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 "gem/i915_gem_internal.h"
10 #include "gt/intel_gt_regs.h"
11 
12 #include "i915_cmd_parser.h"
13 #include "i915_drv.h"
14 #include "intel_breadcrumbs.h"
15 #include "intel_context.h"
16 #include "intel_engine.h"
17 #include "intel_engine_pm.h"
18 #include "intel_engine_regs.h"
19 #include "intel_engine_user.h"
20 #include "intel_execlists_submission.h"
21 #include "intel_gt.h"
22 #include "intel_gt_requests.h"
23 #include "intel_gt_pm.h"
24 #include "intel_lrc.h"
25 #include "intel_lrc_reg.h"
26 #include "intel_reset.h"
27 #include "intel_ring.h"
28 #include "uc/intel_guc_submission.h"
29 
30 /* Haswell does have the CXT_SIZE register however it does not appear to be
31  * valid. Now, docs explain in dwords what is in the context object. The full
32  * size is 70720 bytes, however, the power context and execlist context will
33  * never be saved (power context is stored elsewhere, and execlists don't work
34  * on HSW) - so the final size, including the extra state required for the
35  * Resource Streamer, is 66944 bytes, which rounds to 17 pages.
36  */
37 #define HSW_CXT_TOTAL_SIZE		(17 * PAGE_SIZE)
38 
39 #define DEFAULT_LR_CONTEXT_RENDER_SIZE	(22 * PAGE_SIZE)
40 #define GEN8_LR_CONTEXT_RENDER_SIZE	(20 * PAGE_SIZE)
41 #define GEN9_LR_CONTEXT_RENDER_SIZE	(22 * PAGE_SIZE)
42 #define GEN11_LR_CONTEXT_RENDER_SIZE	(14 * PAGE_SIZE)
43 
44 #define GEN8_LR_CONTEXT_OTHER_SIZE	( 2 * PAGE_SIZE)
45 
46 #define MAX_MMIO_BASES 3
47 struct engine_info {
48 	u8 class;
49 	u8 instance;
50 	/* mmio bases table *must* be sorted in reverse graphics_ver order */
51 	struct engine_mmio_base {
52 		u32 graphics_ver : 8;
53 		u32 base : 24;
54 	} mmio_bases[MAX_MMIO_BASES];
55 };
56 
57 static const struct engine_info intel_engines[] = {
58 	[RCS0] = {
59 		.class = RENDER_CLASS,
60 		.instance = 0,
61 		.mmio_bases = {
62 			{ .graphics_ver = 1, .base = RENDER_RING_BASE }
63 		},
64 	},
65 	[BCS0] = {
66 		.class = COPY_ENGINE_CLASS,
67 		.instance = 0,
68 		.mmio_bases = {
69 			{ .graphics_ver = 6, .base = BLT_RING_BASE }
70 		},
71 	},
72 	[VCS0] = {
73 		.class = VIDEO_DECODE_CLASS,
74 		.instance = 0,
75 		.mmio_bases = {
76 			{ .graphics_ver = 11, .base = GEN11_BSD_RING_BASE },
77 			{ .graphics_ver = 6, .base = GEN6_BSD_RING_BASE },
78 			{ .graphics_ver = 4, .base = BSD_RING_BASE }
79 		},
80 	},
81 	[VCS1] = {
82 		.class = VIDEO_DECODE_CLASS,
83 		.instance = 1,
84 		.mmio_bases = {
85 			{ .graphics_ver = 11, .base = GEN11_BSD2_RING_BASE },
86 			{ .graphics_ver = 8, .base = GEN8_BSD2_RING_BASE }
87 		},
88 	},
89 	[VCS2] = {
90 		.class = VIDEO_DECODE_CLASS,
91 		.instance = 2,
92 		.mmio_bases = {
93 			{ .graphics_ver = 11, .base = GEN11_BSD3_RING_BASE }
94 		},
95 	},
96 	[VCS3] = {
97 		.class = VIDEO_DECODE_CLASS,
98 		.instance = 3,
99 		.mmio_bases = {
100 			{ .graphics_ver = 11, .base = GEN11_BSD4_RING_BASE }
101 		},
102 	},
103 	[VCS4] = {
104 		.class = VIDEO_DECODE_CLASS,
105 		.instance = 4,
106 		.mmio_bases = {
107 			{ .graphics_ver = 12, .base = XEHP_BSD5_RING_BASE }
108 		},
109 	},
110 	[VCS5] = {
111 		.class = VIDEO_DECODE_CLASS,
112 		.instance = 5,
113 		.mmio_bases = {
114 			{ .graphics_ver = 12, .base = XEHP_BSD6_RING_BASE }
115 		},
116 	},
117 	[VCS6] = {
118 		.class = VIDEO_DECODE_CLASS,
119 		.instance = 6,
120 		.mmio_bases = {
121 			{ .graphics_ver = 12, .base = XEHP_BSD7_RING_BASE }
122 		},
123 	},
124 	[VCS7] = {
125 		.class = VIDEO_DECODE_CLASS,
126 		.instance = 7,
127 		.mmio_bases = {
128 			{ .graphics_ver = 12, .base = XEHP_BSD8_RING_BASE }
129 		},
130 	},
131 	[VECS0] = {
132 		.class = VIDEO_ENHANCEMENT_CLASS,
133 		.instance = 0,
134 		.mmio_bases = {
135 			{ .graphics_ver = 11, .base = GEN11_VEBOX_RING_BASE },
136 			{ .graphics_ver = 7, .base = VEBOX_RING_BASE }
137 		},
138 	},
139 	[VECS1] = {
140 		.class = VIDEO_ENHANCEMENT_CLASS,
141 		.instance = 1,
142 		.mmio_bases = {
143 			{ .graphics_ver = 11, .base = GEN11_VEBOX2_RING_BASE }
144 		},
145 	},
146 	[VECS2] = {
147 		.class = VIDEO_ENHANCEMENT_CLASS,
148 		.instance = 2,
149 		.mmio_bases = {
150 			{ .graphics_ver = 12, .base = XEHP_VEBOX3_RING_BASE }
151 		},
152 	},
153 	[VECS3] = {
154 		.class = VIDEO_ENHANCEMENT_CLASS,
155 		.instance = 3,
156 		.mmio_bases = {
157 			{ .graphics_ver = 12, .base = XEHP_VEBOX4_RING_BASE }
158 		},
159 	},
160 	[CCS0] = {
161 		.class = COMPUTE_CLASS,
162 		.instance = 0,
163 		.mmio_bases = {
164 			{ .graphics_ver = 12, .base = GEN12_COMPUTE0_RING_BASE }
165 		}
166 	},
167 	[CCS1] = {
168 		.class = COMPUTE_CLASS,
169 		.instance = 1,
170 		.mmio_bases = {
171 			{ .graphics_ver = 12, .base = GEN12_COMPUTE1_RING_BASE }
172 		}
173 	},
174 	[CCS2] = {
175 		.class = COMPUTE_CLASS,
176 		.instance = 2,
177 		.mmio_bases = {
178 			{ .graphics_ver = 12, .base = GEN12_COMPUTE2_RING_BASE }
179 		}
180 	},
181 	[CCS3] = {
182 		.class = COMPUTE_CLASS,
183 		.instance = 3,
184 		.mmio_bases = {
185 			{ .graphics_ver = 12, .base = GEN12_COMPUTE3_RING_BASE }
186 		}
187 	},
188 };
189 
190 /**
191  * intel_engine_context_size() - return the size of the context for an engine
192  * @gt: the gt
193  * @class: engine class
194  *
195  * Each engine class may require a different amount of space for a context
196  * image.
197  *
198  * Return: size (in bytes) of an engine class specific context image
199  *
200  * Note: this size includes the HWSP, which is part of the context image
201  * in LRC mode, but does not include the "shared data page" used with
202  * GuC submission. The caller should account for this if using the GuC.
203  */
204 u32 intel_engine_context_size(struct intel_gt *gt, u8 class)
205 {
206 	struct intel_uncore *uncore = gt->uncore;
207 	u32 cxt_size;
208 
209 	BUILD_BUG_ON(I915_GTT_PAGE_SIZE != PAGE_SIZE);
210 
211 	switch (class) {
212 	case COMPUTE_CLASS:
213 		fallthrough;
214 	case RENDER_CLASS:
215 		switch (GRAPHICS_VER(gt->i915)) {
216 		default:
217 			MISSING_CASE(GRAPHICS_VER(gt->i915));
218 			return DEFAULT_LR_CONTEXT_RENDER_SIZE;
219 		case 12:
220 		case 11:
221 			return GEN11_LR_CONTEXT_RENDER_SIZE;
222 		case 9:
223 			return GEN9_LR_CONTEXT_RENDER_SIZE;
224 		case 8:
225 			return GEN8_LR_CONTEXT_RENDER_SIZE;
226 		case 7:
227 			if (IS_HASWELL(gt->i915))
228 				return HSW_CXT_TOTAL_SIZE;
229 
230 			cxt_size = intel_uncore_read(uncore, GEN7_CXT_SIZE);
231 			return round_up(GEN7_CXT_TOTAL_SIZE(cxt_size) * 64,
232 					PAGE_SIZE);
233 		case 6:
234 			cxt_size = intel_uncore_read(uncore, CXT_SIZE);
235 			return round_up(GEN6_CXT_TOTAL_SIZE(cxt_size) * 64,
236 					PAGE_SIZE);
237 		case 5:
238 		case 4:
239 			/*
240 			 * There is a discrepancy here between the size reported
241 			 * by the register and the size of the context layout
242 			 * in the docs. Both are described as authorative!
243 			 *
244 			 * The discrepancy is on the order of a few cachelines,
245 			 * but the total is under one page (4k), which is our
246 			 * minimum allocation anyway so it should all come
247 			 * out in the wash.
248 			 */
249 			cxt_size = intel_uncore_read(uncore, CXT_SIZE) + 1;
250 			drm_dbg(&gt->i915->drm,
251 				"graphics_ver = %d CXT_SIZE = %d bytes [0x%08x]\n",
252 				GRAPHICS_VER(gt->i915), cxt_size * 64,
253 				cxt_size - 1);
254 			return round_up(cxt_size * 64, PAGE_SIZE);
255 		case 3:
256 		case 2:
257 		/* For the special day when i810 gets merged. */
258 		case 1:
259 			return 0;
260 		}
261 		break;
262 	default:
263 		MISSING_CASE(class);
264 		fallthrough;
265 	case VIDEO_DECODE_CLASS:
266 	case VIDEO_ENHANCEMENT_CLASS:
267 	case COPY_ENGINE_CLASS:
268 		if (GRAPHICS_VER(gt->i915) < 8)
269 			return 0;
270 		return GEN8_LR_CONTEXT_OTHER_SIZE;
271 	}
272 }
273 
274 static u32 __engine_mmio_base(struct drm_i915_private *i915,
275 			      const struct engine_mmio_base *bases)
276 {
277 	int i;
278 
279 	for (i = 0; i < MAX_MMIO_BASES; i++)
280 		if (GRAPHICS_VER(i915) >= bases[i].graphics_ver)
281 			break;
282 
283 	GEM_BUG_ON(i == MAX_MMIO_BASES);
284 	GEM_BUG_ON(!bases[i].base);
285 
286 	return bases[i].base;
287 }
288 
289 static void __sprint_engine_name(struct intel_engine_cs *engine)
290 {
291 	/*
292 	 * Before we know what the uABI name for this engine will be,
293 	 * we still would like to keep track of this engine in the debug logs.
294 	 * We throw in a ' here as a reminder that this isn't its final name.
295 	 */
296 	GEM_WARN_ON(snprintf(engine->name, sizeof(engine->name), "%s'%u",
297 			     intel_engine_class_repr(engine->class),
298 			     engine->instance) >= sizeof(engine->name));
299 }
300 
301 void intel_engine_set_hwsp_writemask(struct intel_engine_cs *engine, u32 mask)
302 {
303 	/*
304 	 * Though they added more rings on g4x/ilk, they did not add
305 	 * per-engine HWSTAM until gen6.
306 	 */
307 	if (GRAPHICS_VER(engine->i915) < 6 && engine->class != RENDER_CLASS)
308 		return;
309 
310 	if (GRAPHICS_VER(engine->i915) >= 3)
311 		ENGINE_WRITE(engine, RING_HWSTAM, mask);
312 	else
313 		ENGINE_WRITE16(engine, RING_HWSTAM, mask);
314 }
315 
316 static void intel_engine_sanitize_mmio(struct intel_engine_cs *engine)
317 {
318 	/* Mask off all writes into the unknown HWSP */
319 	intel_engine_set_hwsp_writemask(engine, ~0u);
320 }
321 
322 static void nop_irq_handler(struct intel_engine_cs *engine, u16 iir)
323 {
324 	GEM_DEBUG_WARN_ON(iir);
325 }
326 
327 static u32 get_reset_domain(u8 ver, enum intel_engine_id id)
328 {
329 	u32 reset_domain;
330 
331 	if (ver >= 11) {
332 		static const u32 engine_reset_domains[] = {
333 			[RCS0]  = GEN11_GRDOM_RENDER,
334 			[BCS0]  = GEN11_GRDOM_BLT,
335 			[VCS0]  = GEN11_GRDOM_MEDIA,
336 			[VCS1]  = GEN11_GRDOM_MEDIA2,
337 			[VCS2]  = GEN11_GRDOM_MEDIA3,
338 			[VCS3]  = GEN11_GRDOM_MEDIA4,
339 			[VCS4]  = GEN11_GRDOM_MEDIA5,
340 			[VCS5]  = GEN11_GRDOM_MEDIA6,
341 			[VCS6]  = GEN11_GRDOM_MEDIA7,
342 			[VCS7]  = GEN11_GRDOM_MEDIA8,
343 			[VECS0] = GEN11_GRDOM_VECS,
344 			[VECS1] = GEN11_GRDOM_VECS2,
345 			[VECS2] = GEN11_GRDOM_VECS3,
346 			[VECS3] = GEN11_GRDOM_VECS4,
347 			[CCS0]  = GEN11_GRDOM_RENDER,
348 			[CCS1]  = GEN11_GRDOM_RENDER,
349 			[CCS2]  = GEN11_GRDOM_RENDER,
350 			[CCS3]  = GEN11_GRDOM_RENDER,
351 		};
352 		GEM_BUG_ON(id >= ARRAY_SIZE(engine_reset_domains) ||
353 			   !engine_reset_domains[id]);
354 		reset_domain = engine_reset_domains[id];
355 	} else {
356 		static const u32 engine_reset_domains[] = {
357 			[RCS0]  = GEN6_GRDOM_RENDER,
358 			[BCS0]  = GEN6_GRDOM_BLT,
359 			[VCS0]  = GEN6_GRDOM_MEDIA,
360 			[VCS1]  = GEN8_GRDOM_MEDIA2,
361 			[VECS0] = GEN6_GRDOM_VECS,
362 		};
363 		GEM_BUG_ON(id >= ARRAY_SIZE(engine_reset_domains) ||
364 			   !engine_reset_domains[id]);
365 		reset_domain = engine_reset_domains[id];
366 	}
367 
368 	return reset_domain;
369 }
370 
371 static int intel_engine_setup(struct intel_gt *gt, enum intel_engine_id id,
372 			      u8 logical_instance)
373 {
374 	const struct engine_info *info = &intel_engines[id];
375 	struct drm_i915_private *i915 = gt->i915;
376 	struct intel_engine_cs *engine;
377 	u8 guc_class;
378 
379 	BUILD_BUG_ON(MAX_ENGINE_CLASS >= BIT(GEN11_ENGINE_CLASS_WIDTH));
380 	BUILD_BUG_ON(MAX_ENGINE_INSTANCE >= BIT(GEN11_ENGINE_INSTANCE_WIDTH));
381 	BUILD_BUG_ON(I915_MAX_VCS > (MAX_ENGINE_INSTANCE + 1));
382 	BUILD_BUG_ON(I915_MAX_VECS > (MAX_ENGINE_INSTANCE + 1));
383 
384 	if (GEM_DEBUG_WARN_ON(id >= ARRAY_SIZE(gt->engine)))
385 		return -EINVAL;
386 
387 	if (GEM_DEBUG_WARN_ON(info->class > MAX_ENGINE_CLASS))
388 		return -EINVAL;
389 
390 	if (GEM_DEBUG_WARN_ON(info->instance > MAX_ENGINE_INSTANCE))
391 		return -EINVAL;
392 
393 	if (GEM_DEBUG_WARN_ON(gt->engine_class[info->class][info->instance]))
394 		return -EINVAL;
395 
396 	engine = kzalloc(sizeof(*engine), GFP_KERNEL);
397 	if (!engine)
398 		return -ENOMEM;
399 
400 	BUILD_BUG_ON(BITS_PER_TYPE(engine->mask) < I915_NUM_ENGINES);
401 
402 	INIT_LIST_HEAD(&engine->pinned_contexts_list);
403 	engine->id = id;
404 	engine->legacy_idx = INVALID_ENGINE;
405 	engine->mask = BIT(id);
406 	engine->reset_domain = get_reset_domain(GRAPHICS_VER(gt->i915),
407 						id);
408 	engine->i915 = i915;
409 	engine->gt = gt;
410 	engine->uncore = gt->uncore;
411 	guc_class = engine_class_to_guc_class(info->class);
412 	engine->guc_id = MAKE_GUC_ID(guc_class, info->instance);
413 	engine->mmio_base = __engine_mmio_base(i915, info->mmio_bases);
414 
415 	engine->irq_handler = nop_irq_handler;
416 
417 	engine->class = info->class;
418 	engine->instance = info->instance;
419 	engine->logical_mask = BIT(logical_instance);
420 	__sprint_engine_name(engine);
421 
422 	engine->props.heartbeat_interval_ms =
423 		CONFIG_DRM_I915_HEARTBEAT_INTERVAL;
424 	engine->props.max_busywait_duration_ns =
425 		CONFIG_DRM_I915_MAX_REQUEST_BUSYWAIT;
426 	engine->props.preempt_timeout_ms =
427 		CONFIG_DRM_I915_PREEMPT_TIMEOUT;
428 	engine->props.stop_timeout_ms =
429 		CONFIG_DRM_I915_STOP_TIMEOUT;
430 	engine->props.timeslice_duration_ms =
431 		CONFIG_DRM_I915_TIMESLICE_DURATION;
432 
433 	/* Override to uninterruptible for OpenCL workloads. */
434 	if (GRAPHICS_VER(i915) == 12 && engine->class == RENDER_CLASS)
435 		engine->props.preempt_timeout_ms = 0;
436 
437 	/* features common between engines sharing EUs */
438 	if (engine->class == RENDER_CLASS || engine->class == COMPUTE_CLASS) {
439 		engine->flags |= I915_ENGINE_HAS_RCS_REG_STATE;
440 		engine->flags |= I915_ENGINE_HAS_EU_PRIORITY;
441 	}
442 
443 	engine->defaults = engine->props; /* never to change again */
444 
445 	engine->context_size = intel_engine_context_size(gt, engine->class);
446 	if (WARN_ON(engine->context_size > BIT(20)))
447 		engine->context_size = 0;
448 	if (engine->context_size)
449 		DRIVER_CAPS(i915)->has_logical_contexts = true;
450 
451 	ewma__engine_latency_init(&engine->latency);
452 	seqcount_init(&engine->stats.execlists.lock);
453 
454 	ATOMIC_INIT_NOTIFIER_HEAD(&engine->context_status_notifier);
455 
456 	/* Scrub mmio state on takeover */
457 	intel_engine_sanitize_mmio(engine);
458 
459 	gt->engine_class[info->class][info->instance] = engine;
460 	gt->engine[id] = engine;
461 
462 	return 0;
463 }
464 
465 static void __setup_engine_capabilities(struct intel_engine_cs *engine)
466 {
467 	struct drm_i915_private *i915 = engine->i915;
468 
469 	if (engine->class == VIDEO_DECODE_CLASS) {
470 		/*
471 		 * HEVC support is present on first engine instance
472 		 * before Gen11 and on all instances afterwards.
473 		 */
474 		if (GRAPHICS_VER(i915) >= 11 ||
475 		    (GRAPHICS_VER(i915) >= 9 && engine->instance == 0))
476 			engine->uabi_capabilities |=
477 				I915_VIDEO_CLASS_CAPABILITY_HEVC;
478 
479 		/*
480 		 * SFC block is present only on even logical engine
481 		 * instances.
482 		 */
483 		if ((GRAPHICS_VER(i915) >= 11 &&
484 		     (engine->gt->info.vdbox_sfc_access &
485 		      BIT(engine->instance))) ||
486 		    (GRAPHICS_VER(i915) >= 9 && engine->instance == 0))
487 			engine->uabi_capabilities |=
488 				I915_VIDEO_AND_ENHANCE_CLASS_CAPABILITY_SFC;
489 	} else if (engine->class == VIDEO_ENHANCEMENT_CLASS) {
490 		if (GRAPHICS_VER(i915) >= 9 &&
491 		    engine->gt->info.sfc_mask & BIT(engine->instance))
492 			engine->uabi_capabilities |=
493 				I915_VIDEO_AND_ENHANCE_CLASS_CAPABILITY_SFC;
494 	}
495 }
496 
497 static void intel_setup_engine_capabilities(struct intel_gt *gt)
498 {
499 	struct intel_engine_cs *engine;
500 	enum intel_engine_id id;
501 
502 	for_each_engine(engine, gt, id)
503 		__setup_engine_capabilities(engine);
504 }
505 
506 /**
507  * intel_engines_release() - free the resources allocated for Command Streamers
508  * @gt: pointer to struct intel_gt
509  */
510 void intel_engines_release(struct intel_gt *gt)
511 {
512 	struct intel_engine_cs *engine;
513 	enum intel_engine_id id;
514 
515 	/*
516 	 * Before we release the resources held by engine, we must be certain
517 	 * that the HW is no longer accessing them -- having the GPU scribble
518 	 * to or read from a page being used for something else causes no end
519 	 * of fun.
520 	 *
521 	 * The GPU should be reset by this point, but assume the worst just
522 	 * in case we aborted before completely initialising the engines.
523 	 */
524 	GEM_BUG_ON(intel_gt_pm_is_awake(gt));
525 	if (!INTEL_INFO(gt->i915)->gpu_reset_clobbers_display)
526 		__intel_gt_reset(gt, ALL_ENGINES);
527 
528 	/* Decouple the backend; but keep the layout for late GPU resets */
529 	for_each_engine(engine, gt, id) {
530 		if (!engine->release)
531 			continue;
532 
533 		intel_wakeref_wait_for_idle(&engine->wakeref);
534 		GEM_BUG_ON(intel_engine_pm_is_awake(engine));
535 
536 		engine->release(engine);
537 		engine->release = NULL;
538 
539 		memset(&engine->reset, 0, sizeof(engine->reset));
540 	}
541 }
542 
543 void intel_engine_free_request_pool(struct intel_engine_cs *engine)
544 {
545 	if (!engine->request_pool)
546 		return;
547 
548 	kmem_cache_free(i915_request_slab_cache(), engine->request_pool);
549 }
550 
551 void intel_engines_free(struct intel_gt *gt)
552 {
553 	struct intel_engine_cs *engine;
554 	enum intel_engine_id id;
555 
556 	/* Free the requests! dma-resv keeps fences around for an eternity */
557 	rcu_barrier();
558 
559 	for_each_engine(engine, gt, id) {
560 		intel_engine_free_request_pool(engine);
561 		kfree(engine);
562 		gt->engine[id] = NULL;
563 	}
564 }
565 
566 static
567 bool gen11_vdbox_has_sfc(struct intel_gt *gt,
568 			 unsigned int physical_vdbox,
569 			 unsigned int logical_vdbox, u16 vdbox_mask)
570 {
571 	struct drm_i915_private *i915 = gt->i915;
572 
573 	/*
574 	 * In Gen11, only even numbered logical VDBOXes are hooked
575 	 * up to an SFC (Scaler & Format Converter) unit.
576 	 * In Gen12, Even numbered physical instance always are connected
577 	 * to an SFC. Odd numbered physical instances have SFC only if
578 	 * previous even instance is fused off.
579 	 *
580 	 * Starting with Xe_HP, there's also a dedicated SFC_ENABLE field
581 	 * in the fuse register that tells us whether a specific SFC is present.
582 	 */
583 	if ((gt->info.sfc_mask & BIT(physical_vdbox / 2)) == 0)
584 		return false;
585 	else if (GRAPHICS_VER(i915) == 12)
586 		return (physical_vdbox % 2 == 0) ||
587 			!(BIT(physical_vdbox - 1) & vdbox_mask);
588 	else if (GRAPHICS_VER(i915) == 11)
589 		return logical_vdbox % 2 == 0;
590 
591 	MISSING_CASE(GRAPHICS_VER(i915));
592 	return false;
593 }
594 
595 static void engine_mask_apply_compute_fuses(struct intel_gt *gt)
596 {
597 	struct drm_i915_private *i915 = gt->i915;
598 	struct intel_gt_info *info = &gt->info;
599 	int ss_per_ccs = info->sseu.max_subslices / I915_MAX_CCS;
600 	unsigned long ccs_mask;
601 	unsigned int i;
602 
603 	if (GRAPHICS_VER_FULL(i915) < IP_VER(12, 50))
604 		return;
605 
606 	ccs_mask = intel_slicemask_from_dssmask(intel_sseu_get_compute_subslices(&info->sseu),
607 						ss_per_ccs);
608 	/*
609 	 * If all DSS in a quadrant are fused off, the corresponding CCS
610 	 * engine is not available for use.
611 	 */
612 	for_each_clear_bit(i, &ccs_mask, I915_MAX_CCS) {
613 		info->engine_mask &= ~BIT(_CCS(i));
614 		drm_dbg(&i915->drm, "ccs%u fused off\n", i);
615 	}
616 }
617 
618 /*
619  * Determine which engines are fused off in our particular hardware.
620  * Note that we have a catch-22 situation where we need to be able to access
621  * the blitter forcewake domain to read the engine fuses, but at the same time
622  * we need to know which engines are available on the system to know which
623  * forcewake domains are present. We solve this by intializing the forcewake
624  * domains based on the full engine mask in the platform capabilities before
625  * calling this function and pruning the domains for fused-off engines
626  * afterwards.
627  */
628 static intel_engine_mask_t init_engine_mask(struct intel_gt *gt)
629 {
630 	struct drm_i915_private *i915 = gt->i915;
631 	struct intel_gt_info *info = &gt->info;
632 	struct intel_uncore *uncore = gt->uncore;
633 	unsigned int logical_vdbox = 0;
634 	unsigned int i;
635 	u32 media_fuse, fuse1;
636 	u16 vdbox_mask;
637 	u16 vebox_mask;
638 
639 	info->engine_mask = INTEL_INFO(i915)->platform_engine_mask;
640 
641 	if (GRAPHICS_VER(i915) < 11)
642 		return info->engine_mask;
643 
644 	/*
645 	 * On newer platforms the fusing register is called 'enable' and has
646 	 * enable semantics, while on older platforms it is called 'disable'
647 	 * and bits have disable semantices.
648 	 */
649 	media_fuse = intel_uncore_read(uncore, GEN11_GT_VEBOX_VDBOX_DISABLE);
650 	if (GRAPHICS_VER_FULL(i915) < IP_VER(12, 50))
651 		media_fuse = ~media_fuse;
652 
653 	vdbox_mask = media_fuse & GEN11_GT_VDBOX_DISABLE_MASK;
654 	vebox_mask = (media_fuse & GEN11_GT_VEBOX_DISABLE_MASK) >>
655 		      GEN11_GT_VEBOX_DISABLE_SHIFT;
656 
657 	if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 50)) {
658 		fuse1 = intel_uncore_read(uncore, HSW_PAVP_FUSE1);
659 		gt->info.sfc_mask = REG_FIELD_GET(XEHP_SFC_ENABLE_MASK, fuse1);
660 	} else {
661 		gt->info.sfc_mask = ~0;
662 	}
663 
664 	for (i = 0; i < I915_MAX_VCS; i++) {
665 		if (!HAS_ENGINE(gt, _VCS(i))) {
666 			vdbox_mask &= ~BIT(i);
667 			continue;
668 		}
669 
670 		if (!(BIT(i) & vdbox_mask)) {
671 			info->engine_mask &= ~BIT(_VCS(i));
672 			drm_dbg(&i915->drm, "vcs%u fused off\n", i);
673 			continue;
674 		}
675 
676 		if (gen11_vdbox_has_sfc(gt, i, logical_vdbox, vdbox_mask))
677 			gt->info.vdbox_sfc_access |= BIT(i);
678 		logical_vdbox++;
679 	}
680 	drm_dbg(&i915->drm, "vdbox enable: %04x, instances: %04lx\n",
681 		vdbox_mask, VDBOX_MASK(gt));
682 	GEM_BUG_ON(vdbox_mask != VDBOX_MASK(gt));
683 
684 	for (i = 0; i < I915_MAX_VECS; i++) {
685 		if (!HAS_ENGINE(gt, _VECS(i))) {
686 			vebox_mask &= ~BIT(i);
687 			continue;
688 		}
689 
690 		if (!(BIT(i) & vebox_mask)) {
691 			info->engine_mask &= ~BIT(_VECS(i));
692 			drm_dbg(&i915->drm, "vecs%u fused off\n", i);
693 		}
694 	}
695 	drm_dbg(&i915->drm, "vebox enable: %04x, instances: %04lx\n",
696 		vebox_mask, VEBOX_MASK(gt));
697 	GEM_BUG_ON(vebox_mask != VEBOX_MASK(gt));
698 
699 	engine_mask_apply_compute_fuses(gt);
700 
701 	return info->engine_mask;
702 }
703 
704 static void populate_logical_ids(struct intel_gt *gt, u8 *logical_ids,
705 				 u8 class, const u8 *map, u8 num_instances)
706 {
707 	int i, j;
708 	u8 current_logical_id = 0;
709 
710 	for (j = 0; j < num_instances; ++j) {
711 		for (i = 0; i < ARRAY_SIZE(intel_engines); ++i) {
712 			if (!HAS_ENGINE(gt, i) ||
713 			    intel_engines[i].class != class)
714 				continue;
715 
716 			if (intel_engines[i].instance == map[j]) {
717 				logical_ids[intel_engines[i].instance] =
718 					current_logical_id++;
719 				break;
720 			}
721 		}
722 	}
723 }
724 
725 static void setup_logical_ids(struct intel_gt *gt, u8 *logical_ids, u8 class)
726 {
727 	int i;
728 	u8 map[MAX_ENGINE_INSTANCE + 1];
729 
730 	for (i = 0; i < MAX_ENGINE_INSTANCE + 1; ++i)
731 		map[i] = i;
732 	populate_logical_ids(gt, logical_ids, class, map, ARRAY_SIZE(map));
733 }
734 
735 /**
736  * intel_engines_init_mmio() - allocate and prepare the Engine Command Streamers
737  * @gt: pointer to struct intel_gt
738  *
739  * Return: non-zero if the initialization failed.
740  */
741 int intel_engines_init_mmio(struct intel_gt *gt)
742 {
743 	struct drm_i915_private *i915 = gt->i915;
744 	const unsigned int engine_mask = init_engine_mask(gt);
745 	unsigned int mask = 0;
746 	unsigned int i, class;
747 	u8 logical_ids[MAX_ENGINE_INSTANCE + 1];
748 	int err;
749 
750 	drm_WARN_ON(&i915->drm, engine_mask == 0);
751 	drm_WARN_ON(&i915->drm, engine_mask &
752 		    GENMASK(BITS_PER_TYPE(mask) - 1, I915_NUM_ENGINES));
753 
754 	if (i915_inject_probe_failure(i915))
755 		return -ENODEV;
756 
757 	for (class = 0; class < MAX_ENGINE_CLASS + 1; ++class) {
758 		setup_logical_ids(gt, logical_ids, class);
759 
760 		for (i = 0; i < ARRAY_SIZE(intel_engines); ++i) {
761 			u8 instance = intel_engines[i].instance;
762 
763 			if (intel_engines[i].class != class ||
764 			    !HAS_ENGINE(gt, i))
765 				continue;
766 
767 			err = intel_engine_setup(gt, i,
768 						 logical_ids[instance]);
769 			if (err)
770 				goto cleanup;
771 
772 			mask |= BIT(i);
773 		}
774 	}
775 
776 	/*
777 	 * Catch failures to update intel_engines table when the new engines
778 	 * are added to the driver by a warning and disabling the forgotten
779 	 * engines.
780 	 */
781 	if (drm_WARN_ON(&i915->drm, mask != engine_mask))
782 		gt->info.engine_mask = mask;
783 
784 	gt->info.num_engines = hweight32(mask);
785 
786 	intel_gt_check_and_clear_faults(gt);
787 
788 	intel_setup_engine_capabilities(gt);
789 
790 	intel_uncore_prune_engine_fw_domains(gt->uncore, gt);
791 
792 	return 0;
793 
794 cleanup:
795 	intel_engines_free(gt);
796 	return err;
797 }
798 
799 void intel_engine_init_execlists(struct intel_engine_cs *engine)
800 {
801 	struct intel_engine_execlists * const execlists = &engine->execlists;
802 
803 	execlists->port_mask = 1;
804 	GEM_BUG_ON(!is_power_of_2(execlists_num_ports(execlists)));
805 	GEM_BUG_ON(execlists_num_ports(execlists) > EXECLIST_MAX_PORTS);
806 
807 	memset(execlists->pending, 0, sizeof(execlists->pending));
808 	execlists->active =
809 		memset(execlists->inflight, 0, sizeof(execlists->inflight));
810 }
811 
812 static void cleanup_status_page(struct intel_engine_cs *engine)
813 {
814 	struct i915_vma *vma;
815 
816 	/* Prevent writes into HWSP after returning the page to the system */
817 	intel_engine_set_hwsp_writemask(engine, ~0u);
818 
819 	vma = fetch_and_zero(&engine->status_page.vma);
820 	if (!vma)
821 		return;
822 
823 	if (!HWS_NEEDS_PHYSICAL(engine->i915))
824 		i915_vma_unpin(vma);
825 
826 	i915_gem_object_unpin_map(vma->obj);
827 	i915_gem_object_put(vma->obj);
828 }
829 
830 static int pin_ggtt_status_page(struct intel_engine_cs *engine,
831 				struct i915_gem_ww_ctx *ww,
832 				struct i915_vma *vma)
833 {
834 	unsigned int flags;
835 
836 	if (!HAS_LLC(engine->i915) && i915_ggtt_has_aperture(engine->gt->ggtt))
837 		/*
838 		 * On g33, we cannot place HWS above 256MiB, so
839 		 * restrict its pinning to the low mappable arena.
840 		 * Though this restriction is not documented for
841 		 * gen4, gen5, or byt, they also behave similarly
842 		 * and hang if the HWS is placed at the top of the
843 		 * GTT. To generalise, it appears that all !llc
844 		 * platforms have issues with us placing the HWS
845 		 * above the mappable region (even though we never
846 		 * actually map it).
847 		 */
848 		flags = PIN_MAPPABLE;
849 	else
850 		flags = PIN_HIGH;
851 
852 	return i915_ggtt_pin(vma, ww, 0, flags);
853 }
854 
855 static int init_status_page(struct intel_engine_cs *engine)
856 {
857 	struct drm_i915_gem_object *obj;
858 	struct i915_gem_ww_ctx ww;
859 	struct i915_vma *vma;
860 	void *vaddr;
861 	int ret;
862 
863 	INIT_LIST_HEAD(&engine->status_page.timelines);
864 
865 	/*
866 	 * Though the HWS register does support 36bit addresses, historically
867 	 * we have had hangs and corruption reported due to wild writes if
868 	 * the HWS is placed above 4G. We only allow objects to be allocated
869 	 * in GFP_DMA32 for i965, and no earlier physical address users had
870 	 * access to more than 4G.
871 	 */
872 	obj = i915_gem_object_create_internal(engine->i915, PAGE_SIZE);
873 	if (IS_ERR(obj)) {
874 		drm_err(&engine->i915->drm,
875 			"Failed to allocate status page\n");
876 		return PTR_ERR(obj);
877 	}
878 
879 	i915_gem_object_set_cache_coherency(obj, I915_CACHE_LLC);
880 
881 	vma = i915_vma_instance(obj, &engine->gt->ggtt->vm, NULL);
882 	if (IS_ERR(vma)) {
883 		ret = PTR_ERR(vma);
884 		goto err_put;
885 	}
886 
887 	i915_gem_ww_ctx_init(&ww, true);
888 retry:
889 	ret = i915_gem_object_lock(obj, &ww);
890 	if (!ret && !HWS_NEEDS_PHYSICAL(engine->i915))
891 		ret = pin_ggtt_status_page(engine, &ww, vma);
892 	if (ret)
893 		goto err;
894 
895 	vaddr = i915_gem_object_pin_map(obj, I915_MAP_WB);
896 	if (IS_ERR(vaddr)) {
897 		ret = PTR_ERR(vaddr);
898 		goto err_unpin;
899 	}
900 
901 	engine->status_page.addr = memset(vaddr, 0, PAGE_SIZE);
902 	engine->status_page.vma = vma;
903 
904 err_unpin:
905 	if (ret)
906 		i915_vma_unpin(vma);
907 err:
908 	if (ret == -EDEADLK) {
909 		ret = i915_gem_ww_ctx_backoff(&ww);
910 		if (!ret)
911 			goto retry;
912 	}
913 	i915_gem_ww_ctx_fini(&ww);
914 err_put:
915 	if (ret)
916 		i915_gem_object_put(obj);
917 	return ret;
918 }
919 
920 static int engine_setup_common(struct intel_engine_cs *engine)
921 {
922 	int err;
923 
924 	init_llist_head(&engine->barrier_tasks);
925 
926 	err = init_status_page(engine);
927 	if (err)
928 		return err;
929 
930 	engine->breadcrumbs = intel_breadcrumbs_create(engine);
931 	if (!engine->breadcrumbs) {
932 		err = -ENOMEM;
933 		goto err_status;
934 	}
935 
936 	engine->sched_engine = i915_sched_engine_create(ENGINE_PHYSICAL);
937 	if (!engine->sched_engine) {
938 		err = -ENOMEM;
939 		goto err_sched_engine;
940 	}
941 	engine->sched_engine->private_data = engine;
942 
943 	err = intel_engine_init_cmd_parser(engine);
944 	if (err)
945 		goto err_cmd_parser;
946 
947 	intel_engine_init_execlists(engine);
948 	intel_engine_init__pm(engine);
949 	intel_engine_init_retire(engine);
950 
951 	/* Use the whole device by default */
952 	engine->sseu =
953 		intel_sseu_from_device_info(&engine->gt->info.sseu);
954 
955 	intel_engine_init_workarounds(engine);
956 	intel_engine_init_whitelist(engine);
957 	intel_engine_init_ctx_wa(engine);
958 
959 	if (GRAPHICS_VER(engine->i915) >= 12)
960 		engine->flags |= I915_ENGINE_HAS_RELATIVE_MMIO;
961 
962 	return 0;
963 
964 err_cmd_parser:
965 	i915_sched_engine_put(engine->sched_engine);
966 err_sched_engine:
967 	intel_breadcrumbs_put(engine->breadcrumbs);
968 err_status:
969 	cleanup_status_page(engine);
970 	return err;
971 }
972 
973 struct measure_breadcrumb {
974 	struct i915_request rq;
975 	struct intel_ring ring;
976 	u32 cs[2048];
977 };
978 
979 static int measure_breadcrumb_dw(struct intel_context *ce)
980 {
981 	struct intel_engine_cs *engine = ce->engine;
982 	struct measure_breadcrumb *frame;
983 	int dw;
984 
985 	GEM_BUG_ON(!engine->gt->scratch);
986 
987 	frame = kzalloc(sizeof(*frame), GFP_KERNEL);
988 	if (!frame)
989 		return -ENOMEM;
990 
991 	frame->rq.engine = engine;
992 	frame->rq.context = ce;
993 	rcu_assign_pointer(frame->rq.timeline, ce->timeline);
994 	frame->rq.hwsp_seqno = ce->timeline->hwsp_seqno;
995 
996 	frame->ring.vaddr = frame->cs;
997 	frame->ring.size = sizeof(frame->cs);
998 	frame->ring.wrap =
999 		BITS_PER_TYPE(frame->ring.size) - ilog2(frame->ring.size);
1000 	frame->ring.effective_size = frame->ring.size;
1001 	intel_ring_update_space(&frame->ring);
1002 	frame->rq.ring = &frame->ring;
1003 
1004 	mutex_lock(&ce->timeline->mutex);
1005 	spin_lock_irq(&engine->sched_engine->lock);
1006 
1007 	dw = engine->emit_fini_breadcrumb(&frame->rq, frame->cs) - frame->cs;
1008 
1009 	spin_unlock_irq(&engine->sched_engine->lock);
1010 	mutex_unlock(&ce->timeline->mutex);
1011 
1012 	GEM_BUG_ON(dw & 1); /* RING_TAIL must be qword aligned */
1013 
1014 	kfree(frame);
1015 	return dw;
1016 }
1017 
1018 struct intel_context *
1019 intel_engine_create_pinned_context(struct intel_engine_cs *engine,
1020 				   struct i915_address_space *vm,
1021 				   unsigned int ring_size,
1022 				   unsigned int hwsp,
1023 				   struct lock_class_key *key,
1024 				   const char *name)
1025 {
1026 	struct intel_context *ce;
1027 	int err;
1028 
1029 	ce = intel_context_create(engine);
1030 	if (IS_ERR(ce))
1031 		return ce;
1032 
1033 	__set_bit(CONTEXT_BARRIER_BIT, &ce->flags);
1034 	ce->timeline = page_pack_bits(NULL, hwsp);
1035 	ce->ring = NULL;
1036 	ce->ring_size = ring_size;
1037 
1038 	i915_vm_put(ce->vm);
1039 	ce->vm = i915_vm_get(vm);
1040 
1041 	err = intel_context_pin(ce); /* perma-pin so it is always available */
1042 	if (err) {
1043 		intel_context_put(ce);
1044 		return ERR_PTR(err);
1045 	}
1046 
1047 	list_add_tail(&ce->pinned_contexts_link, &engine->pinned_contexts_list);
1048 
1049 	/*
1050 	 * Give our perma-pinned kernel timelines a separate lockdep class,
1051 	 * so that we can use them from within the normal user timelines
1052 	 * should we need to inject GPU operations during their request
1053 	 * construction.
1054 	 */
1055 	lockdep_set_class_and_name(&ce->timeline->mutex, key, name);
1056 
1057 	return ce;
1058 }
1059 
1060 void intel_engine_destroy_pinned_context(struct intel_context *ce)
1061 {
1062 	struct intel_engine_cs *engine = ce->engine;
1063 	struct i915_vma *hwsp = engine->status_page.vma;
1064 
1065 	GEM_BUG_ON(ce->timeline->hwsp_ggtt != hwsp);
1066 
1067 	mutex_lock(&hwsp->vm->mutex);
1068 	list_del(&ce->timeline->engine_link);
1069 	mutex_unlock(&hwsp->vm->mutex);
1070 
1071 	list_del(&ce->pinned_contexts_link);
1072 	intel_context_unpin(ce);
1073 	intel_context_put(ce);
1074 }
1075 
1076 static struct intel_context *
1077 create_kernel_context(struct intel_engine_cs *engine)
1078 {
1079 	static struct lock_class_key kernel;
1080 
1081 	return intel_engine_create_pinned_context(engine, engine->gt->vm, SZ_4K,
1082 						  I915_GEM_HWS_SEQNO_ADDR,
1083 						  &kernel, "kernel_context");
1084 }
1085 
1086 /**
1087  * intel_engines_init_common - initialize cengine state which might require hw access
1088  * @engine: Engine to initialize.
1089  *
1090  * Initializes @engine@ structure members shared between legacy and execlists
1091  * submission modes which do require hardware access.
1092  *
1093  * Typcally done at later stages of submission mode specific engine setup.
1094  *
1095  * Returns zero on success or an error code on failure.
1096  */
1097 static int engine_init_common(struct intel_engine_cs *engine)
1098 {
1099 	struct intel_context *ce;
1100 	int ret;
1101 
1102 	engine->set_default_submission(engine);
1103 
1104 	/*
1105 	 * We may need to do things with the shrinker which
1106 	 * require us to immediately switch back to the default
1107 	 * context. This can cause a problem as pinning the
1108 	 * default context also requires GTT space which may not
1109 	 * be available. To avoid this we always pin the default
1110 	 * context.
1111 	 */
1112 	ce = create_kernel_context(engine);
1113 	if (IS_ERR(ce))
1114 		return PTR_ERR(ce);
1115 
1116 	ret = measure_breadcrumb_dw(ce);
1117 	if (ret < 0)
1118 		goto err_context;
1119 
1120 	engine->emit_fini_breadcrumb_dw = ret;
1121 	engine->kernel_context = ce;
1122 
1123 	return 0;
1124 
1125 err_context:
1126 	intel_engine_destroy_pinned_context(ce);
1127 	return ret;
1128 }
1129 
1130 int intel_engines_init(struct intel_gt *gt)
1131 {
1132 	int (*setup)(struct intel_engine_cs *engine);
1133 	struct intel_engine_cs *engine;
1134 	enum intel_engine_id id;
1135 	int err;
1136 
1137 	if (intel_uc_uses_guc_submission(&gt->uc)) {
1138 		gt->submission_method = INTEL_SUBMISSION_GUC;
1139 		setup = intel_guc_submission_setup;
1140 	} else if (HAS_EXECLISTS(gt->i915)) {
1141 		gt->submission_method = INTEL_SUBMISSION_ELSP;
1142 		setup = intel_execlists_submission_setup;
1143 	} else {
1144 		gt->submission_method = INTEL_SUBMISSION_RING;
1145 		setup = intel_ring_submission_setup;
1146 	}
1147 
1148 	for_each_engine(engine, gt, id) {
1149 		err = engine_setup_common(engine);
1150 		if (err)
1151 			return err;
1152 
1153 		err = setup(engine);
1154 		if (err)
1155 			return err;
1156 
1157 		err = engine_init_common(engine);
1158 		if (err)
1159 			return err;
1160 
1161 		intel_engine_add_user(engine);
1162 	}
1163 
1164 	return 0;
1165 }
1166 
1167 /**
1168  * intel_engines_cleanup_common - cleans up the engine state created by
1169  *                                the common initiailizers.
1170  * @engine: Engine to cleanup.
1171  *
1172  * This cleans up everything created by the common helpers.
1173  */
1174 void intel_engine_cleanup_common(struct intel_engine_cs *engine)
1175 {
1176 	GEM_BUG_ON(!list_empty(&engine->sched_engine->requests));
1177 
1178 	i915_sched_engine_put(engine->sched_engine);
1179 	intel_breadcrumbs_put(engine->breadcrumbs);
1180 
1181 	intel_engine_fini_retire(engine);
1182 	intel_engine_cleanup_cmd_parser(engine);
1183 
1184 	if (engine->default_state)
1185 		fput(engine->default_state);
1186 
1187 	if (engine->kernel_context)
1188 		intel_engine_destroy_pinned_context(engine->kernel_context);
1189 
1190 	GEM_BUG_ON(!llist_empty(&engine->barrier_tasks));
1191 	cleanup_status_page(engine);
1192 
1193 	intel_wa_list_free(&engine->ctx_wa_list);
1194 	intel_wa_list_free(&engine->wa_list);
1195 	intel_wa_list_free(&engine->whitelist);
1196 }
1197 
1198 /**
1199  * intel_engine_resume - re-initializes the HW state of the engine
1200  * @engine: Engine to resume.
1201  *
1202  * Returns zero on success or an error code on failure.
1203  */
1204 int intel_engine_resume(struct intel_engine_cs *engine)
1205 {
1206 	intel_engine_apply_workarounds(engine);
1207 	intel_engine_apply_whitelist(engine);
1208 
1209 	return engine->resume(engine);
1210 }
1211 
1212 u64 intel_engine_get_active_head(const struct intel_engine_cs *engine)
1213 {
1214 	struct drm_i915_private *i915 = engine->i915;
1215 
1216 	u64 acthd;
1217 
1218 	if (GRAPHICS_VER(i915) >= 8)
1219 		acthd = ENGINE_READ64(engine, RING_ACTHD, RING_ACTHD_UDW);
1220 	else if (GRAPHICS_VER(i915) >= 4)
1221 		acthd = ENGINE_READ(engine, RING_ACTHD);
1222 	else
1223 		acthd = ENGINE_READ(engine, ACTHD);
1224 
1225 	return acthd;
1226 }
1227 
1228 u64 intel_engine_get_last_batch_head(const struct intel_engine_cs *engine)
1229 {
1230 	u64 bbaddr;
1231 
1232 	if (GRAPHICS_VER(engine->i915) >= 8)
1233 		bbaddr = ENGINE_READ64(engine, RING_BBADDR, RING_BBADDR_UDW);
1234 	else
1235 		bbaddr = ENGINE_READ(engine, RING_BBADDR);
1236 
1237 	return bbaddr;
1238 }
1239 
1240 static unsigned long stop_timeout(const struct intel_engine_cs *engine)
1241 {
1242 	if (in_atomic() || irqs_disabled()) /* inside atomic preempt-reset? */
1243 		return 0;
1244 
1245 	/*
1246 	 * If we are doing a normal GPU reset, we can take our time and allow
1247 	 * the engine to quiesce. We've stopped submission to the engine, and
1248 	 * if we wait long enough an innocent context should complete and
1249 	 * leave the engine idle. So they should not be caught unaware by
1250 	 * the forthcoming GPU reset (which usually follows the stop_cs)!
1251 	 */
1252 	return READ_ONCE(engine->props.stop_timeout_ms);
1253 }
1254 
1255 static int __intel_engine_stop_cs(struct intel_engine_cs *engine,
1256 				  int fast_timeout_us,
1257 				  int slow_timeout_ms)
1258 {
1259 	struct intel_uncore *uncore = engine->uncore;
1260 	const i915_reg_t mode = RING_MI_MODE(engine->mmio_base);
1261 	int err;
1262 
1263 	intel_uncore_write_fw(uncore, mode, _MASKED_BIT_ENABLE(STOP_RING));
1264 	err = __intel_wait_for_register_fw(engine->uncore, mode,
1265 					   MODE_IDLE, MODE_IDLE,
1266 					   fast_timeout_us,
1267 					   slow_timeout_ms,
1268 					   NULL);
1269 
1270 	/* A final mmio read to let GPU writes be hopefully flushed to memory */
1271 	intel_uncore_posting_read_fw(uncore, mode);
1272 	return err;
1273 }
1274 
1275 int intel_engine_stop_cs(struct intel_engine_cs *engine)
1276 {
1277 	int err = 0;
1278 
1279 	if (GRAPHICS_VER(engine->i915) < 3)
1280 		return -ENODEV;
1281 
1282 	ENGINE_TRACE(engine, "\n");
1283 	if (__intel_engine_stop_cs(engine, 1000, stop_timeout(engine))) {
1284 		ENGINE_TRACE(engine,
1285 			     "timed out on STOP_RING -> IDLE; HEAD:%04x, TAIL:%04x\n",
1286 			     ENGINE_READ_FW(engine, RING_HEAD) & HEAD_ADDR,
1287 			     ENGINE_READ_FW(engine, RING_TAIL) & TAIL_ADDR);
1288 
1289 		/*
1290 		 * Sometimes we observe that the idle flag is not
1291 		 * set even though the ring is empty. So double
1292 		 * check before giving up.
1293 		 */
1294 		if ((ENGINE_READ_FW(engine, RING_HEAD) & HEAD_ADDR) !=
1295 		    (ENGINE_READ_FW(engine, RING_TAIL) & TAIL_ADDR))
1296 			err = -ETIMEDOUT;
1297 	}
1298 
1299 	return err;
1300 }
1301 
1302 void intel_engine_cancel_stop_cs(struct intel_engine_cs *engine)
1303 {
1304 	ENGINE_TRACE(engine, "\n");
1305 
1306 	ENGINE_WRITE_FW(engine, RING_MI_MODE, _MASKED_BIT_DISABLE(STOP_RING));
1307 }
1308 
1309 static u32
1310 read_subslice_reg(const struct intel_engine_cs *engine,
1311 		  int slice, int subslice, i915_reg_t reg)
1312 {
1313 	return intel_uncore_read_with_mcr_steering(engine->uncore, reg,
1314 						   slice, subslice);
1315 }
1316 
1317 /* NB: please notice the memset */
1318 void intel_engine_get_instdone(const struct intel_engine_cs *engine,
1319 			       struct intel_instdone *instdone)
1320 {
1321 	struct drm_i915_private *i915 = engine->i915;
1322 	const struct sseu_dev_info *sseu = &engine->gt->info.sseu;
1323 	struct intel_uncore *uncore = engine->uncore;
1324 	u32 mmio_base = engine->mmio_base;
1325 	int slice;
1326 	int subslice;
1327 	int iter;
1328 
1329 	memset(instdone, 0, sizeof(*instdone));
1330 
1331 	if (GRAPHICS_VER(i915) >= 8) {
1332 		instdone->instdone =
1333 			intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
1334 
1335 		if (engine->id != RCS0)
1336 			return;
1337 
1338 		instdone->slice_common =
1339 			intel_uncore_read(uncore, GEN7_SC_INSTDONE);
1340 		if (GRAPHICS_VER(i915) >= 12) {
1341 			instdone->slice_common_extra[0] =
1342 				intel_uncore_read(uncore, GEN12_SC_INSTDONE_EXTRA);
1343 			instdone->slice_common_extra[1] =
1344 				intel_uncore_read(uncore, GEN12_SC_INSTDONE_EXTRA2);
1345 		}
1346 
1347 		if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 50)) {
1348 			for_each_instdone_gslice_dss_xehp(i915, sseu, iter, slice, subslice) {
1349 				instdone->sampler[slice][subslice] =
1350 					read_subslice_reg(engine, slice, subslice,
1351 							  GEN7_SAMPLER_INSTDONE);
1352 				instdone->row[slice][subslice] =
1353 					read_subslice_reg(engine, slice, subslice,
1354 							  GEN7_ROW_INSTDONE);
1355 			}
1356 		} else {
1357 			for_each_instdone_slice_subslice(i915, sseu, slice, subslice) {
1358 				instdone->sampler[slice][subslice] =
1359 					read_subslice_reg(engine, slice, subslice,
1360 							  GEN7_SAMPLER_INSTDONE);
1361 				instdone->row[slice][subslice] =
1362 					read_subslice_reg(engine, slice, subslice,
1363 							  GEN7_ROW_INSTDONE);
1364 			}
1365 		}
1366 
1367 		if (GRAPHICS_VER_FULL(i915) >= IP_VER(12, 55)) {
1368 			for_each_instdone_gslice_dss_xehp(i915, sseu, iter, slice, subslice)
1369 				instdone->geom_svg[slice][subslice] =
1370 					read_subslice_reg(engine, slice, subslice,
1371 							  XEHPG_INSTDONE_GEOM_SVG);
1372 		}
1373 	} else if (GRAPHICS_VER(i915) >= 7) {
1374 		instdone->instdone =
1375 			intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
1376 
1377 		if (engine->id != RCS0)
1378 			return;
1379 
1380 		instdone->slice_common =
1381 			intel_uncore_read(uncore, GEN7_SC_INSTDONE);
1382 		instdone->sampler[0][0] =
1383 			intel_uncore_read(uncore, GEN7_SAMPLER_INSTDONE);
1384 		instdone->row[0][0] =
1385 			intel_uncore_read(uncore, GEN7_ROW_INSTDONE);
1386 	} else if (GRAPHICS_VER(i915) >= 4) {
1387 		instdone->instdone =
1388 			intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
1389 		if (engine->id == RCS0)
1390 			/* HACK: Using the wrong struct member */
1391 			instdone->slice_common =
1392 				intel_uncore_read(uncore, GEN4_INSTDONE1);
1393 	} else {
1394 		instdone->instdone = intel_uncore_read(uncore, GEN2_INSTDONE);
1395 	}
1396 }
1397 
1398 static bool ring_is_idle(struct intel_engine_cs *engine)
1399 {
1400 	bool idle = true;
1401 
1402 	if (I915_SELFTEST_ONLY(!engine->mmio_base))
1403 		return true;
1404 
1405 	if (!intel_engine_pm_get_if_awake(engine))
1406 		return true;
1407 
1408 	/* First check that no commands are left in the ring */
1409 	if ((ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR) !=
1410 	    (ENGINE_READ(engine, RING_TAIL) & TAIL_ADDR))
1411 		idle = false;
1412 
1413 	/* No bit for gen2, so assume the CS parser is idle */
1414 	if (GRAPHICS_VER(engine->i915) > 2 &&
1415 	    !(ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE))
1416 		idle = false;
1417 
1418 	intel_engine_pm_put(engine);
1419 
1420 	return idle;
1421 }
1422 
1423 void __intel_engine_flush_submission(struct intel_engine_cs *engine, bool sync)
1424 {
1425 	struct tasklet_struct *t = &engine->sched_engine->tasklet;
1426 
1427 	if (!t->callback)
1428 		return;
1429 
1430 	local_bh_disable();
1431 	if (tasklet_trylock(t)) {
1432 		/* Must wait for any GPU reset in progress. */
1433 		if (__tasklet_is_enabled(t))
1434 			t->callback(t);
1435 		tasklet_unlock(t);
1436 	}
1437 	local_bh_enable();
1438 
1439 	/* Synchronise and wait for the tasklet on another CPU */
1440 	if (sync)
1441 		tasklet_unlock_wait(t);
1442 }
1443 
1444 /**
1445  * intel_engine_is_idle() - Report if the engine has finished process all work
1446  * @engine: the intel_engine_cs
1447  *
1448  * Return true if there are no requests pending, nothing left to be submitted
1449  * to hardware, and that the engine is idle.
1450  */
1451 bool intel_engine_is_idle(struct intel_engine_cs *engine)
1452 {
1453 	/* More white lies, if wedged, hw state is inconsistent */
1454 	if (intel_gt_is_wedged(engine->gt))
1455 		return true;
1456 
1457 	if (!intel_engine_pm_is_awake(engine))
1458 		return true;
1459 
1460 	/* Waiting to drain ELSP? */
1461 	intel_synchronize_hardirq(engine->i915);
1462 	intel_engine_flush_submission(engine);
1463 
1464 	/* ELSP is empty, but there are ready requests? E.g. after reset */
1465 	if (!i915_sched_engine_is_empty(engine->sched_engine))
1466 		return false;
1467 
1468 	/* Ring stopped? */
1469 	return ring_is_idle(engine);
1470 }
1471 
1472 bool intel_engines_are_idle(struct intel_gt *gt)
1473 {
1474 	struct intel_engine_cs *engine;
1475 	enum intel_engine_id id;
1476 
1477 	/*
1478 	 * If the driver is wedged, HW state may be very inconsistent and
1479 	 * report that it is still busy, even though we have stopped using it.
1480 	 */
1481 	if (intel_gt_is_wedged(gt))
1482 		return true;
1483 
1484 	/* Already parked (and passed an idleness test); must still be idle */
1485 	if (!READ_ONCE(gt->awake))
1486 		return true;
1487 
1488 	for_each_engine(engine, gt, id) {
1489 		if (!intel_engine_is_idle(engine))
1490 			return false;
1491 	}
1492 
1493 	return true;
1494 }
1495 
1496 bool intel_engine_irq_enable(struct intel_engine_cs *engine)
1497 {
1498 	if (!engine->irq_enable)
1499 		return false;
1500 
1501 	/* Caller disables interrupts */
1502 	spin_lock(&engine->gt->irq_lock);
1503 	engine->irq_enable(engine);
1504 	spin_unlock(&engine->gt->irq_lock);
1505 
1506 	return true;
1507 }
1508 
1509 void intel_engine_irq_disable(struct intel_engine_cs *engine)
1510 {
1511 	if (!engine->irq_disable)
1512 		return;
1513 
1514 	/* Caller disables interrupts */
1515 	spin_lock(&engine->gt->irq_lock);
1516 	engine->irq_disable(engine);
1517 	spin_unlock(&engine->gt->irq_lock);
1518 }
1519 
1520 void intel_engines_reset_default_submission(struct intel_gt *gt)
1521 {
1522 	struct intel_engine_cs *engine;
1523 	enum intel_engine_id id;
1524 
1525 	for_each_engine(engine, gt, id) {
1526 		if (engine->sanitize)
1527 			engine->sanitize(engine);
1528 
1529 		engine->set_default_submission(engine);
1530 	}
1531 }
1532 
1533 bool intel_engine_can_store_dword(struct intel_engine_cs *engine)
1534 {
1535 	switch (GRAPHICS_VER(engine->i915)) {
1536 	case 2:
1537 		return false; /* uses physical not virtual addresses */
1538 	case 3:
1539 		/* maybe only uses physical not virtual addresses */
1540 		return !(IS_I915G(engine->i915) || IS_I915GM(engine->i915));
1541 	case 4:
1542 		return !IS_I965G(engine->i915); /* who knows! */
1543 	case 6:
1544 		return engine->class != VIDEO_DECODE_CLASS; /* b0rked */
1545 	default:
1546 		return true;
1547 	}
1548 }
1549 
1550 static struct intel_timeline *get_timeline(struct i915_request *rq)
1551 {
1552 	struct intel_timeline *tl;
1553 
1554 	/*
1555 	 * Even though we are holding the engine->sched_engine->lock here, there
1556 	 * is no control over the submission queue per-se and we are
1557 	 * inspecting the active state at a random point in time, with an
1558 	 * unknown queue. Play safe and make sure the timeline remains valid.
1559 	 * (Only being used for pretty printing, one extra kref shouldn't
1560 	 * cause a camel stampede!)
1561 	 */
1562 	rcu_read_lock();
1563 	tl = rcu_dereference(rq->timeline);
1564 	if (!kref_get_unless_zero(&tl->kref))
1565 		tl = NULL;
1566 	rcu_read_unlock();
1567 
1568 	return tl;
1569 }
1570 
1571 static int print_ring(char *buf, int sz, struct i915_request *rq)
1572 {
1573 	int len = 0;
1574 
1575 	if (!i915_request_signaled(rq)) {
1576 		struct intel_timeline *tl = get_timeline(rq);
1577 
1578 		len = scnprintf(buf, sz,
1579 				"ring:{start:%08x, hwsp:%08x, seqno:%08x, runtime:%llums}, ",
1580 				i915_ggtt_offset(rq->ring->vma),
1581 				tl ? tl->hwsp_offset : 0,
1582 				hwsp_seqno(rq),
1583 				DIV_ROUND_CLOSEST_ULL(intel_context_get_total_runtime_ns(rq->context),
1584 						      1000 * 1000));
1585 
1586 		if (tl)
1587 			intel_timeline_put(tl);
1588 	}
1589 
1590 	return len;
1591 }
1592 
1593 static void hexdump(struct drm_printer *m, const void *buf, size_t len)
1594 {
1595 	const size_t rowsize = 8 * sizeof(u32);
1596 	const void *prev = NULL;
1597 	bool skip = false;
1598 	size_t pos;
1599 
1600 	for (pos = 0; pos < len; pos += rowsize) {
1601 		char line[128];
1602 
1603 		if (prev && !memcmp(prev, buf + pos, rowsize)) {
1604 			if (!skip) {
1605 				drm_printf(m, "*\n");
1606 				skip = true;
1607 			}
1608 			continue;
1609 		}
1610 
1611 		WARN_ON_ONCE(hex_dump_to_buffer(buf + pos, len - pos,
1612 						rowsize, sizeof(u32),
1613 						line, sizeof(line),
1614 						false) >= sizeof(line));
1615 		drm_printf(m, "[%04zx] %s\n", pos, line);
1616 
1617 		prev = buf + pos;
1618 		skip = false;
1619 	}
1620 }
1621 
1622 static const char *repr_timer(const struct timer_list *t)
1623 {
1624 	if (!READ_ONCE(t->expires))
1625 		return "inactive";
1626 
1627 	if (timer_pending(t))
1628 		return "active";
1629 
1630 	return "expired";
1631 }
1632 
1633 static void intel_engine_print_registers(struct intel_engine_cs *engine,
1634 					 struct drm_printer *m)
1635 {
1636 	struct drm_i915_private *dev_priv = engine->i915;
1637 	struct intel_engine_execlists * const execlists = &engine->execlists;
1638 	u64 addr;
1639 
1640 	if (engine->id == RENDER_CLASS && IS_GRAPHICS_VER(dev_priv, 4, 7))
1641 		drm_printf(m, "\tCCID: 0x%08x\n", ENGINE_READ(engine, CCID));
1642 	if (HAS_EXECLISTS(dev_priv)) {
1643 		drm_printf(m, "\tEL_STAT_HI: 0x%08x\n",
1644 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_HI));
1645 		drm_printf(m, "\tEL_STAT_LO: 0x%08x\n",
1646 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_LO));
1647 	}
1648 	drm_printf(m, "\tRING_START: 0x%08x\n",
1649 		   ENGINE_READ(engine, RING_START));
1650 	drm_printf(m, "\tRING_HEAD:  0x%08x\n",
1651 		   ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR);
1652 	drm_printf(m, "\tRING_TAIL:  0x%08x\n",
1653 		   ENGINE_READ(engine, RING_TAIL) & TAIL_ADDR);
1654 	drm_printf(m, "\tRING_CTL:   0x%08x%s\n",
1655 		   ENGINE_READ(engine, RING_CTL),
1656 		   ENGINE_READ(engine, RING_CTL) & (RING_WAIT | RING_WAIT_SEMAPHORE) ? " [waiting]" : "");
1657 	if (GRAPHICS_VER(engine->i915) > 2) {
1658 		drm_printf(m, "\tRING_MODE:  0x%08x%s\n",
1659 			   ENGINE_READ(engine, RING_MI_MODE),
1660 			   ENGINE_READ(engine, RING_MI_MODE) & (MODE_IDLE) ? " [idle]" : "");
1661 	}
1662 
1663 	if (GRAPHICS_VER(dev_priv) >= 6) {
1664 		drm_printf(m, "\tRING_IMR:   0x%08x\n",
1665 			   ENGINE_READ(engine, RING_IMR));
1666 		drm_printf(m, "\tRING_ESR:   0x%08x\n",
1667 			   ENGINE_READ(engine, RING_ESR));
1668 		drm_printf(m, "\tRING_EMR:   0x%08x\n",
1669 			   ENGINE_READ(engine, RING_EMR));
1670 		drm_printf(m, "\tRING_EIR:   0x%08x\n",
1671 			   ENGINE_READ(engine, RING_EIR));
1672 	}
1673 
1674 	addr = intel_engine_get_active_head(engine);
1675 	drm_printf(m, "\tACTHD:  0x%08x_%08x\n",
1676 		   upper_32_bits(addr), lower_32_bits(addr));
1677 	addr = intel_engine_get_last_batch_head(engine);
1678 	drm_printf(m, "\tBBADDR: 0x%08x_%08x\n",
1679 		   upper_32_bits(addr), lower_32_bits(addr));
1680 	if (GRAPHICS_VER(dev_priv) >= 8)
1681 		addr = ENGINE_READ64(engine, RING_DMA_FADD, RING_DMA_FADD_UDW);
1682 	else if (GRAPHICS_VER(dev_priv) >= 4)
1683 		addr = ENGINE_READ(engine, RING_DMA_FADD);
1684 	else
1685 		addr = ENGINE_READ(engine, DMA_FADD_I8XX);
1686 	drm_printf(m, "\tDMA_FADDR: 0x%08x_%08x\n",
1687 		   upper_32_bits(addr), lower_32_bits(addr));
1688 	if (GRAPHICS_VER(dev_priv) >= 4) {
1689 		drm_printf(m, "\tIPEIR: 0x%08x\n",
1690 			   ENGINE_READ(engine, RING_IPEIR));
1691 		drm_printf(m, "\tIPEHR: 0x%08x\n",
1692 			   ENGINE_READ(engine, RING_IPEHR));
1693 	} else {
1694 		drm_printf(m, "\tIPEIR: 0x%08x\n", ENGINE_READ(engine, IPEIR));
1695 		drm_printf(m, "\tIPEHR: 0x%08x\n", ENGINE_READ(engine, IPEHR));
1696 	}
1697 
1698 	if (intel_engine_uses_guc(engine)) {
1699 		/* nothing to print yet */
1700 	} else if (HAS_EXECLISTS(dev_priv)) {
1701 		struct i915_request * const *port, *rq;
1702 		const u32 *hws =
1703 			&engine->status_page.addr[I915_HWS_CSB_BUF0_INDEX];
1704 		const u8 num_entries = execlists->csb_size;
1705 		unsigned int idx;
1706 		u8 read, write;
1707 
1708 		drm_printf(m, "\tExeclist tasklet queued? %s (%s), preempt? %s, timeslice? %s\n",
1709 			   yesno(test_bit(TASKLET_STATE_SCHED,
1710 					  &engine->sched_engine->tasklet.state)),
1711 			   enableddisabled(!atomic_read(&engine->sched_engine->tasklet.count)),
1712 			   repr_timer(&engine->execlists.preempt),
1713 			   repr_timer(&engine->execlists.timer));
1714 
1715 		read = execlists->csb_head;
1716 		write = READ_ONCE(*execlists->csb_write);
1717 
1718 		drm_printf(m, "\tExeclist status: 0x%08x %08x; CSB read:%d, write:%d, entries:%d\n",
1719 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_LO),
1720 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_HI),
1721 			   read, write, num_entries);
1722 
1723 		if (read >= num_entries)
1724 			read = 0;
1725 		if (write >= num_entries)
1726 			write = 0;
1727 		if (read > write)
1728 			write += num_entries;
1729 		while (read < write) {
1730 			idx = ++read % num_entries;
1731 			drm_printf(m, "\tExeclist CSB[%d]: 0x%08x, context: %d\n",
1732 				   idx, hws[idx * 2], hws[idx * 2 + 1]);
1733 		}
1734 
1735 		i915_sched_engine_active_lock_bh(engine->sched_engine);
1736 		rcu_read_lock();
1737 		for (port = execlists->active; (rq = *port); port++) {
1738 			char hdr[160];
1739 			int len;
1740 
1741 			len = scnprintf(hdr, sizeof(hdr),
1742 					"\t\tActive[%d]:  ccid:%08x%s%s, ",
1743 					(int)(port - execlists->active),
1744 					rq->context->lrc.ccid,
1745 					intel_context_is_closed(rq->context) ? "!" : "",
1746 					intel_context_is_banned(rq->context) ? "*" : "");
1747 			len += print_ring(hdr + len, sizeof(hdr) - len, rq);
1748 			scnprintf(hdr + len, sizeof(hdr) - len, "rq: ");
1749 			i915_request_show(m, rq, hdr, 0);
1750 		}
1751 		for (port = execlists->pending; (rq = *port); port++) {
1752 			char hdr[160];
1753 			int len;
1754 
1755 			len = scnprintf(hdr, sizeof(hdr),
1756 					"\t\tPending[%d]: ccid:%08x%s%s, ",
1757 					(int)(port - execlists->pending),
1758 					rq->context->lrc.ccid,
1759 					intel_context_is_closed(rq->context) ? "!" : "",
1760 					intel_context_is_banned(rq->context) ? "*" : "");
1761 			len += print_ring(hdr + len, sizeof(hdr) - len, rq);
1762 			scnprintf(hdr + len, sizeof(hdr) - len, "rq: ");
1763 			i915_request_show(m, rq, hdr, 0);
1764 		}
1765 		rcu_read_unlock();
1766 		i915_sched_engine_active_unlock_bh(engine->sched_engine);
1767 	} else if (GRAPHICS_VER(dev_priv) > 6) {
1768 		drm_printf(m, "\tPP_DIR_BASE: 0x%08x\n",
1769 			   ENGINE_READ(engine, RING_PP_DIR_BASE));
1770 		drm_printf(m, "\tPP_DIR_BASE_READ: 0x%08x\n",
1771 			   ENGINE_READ(engine, RING_PP_DIR_BASE_READ));
1772 		drm_printf(m, "\tPP_DIR_DCLV: 0x%08x\n",
1773 			   ENGINE_READ(engine, RING_PP_DIR_DCLV));
1774 	}
1775 }
1776 
1777 static void print_request_ring(struct drm_printer *m, struct i915_request *rq)
1778 {
1779 	struct i915_vma_resource *vma_res = rq->batch_res;
1780 	void *ring;
1781 	int size;
1782 
1783 	drm_printf(m,
1784 		   "[head %04x, postfix %04x, tail %04x, batch 0x%08x_%08x]:\n",
1785 		   rq->head, rq->postfix, rq->tail,
1786 		   vma_res ? upper_32_bits(vma_res->start) : ~0u,
1787 		   vma_res ? lower_32_bits(vma_res->start) : ~0u);
1788 
1789 	size = rq->tail - rq->head;
1790 	if (rq->tail < rq->head)
1791 		size += rq->ring->size;
1792 
1793 	ring = kmalloc(size, GFP_ATOMIC);
1794 	if (ring) {
1795 		const void *vaddr = rq->ring->vaddr;
1796 		unsigned int head = rq->head;
1797 		unsigned int len = 0;
1798 
1799 		if (rq->tail < head) {
1800 			len = rq->ring->size - head;
1801 			memcpy(ring, vaddr + head, len);
1802 			head = 0;
1803 		}
1804 		memcpy(ring + len, vaddr + head, size - len);
1805 
1806 		hexdump(m, ring, size);
1807 		kfree(ring);
1808 	}
1809 }
1810 
1811 static unsigned long list_count(struct list_head *list)
1812 {
1813 	struct list_head *pos;
1814 	unsigned long count = 0;
1815 
1816 	list_for_each(pos, list)
1817 		count++;
1818 
1819 	return count;
1820 }
1821 
1822 static unsigned long read_ul(void *p, size_t x)
1823 {
1824 	return *(unsigned long *)(p + x);
1825 }
1826 
1827 static void print_properties(struct intel_engine_cs *engine,
1828 			     struct drm_printer *m)
1829 {
1830 	static const struct pmap {
1831 		size_t offset;
1832 		const char *name;
1833 	} props[] = {
1834 #define P(x) { \
1835 	.offset = offsetof(typeof(engine->props), x), \
1836 	.name = #x \
1837 }
1838 		P(heartbeat_interval_ms),
1839 		P(max_busywait_duration_ns),
1840 		P(preempt_timeout_ms),
1841 		P(stop_timeout_ms),
1842 		P(timeslice_duration_ms),
1843 
1844 		{},
1845 #undef P
1846 	};
1847 	const struct pmap *p;
1848 
1849 	drm_printf(m, "\tProperties:\n");
1850 	for (p = props; p->name; p++)
1851 		drm_printf(m, "\t\t%s: %lu [default %lu]\n",
1852 			   p->name,
1853 			   read_ul(&engine->props, p->offset),
1854 			   read_ul(&engine->defaults, p->offset));
1855 }
1856 
1857 static void engine_dump_request(struct i915_request *rq, struct drm_printer *m, const char *msg)
1858 {
1859 	struct intel_timeline *tl = get_timeline(rq);
1860 
1861 	i915_request_show(m, rq, msg, 0);
1862 
1863 	drm_printf(m, "\t\tring->start:  0x%08x\n",
1864 		   i915_ggtt_offset(rq->ring->vma));
1865 	drm_printf(m, "\t\tring->head:   0x%08x\n",
1866 		   rq->ring->head);
1867 	drm_printf(m, "\t\tring->tail:   0x%08x\n",
1868 		   rq->ring->tail);
1869 	drm_printf(m, "\t\tring->emit:   0x%08x\n",
1870 		   rq->ring->emit);
1871 	drm_printf(m, "\t\tring->space:  0x%08x\n",
1872 		   rq->ring->space);
1873 
1874 	if (tl) {
1875 		drm_printf(m, "\t\tring->hwsp:   0x%08x\n",
1876 			   tl->hwsp_offset);
1877 		intel_timeline_put(tl);
1878 	}
1879 
1880 	print_request_ring(m, rq);
1881 
1882 	if (rq->context->lrc_reg_state) {
1883 		drm_printf(m, "Logical Ring Context:\n");
1884 		hexdump(m, rq->context->lrc_reg_state, PAGE_SIZE);
1885 	}
1886 }
1887 
1888 void intel_engine_dump_active_requests(struct list_head *requests,
1889 				       struct i915_request *hung_rq,
1890 				       struct drm_printer *m)
1891 {
1892 	struct i915_request *rq;
1893 	const char *msg;
1894 	enum i915_request_state state;
1895 
1896 	list_for_each_entry(rq, requests, sched.link) {
1897 		if (rq == hung_rq)
1898 			continue;
1899 
1900 		state = i915_test_request_state(rq);
1901 		if (state < I915_REQUEST_QUEUED)
1902 			continue;
1903 
1904 		if (state == I915_REQUEST_ACTIVE)
1905 			msg = "\t\tactive on engine";
1906 		else
1907 			msg = "\t\tactive in queue";
1908 
1909 		engine_dump_request(rq, m, msg);
1910 	}
1911 }
1912 
1913 static void engine_dump_active_requests(struct intel_engine_cs *engine, struct drm_printer *m)
1914 {
1915 	struct i915_request *hung_rq = NULL;
1916 	struct intel_context *ce;
1917 	bool guc;
1918 
1919 	/*
1920 	 * No need for an engine->irq_seqno_barrier() before the seqno reads.
1921 	 * The GPU is still running so requests are still executing and any
1922 	 * hardware reads will be out of date by the time they are reported.
1923 	 * But the intention here is just to report an instantaneous snapshot
1924 	 * so that's fine.
1925 	 */
1926 	lockdep_assert_held(&engine->sched_engine->lock);
1927 
1928 	drm_printf(m, "\tRequests:\n");
1929 
1930 	guc = intel_uc_uses_guc_submission(&engine->gt->uc);
1931 	if (guc) {
1932 		ce = intel_engine_get_hung_context(engine);
1933 		if (ce)
1934 			hung_rq = intel_context_find_active_request(ce);
1935 	} else {
1936 		hung_rq = intel_engine_execlist_find_hung_request(engine);
1937 	}
1938 
1939 	if (hung_rq)
1940 		engine_dump_request(hung_rq, m, "\t\thung");
1941 
1942 	if (guc)
1943 		intel_guc_dump_active_requests(engine, hung_rq, m);
1944 	else
1945 		intel_engine_dump_active_requests(&engine->sched_engine->requests,
1946 						  hung_rq, m);
1947 }
1948 
1949 void intel_engine_dump(struct intel_engine_cs *engine,
1950 		       struct drm_printer *m,
1951 		       const char *header, ...)
1952 {
1953 	struct i915_gpu_error * const error = &engine->i915->gpu_error;
1954 	struct i915_request *rq;
1955 	intel_wakeref_t wakeref;
1956 	unsigned long flags;
1957 	ktime_t dummy;
1958 
1959 	if (header) {
1960 		va_list ap;
1961 
1962 		va_start(ap, header);
1963 		drm_vprintf(m, header, &ap);
1964 		va_end(ap);
1965 	}
1966 
1967 	if (intel_gt_is_wedged(engine->gt))
1968 		drm_printf(m, "*** WEDGED ***\n");
1969 
1970 	drm_printf(m, "\tAwake? %d\n", atomic_read(&engine->wakeref.count));
1971 	drm_printf(m, "\tBarriers?: %s\n",
1972 		   yesno(!llist_empty(&engine->barrier_tasks)));
1973 	drm_printf(m, "\tLatency: %luus\n",
1974 		   ewma__engine_latency_read(&engine->latency));
1975 	if (intel_engine_supports_stats(engine))
1976 		drm_printf(m, "\tRuntime: %llums\n",
1977 			   ktime_to_ms(intel_engine_get_busy_time(engine,
1978 								  &dummy)));
1979 	drm_printf(m, "\tForcewake: %x domains, %d active\n",
1980 		   engine->fw_domain, READ_ONCE(engine->fw_active));
1981 
1982 	rcu_read_lock();
1983 	rq = READ_ONCE(engine->heartbeat.systole);
1984 	if (rq)
1985 		drm_printf(m, "\tHeartbeat: %d ms ago\n",
1986 			   jiffies_to_msecs(jiffies - rq->emitted_jiffies));
1987 	rcu_read_unlock();
1988 	drm_printf(m, "\tReset count: %d (global %d)\n",
1989 		   i915_reset_engine_count(error, engine),
1990 		   i915_reset_count(error));
1991 	print_properties(engine, m);
1992 
1993 	spin_lock_irqsave(&engine->sched_engine->lock, flags);
1994 	engine_dump_active_requests(engine, m);
1995 
1996 	drm_printf(m, "\tOn hold?: %lu\n",
1997 		   list_count(&engine->sched_engine->hold));
1998 	spin_unlock_irqrestore(&engine->sched_engine->lock, flags);
1999 
2000 	drm_printf(m, "\tMMIO base:  0x%08x\n", engine->mmio_base);
2001 	wakeref = intel_runtime_pm_get_if_in_use(engine->uncore->rpm);
2002 	if (wakeref) {
2003 		intel_engine_print_registers(engine, m);
2004 		intel_runtime_pm_put(engine->uncore->rpm, wakeref);
2005 	} else {
2006 		drm_printf(m, "\tDevice is asleep; skipping register dump\n");
2007 	}
2008 
2009 	intel_execlists_show_requests(engine, m, i915_request_show, 8);
2010 
2011 	drm_printf(m, "HWSP:\n");
2012 	hexdump(m, engine->status_page.addr, PAGE_SIZE);
2013 
2014 	drm_printf(m, "Idle? %s\n", yesno(intel_engine_is_idle(engine)));
2015 
2016 	intel_engine_print_breadcrumbs(engine, m);
2017 }
2018 
2019 /**
2020  * intel_engine_get_busy_time() - Return current accumulated engine busyness
2021  * @engine: engine to report on
2022  * @now: monotonic timestamp of sampling
2023  *
2024  * Returns accumulated time @engine was busy since engine stats were enabled.
2025  */
2026 ktime_t intel_engine_get_busy_time(struct intel_engine_cs *engine, ktime_t *now)
2027 {
2028 	return engine->busyness(engine, now);
2029 }
2030 
2031 struct intel_context *
2032 intel_engine_create_virtual(struct intel_engine_cs **siblings,
2033 			    unsigned int count, unsigned long flags)
2034 {
2035 	if (count == 0)
2036 		return ERR_PTR(-EINVAL);
2037 
2038 	if (count == 1 && !(flags & FORCE_VIRTUAL))
2039 		return intel_context_create(siblings[0]);
2040 
2041 	GEM_BUG_ON(!siblings[0]->cops->create_virtual);
2042 	return siblings[0]->cops->create_virtual(siblings, count, flags);
2043 }
2044 
2045 struct i915_request *
2046 intel_engine_execlist_find_hung_request(struct intel_engine_cs *engine)
2047 {
2048 	struct i915_request *request, *active = NULL;
2049 
2050 	/*
2051 	 * This search does not work in GuC submission mode. However, the GuC
2052 	 * will report the hanging context directly to the driver itself. So
2053 	 * the driver should never get here when in GuC mode.
2054 	 */
2055 	GEM_BUG_ON(intel_uc_uses_guc_submission(&engine->gt->uc));
2056 
2057 	/*
2058 	 * We are called by the error capture, reset and to dump engine
2059 	 * state at random points in time. In particular, note that neither is
2060 	 * crucially ordered with an interrupt. After a hang, the GPU is dead
2061 	 * and we assume that no more writes can happen (we waited long enough
2062 	 * for all writes that were in transaction to be flushed) - adding an
2063 	 * extra delay for a recent interrupt is pointless. Hence, we do
2064 	 * not need an engine->irq_seqno_barrier() before the seqno reads.
2065 	 * At all other times, we must assume the GPU is still running, but
2066 	 * we only care about the snapshot of this moment.
2067 	 */
2068 	lockdep_assert_held(&engine->sched_engine->lock);
2069 
2070 	rcu_read_lock();
2071 	request = execlists_active(&engine->execlists);
2072 	if (request) {
2073 		struct intel_timeline *tl = request->context->timeline;
2074 
2075 		list_for_each_entry_from_reverse(request, &tl->requests, link) {
2076 			if (__i915_request_is_complete(request))
2077 				break;
2078 
2079 			active = request;
2080 		}
2081 	}
2082 	rcu_read_unlock();
2083 	if (active)
2084 		return active;
2085 
2086 	list_for_each_entry(request, &engine->sched_engine->requests,
2087 			    sched.link) {
2088 		if (i915_test_request_state(request) != I915_REQUEST_ACTIVE)
2089 			continue;
2090 
2091 		active = request;
2092 		break;
2093 	}
2094 
2095 	return active;
2096 }
2097 
2098 void xehp_enable_ccs_engines(struct intel_engine_cs *engine)
2099 {
2100 	/*
2101 	 * If there are any non-fused-off CCS engines, we need to enable CCS
2102 	 * support in the RCU_MODE register.  This only needs to be done once,
2103 	 * so for simplicity we'll take care of this in the RCS engine's
2104 	 * resume handler; since the RCS and all CCS engines belong to the
2105 	 * same reset domain and are reset together, this will also take care
2106 	 * of re-applying the setting after i915-triggered resets.
2107 	 */
2108 	if (!CCS_MASK(engine->gt))
2109 		return;
2110 
2111 	intel_uncore_write(engine->uncore, GEN12_RCU_MODE,
2112 			   _MASKED_BIT_ENABLE(GEN12_RCU_MODE_CCS_ENABLE));
2113 }
2114 
2115 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
2116 #include "mock_engine.c"
2117 #include "selftest_engine.c"
2118 #include "selftest_engine_cs.c"
2119 #endif
2120