine: Engine to cleanup.
1092  *
1093  * This cleans up everything created by the common helpers.
1094  */
1095 void intel_engine_cleanup_common(struct intel_engine_cs *engine)
1096 {
1097 	GEM_BUG_ON(!list_empty(&engine->sched_engine->requests));
1098 
1099 	i915_sched_engine_put(engine->sched_engine);
1100 	intel_breadcrumbs_put(engine->breadcrumbs);
1101 
1102 	intel_engine_fini_retire(engine);
1103 	intel_engine_cleanup_cmd_parser(engine);
1104 
1105 	if (engine->default_state)
1106 		fput(engine->default_state);
1107 
1108 	if (engine->kernel_context)
1109 		intel_engine_destroy_pinned_context(engine->kernel_context);
1110 
1111 	GEM_BUG_ON(!llist_empty(&engine->barrier_tasks));
1112 	cleanup_status_page(engine);
1113 
1114 	intel_wa_list_free(&engine->ctx_wa_list);
1115 	intel_wa_list_free(&engine->wa_list);
1116 	intel_wa_list_free(&engine->whitelist);
1117 }
1118 
1119 /**
1120  * intel_engine_resume - re-initializes the HW state of the engine
1121  * @engine: Engine to resume.
1122  *
1123  * Returns zero on success or an error code on failure.
1124  */
1125 int intel_engine_resume(struct intel_engine_cs *engine)
1126 {
1127 	intel_engine_apply_workarounds(engine);
1128 	intel_engine_apply_whitelist(engine);
1129 
1130 	return engine->resume(engine);
1131 }
1132 
1133 u64 intel_engine_get_active_head(const struct intel_engine_cs *engine)
1134 {
1135 	struct drm_i915_private *i915 = engine->i915;
1136 
1137 	u64 acthd;
1138 
1139 	if (GRAPHICS_VER(i915) >= 8)
1140 		acthd = ENGINE_READ64(engine, RING_ACTHD, RING_ACTHD_UDW);
1141 	else if (GRAPHICS_VER(i915) >= 4)
1142 		acthd = ENGINE_READ(engine, RING_ACTHD);
1143 	else
1144 		acthd = ENGINE_READ(engine, ACTHD);
1145 
1146 	return acthd;
1147 }
1148 
1149 u64 intel_engine_get_last_batch_head(const struct intel_engine_cs *engine)
1150 {
1151 	u64 bbaddr;
1152 
1153 	if (GRAPHICS_VER(engine->i915) >= 8)
1154 		bbaddr = ENGINE_READ64(engine, RING_BBADDR, RING_BBADDR_UDW);
1155 	else
1156 		bbaddr = ENGINE_READ(engine, RING_BBADDR);
1157 
1158 	return bbaddr;
1159 }
1160 
1161 static unsigned long stop_timeout(const struct intel_engine_cs *engine)
1162 {
1163 	if (in_atomic() || irqs_disabled()) /* inside atomic preempt-reset? */
1164 		return 0;
1165 
1166 	/*
1167 	 * If we are doing a normal GPU reset, we can take our time and allow
1168 	 * the engine to quiesce. We've stopped submission to the engine, and
1169 	 * if we wait long enough an innocent context should complete and
1170 	 * leave the engine idle. So they should not be caught unaware by
1171 	 * the forthcoming GPU reset (which usually follows the stop_cs)!
1172 	 */
1173 	return READ_ONCE(engine->props.stop_timeout_ms);
1174 }
1175 
1176 static int __intel_engine_stop_cs(struct intel_engine_cs *engine,
1177 				  int fast_timeout_us,
1178 				  int slow_timeout_ms)
1179 {
1180 	struct intel_uncore *uncore = engine->uncore;
1181 	const i915_reg_t mode = RING_MI_MODE(engine->mmio_base);
1182 	int err;
1183 
1184 	intel_uncore_write_fw(uncore, mode, _MASKED_BIT_ENABLE(STOP_RING));
1185 	err = __intel_wait_for_register_fw(engine->uncore, mode,
1186 					   MODE_IDLE, MODE_IDLE,
1187 					   fast_timeout_us,
1188 					   slow_timeout_ms,
1189 					   NULL);
1190 
1191 	/* A final mmio read to let GPU writes be hopefully flushed to memory */
1192 	intel_uncore_posting_read_fw(uncore, mode);
1193 	return err;
1194 }
1195 
1196 int intel_engine_stop_cs(struct intel_engine_cs *engine)
1197 {
1198 	int err = 0;
1199 
1200 	if (GRAPHICS_VER(engine->i915) < 3)
1201 		return -ENODEV;
1202 
1203 	ENGINE_TRACE(engine, "\n");
1204 	if (__intel_engine_stop_cs(engine, 1000, stop_timeout(engine))) {
1205 		ENGINE_TRACE(engine,
1206 			     "timed out on STOP_RING -> IDLE; HEAD:%04x, TAIL:%04x\n",
1207 			     ENGINE_READ_FW(engine, RING_HEAD) & HEAD_ADDR,
1208 			     ENGINE_READ_FW(engine, RING_TAIL) & TAIL_ADDR);
1209 
1210 		/*
1211 		 * Sometimes we observe that the idle flag is not
1212 		 * set even though the ring is empty. So double
1213 		 * check before giving up.
1214 		 */
1215 		if ((ENGINE_READ_FW(engine, RING_HEAD) & HEAD_ADDR) !=
1216 		    (ENGINE_READ_FW(engine, RING_TAIL) & TAIL_ADDR))
1217 			err = -ETIMEDOUT;
1218 	}
1219 
1220 	return err;
1221 }
1222 
1223 void intel_engine_cancel_stop_cs(struct intel_engine_cs *engine)
1224 {
1225 	ENGINE_TRACE(engine, "\n");
1226 
1227 	ENGINE_WRITE_FW(engine, RING_MI_MODE, _MASKED_BIT_DISABLE(STOP_RING));
1228 }
1229 
1230 const char *i915_cache_level_str(struct drm_i915_private *i915, int type)
1231 {
1232 	switch (type) {
1233 	case I915_CACHE_NONE: return " uncached";
1234 	case I915_CACHE_LLC: return HAS_LLC(i915) ? " LLC" : " snooped";
1235 	case I915_CACHE_L3_LLC: return " L3+LLC";
1236 	case I915_CACHE_WT: return " WT";
1237 	default: return "";
1238 	}
1239 }
1240 
1241 static u32
1242 read_subslice_reg(const struct intel_engine_cs *engine,
1243 		  int slice, int subslice, i915_reg_t reg)
1244 {
1245 	return intel_uncore_read_with_mcr_steering(engine->uncore, reg,
1246 						   slice, subslice);
1247 }
1248 
1249 /* NB: please notice the memset */
1250 void intel_engine_get_instdone(const struct intel_engine_cs *engine,
1251 			       struct intel_instdone *instdone)
1252 {
1253 	struct drm_i915_private *i915 = engine->i915;
1254 	const struct sseu_dev_info *sseu = &engine->gt->info.sseu;
1255 	struct intel_uncore *uncore = engine->uncore;
1256 	u32 mmio_base = engine->mmio_base;
1257 	int slice;
1258 	int subslice;
1259 	int iter;
1260 
1261 	memset(instdone, 0, sizeof(*instdone));
1262 
1263 	if (GRAPHICS_VER(i915) >= 8) {
1264 		instdone->instdone =
1265 			intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
1266 
1267 		if (engine->id != RCS0)
1268 			return;
1269 
1270 		instdone->slice_common =
1271 			intel_uncore_read(uncore, GEN7_SC_INSTDONE);
1272 		if (GRAPHICS_VER(i915) >= 12) {
1273 			instdone->slice_common_extra[0] =
1274 				intel_uncore_read(uncore, GEN12_SC_INSTDONE_EXTRA);
1275 			instdone->slice_common_extra[1] =
1276 				intel_uncore_read(uncore, GEN12_SC_INSTDONE_EXTRA2);
1277 		}
1278 
1279 		if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 50)) {
1280 			for_each_instdone_gslice_dss_xehp(i915, sseu, iter, slice, subslice) {
1281 				instdone->sampler[slice][subslice] =
1282 					read_subslice_reg(engine, slice, subslice,
1283 							  GEN7_SAMPLER_INSTDONE);
1284 				instdone->row[slice][subslice] =
1285 					read_subslice_reg(engine, slice, subslice,
1286 							  GEN7_ROW_INSTDONE);
1287 			}
1288 		} else {
1289 			for_each_instdone_slice_subslice(i915, sseu, slice, subslice) {
1290 				instdone->sampler[slice][subslice] =
1291 					read_subslice_reg(engine, slice, subslice,
1292 							  GEN7_SAMPLER_INSTDONE);
1293 				instdone->row[slice][subslice] =
1294 					read_subslice_reg(engine, slice, subslice,
1295 							  GEN7_ROW_INSTDONE);
1296 			}
1297 		}
1298 
1299 		if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 55)) {
1300 			for_each_instdone_gslice_dss_xehp(i915, sseu, iter, slice, subslice)
1301 				instdone->geom_svg[slice][subslice] =
1302 					read_subslice_reg(engine, slice, subslice,
1303 							  XEHPG_INSTDONE_GEOM_SVG);
1304 		}
1305 	} else if (GRAPHICS_VER(i915) >= 7) {
1306 		instdone->instdone =
1307 			intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
1308 
1309 		if (engine->id != RCS0)
1310 			return;
1311 
1312 		instdone->slice_common =
1313 			intel_uncore_read(uncore, GEN7_SC_INSTDONE);
1314 		instdone->sampler[0][0] =
1315 			intel_uncore_read(uncore, GEN7_SAMPLER_INSTDONE);
1316 		instdone->row[0][0] =
1317 			intel_uncore_read(uncore, GEN7_ROW_INSTDONE);
1318 	} else if (GRAPHICS_VER(i915) >= 4) {
1319 		instdone->instdone =
1320 			intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
1321 		if (engine->id == RCS0)
1322 			/* HACK: Using the wrong struct member */
1323 			instdone->slice_common =
1324 				intel_uncore_read(uncore, GEN4_INSTDONE1);
1325 	} else {
1326 		instdone->instdone = intel_uncore_read(uncore, GEN2_INSTDONE);
1327 	}
1328 }
1329 
1330 static bool ring_is_idle(struct intel_engine_cs *engine)
1331 {
1332 	bool idle = true;
1333 
1334 	if (I915_SELFTEST_ONLY(!engine->mmio_base))
1335 		return true;
1336 
1337 	if (!intel_engine_pm_get_if_awake(engine))
1338 		return true;
1339 
1340 	/* First check that no commands are left in the ring */
1341 	if ((ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR) !=
1342 	    (ENGINE_READ(engine, RING_TAIL) & TAIL_ADDR))
1343 		idle = false;
1344 
1345 	/* No bit for gen2, so assume the CS parser is idle */
1346 	if (GRAPHICS_VER(engine->i915) > 2 &&
1347 	    !(ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE))
1348 		idle = false;
1349 
1350 	intel_engine_pm_put(engine);
1351 
1352 	return idle;
1353 }
1354 
1355 void __intel_engine_flush_submission(struct intel_engine_cs *engine, bool sync)
1356 {
1357 	struct tasklet_struct *t = &engine->sched_engine->tasklet;
1358 
1359 	if (!t->callback)
1360 		return;
1361 
1362 	local_bh_disable();
1363 	if (tasklet_trylock(t)) {
1364 		/* Must wait for any GPU reset in progress. */
1365 		if (__tasklet_is_enabled(t))
1366 			t->callback(t);
1367 		tasklet_unlock(t);
1368 	}
1369 	local_bh_enable();
1370 
1371 	/* Synchronise and wait for the tasklet on another CPU */
1372 	if (sync)
1373 		tasklet_unlock_wait(t);
1374 }
1375 
1376 /**
1377  * intel_engine_is_idle() - Report if the engine has finished process all work
1378  * @engine: the intel_engine_cs
1379  *
1380  * Return true if there are no requests pending, nothing left to be submitted
1381  * to hardware, and that the engine is idle.
1382  */
1383 bool intel_engine_is_idle(struct intel_engine_cs *engine)
1384 {
1385 	/* More white lies, if wedged, hw state is inconsistent */
1386 	if (intel_gt_is_wedged(engine->gt))
1387 		return true;
1388 
1389 	if (!intel_engine_pm_is_awake(engine))
1390 		return true;
1391 
1392 	/* Waiting to drain ELSP? */
1393 	intel_synchronize_hardirq(engine->i915);
1394 	intel_engine_flush_submission(engine);
1395 
1396 	/* ELSP is empty, but there are ready requests? E.g. after reset */
1397 	if (!i915_sched_engine_is_empty(engine->sched_engine))
1398 		return false;
1399 
1400 	/* Ring stopped? */
1401 	return ring_is_idle(engine);
1402 }
1403 
1404 bool intel_engines_are_idle(struct intel_gt *gt)
1405 {
1406 	struct intel_engine_cs *engine;
1407 	enum intel_engine_id id;
1408 
1409 	/*
1410 	 * If the driver is wedged, HW state may be very inconsistent and
1411 	 * report that it is still busy, even though we have stopped using it.
1412 	 */
1413 	if (intel_gt_is_wedged(gt))
1414 		return true;
1415 
1416 	/* Already parked (and passed an idleness test); must still be idle */
1417 	if (!READ_ONCE(gt->awake))
1418 		return true;
1419 
1420 	for_each_engine(engine, gt, id) {
1421 		if (!intel_engine_is_idle(engine))
1422 			return false;
1423 	}
1424 
1425 	return true;
1426 }
1427 
1428 bool intel_engine_irq_enable(struct intel_engine_cs *engine)
1429 {
1430 	if (!engine->irq_enable)
1431 		return false;
1432 
1433 	/* Caller disables interrupts */
1434 	spin_lock(&engine->gt->irq_lock);
1435 	engine->irq_enable(engine);
1436 	spin_unlock(&engine->gt->irq_lock);
1437 
1438 	return true;
1439 }
1440 
1441 void intel_engine_irq_disable(struct intel_engine_cs *engine)
1442 {
1443 	if (!engine->irq_disable)
1444 		return;
1445 
1446 	/* Caller disables interrupts */
1447 	spin_lock(&engine->gt->irq_lock);
1448 	engine->irq_disable(engine);
1449 	spin_unlock(&engine->gt->irq_lock);
1450 }
1451 
1452 void intel_engines_reset_default_submission(struct intel_gt *gt)
1453 {
1454 	struct intel_engine_cs *engine;
1455 	enum intel_engine_id id;
1456 
1457 	for_each_engine(engine, gt, id) {
1458 		if (engine->sanitize)
1459 			engine->sanitize(engine);
1460 
1461 		engine->set_default_submission(engine);
1462 	}
1463 }
1464 
1465 bool intel_engine_can_store_dword(struct intel_engine_cs *engine)
1466 {
1467 	switch (GRAPHICS_VER(engine->i915)) {
1468 	case 2:
1469 		return false; /* uses physical not virtual addresses */
1470 	case 3:
1471 		/* maybe only uses physical not virtual addresses */
1472 		return !(IS_I915G(engine->i915) || IS_I915GM(engine->i915));
1473 	case 4:
1474 		return !IS_I965G(engine->i915); /* who knows! */
1475 	case 6:
1476 		return engine->class != VIDEO_DECODE_CLASS; /* b0rked */
1477 	default:
1478 		return true;
1479 	}
1480 }
1481 
1482 static struct intel_timeline *get_timeline(struct i915_request *rq)
1483 {
1484 	struct intel_timeline *tl;
1485 
1486 	/*
1487 	 * Even though we are holding the engine->sched_engine->lock here, there
1488 	 * is no control over the submission queue per-se and we are
1489 	 * inspecting the active state at a random point in time, with an
1490 	 * unknown queue. Play safe and make sure the timeline remains valid.
1491 	 * (Only being used for pretty printing, one extra kref shouldn't
1492 	 * cause a camel stampede!)
1493 	 */
1494 	rcu_read_lock();
1495 	tl = rcu_dereference(rq->timeline);
1496 	if (!kref_get_unless_zero(&tl->kref))
1497 		tl = NULL;
1498 	rcu_read_unlock();
1499 
1500 	return tl;
1501 }
1502 
1503 static int print_ring(char *buf, int sz, struct i915_request *rq)
1504 {
1505 	int len = 0;
1506 
1507 	if (!i915_request_signaled(rq)) {
1508 		struct intel_timeline *tl = get_timeline(rq);
1509 
1510 		len = scnprintf(buf, sz,
1511 				"ring:{start:%08x, hwsp:%08x, seqno:%08x, runtime:%llums}, ",
1512 				i915_ggtt_offset(rq->ring->vma),
1513 				tl ? tl->hwsp_offset : 0,
1514 				hwsp_seqno(rq),
1515 				DIV_ROUND_CLOSEST_ULL(intel_context_get_total_runtime_ns(rq->context),
1516 						      1000 * 1000));
1517 
1518 		if (tl)
1519 			intel_timeline_put(tl);
1520 	}
1521 
1522 	return len;
1523 }
1524 
1525 static void hexdump(struct drm_printer *m, const void *buf, size_t len)
1526 {
1527 	const size_t rowsize = 8 * sizeof(u32);
1528 	const void *prev = NULL;
1529 	bool skip = false;
1530 	size_t pos;
1531 
1532 	for (pos = 0; pos < len; pos += rowsize) {
1533 		char line[128];
1534 
1535 		if (prev && !memcmp(prev, buf + pos, rowsize)) {
1536 			if (!skip) {
1537 				drm_printf(m, "*\n");
1538 				skip = true;
1539 			}
1540 			continue;
1541 		}
1542 
1543 		WARN_ON_ONCE(hex_dump_to_buffer(buf + pos, len - pos,
1544 						rowsize, sizeof(u32),
1545 						line, sizeof(line),
1546 						false) >= sizeof(line));
1547 		drm_printf(m, "[%04zx] %s\n", pos, line);
1548 
1549 		prev = buf + pos;
1550 		skip = false;
1551 	}
1552 }
1553 
1554 static const char *repr_timer(const struct timer_list *t)
1555 {
1556 	if (!READ_ONCE(t->expires))
1557 		return "inactive";
1558 
1559 	if (timer_pending(t))
1560 		return "active";
1561 
1562 	return "expired";
1563 }
1564 
1565 static void intel_engine_print_registers(struct intel_engine_cs *engine,
1566 					 struct drm_printer *m)
1567 {
1568 	struct drm_i915_private *dev_priv = engine->i915;
1569 	struct intel_engine_execlists * const execlists = &engine->execlists;
1570 	u64 addr;
1571 
1572 	if (engine->id == RENDER_CLASS && IS_GRAPHICS_VER(dev_priv, 4, 7))
1573 		drm_printf(m, "\tCCID: 0x%08x\n", ENGINE_READ(engine, CCID));
1574 	if (HAS_EXECLISTS(dev_priv)) {
1575 		drm_printf(m, "\tEL_STAT_HI: 0x%08x\n",
1576 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_HI));
1577 		drm_printf(m, "\tEL_STAT_LO: 0x%08x\n",
1578 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_LO));
1579 	}
1580 	drm_printf(m, "\tRING_START: 0x%08x\n",
1581 		   ENGINE_READ(engine, RING_START));
1582 	drm_printf(m, "\tRING_HEAD:  0x%08x\n",
1583 		   ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR);
1584 	drm_printf(m, "\tRING_TAIL:  0x%08x\n",
1585 		   ENGINE_READ(engine, RING_TAIL) & TAIL_ADDR);
1586 	drm_printf(m, "\tRING_CTL:   0x%08x%s\n",
1587 		   ENGINE_READ(engine, RING_CTL),
1588 		   ENGINE_READ(engine, RING_CTL) & (RING_WAIT | RING_WAIT_SEMAPHORE) ? " [waiting]" : "");
1589 	if (GRAPHICS_VER(engine->i915) > 2) {
1590 		drm_printf(m, "\tRING_MODE:  0x%08x%s\n",
1591 			   ENGINE_READ(engine, RING_MI_MODE),
1592 			   ENGINE_READ(engine, RING_MI_MODE) & (MODE_IDLE) ? " [idle]" : "");
1593 	}
1594 
1595 	if (GRAPHICS_VER(dev_priv) >= 6) {
1596 		drm_printf(m, "\tRING_IMR:   0x%08x\n",
1597 			   ENGINE_READ(engine, RING_IMR));
1598 		drm_printf(m, "\tRING_ESR:   0x%08x\n",
1599 			   ENGINE_READ(engine, RING_ESR));
1600 		drm_printf(m, "\tRING_EMR:   0x%08x\n",
1601 			   ENGINE_READ(engine, RING_EMR));
1602 		drm_printf(m, "\tRING_EIR:   0x%08x\n",
1603 			   ENGINE_READ(engine, RING_EIR));
1604 	}
1605 
1606 	addr = intel_engine_get_active_head(engine);
1607 	drm_printf(m, "\tACTHD:  0x%08x_%08x\n",
1608 		   upper_32_bits(addr), lower_32_bits(addr));
1609 	addr = intel_engine_get_last_batch_head(engine);
1610 	drm_printf(m, "\tBBADDR: 0x%08x_%08x\n",
1611 		   upper_32_bits(addr), lower_32_bits(addr));
1612 	if (GRAPHICS_VER(dev_priv) >= 8)
1613 		addr = ENGINE_READ64(engine, RING_DMA_FADD, RING_DMA_FADD_UDW);
1614 	else if (GRAPHICS_VER(dev_priv) >= 4)
1615 		addr = ENGINE_READ(engine, RING_DMA_FADD);
1616 	else
1617 		addr = ENGINE_READ(engine, DMA_FADD_I8XX);
1618 	drm_printf(m, "\tDMA_FADDR: 0x%08x_%08x\n",
1619 		   upper_32_bits(addr), lower_32_bits(addr));
1620 	if (GRAPHICS_VER(dev_priv) >= 4) {
1621 		drm_printf(m, "\tIPEIR: 0x%08x\n",
1622 			   ENGINE_READ(engine, RING_IPEIR));
1623 		drm_printf(m, "\tIPEHR: 0x%08x\n",
1624 			   ENGINE_READ(engine, RING_IPEHR));
1625 	} else {
1626 		drm_printf(m, "\tIPEIR: 0x%08x\n", ENGINE_READ(engine, IPEIR));
1627 		drm_printf(m, "\tIPEHR: 0x%08x\n", ENGINE_READ(engine, IPEHR));
1628 	}
1629 
1630 	if (intel_engine_uses_guc(engine)) {
1631 		/* nothing to print yet */
1632 	} else if (HAS_EXECLISTS(dev_priv)) {
1633 		struct i915_request * const *port, *rq;
1634 		const u32 *hws =
1635 			&engine->status_page.addr[I915_HWS_CSB_BUF0_INDEX];
1636 		const u8 num_entries = execlists->csb_size;
1637 		unsigned int idx;
1638 		u8 read, write;
1639 
1640 		drm_printf(m, "\tExeclist tasklet queued? %s (%s), preempt? %s, timeslice? %s\n",
1641 			   yesno(test_bit(TASKLET_STATE_SCHED,
1642 					  &engine->sched_engine->tasklet.state)),
1643 			   enableddisabled(!atomic_read(&engine->sched_engine->tasklet.count)),
1644 			   repr_timer(&engine->execlists.preempt),
1645 			   repr_timer(&engine->execlists.timer));
1646 
1647 		read = execlists->csb_head;
1648 		write = READ_ONCE(*execlists->csb_write);
1649 
1650 		drm_printf(m, "\tExeclist status: 0x%08x %08x; CSB read:%d, write:%d, entries:%d\n",
1651 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_LO),
1652 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_HI),
1653 			   read, write, num_entries);
1654 
1655 		if (read >= num_entries)
1656 			read = 0;
1657 		if (write >= num_entries)
1658 			write = 0;
1659 		if (read > write)
1660 			write += num_entries;
1661 		while (read < write) {
1662 			idx = ++read % num_entries;
1663 			drm_printf(m, "\tExeclist CSB[%d]: 0x%08x, context: %d\n",
1664 				   idx, hws[idx * 2], hws[idx * 2 + 1]);
1665 		}
1666 
1667 		i915_sched_engine_active_lock_bh(engine->sched_engine);
1668 		rcu_read_lock();
1669 		for (port = execlists->active; (rq = *port); port++) {
1670 			char hdr[160];
1671 			int len;
1672 
1673 			len = scnprintf(hdr, sizeof(hdr),
1674 					"\t\tActive[%d]:  ccid:%08x%s%s, ",
1675 					(int)(port - execlists->active),
1676 					rq->context->lrc.ccid,
1677 					intel_context_is_closed(rq->context) ? "!" : "",
1678 					intel_context_is_banned(rq->context) ? "*" : "");
1679 			len += print_ring(hdr + len, sizeof(hdr) - len, rq);
1680 			scnprintf(hdr + len, sizeof(hdr) - len, "rq: ");
1681 			i915_request_show(m, rq, hdr, 0);
1682 		}
1683 		for (port = execlists->pending; (rq = *port); port++) {
1684 			char hdr[160];
1685 			int len;
1686 
1687 			len = scnprintf(hdr, sizeof(hdr),
1688 					"\t\tPending[%d]: ccid:%08x%s%s, ",
1689 					(int)(port - execlists->pending),
1690 					rq->context->lrc.ccid,
1691 					intel_context_is_closed(rq->context) ? "!" : "",
1692 					intel_context_is_banned(rq->context) ? "*" : "");
1693 			len += print_ring(hdr + len, sizeof(hdr) - len, rq);
1694 			scnprintf(hdr + len, sizeof(hdr) - len, "rq: ");
1695 			i915_request_show(m, rq, hdr, 0);
1696 		}
1697 		rcu_read_unlock();
1698 		i915_sched_engine_active_unlock_bh(engine->sched_engine);
1699 	} else if (GRAPHICS_VER(dev_priv) > 6) {
1700 		drm_printf(m, "\tPP_DIR_BASE: 0x%08x\n",
1701 			   ENGINE_READ(engine, RING_PP_DIR_BASE));
1702 		drm_printf(m, "\tPP_DIR_BASE_READ: 0x%08x\n",
1703 			   ENGINE_READ(engine, RING_PP_DIR_BASE_READ));
1704 		drm_printf(m, "\tPP_DIR_DCLV: 0x%08x\n",
1705 			   ENGINE_READ(engine, RING_PP_DIR_DCLV));
1706 	}
1707 }
1708 
1709 static void print_request_ring(struct drm_printer *m, struct i915_request *rq)
1710 {
1711 	struct i915_vma_snapshot *vsnap = &rq->batch_snapshot;
1712 	void *ring;
1713 	int size;
1714 
1715 	if (!i915_vma_snapshot_present(vsnap))
1716 		vsnap = NULL;
1717 
1718 	drm_printf(m,
1719 		   "[head %04x, postfix %04x, tail %04x, batch 0x%08x_%08x]:\n",
1720 		   rq->head, rq->postfix, rq->tail,
1721 		   vsnap ? upper_32_bits(vsnap->gtt_offset) : ~0u,
1722 		   vsnap ? lower_32_bits(vsnap->gtt_offset) : ~0u);
1723 
1724 	size = rq->tail - rq->head;
1725 	if (rq->tail < rq->head)
1726 		size += rq->ring->size;
1727 
1728 	ring = kmalloc(size, GFP_ATOMIC);
1729 	if (ring) {
1730 		const void *vaddr = rq->ring->vaddr;
1731 		unsigned int head = rq->head;
1732 		unsigned int len = 0;
1733 
1734 		if (rq->tail < head) {
1735 			len = rq->ring->size - head;
1736 			memcpy(ring, vaddr + head, len);
1737 			head = 0;
1738 		}
1739 		memcpy(ring + len, vaddr + head, size - len);
1740 
1741 		hexdump(m, ring, size);
1742 		kfree(ring);
1743 	}
1744 }
1745 
1746 static unsigned long list_count(struct list_head *list)
1747 {
1748 	struct list_head *pos;
1749 	unsigned long count = 0;
1750 
1751 	list_for_each(pos, list)
1752 		count++;
1753 
1754 	return count;
1755 }
1756 
1757 static unsigned long read_ul(void *p, size_t x)
1758 {
1759 	return *(unsigned long *)(p + x);
1760 }
1761 
1762 static void print_properties(struct intel_engine_cs *engine,
1763 			     struct drm_printer *m)
1764 {
1765 	static const struct pmap {
1766 		size_t offset;
1767 		const char *name;
1768 	} props[] = {
1769 #define P(x) { \
1770 	.offset = offsetof(typeof(engine->props), x), \
1771 	.name = #x \
1772 }
1773 		P(heartbeat_interval_ms),
1774 		P(max_busywait_duration_ns),
1775 		P(preempt_timeout_ms),
1776 		P(stop_timeout_ms),
1777 		P(timeslice_duration_ms),
1778 
1779 		{},
1780 #undef P
1781 	};
1782 	const struct pmap *p;
1783 
1784 	drm_printf(m, "\tProperties:\n");
1785 	for (p = props; p->name; p++)
1786 		drm_printf(m, "\t\t%s: %lu [default %lu]\n",
1787 			   p->name,
1788 			   read_ul(&engine->props, p->offset),
1789 			   read_ul(&engine->defaults, p->offset));
1790 }
1791 
1792 static void engine_dump_request(struct i915_request *rq, struct drm_printer *m, const char *msg)
1793 {
1794 	struct intel_timeline *tl = get_timeline(rq);
1795 
1796 	i915_request_show(m, rq, msg, 0);
1797 
1798 	drm_printf(m, "\t\tring->start:  0x%08x\n",
1799 		   i915_ggtt_offset(rq->ring->vma));
1800 	drm_printf(m, "\t\tring->head:   0x%08x\n",
1801 		   rq->ring->head);
1802 	drm_printf(m, "\t\tring->tail:   0x%08x\n",
1803 		   rq->ring->tail);
1804 	drm_printf(m, "\t\tring->emit:   0x%08x\n",
1805 		   rq->ring->emit);
1806 	drm_printf(m, "\t\tring->space:  0x%08x\n",
1807 		   rq->ring->space);
1808 
1809 	if (tl) {
1810 		drm_printf(m, "\t\tring->hwsp:   0x%08x\n",
1811 			   tl->hwsp_offset);
1812 		intel_timeline_put(tl);
1813 	}
1814 
1815 	print_request_ring(m, rq);
1816 
1817 	if (rq->context->lrc_reg_state) {
1818 		drm_printf(m, "Logical Ring Context:\n");
1819 		hexdump(m, rq->context->lrc_reg_state, PAGE_SIZE);
1820 	}
1821 }
1822 
1823 void intel_engine_dump_active_requests(struct list_head *requests,
1824 				       struct i915_request *hung_rq,
1825 				       struct drm_printer *m)
1826 {
1827 	struct i915_request *rq;
1828 	const char *msg;
1829 	enum i915_request_state state;
1830 
1831 	list_for_each_entry(rq, requests, sched.link) {
1832 		if (rq == hung_rq)
1833 			continue;
1834 
1835 		state = i915_test_request_state(rq);
1836 		if (state < I915_REQUEST_QUEUED)
1837 			continue;
1838 
1839 		if (state == I915_REQUEST_ACTIVE)
1840 			msg = "\t\tactive on engine";
1841 		else
1842 			msg = "\t\tactive in queue";
1843 
1844 		engine_dump_request(rq, m, msg);
1845 	}
1846 }
1847 
1848 static void engine_dump_active_requests(struct intel_engine_cs *engine, struct drm_printer *m)
1849 {
1850 	struct i915_request *hung_rq = NULL;
1851 	struct intel_context *ce;
1852 	bool guc;
1853 
1854 	/*
1855 	 * No need for an engine->irq_seqno_barrier() before the seqno reads.
1856 	 * The GPU is still running so requests are still executing and any
1857 	 * hardware reads will be out of date by the time they are reported.
1858 	 * But the intention here is just to report an instantaneous snapshot
1859 	 * so that's fine.
1860 	 */
1861 	lockdep_assert_held(&engine->sched_engine->lock);
1862 
1863 	drm_printf(m, "\tRequests:\n");
1864 
1865 	guc = intel_uc_uses_guc_submission(&engine->gt->uc);
1866 	if (guc) {
1867 		ce = intel_engine_get_hung_context(engine);
1868 		if (ce)
1869 			hung_rq = intel_context_find_active_request(ce);
1870 	} else {
1871 		hung_rq = intel_engine_execlist_find_hung_request(engine);
1872 	}
1873 
1874 	if (hung_rq)
1875 		engine_dump_request(hung_rq, m, "\t\thung");
1876 
1877 	if (guc)
1878 		intel_guc_dump_active_requests(engine, hung_rq, m);
1879 	else
1880 		intel_engine_dump_active_requests(&engine->sched_engine->requests,
1881 						  hung_rq, m);
1882 }
1883 
1884 void intel_engine_dump(struct intel_engine_cs *engine,
1885 		       struct drm_printer *m,
1886 		       const char *header, ...)
1887 {
1888 	struct i915_gpu_error * const error = &engine->i915->gpu_error;
1889 	struct i915_request *rq;
1890 	intel_wakeref_t wakeref;
1891 	unsigned long flags;
1892 	ktime_t dummy;
1893 
1894 	if (header) {
1895 		va_list ap;
1896 
1897 		va_start(ap, header);
1898 		drm_vprintf(m, header, &ap);
1899 		va_end(ap);
1900 	}
1901 
1902 	if (intel_gt_is_wedged(engine->gt))
1903 		drm_printf(m, "*** WEDGED ***\n");
1904 
1905 	drm_printf(m, "\tAwake? %d\n", atomic_read(&engine->wakeref.count));
1906 	drm_printf(m, "\tBarriers?: %s\n",
1907 		   yesno(!llist_empty(&engine->barrier_tasks)));
1908 	drm_printf(m, "\tLatency: %luus\n",
1909 		   ewma__engine_latency_read(&engine->latency));
1910 	if (intel_engine_supports_stats(engine))
1911 		drm_printf(m, "\tRuntime: %llums\n",
1912 			   ktime_to_ms(intel_engine_get_busy_time(engine,
1913 								  &dummy)));
1914 	drm_printf(m, "\tForcewake: %x domains, %d active\n",
1915 		   engine->fw_domain, READ_ONCE(engine->fw_active));
1916 
1917 	rcu_read_lock();
1918 	rq = READ_ONCE(engine->heartbeat.systole);
1919 	if (rq)
1920 		drm_printf(m, "\tHeartbeat: %d ms ago\n",
1921 			   jiffies_to_msecs(jiffies - rq->emitted_jiffies));
1922 	rcu_read_unlock();
1923 	drm_printf(m, "\tReset count: %d (global %d)\n",
1924 		   i915_reset_engine_count(error, engine),
1925 		   i915_reset_count(error));
1926 	print_properties(engine, m);
1927 
1928 	spin_lock_irqsave(&engine->sched_engine->lock, flags);
1929 	engine_dump_active_requests(engine, m);
1930 
1931 	drm_printf(m, "\tOn hold?: %lu\n",
1932 		   list_count(&engine->sched_engine->hold));
1933 	spin_unlock_irqrestore(&engine->sched_engine->lock, flags);
1934 
1935 	drm_printf(m, "\tMMIO base:  0x%08x\n", engine->mmio_base);
1936 	wakeref = intel_runtime_pm_get_if_in_use(engine->uncore->rpm);
1937 	if (wakeref) {
1938 		intel_engine_print_registers(engine, m);
1939 		intel_runtime_pm_put(engine->uncore->rpm, wakeref);
1940 	} else {
1941 		drm_printf(m, "\tDevice is asleep; skipping register dump\n");
1942 	}
1943 
1944 	intel_execlists_show_requests(engine, m, i915_request_show, 8);
1945 
1946 	drm_printf(m, "HWSP:\n");
1947 	hexdump(m, engine->status_page.addr, PAGE_SIZE);
1948 
1949 	drm_printf(m, "Idle? %s\n", yesno(intel_engine_is_idle(engine)));
1950 
1951 	intel_engine_print_breadcrumbs(engine, m);
1952 }
1953 
1954 /**
1955  * intel_engine_get_busy_time() - Return current accumulated engine busyness
1956  * @engine: engine to report on
1957  * @now: monotonic timestamp of sampling
1958  *
1959  * Returns accumulated time @engine was busy since engine stats were enabled.
1960  */
1961 ktime_t intel_engine_get_busy_time(struct intel_engine_cs *engine, ktime_t *now)
1962 {
1963 	return engine->busyness(engine, now);
1964 }
1965 
1966 struct intel_context *
1967 intel_engine_create_virtual(struct intel_engine_cs **siblings,
1968 			    unsigned int count, unsigned long flags)
1969 {
1970 	if (count == 0)
1971 		return ERR_PTR(-EINVAL);
1972 
1973 	if (count == 1 && !(flags & FORCE_VIRTUAL))
1974 		return intel_context_create(siblings[0]);
1975 
1976 	GEM_BUG_ON(!siblings[0]->cops->create_virtual);
1977 	return siblings[0]->cops->create_virtual(siblings, count, flags);
1978 }
1979 
1980 struct i915_request *
1981 intel_engine_execlist_find_hung_request(struct intel_engine_cs *engine)
1982 {
1983 	struct i915_request *request, *active = NULL;
1984 
1985 	/*
1986 	 * This search does not work in GuC submission mode. However, the GuC
1987 	 * will report the hanging context directly to the driver itself. So
1988 	 * the driver should never get here when in GuC mode.
1989 	 */
1990 	GEM_BUG_ON(intel_uc_uses_guc_submission(&engine->gt->uc));
1991 
1992 	/*
1993 	 * We are called by the error capture, reset and to dump engine
1994 	 * state at random points in time. In particular, note that neither is
1995 	 * crucially ordered with an interrupt. After a hang, the GPU is dead
1996 	 * and we assume that no more writes can happen (we waited long enough
1997 	 * for all writes that were in transaction to be flushed) - adding an
1998 	 * extra delay for a recent interrupt is pointless. Hence, we do
1999 	 * not need an engine->irq_seqno_barrier() before the seqno reads.
2000 	 * At all other times, we must assume the GPU is still running, but
2001 	 * we only care about the snapshot of this moment.
2002 	 */
2003 	lockdep_assert_held(&engine->sched_engine->lock);
2004 
2005 	rcu_read_lock();
2006 	request = execlists_active(&engine->execlists);
2007 	if (request) {
2008 		struct intel_timeline *tl = request->context->timeline;
2009 
2010 		list_for_each_entry_from_reverse(request, &tl->requests, link) {
2011 			if (__i915_request_is_complete(request))
2012 				break;
2013 
2014 			active = request;
2015 		}
2016 	}
2017 	rcu_read_unlock();
2018 	if (active)
2019 		return active;
2020 
2021 	list_for_each_entry(request, &engine->sched_engine->requests,
2022 			    sched.link) {
2023 		if (i915_test_request_state(request) != I915_REQUEST_ACTIVE)
2024 			continue;
2025 
2026 		active = request;
2027 		break;
2028 	}
2029 
2030 	return active;
2031 }
2032 
2033 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
2034 #include "mock_engine.c"
2035 #include "selftest_engine.c"
2036 #include "selftest_engine_cs.c"
2037 #endif
2038