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 #include "gt/intel_gt_regs.h"
10 
11 #include "i915_cmd_parser.h"
12 #include "i915_drv.h"
13 #include "intel_breadcrumbs.h"
14 #include "intel_context.h"
15 #include "intel_engine.h"
16 #include "intel_engine_pm.h"
17 #include "intel_engine_regs.h"
18 #include "intel_engine_user.h"
19 #include "intel_execlists_submission.h"
20 #include "intel_gt.h"
21 #include "intel_gt_requests.h"
22 #include "intel_gt_pm.h"
23 #include "intel_lrc_reg.h"
24 #include "intel_reset.h"
25 #include "intel_ring.h"
26 #include "uc/intel_guc_submission.h"
27 
28 /* Haswell does have the CXT_SIZE register however it does not appear to be
29  * valid. Now, docs explain in dwords what is in the context object. The full
30  * size is 70720 bytes, however, the power context and execlist context will
31  * never be saved (power context is stored elsewhere, and execlists don't work
32  * on HSW) - so the final size, including the extra state required for the
33  * Resource Streamer, is 66944 bytes, which rounds to 17 pages.
34  */
35 #define HSW_CXT_TOTAL_SIZE		(17 * PAGE_SIZE)
36 
37 #define DEFAULT_LR_CONTEXT_RENDER_SIZE	(22 * PAGE_SIZE)
38 #define GEN8_LR_CONTEXT_RENDER_SIZE	(20 * PAGE_SIZE)
39 #define GEN9_LR_CONTEXT_RENDER_SIZE	(22 * PAGE_SIZE)
40 #define GEN11_LR_CONTEXT_RENDER_SIZE	(14 * PAGE_SIZE)
41 
42 #define GEN8_LR_CONTEXT_OTHER_SIZE	( 2 * PAGE_SIZE)
43 
44 #define MAX_MMIO_BASES 3
45 struct engine_info {
46 	u8 class;
47 	u8 instance;
48 	/* mmio bases table *must* be sorted in reverse graphics_ver order */
49 	struct engine_mmio_base {
50 		u32 graphics_ver : 8;
51 		u32 base : 24;
52 	} mmio_bases[MAX_MMIO_BASES];
53 };
54 
55 static const struct engine_info intel_engines[] = {
56 	[RCS0] = {
57 		.class = RENDER_CLASS,
58 		.instance = 0,
59 		.mmio_bases = {
60 			{ .graphics_ver = 1, .base = RENDER_RING_BASE }
61 		},
62 	},
63 	[BCS0] = {
64 		.class = COPY_ENGINE_CLASS,
65 		.instance = 0,
66 		.mmio_bases = {
67 			{ .graphics_ver = 6, .base = BLT_RING_BASE }
68 		},
69 	},
70 	[VCS0] = {
71 		.class = VIDEO_DECODE_CLASS,
72 		.instance = 0,
73 		.mmio_bases = {
74 			{ .graphics_ver = 11, .base = GEN11_BSD_RING_BASE },
75 			{ .graphics_ver = 6, .base = GEN6_BSD_RING_BASE },
76 			{ .graphics_ver = 4, .base = BSD_RING_BASE }
77 		},
78 	},
79 	[VCS1] = {
80 		.class = VIDEO_DECODE_CLASS,
81 		.instance = 1,
82 		.mmio_bases = {
83 			{ .graphics_ver = 11, .base = GEN11_BSD2_RING_BASE },
84 			{ .graphics_ver = 8, .base = GEN8_BSD2_RING_BASE }
85 		},
86 	},
87 	[VCS2] = {
88 		.class = VIDEO_DECODE_CLASS,
89 		.instance = 2,
90 		.mmio_bases = {
91 			{ .graphics_ver = 11, .base = GEN11_BSD3_RING_BASE }
92 		},
93 	},
94 	[VCS3] = {
95 		.class = VIDEO_DECODE_CLASS,
96 		.instance = 3,
97 		.mmio_bases = {
98 			{ .graphics_ver = 11, .base = GEN11_BSD4_RING_BASE }
99 		},
100 	},
101 	[VCS4] = {
102 		.class = VIDEO_DECODE_CLASS,
103 		.instance = 4,
104 		.mmio_bases = {
105 			{ .graphics_ver = 12, .base = XEHP_BSD5_RING_BASE }
106 		},
107 	},
108 	[VCS5] = {
109 		.class = VIDEO_DECODE_CLASS,
110 		.instance = 5,
111 		.mmio_bases = {
112 			{ .graphics_ver = 12, .base = XEHP_BSD6_RING_BASE }
113 		},
114 	},
115 	[VCS6] = {
116 		.class = VIDEO_DECODE_CLASS,
117 		.instance = 6,
118 		.mmio_bases = {
119 			{ .graphics_ver = 12, .base = XEHP_BSD7_RING_BASE }
120 		},
121 	},
122 	[VCS7] = {
123 		.class = VIDEO_DECODE_CLASS,
124 		.instance = 7,
125 		.mmio_bases = {
126 			{ .graphics_ver = 12, .base = XEHP_BSD8_RING_BASE }
127 		},
128 	},
129 	[VECS0] = {
130 		.class = VIDEO_ENHANCEMENT_CLASS,
131 		.instance = 0,
132 		.mmio_bases = {
133 			{ .graphics_ver = 11, .base = GEN11_VEBOX_RING_BASE },
134 			{ .graphics_ver = 7, .base = VEBOX_RING_BASE }
135 		},
136 	},
137 	[VECS1] = {
138 		.class = VIDEO_ENHANCEMENT_CLASS,
139 		.instance = 1,
140 		.mmio_bases = {
141 			{ .graphics_ver = 11, .base = GEN11_VEBOX2_RING_BASE }
142 		},
143 	},
144 	[VECS2] = {
145 		.class = VIDEO_ENHANCEMENT_CLASS,
146 		.instance = 2,
147 		.mmio_bases = {
148 			{ .graphics_ver = 12, .base = XEHP_VEBOX3_RING_BASE }
149 		},
150 	},
151 	[VECS3] = {
152 		.class = VIDEO_ENHANCEMENT_CLASS,
153 		.instance = 3,
154 		.mmio_bases = {
155 			{ .graphics_ver = 12, .base = XEHP_VEBOX4_RING_BASE }
156 		},
157 	},
158 };
159 
160 /**
161  * intel_engine_context_size() - return the size of the context for an engine
162  * @gt: the gt
163  * @class: engine class
164  *
165  * Each engine class may require a different amount of space for a context
166  * image.
167  *
168  * Return: size (in bytes) of an engine class specific context image
169  *
170  * Note: this size includes the HWSP, which is part of the context image
171  * in LRC mode, but does not include the "shared data page" used with
172  * GuC submission. The caller should account for this if using the GuC.
173  */
174 u32 intel_engine_context_size(struct intel_gt *gt, u8 class)
175 {
176 	struct intel_uncore *uncore = gt->uncore;
177 	u32 cxt_size;
178 
179 	BUILD_BUG_ON(I915_GTT_PAGE_SIZE != PAGE_SIZE);
180 
181 	switch (class) {
182 	case RENDER_CLASS:
183 		switch (GRAPHICS_VER(gt->i915)) {
184 		default:
185 			MISSING_CASE(GRAPHICS_VER(gt->i915));
186 			return DEFAULT_LR_CONTEXT_RENDER_SIZE;
187 		case 12:
188 		case 11:
189 			return GEN11_LR_CONTEXT_RENDER_SIZE;
190 		case 9:
191 			return GEN9_LR_CONTEXT_RENDER_SIZE;
192 		case 8:
193 			return GEN8_LR_CONTEXT_RENDER_SIZE;
194 		case 7:
195 			if (IS_HASWELL(gt->i915))
196 				return HSW_CXT_TOTAL_SIZE;
197 
198 			cxt_size = intel_uncore_read(uncore, GEN7_CXT_SIZE);
199 			return round_up(GEN7_CXT_TOTAL_SIZE(cxt_size) * 64,
200 					PAGE_SIZE);
201 		case 6:
202 			cxt_size = intel_uncore_read(uncore, CXT_SIZE);
203 			return round_up(GEN6_CXT_TOTAL_SIZE(cxt_size) * 64,
204 					PAGE_SIZE);
205 		case 5:
206 		case 4:
207 			/*
208 			 * There is a discrepancy here between the size reported
209 			 * by the register and the size of the context layout
210 			 * in the docs. Both are described as authorative!
211 			 *
212 			 * The discrepancy is on the order of a few cachelines,
213 			 * but the total is under one page (4k), which is our
214 			 * minimum allocation anyway so it should all come
215 			 * out in the wash.
216 			 */
217 			cxt_size = intel_uncore_read(uncore, CXT_SIZE) + 1;
218 			drm_dbg(&gt->i915->drm,
219 				"graphics_ver = %d CXT_SIZE = %d bytes [0x%08x]\n",
220 				GRAPHICS_VER(gt->i915), cxt_size * 64,
221 				cxt_size - 1);
222 			return round_up(cxt_size * 64, PAGE_SIZE);
223 		case 3:
224 		case 2:
225 		/* For the special day when i810 gets merged. */
226 		case 1:
227 			return 0;
228 		}
229 		break;
230 	default:
231 		MISSING_CASE(class);
232 		fallthrough;
233 	case VIDEO_DECODE_CLASS:
234 	case VIDEO_ENHANCEMENT_CLASS:
235 	case COPY_ENGINE_CLASS:
236 		if (GRAPHICS_VER(gt->i915) < 8)
237 			return 0;
238 		return GEN8_LR_CONTEXT_OTHER_SIZE;
239 	}
240 }
241 
242 static u32 __engine_mmio_base(struct drm_i915_private *i915,
243 			      const struct engine_mmio_base *bases)
244 {
245 	int i;
246 
247 	for (i = 0; i < MAX_MMIO_BASES; i++)
248 		if (GRAPHICS_VER(i915) >= bases[i].graphics_ver)
249 			break;
250 
251 	GEM_BUG_ON(i == MAX_MMIO_BASES);
252 	GEM_BUG_ON(!bases[i].base);
253 
254 	return bases[i].base;
255 }
256 
257 static void __sprint_engine_name(struct intel_engine_cs *engine)
258 {
259 	/*
260 	 * Before we know what the uABI name for this engine will be,
261 	 * we still would like to keep track of this engine in the debug logs.
262 	 * We throw in a ' here as a reminder that this isn't its final name.
263 	 */
264 	GEM_WARN_ON(snprintf(engine->name, sizeof(engine->name), "%s'%u",
265 			     intel_engine_class_repr(engine->class),
266 			     engine->instance) >= sizeof(engine->name));
267 }
268 
269 void intel_engine_set_hwsp_writemask(struct intel_engine_cs *engine, u32 mask)
270 {
271 	/*
272 	 * Though they added more rings on g4x/ilk, they did not add
273 	 * per-engine HWSTAM until gen6.
274 	 */
275 	if (GRAPHICS_VER(engine->i915) < 6 && engine->class != RENDER_CLASS)
276 		return;
277 
278 	if (GRAPHICS_VER(engine->i915) >= 3)
279 		ENGINE_WRITE(engine, RING_HWSTAM, mask);
280 	else
281 		ENGINE_WRITE16(engine, RING_HWSTAM, mask);
282 }
283 
284 static void intel_engine_sanitize_mmio(struct intel_engine_cs *engine)
285 {
286 	/* Mask off all writes into the unknown HWSP */
287 	intel_engine_set_hwsp_writemask(engine, ~0u);
288 }
289 
290 static void nop_irq_handler(struct intel_engine_cs *engine, u16 iir)
291 {
292 	GEM_DEBUG_WARN_ON(iir);
293 }
294 
295 static int intel_engine_setup(struct intel_gt *gt, enum intel_engine_id id,
296 			      u8 logical_instance)
297 {
298 	const struct engine_info *info = &intel_engines[id];
299 	struct drm_i915_private *i915 = gt->i915;
300 	struct intel_engine_cs *engine;
301 	u8 guc_class;
302 
303 	BUILD_BUG_ON(MAX_ENGINE_CLASS >= BIT(GEN11_ENGINE_CLASS_WIDTH));
304 	BUILD_BUG_ON(MAX_ENGINE_INSTANCE >= BIT(GEN11_ENGINE_INSTANCE_WIDTH));
305 	BUILD_BUG_ON(I915_MAX_VCS > (MAX_ENGINE_INSTANCE + 1));
306 	BUILD_BUG_ON(I915_MAX_VECS > (MAX_ENGINE_INSTANCE + 1));
307 
308 	if (GEM_DEBUG_WARN_ON(id >= ARRAY_SIZE(gt->engine)))
309 		return -EINVAL;
310 
311 	if (GEM_DEBUG_WARN_ON(info->class > MAX_ENGINE_CLASS))
312 		return -EINVAL;
313 
314 	if (GEM_DEBUG_WARN_ON(info->instance > MAX_ENGINE_INSTANCE))
315 		return -EINVAL;
316 
317 	if (GEM_DEBUG_WARN_ON(gt->engine_class[info->class][info->instance]))
318 		return -EINVAL;
319 
320 	engine = kzalloc(sizeof(*engine), GFP_KERNEL);
321 	if (!engine)
322 		return -ENOMEM;
323 
324 	BUILD_BUG_ON(BITS_PER_TYPE(engine->mask) < I915_NUM_ENGINES);
325 
326 	INIT_LIST_HEAD(&engine->pinned_contexts_list);
327 	engine->id = id;
328 	engine->legacy_idx = INVALID_ENGINE;
329 	engine->mask = BIT(id);
330 	if (GRAPHICS_VER(gt->i915) >= 11) {
331 		static const u32 engine_reset_domains[] = {
332 			[RCS0]  = GEN11_GRDOM_RENDER,
333 			[BCS0]  = GEN11_GRDOM_BLT,
334 			[VCS0]  = GEN11_GRDOM_MEDIA,
335 			[VCS1]  = GEN11_GRDOM_MEDIA2,
336 			[VCS2]  = GEN11_GRDOM_MEDIA3,
337 			[VCS3]  = GEN11_GRDOM_MEDIA4,
338 			[VCS4]  = GEN11_GRDOM_MEDIA5,
339 			[VCS5]  = GEN11_GRDOM_MEDIA6,
340 			[VCS6]  = GEN11_GRDOM_MEDIA7,
341 			[VCS7]  = GEN11_GRDOM_MEDIA8,
342 			[VECS0] = GEN11_GRDOM_VECS,
343 			[VECS1] = GEN11_GRDOM_VECS2,
344 			[VECS2] = GEN11_GRDOM_VECS3,
345 			[VECS3] = GEN11_GRDOM_VECS4,
346 		};
347 		GEM_BUG_ON(id >= ARRAY_SIZE(engine_reset_domains) ||
348 			   !engine_reset_domains[id]);
349 		engine->reset_domain = engine_reset_domains[id];
350 	} else {
351 		static const u32 engine_reset_domains[] = {
352 			[RCS0]  = GEN6_GRDOM_RENDER,
353 			[BCS0]  = GEN6_GRDOM_BLT,
354 			[VCS0]  = GEN6_GRDOM_MEDIA,
355 			[VCS1]  = GEN8_GRDOM_MEDIA2,
356 			[VECS0] = GEN6_GRDOM_VECS,
357 		};
358 		GEM_BUG_ON(id >= ARRAY_SIZE(engine_reset_domains) ||
359 			   !engine_reset_domains[id]);
360 		engine->reset_domain = engine_reset_domains[id];
361 	}
362 	engine->i915 = i915;
363 	engine->gt = gt;
364 	engine->uncore = gt->uncore;
365 	guc_class = engine_class_to_guc_class(info->class);
366 	engine->guc_id = MAKE_GUC_ID(guc_class, info->instance);
367 	engine->mmio_base = __engine_mmio_base(i915, info->mmio_bases);
368 
369 	engine->irq_handler = nop_irq_handler;
370 
371 	engine->class = info->class;
372 	engine->instance = info->instance;
373 	engine->logical_mask = BIT(logical_instance);
374 	__sprint_engine_name(engine);
375 
376 	engine->props.heartbeat_interval_ms =
377 		CONFIG_DRM_I915_HEARTBEAT_INTERVAL;
378 	engine->props.max_busywait_duration_ns =
379 		CONFIG_DRM_I915_MAX_REQUEST_BUSYWAIT;
380 	engine->props.preempt_timeout_ms =
381 		CONFIG_DRM_I915_PREEMPT_TIMEOUT;
382 	engine->props.stop_timeout_ms =
383 		CONFIG_DRM_I915_STOP_TIMEOUT;
384 	engine->props.timeslice_duration_ms =
385 		CONFIG_DRM_I915_TIMESLICE_DURATION;
386 
387 	/* Override to uninterruptible for OpenCL workloads. */
388 	if (GRAPHICS_VER(i915) == 12 && engine->class == RENDER_CLASS)
389 		engine->props.preempt_timeout_ms = 0;
390 
391 	engine->defaults = engine->props; /* never to change again */
392 
393 	engine->context_size = intel_engine_context_size(gt, engine->class);
394 	if (WARN_ON(engine->context_size > BIT(20)))
395 		engine->context_size = 0;
396 	if (engine->context_size)
397 		DRIVER_CAPS(i915)->has_logical_contexts = true;
398 
399 	ewma__engine_latency_init(&engine->latency);
400 	seqcount_init(&engine->stats.execlists.lock);
401 
402 	ATOMIC_INIT_NOTIFIER_HEAD(&engine->context_status_notifier);
403 
404 	/* Scrub mmio state on takeover */
405 	intel_engine_sanitize_mmio(engine);
406 
407 	gt->engine_class[info->class][info->instance] = engine;
408 	gt->engine[id] = engine;
409 
410 	return 0;
411 }
412 
413 static void __setup_engine_capabilities(struct intel_engine_cs *engine)
414 {
415 	struct drm_i915_private *i915 = engine->i915;
416 
417 	if (engine->class == VIDEO_DECODE_CLASS) {
418 		/*
419 		 * HEVC support is present on first engine instance
420 		 * before Gen11 and on all instances afterwards.
421 		 */
422 		if (GRAPHICS_VER(i915) >= 11 ||
423 		    (GRAPHICS_VER(i915) >= 9 && engine->instance == 0))
424 			engine->uabi_capabilities |=
425 				I915_VIDEO_CLASS_CAPABILITY_HEVC;
426 
427 		/*
428 		 * SFC block is present only on even logical engine
429 		 * instances.
430 		 */
431 		if ((GRAPHICS_VER(i915) >= 11 &&
432 		     (engine->gt->info.vdbox_sfc_access &
433 		      BIT(engine->instance))) ||
434 		    (GRAPHICS_VER(i915) >= 9 && engine->instance == 0))
435 			engine->uabi_capabilities |=
436 				I915_VIDEO_AND_ENHANCE_CLASS_CAPABILITY_SFC;
437 	} else if (engine->class == VIDEO_ENHANCEMENT_CLASS) {
438 		if (GRAPHICS_VER(i915) >= 9 &&
439 		    engine->gt->info.sfc_mask & BIT(engine->instance))
440 			engine->uabi_capabilities |=
441 				I915_VIDEO_AND_ENHANCE_CLASS_CAPABILITY_SFC;
442 	}
443 }
444 
445 static void intel_setup_engine_capabilities(struct intel_gt *gt)
446 {
447 	struct intel_engine_cs *engine;
448 	enum intel_engine_id id;
449 
450 	for_each_engine(engine, gt, id)
451 		__setup_engine_capabilities(engine);
452 }
453 
454 /**
455  * intel_engines_release() - free the resources allocated for Command Streamers
456  * @gt: pointer to struct intel_gt
457  */
458 void intel_engines_release(struct intel_gt *gt)
459 {
460 	struct intel_engine_cs *engine;
461 	enum intel_engine_id id;
462 
463 	/*
464 	 * Before we release the resources held by engine, we must be certain
465 	 * that the HW is no longer accessing them -- having the GPU scribble
466 	 * to or read from a page being used for something else causes no end
467 	 * of fun.
468 	 *
469 	 * The GPU should be reset by this point, but assume the worst just
470 	 * in case we aborted before completely initialising the engines.
471 	 */
472 	GEM_BUG_ON(intel_gt_pm_is_awake(gt));
473 	if (!INTEL_INFO(gt->i915)->gpu_reset_clobbers_display)
474 		__intel_gt_reset(gt, ALL_ENGINES);
475 
476 	/* Decouple the backend; but keep the layout for late GPU resets */
477 	for_each_engine(engine, gt, id) {
478 		if (!engine->release)
479 			continue;
480 
481 		intel_wakeref_wait_for_idle(&engine->wakeref);
482 		GEM_BUG_ON(intel_engine_pm_is_awake(engine));
483 
484 		engine->release(engine);
485 		engine->release = NULL;
486 
487 		memset(&engine->reset, 0, sizeof(engine->reset));
488 	}
489 }
490 
491 void intel_engine_free_request_pool(struct intel_engine_cs *engine)
492 {
493 	if (!engine->request_pool)
494 		return;
495 
496 	kmem_cache_free(i915_request_slab_cache(), engine->request_pool);
497 }
498 
499 void intel_engines_free(struct intel_gt *gt)
500 {
501 	struct intel_engine_cs *engine;
502 	enum intel_engine_id id;
503 
504 	/* Free the requests! dma-resv keeps fences around for an eternity */
505 	rcu_barrier();
506 
507 	for_each_engine(engine, gt, id) {
508 		intel_engine_free_request_pool(engine);
509 		kfree(engine);
510 		gt->engine[id] = NULL;
511 	}
512 }
513 
514 static
515 bool gen11_vdbox_has_sfc(struct intel_gt *gt,
516 			 unsigned int physical_vdbox,
517 			 unsigned int logical_vdbox, u16 vdbox_mask)
518 {
519 	struct drm_i915_private *i915 = gt->i915;
520 
521 	/*
522 	 * In Gen11, only even numbered logical VDBOXes are hooked
523 	 * up to an SFC (Scaler & Format Converter) unit.
524 	 * In Gen12, Even numbered physical instance always are connected
525 	 * to an SFC. Odd numbered physical instances have SFC only if
526 	 * previous even instance is fused off.
527 	 *
528 	 * Starting with Xe_HP, there's also a dedicated SFC_ENABLE field
529 	 * in the fuse register that tells us whether a specific SFC is present.
530 	 */
531 	if ((gt->info.sfc_mask & BIT(physical_vdbox / 2)) == 0)
532 		return false;
533 	else if (GRAPHICS_VER(i915) == 12)
534 		return (physical_vdbox % 2 == 0) ||
535 			!(BIT(physical_vdbox - 1) & vdbox_mask);
536 	else if (GRAPHICS_VER(i915) == 11)
537 		return logical_vdbox % 2 == 0;
538 
539 	MISSING_CASE(GRAPHICS_VER(i915));
540 	return false;
541 }
542 
543 /*
544  * Determine which engines are fused off in our particular hardware.
545  * Note that we have a catch-22 situation where we need to be able to access
546  * the blitter forcewake domain to read the engine fuses, but at the same time
547  * we need to know which engines are available on the system to know which
548  * forcewake domains are present. We solve this by intializing the forcewake
549  * domains based on the full engine mask in the platform capabilities before
550  * calling this function and pruning the domains for fused-off engines
551  * afterwards.
552  */
553 static intel_engine_mask_t init_engine_mask(struct intel_gt *gt)
554 {
555 	struct drm_i915_private *i915 = gt->i915;
556 	struct intel_gt_info *info = &gt->info;
557 	struct intel_uncore *uncore = gt->uncore;
558 	unsigned int logical_vdbox = 0;
559 	unsigned int i;
560 	u32 media_fuse, fuse1;
561 	u16 vdbox_mask;
562 	u16 vebox_mask;
563 
564 	info->engine_mask = INTEL_INFO(i915)->platform_engine_mask;
565 
566 	if (GRAPHICS_VER(i915) < 11)
567 		return info->engine_mask;
568 
569 	/*
570 	 * On newer platforms the fusing register is called 'enable' and has
571 	 * enable semantics, while on older platforms it is called 'disable'
572 	 * and bits have disable semantices.
573 	 */
574 	media_fuse = intel_uncore_read(uncore, GEN11_GT_VEBOX_VDBOX_DISABLE);
575 	if (GRAPHICS_VER_FULL(i915) < IP_VER(12, 50))
576 		media_fuse = ~media_fuse;
577 
578 	vdbox_mask = media_fuse & GEN11_GT_VDBOX_DISABLE_MASK;
579 	vebox_mask = (media_fuse & GEN11_GT_VEBOX_DISABLE_MASK) >>
580 		      GEN11_GT_VEBOX_DISABLE_SHIFT;
581 
582 	if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 50)) {
583 		fuse1 = intel_uncore_read(uncore, HSW_PAVP_FUSE1);
584 		gt->info.sfc_mask = REG_FIELD_GET(XEHP_SFC_ENABLE_MASK, fuse1);
585 	} else {
586 		gt->info.sfc_mask = ~0;
587 	}
588 
589 	for (i = 0; i < I915_MAX_VCS; i++) {
590 		if (!HAS_ENGINE(gt, _VCS(i))) {
591 			vdbox_mask &= ~BIT(i);
592 			continue;
593 		}
594 
595 		if (!(BIT(i) & vdbox_mask)) {
596 			info->engine_mask &= ~BIT(_VCS(i));
597 			drm_dbg(&i915->drm, "vcs%u fused off\n", i);
598 			continue;
599 		}
600 
601 		if (gen11_vdbox_has_sfc(gt, i, logical_vdbox, vdbox_mask))
602 			gt->info.vdbox_sfc_access |= BIT(i);
603 		logical_vdbox++;
604 	}
605 	drm_dbg(&i915->drm, "vdbox enable: %04x, instances: %04lx\n",
606 		vdbox_mask, VDBOX_MASK(gt));
607 	GEM_BUG_ON(vdbox_mask != VDBOX_MASK(gt));
608 
609 	for (i = 0; i < I915_MAX_VECS; i++) {
610 		if (!HAS_ENGINE(gt, _VECS(i))) {
611 			vebox_mask &= ~BIT(i);
612 			continue;
613 		}
614 
615 		if (!(BIT(i) & vebox_mask)) {
616 			info->engine_mask &= ~BIT(_VECS(i));
617 			drm_dbg(&i915->drm, "vecs%u fused off\n", i);
618 		}
619 	}
620 	drm_dbg(&i915->drm, "vebox enable: %04x, instances: %04lx\n",
621 		vebox_mask, VEBOX_MASK(gt));
622 	GEM_BUG_ON(vebox_mask != VEBOX_MASK(gt));
623 
624 	return info->engine_mask;
625 }
626 
627 static void populate_logical_ids(struct intel_gt *gt, u8 *logical_ids,
628 				 u8 class, const u8 *map, u8 num_instances)
629 {
630 	int i, j;
631 	u8 current_logical_id = 0;
632 
633 	for (j = 0; j < num_instances; ++j) {
634 		for (i = 0; i < ARRAY_SIZE(intel_engines); ++i) {
635 			if (!HAS_ENGINE(gt, i) ||
636 			    intel_engines[i].class != class)
637 				continue;
638 
639 			if (intel_engines[i].instance == map[j]) {
640 				logical_ids[intel_engines[i].instance] =
641 					current_logical_id++;
642 				break;
643 			}
644 		}
645 	}
646 }
647 
648 static void setup_logical_ids(struct intel_gt *gt, u8 *logical_ids, u8 class)
649 {
650 	int i;
651 	u8 map[MAX_ENGINE_INSTANCE + 1];
652 
653 	for (i = 0; i < MAX_ENGINE_INSTANCE + 1; ++i)
654 		map[i] = i;
655 	populate_logical_ids(gt, logical_ids, class, map, ARRAY_SIZE(map));
656 }
657 
658 /**
659  * intel_engines_init_mmio() - allocate and prepare the Engine Command Streamers
660  * @gt: pointer to struct intel_gt
661  *
662  * Return: non-zero if the initialization failed.
663  */
664 int intel_engines_init_mmio(struct intel_gt *gt)
665 {
666 	struct drm_i915_private *i915 = gt->i915;
667 	const unsigned int engine_mask = init_engine_mask(gt);
668 	unsigned int mask = 0;
669 	unsigned int i, class;
670 	u8 logical_ids[MAX_ENGINE_INSTANCE + 1];
671 	int err;
672 
673 	drm_WARN_ON(&i915->drm, engine_mask == 0);
674 	drm_WARN_ON(&i915->drm, engine_mask &
675 		    GENMASK(BITS_PER_TYPE(mask) - 1, I915_NUM_ENGINES));
676 
677 	if (i915_inject_probe_failure(i915))
678 		return -ENODEV;
679 
680 	for (class = 0; class < MAX_ENGINE_CLASS + 1; ++class) {
681 		setup_logical_ids(gt, logical_ids, class);
682 
683 		for (i = 0; i < ARRAY_SIZE(intel_engines); ++i) {
684 			u8 instance = intel_engines[i].instance;
685 
686 			if (intel_engines[i].class != class ||
687 			    !HAS_ENGINE(gt, i))
688 				continue;
689 
690 			err = intel_engine_setup(gt, i,
691 						 logical_ids[instance]);
692 			if (err)
693 				goto cleanup;
694 
695 			mask |= BIT(i);
696 		}
697 	}
698 
699 	/*
700 	 * Catch failures to update intel_engines table when the new engines
701 	 * are added to the driver by a warning and disabling the forgotten
702 	 * engines.
703 	 */
704 	if (drm_WARN_ON(&i915->drm, mask != engine_mask))
705 		gt->info.engine_mask = mask;
706 
707 	gt->info.num_engines = hweight32(mask);
708 
709 	intel_gt_check_and_clear_faults(gt);
710 
711 	intel_setup_engine_capabilities(gt);
712 
713 	intel_uncore_prune_engine_fw_domains(gt->uncore, gt);
714 
715 	return 0;
716 
717 cleanup:
718 	intel_engines_free(gt);
719 	return err;
720 }
721 
722 void intel_engine_init_execlists(struct intel_engine_cs *engine)
723 {
724 	struct intel_engine_execlists * const execlists = &engine->execlists;
725 
726 	execlists->port_mask = 1;
727 	GEM_BUG_ON(!is_power_of_2(execlists_num_ports(execlists)));
728 	GEM_BUG_ON(execlists_num_ports(execlists) > EXECLIST_MAX_PORTS);
729 
730 	memset(execlists->pending, 0, sizeof(execlists->pending));
731 	execlists->active =
732 		memset(execlists->inflight, 0, sizeof(execlists->inflight));
733 }
734 
735 static void cleanup_status_page(struct intel_engine_cs *engine)
736 {
737 	struct i915_vma *vma;
738 
739 	/* Prevent writes into HWSP after returning the page to the system */
740 	intel_engine_set_hwsp_writemask(engine, ~0u);
741 
742 	vma = fetch_and_zero(&engine->status_page.vma);
743 	if (!vma)
744 		return;
745 
746 	if (!HWS_NEEDS_PHYSICAL(engine->i915))
747 		i915_vma_unpin(vma);
748 
749 	i915_gem_object_unpin_map(vma->obj);
750 	i915_gem_object_put(vma->obj);
751 }
752 
753 static int pin_ggtt_status_page(struct intel_engine_cs *engine,
754 				struct i915_gem_ww_ctx *ww,
755 				struct i915_vma *vma)
756 {
757 	unsigned int flags;
758 
759 	if (!HAS_LLC(engine->i915) && i915_ggtt_has_aperture(engine->gt->ggtt))
760 		/*
761 		 * On g33, we cannot place HWS above 256MiB, so
762 		 * restrict its pinning to the low mappable arena.
763 		 * Though this restriction is not documented for
764 		 * gen4, gen5, or byt, they also behave similarly
765 		 * and hang if the HWS is placed at the top of the
766 		 * GTT. To generalise, it appears that all !llc
767 		 * platforms have issues with us placing the HWS
768 		 * above the mappable region (even though we never
769 		 * actually map it).
770 		 */
771 		flags = PIN_MAPPABLE;
772 	else
773 		flags = PIN_HIGH;
774 
775 	return i915_ggtt_pin(vma, ww, 0, flags);
776 }
777 
778 static int init_status_page(struct intel_engine_cs *engine)
779 {
780 	struct drm_i915_gem_object *obj;
781 	struct i915_gem_ww_ctx ww;
782 	struct i915_vma *vma;
783 	void *vaddr;
784 	int ret;
785 
786 	INIT_LIST_HEAD(&engine->status_page.timelines);
787 
788 	/*
789 	 * Though the HWS register does support 36bit addresses, historically
790 	 * we have had hangs and corruption reported due to wild writes if
791 	 * the HWS is placed above 4G. We only allow objects to be allocated
792 	 * in GFP_DMA32 for i965, and no earlier physical address users had
793 	 * access to more than 4G.
794 	 */
795 	obj = i915_gem_object_create_internal(engine->i915, PAGE_SIZE);
796 	if (IS_ERR(obj)) {
797 		drm_err(&engine->i915->drm,
798 			"Failed to allocate status page\n");
799 		return PTR_ERR(obj);
800 	}
801 
802 	i915_gem_object_set_cache_coherency(obj, I915_CACHE_LLC);
803 
804 	vma = i915_vma_instance(obj, &engine->gt->ggtt->vm, NULL);
805 	if (IS_ERR(vma)) {
806 		ret = PTR_ERR(vma);
807 		goto err_put;
808 	}
809 
810 	i915_gem_ww_ctx_init(&ww, true);
811 retry:
812 	ret = i915_gem_object_lock(obj, &ww);
813 	if (!ret && !HWS_NEEDS_PHYSICAL(engine->i915))
814 		ret = pin_ggtt_status_page(engine, &ww, vma);
815 	if (ret)
816 		goto err;
817 
818 	vaddr = i915_gem_object_pin_map(obj, I915_MAP_WB);
819 	if (IS_ERR(vaddr)) {
820 		ret = PTR_ERR(vaddr);
821 		goto err_unpin;
822 	}
823 
824 	engine->status_page.addr = memset(vaddr, 0, PAGE_SIZE);
825 	engine->status_page.vma = vma;
826 
827 err_unpin:
828 	if (ret)
829 		i915_vma_unpin(vma);
830 err:
831 	if (ret == -EDEADLK) {
832 		ret = i915_gem_ww_ctx_backoff(&ww);
833 		if (!ret)
834 			goto retry;
835 	}
836 	i915_gem_ww_ctx_fini(&ww);
837 err_put:
838 	if (ret)
839 		i915_gem_object_put(obj);
840 	return ret;
841 }
842 
843 static int engine_setup_common(struct intel_engine_cs *engine)
844 {
845 	int err;
846 
847 	init_llist_head(&engine->barrier_tasks);
848 
849 	err = init_status_page(engine);
850 	if (err)
851 		return err;
852 
853 	engine->breadcrumbs = intel_breadcrumbs_create(engine);
854 	if (!engine->breadcrumbs) {
855 		err = -ENOMEM;
856 		goto err_status;
857 	}
858 
859 	engine->sched_engine = i915_sched_engine_create(ENGINE_PHYSICAL);
860 	if (!engine->sched_engine) {
861 		err = -ENOMEM;
862 		goto err_sched_engine;
863 	}
864 	engine->sched_engine->private_data = engine;
865 
866 	err = intel_engine_init_cmd_parser(engine);
867 	if (err)
868 		goto err_cmd_parser;
869 
870 	intel_engine_init_execlists(engine);
871 	intel_engine_init__pm(engine);
872 	intel_engine_init_retire(engine);
873 
874 	/* Use the whole device by default */
875 	engine->sseu =
876 		intel_sseu_from_device_info(&engine->gt->info.sseu);
877 
878 	intel_engine_init_workarounds(engine);
879 	intel_engine_init_whitelist(engine);
880 	intel_engine_init_ctx_wa(engine);
881 
882 	if (GRAPHICS_VER(engine->i915) >= 12)
883 		engine->flags |= I915_ENGINE_HAS_RELATIVE_MMIO;
884 
885 	return 0;
886 
887 err_cmd_parser:
888 	i915_sched_engine_put(engine->sched_engine);
889 err_sched_engine:
890 	intel_breadcrumbs_put(engine->breadcrumbs);
891 err_status:
892 	cleanup_status_page(engine);
893 	return err;
894 }
895 
896 struct measure_breadcrumb {
897 	struct i915_request rq;
898 	struct intel_ring ring;
899 	u32 cs[2048];
900 };
901 
902 static int measure_breadcrumb_dw(struct intel_context *ce)
903 {
904 	struct intel_engine_cs *engine = ce->engine;
905 	struct measure_breadcrumb *frame;
906 	int dw;
907 
908 	GEM_BUG_ON(!engine->gt->scratch);
909 
910 	frame = kzalloc(sizeof(*frame), GFP_KERNEL);
911 	if (!frame)
912 		return -ENOMEM;
913 
914 	frame->rq.engine = engine;
915 	frame->rq.context = ce;
916 	rcu_assign_pointer(frame->rq.timeline, ce->timeline);
917 	frame->rq.hwsp_seqno = ce->timeline->hwsp_seqno;
918 
919 	frame->ring.vaddr = frame->cs;
920 	frame->ring.size = sizeof(frame->cs);
921 	frame->ring.wrap =
922 		BITS_PER_TYPE(frame->ring.size) - ilog2(frame->ring.size);
923 	frame->ring.effective_size = frame->ring.size;
924 	intel_ring_update_space(&frame->ring);
925 	frame->rq.ring = &frame->ring;
926 
927 	mutex_lock(&ce->timeline->mutex);
928 	spin_lock_irq(&engine->sched_engine->lock);
929 
930 	dw = engine->emit_fini_breadcrumb(&frame->rq, frame->cs) - frame->cs;
931 
932 	spin_unlock_irq(&engine->sched_engine->lock);
933 	mutex_unlock(&ce->timeline->mutex);
934 
935 	GEM_BUG_ON(dw & 1); /* RING_TAIL must be qword aligned */
936 
937 	kfree(frame);
938 	return dw;
939 }
940 
941 struct intel_context *
942 intel_engine_create_pinned_context(struct intel_engine_cs *engine,
943 				   struct i915_address_space *vm,
944 				   unsigned int ring_size,
945 				   unsigned int hwsp,
946 				   struct lock_class_key *key,
947 				   const char *name)
948 {
949 	struct intel_context *ce;
950 	int err;
951 
952 	ce = intel_context_create(engine);
953 	if (IS_ERR(ce))
954 		return ce;
955 
956 	__set_bit(CONTEXT_BARRIER_BIT, &ce->flags);
957 	ce->timeline = page_pack_bits(NULL, hwsp);
958 	ce->ring = NULL;
959 	ce->ring_size = ring_size;
960 
961 	i915_vm_put(ce->vm);
962 	ce->vm = i915_vm_get(vm);
963 
964 	err = intel_context_pin(ce); /* perma-pin so it is always available */
965 	if (err) {
966 		intel_context_put(ce);
967 		return ERR_PTR(err);
968 	}
969 
970 	list_add_tail(&ce->pinned_contexts_link, &engine->pinned_contexts_list);
971 
972 	/*
973 	 * Give our perma-pinned kernel timelines a separate lockdep class,
974 	 * so that we can use them from within the normal user timelines
975 	 * should we need to inject GPU operations during their request
976 	 * construction.
977 	 */
978 	lockdep_set_class_and_name(&ce->timeline->mutex, key, name);
979 
980 	return ce;
981 }
982 
983 void intel_engine_destroy_pinned_context(struct intel_context *ce)
984 {
985 	struct intel_engine_cs *engine = ce->engine;
986 	struct i915_vma *hwsp = engine->status_page.vma;
987 
988 	GEM_BUG_ON(ce->timeline->hwsp_ggtt != hwsp);
989 
990 	mutex_lock(&hwsp->vm->mutex);
991 	list_del(&ce->timeline->engine_link);
992 	mutex_unlock(&hwsp->vm->mutex);
993 
994 	list_del(&ce->pinned_contexts_link);
995 	intel_context_unpin(ce);
996 	intel_context_put(ce);
997 }
998 
999 static struct intel_context *
1000 create_kernel_context(struct intel_engine_cs *engine)
1001 {
1002 	static struct lock_class_key kernel;
1003 
1004 	return intel_engine_create_pinned_context(engine, engine->gt->vm, SZ_4K,
1005 						  I915_GEM_HWS_SEQNO_ADDR,
1006 						  &kernel, "kernel_context");
1007 }
1008 
1009 /**
1010  * intel_engines_init_common - initialize cengine state which might require hw access
1011  * @engine: Engine to initialize.
1012  *
1013  * Initializes @engine@ structure members shared between legacy and execlists
1014  * submission modes which do require hardware access.
1015  *
1016  * Typcally done at later stages of submission mode specific engine setup.
1017  *
1018  * Returns zero on success or an error code on failure.
1019  */
1020 static int engine_init_common(struct intel_engine_cs *engine)
1021 {
1022 	struct intel_context *ce;
1023 	int ret;
1024 
1025 	engine->set_default_submission(engine);
1026 
1027 	/*
1028 	 * We may need to do things with the shrinker which
1029 	 * require us to immediately switch back to the default
1030 	 * context. This can cause a problem as pinning the
1031 	 * default context also requires GTT space which may not
1032 	 * be available. To avoid this we always pin the default
1033 	 * context.
1034 	 */
1035 	ce = create_kernel_context(engine);
1036 	if (IS_ERR(ce))
1037 		return PTR_ERR(ce);
1038 
1039 	ret = measure_breadcrumb_dw(ce);
1040 	if (ret < 0)
1041 		goto err_context;
1042 
1043 	engine->emit_fini_breadcrumb_dw = ret;
1044 	engine->kernel_context = ce;
1045 
1046 	return 0;
1047 
1048 err_context:
1049 	intel_engine_destroy_pinned_context(ce);
1050 	return ret;
1051 }
1052 
1053 int intel_engines_init(struct intel_gt *gt)
1054 {
1055 	int (*setup)(struct intel_engine_cs *engine);
1056 	struct intel_engine_cs *engine;
1057 	enum intel_engine_id id;
1058 	int err;
1059 
1060 	if (intel_uc_uses_guc_submission(&gt->uc)) {
1061 		gt->submission_method = INTEL_SUBMISSION_GUC;
1062 		setup = intel_guc_submission_setup;
1063 	} else if (HAS_EXECLISTS(gt->i915)) {
1064 		gt->submission_method = INTEL_SUBMISSION_ELSP;
1065 		setup = intel_execlists_submission_setup;
1066 	} else {
1067 		gt->submission_method = INTEL_SUBMISSION_RING;
1068 		setup = intel_ring_submission_setup;
1069 	}
1070 
1071 	for_each_engine(engine, gt, id) {
1072 		err = engine_setup_common(engine);
1073 		if (err)
1074 			return err;
1075 
1076 		err = setup(engine);
1077 		if (err)
1078 			return err;
1079 
1080 		err = engine_init_common(engine);
1081 		if (err)
1082 			return err;
1083 
1084 		intel_engine_add_user(engine);
1085 	}
1086 
1087 	return 0;
1088 }
1089 
1090 /**
1091  * intel_engines_cleanup_common - cleans up the engine state created by
1092  *                                the common initiailizers.
1093  * @engine: Engine to cleanup.
1094  *
1095  * This cleans up everything created by the common helpers.
1096  */
1097 void intel_engine_cleanup_common(struct intel_engine_cs *engine)
1098 {
1099 	GEM_BUG_ON(!list_empty(&engine->sched_engine->requests));
1100 
1101 	i915_sched_engine_put(engine->sched_engine);
1102 	intel_breadcrumbs_put(engine->breadcrumbs);
1103 
1104 	intel_engine_fini_retire(engine);
1105 	intel_engine_cleanup_cmd_parser(engine);
1106 
1107 	if (engine->default_state)
1108 		fput(engine->default_state);
1109 
1110 	if (engine->kernel_context)
1111 		intel_engine_destroy_pinned_context(engine->kernel_context);
1112 
1113 	GEM_BUG_ON(!llist_empty(&engine->barrier_tasks));
1114 	cleanup_status_page(engine);
1115 
1116 	intel_wa_list_free(&engine->ctx_wa_list);
1117 	intel_wa_list_free(&engine->wa_list);
1118 	intel_wa_list_free(&engine->whitelist);
1119 }
1120 
1121 /**
1122  * intel_engine_resume - re-initializes the HW state of the engine
1123  * @engine: Engine to resume.
1124  *
1125  * Returns zero on success or an error code on failure.
1126  */
1127 int intel_engine_resume(struct intel_engine_cs *engine)
1128 {
1129 	intel_engine_apply_workarounds(engine);
1130 	intel_engine_apply_whitelist(engine);
1131 
1132 	return engine->resume(engine);
1133 }
1134 
1135 u64 intel_engine_get_active_head(const struct intel_engine_cs *engine)
1136 {
1137 	struct drm_i915_private *i915 = engine->i915;
1138 
1139 	u64 acthd;
1140 
1141 	if (GRAPHICS_VER(i915) >= 8)
1142 		acthd = ENGINE_READ64(engine, RING_ACTHD, RING_ACTHD_UDW);
1143 	else if (GRAPHICS_VER(i915) >= 4)
1144 		acthd = ENGINE_READ(engine, RING_ACTHD);
1145 	else
1146 		acthd = ENGINE_READ(engine, ACTHD);
1147 
1148 	return acthd;
1149 }
1150 
1151 u64 intel_engine_get_last_batch_head(const struct intel_engine_cs *engine)
1152 {
1153 	u64 bbaddr;
1154 
1155 	if (GRAPHICS_VER(engine->i915) >= 8)
1156 		bbaddr = ENGINE_READ64(engine, RING_BBADDR, RING_BBADDR_UDW);
1157 	else
1158 		bbaddr = ENGINE_READ(engine, RING_BBADDR);
1159 
1160 	return bbaddr;
1161 }
1162 
1163 static unsigned long stop_timeout(const struct intel_engine_cs *engine)
1164 {
1165 	if (in_atomic() || irqs_disabled()) /* inside atomic preempt-reset? */
1166 		return 0;
1167 
1168 	/*
1169 	 * If we are doing a normal GPU reset, we can take our time and allow
1170 	 * the engine to quiesce. We've stopped submission to the engine, and
1171 	 * if we wait long enough an innocent context should complete and
1172 	 * leave the engine idle. So they should not be caught unaware by
1173 	 * the forthcoming GPU reset (which usually follows the stop_cs)!
1174 	 */
1175 	return READ_ONCE(engine->props.stop_timeout_ms);
1176 }
1177 
1178 static int __intel_engine_stop_cs(struct intel_engine_cs *engine,
1179 				  int fast_timeout_us,
1180 				  int slow_timeout_ms)
1181 {
1182 	struct intel_uncore *uncore = engine->uncore;
1183 	const i915_reg_t mode = RING_MI_MODE(engine->mmio_base);
1184 	int err;
1185 
1186 	intel_uncore_write_fw(uncore, mode, _MASKED_BIT_ENABLE(STOP_RING));
1187 	err = __intel_wait_for_register_fw(engine->uncore, mode,
1188 					   MODE_IDLE, MODE_IDLE,
1189 					   fast_timeout_us,
1190 					   slow_timeout_ms,
1191 					   NULL);
1192 
1193 	/* A final mmio read to let GPU writes be hopefully flushed to memory */
1194 	intel_uncore_posting_read_fw(uncore, mode);
1195 	return err;
1196 }
1197 
1198 int intel_engine_stop_cs(struct intel_engine_cs *engine)
1199 {
1200 	int err = 0;
1201 
1202 	if (GRAPHICS_VER(engine->i915) < 3)
1203 		return -ENODEV;
1204 
1205 	ENGINE_TRACE(engine, "\n");
1206 	if (__intel_engine_stop_cs(engine, 1000, stop_timeout(engine))) {
1207 		ENGINE_TRACE(engine,
1208 			     "timed out on STOP_RING -> IDLE; HEAD:%04x, TAIL:%04x\n",
1209 			     ENGINE_READ_FW(engine, RING_HEAD) & HEAD_ADDR,
1210 			     ENGINE_READ_FW(engine, RING_TAIL) & TAIL_ADDR);
1211 
1212 		/*
1213 		 * Sometimes we observe that the idle flag is not
1214 		 * set even though the ring is empty. So double
1215 		 * check before giving up.
1216 		 */
1217 		if ((ENGINE_READ_FW(engine, RING_HEAD) & HEAD_ADDR) !=
1218 		    (ENGINE_READ_FW(engine, RING_TAIL) & TAIL_ADDR))
1219 			err = -ETIMEDOUT;
1220 	}
1221 
1222 	return err;
1223 }
1224 
1225 void intel_engine_cancel_stop_cs(struct intel_engine_cs *engine)
1226 {
1227 	ENGINE_TRACE(engine, "\n");
1228 
1229 	ENGINE_WRITE_FW(engine, RING_MI_MODE, _MASKED_BIT_DISABLE(STOP_RING));
1230 }
1231 
1232 const char *i915_cache_level_str(struct drm_i915_private *i915, int type)
1233 {
1234 	switch (type) {
1235 	case I915_CACHE_NONE: return " uncached";
1236 	case I915_CACHE_LLC: return HAS_LLC(i915) ? " LLC" : " snooped";
1237 	case I915_CACHE_L3_LLC: return " L3+LLC";
1238 	case I915_CACHE_WT: return " WT";
1239 	default: return "";
1240 	}
1241 }
1242 
1243 static u32
1244 read_subslice_reg(const struct intel_engine_cs *engine,
1245 		  int slice, int subslice, i915_reg_t reg)
1246 {
1247 	return intel_uncore_read_with_mcr_steering(engine->uncore, reg,
1248 						   slice, subslice);
1249 }
1250 
1251 /* NB: please notice the memset */
1252 void intel_engine_get_instdone(const struct intel_engine_cs *engine,
1253 			       struct intel_instdone *instdone)
1254 {
1255 	struct drm_i915_private *i915 = engine->i915;
1256 	const struct sseu_dev_info *sseu = &engine->gt->info.sseu;
1257 	struct intel_uncore *uncore = engine->uncore;
1258 	u32 mmio_base = engine->mmio_base;
1259 	int slice;
1260 	int subslice;
1261 	int iter;
1262 
1263 	memset(instdone, 0, sizeof(*instdone));
1264 
1265 	if (GRAPHICS_VER(i915) >= 8) {
1266 		instdone->instdone =
1267 			intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
1268 
1269 		if (engine->id != RCS0)
1270 			return;
1271 
1272 		instdone->slice_common =
1273 			intel_uncore_read(uncore, GEN7_SC_INSTDONE);
1274 		if (GRAPHICS_VER(i915) >= 12) {
1275 			instdone->slice_common_extra[0] =
1276 				intel_uncore_read(uncore, GEN12_SC_INSTDONE_EXTRA);
1277 			instdone->slice_common_extra[1] =
1278 				intel_uncore_read(uncore, GEN12_SC_INSTDONE_EXTRA2);
1279 		}
1280 
1281 		if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 50)) {
1282 			for_each_instdone_gslice_dss_xehp(i915, sseu, iter, slice, subslice) {
1283 				instdone->sampler[slice][subslice] =
1284 					read_subslice_reg(engine, slice, subslice,
1285 							  GEN7_SAMPLER_INSTDONE);
1286 				instdone->row[slice][subslice] =
1287 					read_subslice_reg(engine, slice, subslice,
1288 							  GEN7_ROW_INSTDONE);
1289 			}
1290 		} else {
1291 			for_each_instdone_slice_subslice(i915, sseu, slice, subslice) {
1292 				instdone->sampler[slice][subslice] =
1293 					read_subslice_reg(engine, slice, subslice,
1294 							  GEN7_SAMPLER_INSTDONE);
1295 				instdone->row[slice][subslice] =
1296 					read_subslice_reg(engine, slice, subslice,
1297 							  GEN7_ROW_INSTDONE);
1298 			}
1299 		}
1300 
1301 		if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 55)) {
1302 			for_each_instdone_gslice_dss_xehp(i915, sseu, iter, slice, subslice)
1303 				instdone->geom_svg[slice][subslice] =
1304 					read_subslice_reg(engine, slice, subslice,
1305 							  XEHPG_INSTDONE_GEOM_SVG);
1306 		}
1307 	} else if (GRAPHICS_VER(i915) >= 7) {
1308 		instdone->instdone =
1309 			intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
1310 
1311 		if (engine->id != RCS0)
1312 			return;
1313 
1314 		instdone->slice_common =
1315 			intel_uncore_read(uncore, GEN7_SC_INSTDONE);
1316 		instdone->sampler[0][0] =
1317 			intel_uncore_read(uncore, GEN7_SAMPLER_INSTDONE);
1318 		instdone->row[0][0] =
1319 			intel_uncore_read(uncore, GEN7_ROW_INSTDONE);
1320 	} else if (GRAPHICS_VER(i915) >= 4) {
1321 		instdone->instdone =
1322 			intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
1323 		if (engine->id == RCS0)
1324 			/* HACK: Using the wrong struct member */
1325 			instdone->slice_common =
1326 				intel_uncore_read(uncore, GEN4_INSTDONE1);
1327 	} else {
1328 		instdone->instdone = intel_uncore_read(uncore, GEN2_INSTDONE);
1329 	}
1330 }
1331 
1332 static bool ring_is_idle(struct intel_engine_cs *engine)
1333 {
1334 	bool idle = true;
1335 
1336 	if (I915_SELFTEST_ONLY(!engine->mmio_base))
1337 		return true;
1338 
1339 	if (!intel_engine_pm_get_if_awake(engine))
1340 		return true;
1341 
1342 	/* First check that no commands are left in the ring */
1343 	if ((ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR) !=
1344 	    (ENGINE_READ(engine, RING_TAIL) & TAIL_ADDR))
1345 		idle = false;
1346 
1347 	/* No bit for gen2, so assume the CS parser is idle */
1348 	if (GRAPHICS_VER(engine->i915) > 2 &&
1349 	    !(ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE))
1350 		idle = false;
1351 
1352 	intel_engine_pm_put(engine);
1353 
1354 	return idle;
1355 }
1356 
1357 void __intel_engine_flush_submission(struct intel_engine_cs *engine, bool sync)
1358 {
1359 	struct tasklet_struct *t = &engine->sched_engine->tasklet;
1360 
1361 	if (!t->callback)
1362 		return;
1363 
1364 	local_bh_disable();
1365 	if (tasklet_trylock(t)) {
1366 		/* Must wait for any GPU reset in progress. */
1367 		if (__tasklet_is_enabled(t))
1368 			t->callback(t);
1369 		tasklet_unlock(t);
1370 	}
1371 	local_bh_enable();
1372 
1373 	/* Synchronise and wait for the tasklet on another CPU */
1374 	if (sync)
1375 		tasklet_unlock_wait(t);
1376 }
1377 
1378 /**
1379  * intel_engine_is_idle() - Report if the engine has finished process all work
1380  * @engine: the intel_engine_cs
1381  *
1382  * Return true if there are no requests pending, nothing left to be submitted
1383  * to hardware, and that the engine is idle.
1384  */
1385 bool intel_engine_is_idle(struct intel_engine_cs *engine)
1386 {
1387 	/* More white lies, if wedged, hw state is inconsistent */
1388 	if (intel_gt_is_wedged(engine->gt))
1389 		return true;
1390 
1391 	if (!intel_engine_pm_is_awake(engine))
1392 		return true;
1393 
1394 	/* Waiting to drain ELSP? */
1395 	intel_synchronize_hardirq(engine->i915);
1396 	intel_engine_flush_submission(engine);
1397 
1398 	/* ELSP is empty, but there are ready requests? E.g. after reset */
1399 	if (!i915_sched_engine_is_empty(engine->sched_engine))
1400 		return false;
1401 
1402 	/* Ring stopped? */
1403 	return ring_is_idle(engine);
1404 }
1405 
1406 bool intel_engines_are_idle(struct intel_gt *gt)
1407 {
1408 	struct intel_engine_cs *engine;
1409 	enum intel_engine_id id;
1410 
1411 	/*
1412 	 * If the driver is wedged, HW state may be very inconsistent and
1413 	 * report that it is still busy, even though we have stopped using it.
1414 	 */
1415 	if (intel_gt_is_wedged(gt))
1416 		return true;
1417 
1418 	/* Already parked (and passed an idleness test); must still be idle */
1419 	if (!READ_ONCE(gt->awake))
1420 		return true;
1421 
1422 	for_each_engine(engine, gt, id) {
1423 		if (!intel_engine_is_idle(engine))
1424 			return false;
1425 	}
1426 
1427 	return true;
1428 }
1429 
1430 bool intel_engine_irq_enable(struct intel_engine_cs *engine)
1431 {
1432 	if (!engine->irq_enable)
1433 		return false;
1434 
1435 	/* Caller disables interrupts */
1436 	spin_lock(&engine->gt->irq_lock);
1437 	engine->irq_enable(engine);
1438 	spin_unlock(&engine->gt->irq_lock);
1439 
1440 	return true;
1441 }
1442 
1443 void intel_engine_irq_disable(struct intel_engine_cs *engine)
1444 {
1445 	if (!engine->irq_disable)
1446 		return;
1447 
1448 	/* Caller disables interrupts */
1449 	spin_lock(&engine->gt->irq_lock);
1450 	engine->irq_disable(engine);
1451 	spin_unlock(&engine->gt->irq_lock);
1452 }
1453 
1454 void intel_engines_reset_default_submission(struct intel_gt *gt)
1455 {
1456 	struct intel_engine_cs *engine;
1457 	enum intel_engine_id id;
1458 
1459 	for_each_engine(engine, gt, id) {
1460 		if (engine->sanitize)
1461 			engine->sanitize(engine);
1462 
1463 		engine->set_default_submission(engine);
1464 	}
1465 }
1466 
1467 bool intel_engine_can_store_dword(struct intel_engine_cs *engine)
1468 {
1469 	switch (GRAPHICS_VER(engine->i915)) {
1470 	case 2:
1471 		return false; /* uses physical not virtual addresses */
1472 	case 3:
1473 		/* maybe only uses physical not virtual addresses */
1474 		return !(IS_I915G(engine->i915) || IS_I915GM(engine->i915));
1475 	case 4:
1476 		return !IS_I965G(engine->i915); /* who knows! */
1477 	case 6:
1478 		return engine->class != VIDEO_DECODE_CLASS; /* b0rked */
1479 	default:
1480 		return true;
1481 	}
1482 }
1483 
1484 static struct intel_timeline *get_timeline(struct i915_request *rq)
1485 {
1486 	struct intel_timeline *tl;
1487 
1488 	/*
1489 	 * Even though we are holding the engine->sched_engine->lock here, there
1490 	 * is no control over the submission queue per-se and we are
1491 	 * inspecting the active state at a random point in time, with an
1492 	 * unknown queue. Play safe and make sure the timeline remains valid.
1493 	 * (Only being used for pretty printing, one extra kref shouldn't
1494 	 * cause a camel stampede!)
1495 	 */
1496 	rcu_read_lock();
1497 	tl = rcu_dereference(rq->timeline);
1498 	if (!kref_get_unless_zero(&tl->kref))
1499 		tl = NULL;
1500 	rcu_read_unlock();
1501 
1502 	return tl;
1503 }
1504 
1505 static int print_ring(char *buf, int sz, struct i915_request *rq)
1506 {
1507 	int len = 0;
1508 
1509 	if (!i915_request_signaled(rq)) {
1510 		struct intel_timeline *tl = get_timeline(rq);
1511 
1512 		len = scnprintf(buf, sz,
1513 				"ring:{start:%08x, hwsp:%08x, seqno:%08x, runtime:%llums}, ",
1514 				i915_ggtt_offset(rq->ring->vma),
1515 				tl ? tl->hwsp_offset : 0,
1516 				hwsp_seqno(rq),
1517 				DIV_ROUND_CLOSEST_ULL(intel_context_get_total_runtime_ns(rq->context),
1518 						      1000 * 1000));
1519 
1520 		if (tl)
1521 			intel_timeline_put(tl);
1522 	}
1523 
1524 	return len;
1525 }
1526 
1527 static void hexdump(struct drm_printer *m, const void *buf, size_t len)
1528 {
1529 	const size_t rowsize = 8 * sizeof(u32);
1530 	const void *prev = NULL;
1531 	bool skip = false;
1532 	size_t pos;
1533 
1534 	for (pos = 0; pos < len; pos += rowsize) {
1535 		char line[128];
1536 
1537 		if (prev && !memcmp(prev, buf + pos, rowsize)) {
1538 			if (!skip) {
1539 				drm_printf(m, "*\n");
1540 				skip = true;
1541 			}
1542 			continue;
1543 		}
1544 
1545 		WARN_ON_ONCE(hex_dump_to_buffer(buf + pos, len - pos,
1546 						rowsize, sizeof(u32),
1547 						line, sizeof(line),
1548 						false) >= sizeof(line));
1549 		drm_printf(m, "[%04zx] %s\n", pos, line);
1550 
1551 		prev = buf + pos;
1552 		skip = false;
1553 	}
1554 }
1555 
1556 static const char *repr_timer(const struct timer_list *t)
1557 {
1558 	if (!READ_ONCE(t->expires))
1559 		return "inactive";
1560 
1561 	if (timer_pending(t))
1562 		return "active";
1563 
1564 	return "expired";
1565 }
1566 
1567 static void intel_engine_print_registers(struct intel_engine_cs *engine,
1568 					 struct drm_printer *m)
1569 {
1570 	struct drm_i915_private *dev_priv = engine->i915;
1571 	struct intel_engine_execlists * const execlists = &engine->execlists;
1572 	u64 addr;
1573 
1574 	if (engine->id == RENDER_CLASS && IS_GRAPHICS_VER(dev_priv, 4, 7))
1575 		drm_printf(m, "\tCCID: 0x%08x\n", ENGINE_READ(engine, CCID));
1576 	if (HAS_EXECLISTS(dev_priv)) {
1577 		drm_printf(m, "\tEL_STAT_HI: 0x%08x\n",
1578 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_HI));
1579 		drm_printf(m, "\tEL_STAT_LO: 0x%08x\n",
1580 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_LO));
1581 	}
1582 	drm_printf(m, "\tRING_START: 0x%08x\n",
1583 		   ENGINE_READ(engine, RING_START));
1584 	drm_printf(m, "\tRING_HEAD:  0x%08x\n",
1585 		   ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR);
1586 	drm_printf(m, "\tRING_TAIL:  0x%08x\n",
1587 		   ENGINE_READ(engine, RING_TAIL) & TAIL_ADDR);
1588 	drm_printf(m, "\tRING_CTL:   0x%08x%s\n",
1589 		   ENGINE_READ(engine, RING_CTL),
1590 		   ENGINE_READ(engine, RING_CTL) & (RING_WAIT | RING_WAIT_SEMAPHORE) ? " [waiting]" : "");
1591 	if (GRAPHICS_VER(engine->i915) > 2) {
1592 		drm_printf(m, "\tRING_MODE:  0x%08x%s\n",
1593 			   ENGINE_READ(engine, RING_MI_MODE),
1594 			   ENGINE_READ(engine, RING_MI_MODE) & (MODE_IDLE) ? " [idle]" : "");
1595 	}
1596 
1597 	if (GRAPHICS_VER(dev_priv) >= 6) {
1598 		drm_printf(m, "\tRING_IMR:   0x%08x\n",
1599 			   ENGINE_READ(engine, RING_IMR));
1600 		drm_printf(m, "\tRING_ESR:   0x%08x\n",
1601 			   ENGINE_READ(engine, RING_ESR));
1602 		drm_printf(m, "\tRING_EMR:   0x%08x\n",
1603 			   ENGINE_READ(engine, RING_EMR));
1604 		drm_printf(m, "\tRING_EIR:   0x%08x\n",
1605 			   ENGINE_READ(engine, RING_EIR));
1606 	}
1607 
1608 	addr = intel_engine_get_active_head(engine);
1609 	drm_printf(m, "\tACTHD:  0x%08x_%08x\n",
1610 		   upper_32_bits(addr), lower_32_bits(addr));
1611 	addr = intel_engine_get_last_batch_head(engine);
1612 	drm_printf(m, "\tBBADDR: 0x%08x_%08x\n",
1613 		   upper_32_bits(addr), lower_32_bits(addr));
1614 	if (GRAPHICS_VER(dev_priv) >= 8)
1615 		addr = ENGINE_READ64(engine, RING_DMA_FADD, RING_DMA_FADD_UDW);
1616 	else if (GRAPHICS_VER(dev_priv) >= 4)
1617 		addr = ENGINE_READ(engine, RING_DMA_FADD);
1618 	else
1619 		addr = ENGINE_READ(engine, DMA_FADD_I8XX);
1620 	drm_printf(m, "\tDMA_FADDR: 0x%08x_%08x\n",
1621 		   upper_32_bits(addr), lower_32_bits(addr));
1622 	if (GRAPHICS_VER(dev_priv) >= 4) {
1623 		drm_printf(m, "\tIPEIR: 0x%08x\n",
1624 			   ENGINE_READ(engine, RING_IPEIR));
1625 		drm_printf(m, "\tIPEHR: 0x%08x\n",
1626 			   ENGINE_READ(engine, RING_IPEHR));
1627 	} else {
1628 		drm_printf(m, "\tIPEIR: 0x%08x\n", ENGINE_READ(engine, IPEIR));
1629 		drm_printf(m, "\tIPEHR: 0x%08x\n", ENGINE_READ(engine, IPEHR));
1630 	}
1631 
1632 	if (intel_engine_uses_guc(engine)) {
1633 		/* nothing to print yet */
1634 	} else if (HAS_EXECLISTS(dev_priv)) {
1635 		struct i915_request * const *port, *rq;
1636 		const u32 *hws =
1637 			&engine->status_page.addr[I915_HWS_CSB_BUF0_INDEX];
1638 		const u8 num_entries = execlists->csb_size;
1639 		unsigned int idx;
1640 		u8 read, write;
1641 
1642 		drm_printf(m, "\tExeclist tasklet queued? %s (%s), preempt? %s, timeslice? %s\n",
1643 			   yesno(test_bit(TASKLET_STATE_SCHED,
1644 					  &engine->sched_engine->tasklet.state)),
1645 			   enableddisabled(!atomic_read(&engine->sched_engine->tasklet.count)),
1646 			   repr_timer(&engine->execlists.preempt),
1647 			   repr_timer(&engine->execlists.timer));
1648 
1649 		read = execlists->csb_head;
1650 		write = READ_ONCE(*execlists->csb_write);
1651 
1652 		drm_printf(m, "\tExeclist status: 0x%08x %08x; CSB read:%d, write:%d, entries:%d\n",
1653 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_LO),
1654 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_HI),
1655 			   read, write, num_entries);
1656 
1657 		if (read >= num_entries)
1658 			read = 0;
1659 		if (write >= num_entries)
1660 			write = 0;
1661 		if (read > write)
1662 			write += num_entries;
1663 		while (read < write) {
1664 			idx = ++read % num_entries;
1665 			drm_printf(m, "\tExeclist CSB[%d]: 0x%08x, context: %d\n",
1666 				   idx, hws[idx * 2], hws[idx * 2 + 1]);
1667 		}
1668 
1669 		i915_sched_engine_active_lock_bh(engine->sched_engine);
1670 		rcu_read_lock();
1671 		for (port = execlists->active; (rq = *port); port++) {
1672 			char hdr[160];
1673 			int len;
1674 
1675 			len = scnprintf(hdr, sizeof(hdr),
1676 					"\t\tActive[%d]:  ccid:%08x%s%s, ",
1677 					(int)(port - execlists->active),
1678 					rq->context->lrc.ccid,
1679 					intel_context_is_closed(rq->context) ? "!" : "",
1680 					intel_context_is_banned(rq->context) ? "*" : "");
1681 			len += print_ring(hdr + len, sizeof(hdr) - len, rq);
1682 			scnprintf(hdr + len, sizeof(hdr) - len, "rq: ");
1683 			i915_request_show(m, rq, hdr, 0);
1684 		}
1685 		for (port = execlists->pending; (rq = *port); port++) {
1686 			char hdr[160];
1687 			int len;
1688 
1689 			len = scnprintf(hdr, sizeof(hdr),
1690 					"\t\tPending[%d]: ccid:%08x%s%s, ",
1691 					(int)(port - execlists->pending),
1692 					rq->context->lrc.ccid,
1693 					intel_context_is_closed(rq->context) ? "!" : "",
1694 					intel_context_is_banned(rq->context) ? "*" : "");
1695 			len += print_ring(hdr + len, sizeof(hdr) - len, rq);
1696 			scnprintf(hdr + len, sizeof(hdr) - len, "rq: ");
1697 			i915_request_show(m, rq, hdr, 0);
1698 		}
1699 		rcu_read_unlock();
1700 		i915_sched_engine_active_unlock_bh(engine->sched_engine);
1701 	} else if (GRAPHICS_VER(dev_priv) > 6) {
1702 		drm_printf(m, "\tPP_DIR_BASE: 0x%08x\n",
1703 			   ENGINE_READ(engine, RING_PP_DIR_BASE));
1704 		drm_printf(m, "\tPP_DIR_BASE_READ: 0x%08x\n",
1705 			   ENGINE_READ(engine, RING_PP_DIR_BASE_READ));
1706 		drm_printf(m, "\tPP_DIR_DCLV: 0x%08x\n",
1707 			   ENGINE_READ(engine, RING_PP_DIR_DCLV));
1708 	}
1709 }
1710 
1711 static void print_request_ring(struct drm_printer *m, struct i915_request *rq)
1712 {
1713 	struct i915_vma_snapshot *vsnap = &rq->batch_snapshot;
1714 	void *ring;
1715 	int size;
1716 
1717 	if (!i915_vma_snapshot_present(vsnap))
1718 		vsnap = NULL;
1719 
1720 	drm_printf(m,
1721 		   "[head %04x, postfix %04x, tail %04x, batch 0x%08x_%08x]:\n",
1722 		   rq->head, rq->postfix, rq->tail,
1723 		   vsnap ? upper_32_bits(vsnap->gtt_offset) : ~0u,
1724 		   vsnap ? lower_32_bits(vsnap->gtt_offset) : ~0u);
1725 
1726 	size = rq->tail - rq->head;
1727 	if (rq->tail < rq->head)
1728 		size += rq->ring->size;
1729 
1730 	ring = kmalloc(size, GFP_ATOMIC);
1731 	if (ring) {
1732 		const void *vaddr = rq->ring->vaddr;
1733 		unsigned int head = rq->head;
1734 		unsigned int len = 0;
1735 
1736 		if (rq->tail < head) {
1737 			len = rq->ring->size - head;
1738 			memcpy(ring, vaddr + head, len);
1739 			head = 0;
1740 		}
1741 		memcpy(ring + len, vaddr + head, size - len);
1742 
1743 		hexdump(m, ring, size);
1744 		kfree(ring);
1745 	}
1746 }
1747 
1748 static unsigned long list_count(struct list_head *list)
1749 {
1750 	struct list_head *pos;
1751 	unsigned long count = 0;
1752 
1753 	list_for_each(pos, list)
1754 		count++;
1755 
1756 	return count;
1757 }
1758 
1759 static unsigned long read_ul(void *p, size_t x)
1760 {
1761 	return *(unsigned long *)(p + x);
1762 }
1763 
1764 static void print_properties(struct intel_engine_cs *engine,
1765 			     struct drm_printer *m)
1766 {
1767 	static const struct pmap {
1768 		size_t offset;
1769 		const char *name;
1770 	} props[] = {
1771 #define P(x) { \
1772 	.offset = offsetof(typeof(engine->props), x), \
1773 	.name = #x \
1774 }
1775 		P(heartbeat_interval_ms),
1776 		P(max_busywait_duration_ns),
1777 		P(preempt_timeout_ms),
1778 		P(stop_timeout_ms),
1779 		P(timeslice_duration_ms),
1780 
1781 		{},
1782 #undef P
1783 	};
1784 	const struct pmap *p;
1785 
1786 	drm_printf(m, "\tProperties:\n");
1787 	for (p = props; p->name; p++)
1788 		drm_printf(m, "\t\t%s: %lu [default %lu]\n",
1789 			   p->name,
1790 			   read_ul(&engine->props, p->offset),
1791 			   read_ul(&engine->defaults, p->offset));
1792 }
1793 
1794 static void engine_dump_request(struct i915_request *rq, struct drm_printer *m, const char *msg)
1795 {
1796 	struct intel_timeline *tl = get_timeline(rq);
1797 
1798 	i915_request_show(m, rq, msg, 0);
1799 
1800 	drm_printf(m, "\t\tring->start:  0x%08x\n",
1801 		   i915_ggtt_offset(rq->ring->vma));
1802 	drm_printf(m, "\t\tring->head:   0x%08x\n",
1803 		   rq->ring->head);
1804 	drm_printf(m, "\t\tring->tail:   0x%08x\n",
1805 		   rq->ring->tail);
1806 	drm_printf(m, "\t\tring->emit:   0x%08x\n",
1807 		   rq->ring->emit);
1808 	drm_printf(m, "\t\tring->space:  0x%08x\n",
1809 		   rq->ring->space);
1810 
1811 	if (tl) {
1812 		drm_printf(m, "\t\tring->hwsp:   0x%08x\n",
1813 			   tl->hwsp_offset);
1814 		intel_timeline_put(tl);
1815 	}
1816 
1817 	print_request_ring(m, rq);
1818 
1819 	if (rq->context->lrc_reg_state) {
1820 		drm_printf(m, "Logical Ring Context:\n");
1821 		hexdump(m, rq->context->lrc_reg_state, PAGE_SIZE);
1822 	}
1823 }
1824 
1825 void intel_engine_dump_active_requests(struct list_head *requests,
1826 				       struct i915_request *hung_rq,
1827 				       struct drm_printer *m)
1828 {
1829 	struct i915_request *rq;
1830 	const char *msg;
1831 	enum i915_request_state state;
1832 
1833 	list_for_each_entry(rq, requests, sched.link) {
1834 		if (rq == hung_rq)
1835 			continue;
1836 
1837 		state = i915_test_request_state(rq);
1838 		if (state < I915_REQUEST_QUEUED)
1839 			continue;
1840 
1841 		if (state == I915_REQUEST_ACTIVE)
1842 			msg = "\t\tactive on engine";
1843 		else
1844 			msg = "\t\tactive in queue";
1845 
1846 		engine_dump_request(rq, m, msg);
1847 	}
1848 }
1849 
1850 static void engine_dump_active_requests(struct intel_engine_cs *engine, struct drm_printer *m)
1851 {
1852 	struct i915_request *hung_rq = NULL;
1853 	struct intel_context *ce;
1854 	bool guc;
1855 
1856 	/*
1857 	 * No need for an engine->irq_seqno_barrier() before the seqno reads.
1858 	 * The GPU is still running so requests are still executing and any
1859 	 * hardware reads will be out of date by the time they are reported.
1860 	 * But the intention here is just to report an instantaneous snapshot
1861 	 * so that's fine.
1862 	 */
1863 	lockdep_assert_held(&engine->sched_engine->lock);
1864 
1865 	drm_printf(m, "\tRequests:\n");
1866 
1867 	guc = intel_uc_uses_guc_submission(&engine->gt->uc);
1868 	if (guc) {
1869 		ce = intel_engine_get_hung_context(engine);
1870 		if (ce)
1871 			hung_rq = intel_context_find_active_request(ce);
1872 	} else {
1873 		hung_rq = intel_engine_execlist_find_hung_request(engine);
1874 	}
1875 
1876 	if (hung_rq)
1877 		engine_dump_request(hung_rq, m, "\t\thung");
1878 
1879 	if (guc)
1880 		intel_guc_dump_active_requests(engine, hung_rq, m);
1881 	else
1882 		intel_engine_dump_active_requests(&engine->sched_engine->requests,
1883 						  hung_rq, m);
1884 }
1885 
1886 void intel_engine_dump(struct intel_engine_cs *engine,
1887 		       struct drm_printer *m,
1888 		       const char *header, ...)
1889 {
1890 	struct i915_gpu_error * const error = &engine->i915->gpu_error;
1891 	struct i915_request *rq;
1892 	intel_wakeref_t wakeref;
1893 	unsigned long flags;
1894 	ktime_t dummy;
1895 
1896 	if (header) {
1897 		va_list ap;
1898 
1899 		va_start(ap, header);
1900 		drm_vprintf(m, header, &ap);
1901 		va_end(ap);
1902 	}
1903 
1904 	if (intel_gt_is_wedged(engine->gt))
1905 		drm_printf(m, "*** WEDGED ***\n");
1906 
1907 	drm_printf(m, "\tAwake? %d\n", atomic_read(&engine->wakeref.count));
1908 	drm_printf(m, "\tBarriers?: %s\n",
1909 		   yesno(!llist_empty(&engine->barrier_tasks)));
1910 	drm_printf(m, "\tLatency: %luus\n",
1911 		   ewma__engine_latency_read(&engine->latency));
1912 	if (intel_engine_supports_stats(engine))
1913 		drm_printf(m, "\tRuntime: %llums\n",
1914 			   ktime_to_ms(intel_engine_get_busy_time(engine,
1915 								  &dummy)));
1916 	drm_printf(m, "\tForcewake: %x domains, %d active\n",
1917 		   engine->fw_domain, READ_ONCE(engine->fw_active));
1918 
1919 	rcu_read_lock();
1920 	rq = READ_ONCE(engine->heartbeat.systole);
1921 	if (rq)
1922 		drm_printf(m, "\tHeartbeat: %d ms ago\n",
1923 			   jiffies_to_msecs(jiffies - rq->emitted_jiffies));
1924 	rcu_read_unlock();
1925 	drm_printf(m, "\tReset count: %d (global %d)\n",
1926 		   i915_reset_engine_count(error, engine),
1927 		   i915_reset_count(error));
1928 	print_properties(engine, m);
1929 
1930 	spin_lock_irqsave(&engine->sched_engine->lock, flags);
1931 	engine_dump_active_requests(engine, m);
1932 
1933 	drm_printf(m, "\tOn hold?: %lu\n",
1934 		   list_count(&engine->sched_engine->hold));
1935 	spin_unlock_irqrestore(&engine->sched_engine->lock, flags);
1936 
1937 	drm_printf(m, "\tMMIO base:  0x%08x\n", engine->mmio_base);
1938 	wakeref = intel_runtime_pm_get_if_in_use(engine->uncore->rpm);
1939 	if (wakeref) {
1940 		intel_engine_print_registers(engine, m);
1941 		intel_runtime_pm_put(engine->uncore->rpm, wakeref);
1942 	} else {
1943 		drm_printf(m, "\tDevice is asleep; skipping register dump\n");
1944 	}
1945 
1946 	intel_execlists_show_requests(engine, m, i915_request_show, 8);
1947 
1948 	drm_printf(m, "HWSP:\n");
1949 	hexdump(m, engine->status_page.addr, PAGE_SIZE);
1950 
1951 	drm_printf(m, "Idle? %s\n", yesno(intel_engine_is_idle(engine)));
1952 
1953 	intel_engine_print_breadcrumbs(engine, m);
1954 }
1955 
1956 /**
1957  * intel_engine_get_busy_time() - Return current accumulated engine busyness
1958  * @engine: engine to report on
1959  * @now: monotonic timestamp of sampling
1960  *
1961  * Returns accumulated time @engine was busy since engine stats were enabled.
1962  */
1963 ktime_t intel_engine_get_busy_time(struct intel_engine_cs *engine, ktime_t *now)
1964 {
1965 	return engine->busyness(engine, now);
1966 }
1967 
1968 struct intel_context *
1969 intel_engine_create_virtual(struct intel_engine_cs **siblings,
1970 			    unsigned int count, unsigned long flags)
1971 {
1972 	if (count == 0)
1973 		return ERR_PTR(-EINVAL);
1974 
1975 	if (count == 1 && !(flags & FORCE_VIRTUAL))
1976 		return intel_context_create(siblings[0]);
1977 
1978 	GEM_BUG_ON(!siblings[0]->cops->create_virtual);
1979 	return siblings[0]->cops->create_virtual(siblings, count, flags);
1980 }
1981 
1982 struct i915_request *
1983 intel_engine_execlist_find_hung_request(struct intel_engine_cs *engine)
1984 {
1985 	struct i915_request *request, *active = NULL;
1986 
1987 	/*
1988 	 * This search does not work in GuC submission mode. However, the GuC
1989 	 * will report the hanging context directly to the driver itself. So
1990 	 * the driver should never get here when in GuC mode.
1991 	 */
1992 	GEM_BUG_ON(intel_uc_uses_guc_submission(&engine->gt->uc));
1993 
1994 	/*
1995 	 * We are called by the error capture, reset and to dump engine
1996 	 * state at random points in time. In particular, note that neither is
1997 	 * crucially ordered with an interrupt. After a hang, the GPU is dead
1998 	 * and we assume that no more writes can happen (we waited long enough
1999 	 * for all writes that were in transaction to be flushed) - adding an
2000 	 * extra delay for a recent interrupt is pointless. Hence, we do
2001 	 * not need an engine->irq_seqno_barrier() before the seqno reads.
2002 	 * At all other times, we must assume the GPU is still running, but
2003 	 * we only care about the snapshot of this moment.
2004 	 */
2005 	lockdep_assert_held(&engine->sched_engine->lock);
2006 
2007 	rcu_read_lock();
2008 	request = execlists_active(&engine->execlists);
2009 	if (request) {
2010 		struct intel_timeline *tl = request->context->timeline;
2011 
2012 		list_for_each_entry_from_reverse(request, &tl->requests, link) {
2013 			if (__i915_request_is_complete(request))
2014 				break;
2015 
2016 			active = request;
2017 		}
2018 	}
2019 	rcu_read_unlock();
2020 	if (active)
2021 		return active;
2022 
2023 	list_for_each_entry(request, &engine->sched_engine->requests,
2024 			    sched.link) {
2025 		if (i915_test_request_state(request) != I915_REQUEST_ACTIVE)
2026 			continue;
2027 
2028 		active = request;
2029 		break;
2030 	}
2031 
2032 	return active;
2033 }
2034 
2035 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
2036 #include "mock_engine.c"
2037 #include "selftest_engine.c"
2038 #include "selftest_engine_cs.c"
2039 #endif
2040