1 /*
2  * Copyright © 2016 Intel Corporation
3  *
4  * Permission is hereby granted, free of charge, to any person obtaining a
5  * copy of this software and associated documentation files (the "Software"),
6  * to deal in the Software without restriction, including without limitation
7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
8  * and/or sell copies of the Software, and to permit persons to whom the
9  * Software is furnished to do so, subject to the following conditions:
10  *
11  * The above copyright notice and this permission notice (including the next
12  * paragraph) shall be included in all copies or substantial portions of the
13  * Software.
14  *
15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
21  * IN THE SOFTWARE.
22  *
23  */
24 
25 #include <drm/drm_print.h>
26 
27 #include "gem/i915_gem_context.h"
28 
29 #include "i915_drv.h"
30 
31 #include "intel_engine.h"
32 #include "intel_engine_pm.h"
33 #include "intel_context.h"
34 #include "intel_lrc.h"
35 #include "intel_reset.h"
36 
37 /* Haswell does have the CXT_SIZE register however it does not appear to be
38  * valid. Now, docs explain in dwords what is in the context object. The full
39  * size is 70720 bytes, however, the power context and execlist context will
40  * never be saved (power context is stored elsewhere, and execlists don't work
41  * on HSW) - so the final size, including the extra state required for the
42  * Resource Streamer, is 66944 bytes, which rounds to 17 pages.
43  */
44 #define HSW_CXT_TOTAL_SIZE		(17 * PAGE_SIZE)
45 
46 #define DEFAULT_LR_CONTEXT_RENDER_SIZE	(22 * PAGE_SIZE)
47 #define GEN8_LR_CONTEXT_RENDER_SIZE	(20 * PAGE_SIZE)
48 #define GEN9_LR_CONTEXT_RENDER_SIZE	(22 * PAGE_SIZE)
49 #define GEN10_LR_CONTEXT_RENDER_SIZE	(18 * PAGE_SIZE)
50 #define GEN11_LR_CONTEXT_RENDER_SIZE	(14 * PAGE_SIZE)
51 
52 #define GEN8_LR_CONTEXT_OTHER_SIZE	( 2 * PAGE_SIZE)
53 
54 struct engine_class_info {
55 	const char *name;
56 	u8 uabi_class;
57 };
58 
59 static const struct engine_class_info intel_engine_classes[] = {
60 	[RENDER_CLASS] = {
61 		.name = "rcs",
62 		.uabi_class = I915_ENGINE_CLASS_RENDER,
63 	},
64 	[COPY_ENGINE_CLASS] = {
65 		.name = "bcs",
66 		.uabi_class = I915_ENGINE_CLASS_COPY,
67 	},
68 	[VIDEO_DECODE_CLASS] = {
69 		.name = "vcs",
70 		.uabi_class = I915_ENGINE_CLASS_VIDEO,
71 	},
72 	[VIDEO_ENHANCEMENT_CLASS] = {
73 		.name = "vecs",
74 		.uabi_class = I915_ENGINE_CLASS_VIDEO_ENHANCE,
75 	},
76 };
77 
78 #define MAX_MMIO_BASES 3
79 struct engine_info {
80 	unsigned int hw_id;
81 	u8 class;
82 	u8 instance;
83 	/* mmio bases table *must* be sorted in reverse gen order */
84 	struct engine_mmio_base {
85 		u32 gen : 8;
86 		u32 base : 24;
87 	} mmio_bases[MAX_MMIO_BASES];
88 };
89 
90 static const struct engine_info intel_engines[] = {
91 	[RCS0] = {
92 		.hw_id = RCS0_HW,
93 		.class = RENDER_CLASS,
94 		.instance = 0,
95 		.mmio_bases = {
96 			{ .gen = 1, .base = RENDER_RING_BASE }
97 		},
98 	},
99 	[BCS0] = {
100 		.hw_id = BCS0_HW,
101 		.class = COPY_ENGINE_CLASS,
102 		.instance = 0,
103 		.mmio_bases = {
104 			{ .gen = 6, .base = BLT_RING_BASE }
105 		},
106 	},
107 	[VCS0] = {
108 		.hw_id = VCS0_HW,
109 		.class = VIDEO_DECODE_CLASS,
110 		.instance = 0,
111 		.mmio_bases = {
112 			{ .gen = 11, .base = GEN11_BSD_RING_BASE },
113 			{ .gen = 6, .base = GEN6_BSD_RING_BASE },
114 			{ .gen = 4, .base = BSD_RING_BASE }
115 		},
116 	},
117 	[VCS1] = {
118 		.hw_id = VCS1_HW,
119 		.class = VIDEO_DECODE_CLASS,
120 		.instance = 1,
121 		.mmio_bases = {
122 			{ .gen = 11, .base = GEN11_BSD2_RING_BASE },
123 			{ .gen = 8, .base = GEN8_BSD2_RING_BASE }
124 		},
125 	},
126 	[VCS2] = {
127 		.hw_id = VCS2_HW,
128 		.class = VIDEO_DECODE_CLASS,
129 		.instance = 2,
130 		.mmio_bases = {
131 			{ .gen = 11, .base = GEN11_BSD3_RING_BASE }
132 		},
133 	},
134 	[VCS3] = {
135 		.hw_id = VCS3_HW,
136 		.class = VIDEO_DECODE_CLASS,
137 		.instance = 3,
138 		.mmio_bases = {
139 			{ .gen = 11, .base = GEN11_BSD4_RING_BASE }
140 		},
141 	},
142 	[VECS0] = {
143 		.hw_id = VECS0_HW,
144 		.class = VIDEO_ENHANCEMENT_CLASS,
145 		.instance = 0,
146 		.mmio_bases = {
147 			{ .gen = 11, .base = GEN11_VEBOX_RING_BASE },
148 			{ .gen = 7, .base = VEBOX_RING_BASE }
149 		},
150 	},
151 	[VECS1] = {
152 		.hw_id = VECS1_HW,
153 		.class = VIDEO_ENHANCEMENT_CLASS,
154 		.instance = 1,
155 		.mmio_bases = {
156 			{ .gen = 11, .base = GEN11_VEBOX2_RING_BASE }
157 		},
158 	},
159 };
160 
161 /**
162  * intel_engine_context_size() - return the size of the context for an engine
163  * @dev_priv: i915 device private
164  * @class: engine class
165  *
166  * Each engine class may require a different amount of space for a context
167  * image.
168  *
169  * Return: size (in bytes) of an engine class specific context image
170  *
171  * Note: this size includes the HWSP, which is part of the context image
172  * in LRC mode, but does not include the "shared data page" used with
173  * GuC submission. The caller should account for this if using the GuC.
174  */
175 u32 intel_engine_context_size(struct drm_i915_private *dev_priv, u8 class)
176 {
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 (INTEL_GEN(dev_priv)) {
184 		default:
185 			MISSING_CASE(INTEL_GEN(dev_priv));
186 			return DEFAULT_LR_CONTEXT_RENDER_SIZE;
187 		case 11:
188 			return GEN11_LR_CONTEXT_RENDER_SIZE;
189 		case 10:
190 			return GEN10_LR_CONTEXT_RENDER_SIZE;
191 		case 9:
192 			return GEN9_LR_CONTEXT_RENDER_SIZE;
193 		case 8:
194 			return GEN8_LR_CONTEXT_RENDER_SIZE;
195 		case 7:
196 			if (IS_HASWELL(dev_priv))
197 				return HSW_CXT_TOTAL_SIZE;
198 
199 			cxt_size = I915_READ(GEN7_CXT_SIZE);
200 			return round_up(GEN7_CXT_TOTAL_SIZE(cxt_size) * 64,
201 					PAGE_SIZE);
202 		case 6:
203 			cxt_size = I915_READ(CXT_SIZE);
204 			return round_up(GEN6_CXT_TOTAL_SIZE(cxt_size) * 64,
205 					PAGE_SIZE);
206 		case 5:
207 		case 4:
208 			/*
209 			 * There is a discrepancy here between the size reported
210 			 * by the register and the size of the context layout
211 			 * in the docs. Both are described as authorative!
212 			 *
213 			 * The discrepancy is on the order of a few cachelines,
214 			 * but the total is under one page (4k), which is our
215 			 * minimum allocation anyway so it should all come
216 			 * out in the wash.
217 			 */
218 			cxt_size = I915_READ(CXT_SIZE) + 1;
219 			DRM_DEBUG_DRIVER("gen%d CXT_SIZE = %d bytes [0x%08x]\n",
220 					 INTEL_GEN(dev_priv),
221 					 cxt_size * 64,
222 					 cxt_size - 1);
223 			return round_up(cxt_size * 64, PAGE_SIZE);
224 		case 3:
225 		case 2:
226 		/* For the special day when i810 gets merged. */
227 		case 1:
228 			return 0;
229 		}
230 		break;
231 	default:
232 		MISSING_CASE(class);
233 		/* fall through */
234 	case VIDEO_DECODE_CLASS:
235 	case VIDEO_ENHANCEMENT_CLASS:
236 	case COPY_ENGINE_CLASS:
237 		if (INTEL_GEN(dev_priv) < 8)
238 			return 0;
239 		return GEN8_LR_CONTEXT_OTHER_SIZE;
240 	}
241 }
242 
243 static u32 __engine_mmio_base(struct drm_i915_private *i915,
244 			      const struct engine_mmio_base *bases)
245 {
246 	int i;
247 
248 	for (i = 0; i < MAX_MMIO_BASES; i++)
249 		if (INTEL_GEN(i915) >= bases[i].gen)
250 			break;
251 
252 	GEM_BUG_ON(i == MAX_MMIO_BASES);
253 	GEM_BUG_ON(!bases[i].base);
254 
255 	return bases[i].base;
256 }
257 
258 static void __sprint_engine_name(char *name, const struct engine_info *info)
259 {
260 	WARN_ON(snprintf(name, INTEL_ENGINE_CS_MAX_NAME, "%s%u",
261 			 intel_engine_classes[info->class].name,
262 			 info->instance) >= INTEL_ENGINE_CS_MAX_NAME);
263 }
264 
265 void intel_engine_set_hwsp_writemask(struct intel_engine_cs *engine, u32 mask)
266 {
267 	/*
268 	 * Though they added more rings on g4x/ilk, they did not add
269 	 * per-engine HWSTAM until gen6.
270 	 */
271 	if (INTEL_GEN(engine->i915) < 6 && engine->class != RENDER_CLASS)
272 		return;
273 
274 	if (INTEL_GEN(engine->i915) >= 3)
275 		ENGINE_WRITE(engine, RING_HWSTAM, mask);
276 	else
277 		ENGINE_WRITE16(engine, RING_HWSTAM, mask);
278 }
279 
280 static void intel_engine_sanitize_mmio(struct intel_engine_cs *engine)
281 {
282 	/* Mask off all writes into the unknown HWSP */
283 	intel_engine_set_hwsp_writemask(engine, ~0u);
284 }
285 
286 static int
287 intel_engine_setup(struct drm_i915_private *dev_priv,
288 		   enum intel_engine_id id)
289 {
290 	const struct engine_info *info = &intel_engines[id];
291 	struct intel_engine_cs *engine;
292 
293 	GEM_BUG_ON(info->class >= ARRAY_SIZE(intel_engine_classes));
294 
295 	BUILD_BUG_ON(MAX_ENGINE_CLASS >= BIT(GEN11_ENGINE_CLASS_WIDTH));
296 	BUILD_BUG_ON(MAX_ENGINE_INSTANCE >= BIT(GEN11_ENGINE_INSTANCE_WIDTH));
297 
298 	if (GEM_DEBUG_WARN_ON(info->class > MAX_ENGINE_CLASS))
299 		return -EINVAL;
300 
301 	if (GEM_DEBUG_WARN_ON(info->instance > MAX_ENGINE_INSTANCE))
302 		return -EINVAL;
303 
304 	if (GEM_DEBUG_WARN_ON(dev_priv->engine_class[info->class][info->instance]))
305 		return -EINVAL;
306 
307 	GEM_BUG_ON(dev_priv->engine[id]);
308 	engine = kzalloc(sizeof(*engine), GFP_KERNEL);
309 	if (!engine)
310 		return -ENOMEM;
311 
312 	BUILD_BUG_ON(BITS_PER_TYPE(engine->mask) < I915_NUM_ENGINES);
313 
314 	engine->id = id;
315 	engine->mask = BIT(id);
316 	engine->i915 = dev_priv;
317 	engine->uncore = &dev_priv->uncore;
318 	__sprint_engine_name(engine->name, info);
319 	engine->hw_id = engine->guc_id = info->hw_id;
320 	engine->mmio_base = __engine_mmio_base(dev_priv, info->mmio_bases);
321 	engine->class = info->class;
322 	engine->instance = info->instance;
323 
324 	/*
325 	 * To be overridden by the backend on setup. However to facilitate
326 	 * cleanup on error during setup, we always provide the destroy vfunc.
327 	 */
328 	engine->destroy = (typeof(engine->destroy))kfree;
329 
330 	engine->uabi_class = intel_engine_classes[info->class].uabi_class;
331 
332 	engine->context_size = intel_engine_context_size(dev_priv,
333 							 engine->class);
334 	if (WARN_ON(engine->context_size > BIT(20)))
335 		engine->context_size = 0;
336 	if (engine->context_size)
337 		DRIVER_CAPS(dev_priv)->has_logical_contexts = true;
338 
339 	/* Nothing to do here, execute in order of dependencies */
340 	engine->schedule = NULL;
341 
342 	seqlock_init(&engine->stats.lock);
343 
344 	ATOMIC_INIT_NOTIFIER_HEAD(&engine->context_status_notifier);
345 
346 	/* Scrub mmio state on takeover */
347 	intel_engine_sanitize_mmio(engine);
348 
349 	dev_priv->engine_class[info->class][info->instance] = engine;
350 	dev_priv->engine[id] = engine;
351 	return 0;
352 }
353 
354 static void __setup_engine_capabilities(struct intel_engine_cs *engine)
355 {
356 	struct drm_i915_private *i915 = engine->i915;
357 
358 	if (engine->class == VIDEO_DECODE_CLASS) {
359 		/*
360 		 * HEVC support is present on first engine instance
361 		 * before Gen11 and on all instances afterwards.
362 		 */
363 		if (INTEL_GEN(i915) >= 11 ||
364 		    (INTEL_GEN(i915) >= 9 && engine->instance == 0))
365 			engine->uabi_capabilities |=
366 				I915_VIDEO_CLASS_CAPABILITY_HEVC;
367 
368 		/*
369 		 * SFC block is present only on even logical engine
370 		 * instances.
371 		 */
372 		if ((INTEL_GEN(i915) >= 11 &&
373 		     RUNTIME_INFO(i915)->vdbox_sfc_access & engine->mask) ||
374 		    (INTEL_GEN(i915) >= 9 && engine->instance == 0))
375 			engine->uabi_capabilities |=
376 				I915_VIDEO_AND_ENHANCE_CLASS_CAPABILITY_SFC;
377 	} else if (engine->class == VIDEO_ENHANCEMENT_CLASS) {
378 		if (INTEL_GEN(i915) >= 9)
379 			engine->uabi_capabilities |=
380 				I915_VIDEO_AND_ENHANCE_CLASS_CAPABILITY_SFC;
381 	}
382 }
383 
384 static void intel_setup_engine_capabilities(struct drm_i915_private *i915)
385 {
386 	struct intel_engine_cs *engine;
387 	enum intel_engine_id id;
388 
389 	for_each_engine(engine, i915, id)
390 		__setup_engine_capabilities(engine);
391 }
392 
393 /**
394  * intel_engines_cleanup() - free the resources allocated for Command Streamers
395  * @i915: the i915 devic
396  */
397 void intel_engines_cleanup(struct drm_i915_private *i915)
398 {
399 	struct intel_engine_cs *engine;
400 	enum intel_engine_id id;
401 
402 	for_each_engine(engine, i915, id) {
403 		engine->destroy(engine);
404 		i915->engine[id] = NULL;
405 	}
406 }
407 
408 /**
409  * intel_engines_init_mmio() - allocate and prepare the Engine Command Streamers
410  * @i915: the i915 device
411  *
412  * Return: non-zero if the initialization failed.
413  */
414 int intel_engines_init_mmio(struct drm_i915_private *i915)
415 {
416 	struct intel_device_info *device_info = mkwrite_device_info(i915);
417 	const unsigned int engine_mask = INTEL_INFO(i915)->engine_mask;
418 	unsigned int mask = 0;
419 	unsigned int i;
420 	int err;
421 
422 	WARN_ON(engine_mask == 0);
423 	WARN_ON(engine_mask &
424 		GENMASK(BITS_PER_TYPE(mask) - 1, I915_NUM_ENGINES));
425 
426 	if (i915_inject_load_failure())
427 		return -ENODEV;
428 
429 	for (i = 0; i < ARRAY_SIZE(intel_engines); i++) {
430 		if (!HAS_ENGINE(i915, i))
431 			continue;
432 
433 		err = intel_engine_setup(i915, i);
434 		if (err)
435 			goto cleanup;
436 
437 		mask |= BIT(i);
438 	}
439 
440 	/*
441 	 * Catch failures to update intel_engines table when the new engines
442 	 * are added to the driver by a warning and disabling the forgotten
443 	 * engines.
444 	 */
445 	if (WARN_ON(mask != engine_mask))
446 		device_info->engine_mask = mask;
447 
448 	/* We always presume we have at least RCS available for later probing */
449 	if (WARN_ON(!HAS_ENGINE(i915, RCS0))) {
450 		err = -ENODEV;
451 		goto cleanup;
452 	}
453 
454 	RUNTIME_INFO(i915)->num_engines = hweight32(mask);
455 
456 	i915_check_and_clear_faults(i915);
457 
458 	intel_setup_engine_capabilities(i915);
459 
460 	return 0;
461 
462 cleanup:
463 	intel_engines_cleanup(i915);
464 	return err;
465 }
466 
467 /**
468  * intel_engines_init() - init the Engine Command Streamers
469  * @i915: i915 device private
470  *
471  * Return: non-zero if the initialization failed.
472  */
473 int intel_engines_init(struct drm_i915_private *i915)
474 {
475 	int (*init)(struct intel_engine_cs *engine);
476 	struct intel_engine_cs *engine;
477 	enum intel_engine_id id;
478 	int err;
479 
480 	if (HAS_EXECLISTS(i915))
481 		init = intel_execlists_submission_init;
482 	else
483 		init = intel_ring_submission_init;
484 
485 	for_each_engine(engine, i915, id) {
486 		err = init(engine);
487 		if (err)
488 			goto cleanup;
489 	}
490 
491 	return 0;
492 
493 cleanup:
494 	intel_engines_cleanup(i915);
495 	return err;
496 }
497 
498 static void intel_engine_init_batch_pool(struct intel_engine_cs *engine)
499 {
500 	i915_gem_batch_pool_init(&engine->batch_pool, engine);
501 }
502 
503 void intel_engine_init_execlists(struct intel_engine_cs *engine)
504 {
505 	struct intel_engine_execlists * const execlists = &engine->execlists;
506 
507 	execlists->port_mask = 1;
508 	GEM_BUG_ON(!is_power_of_2(execlists_num_ports(execlists)));
509 	GEM_BUG_ON(execlists_num_ports(execlists) > EXECLIST_MAX_PORTS);
510 
511 	execlists->queue_priority_hint = INT_MIN;
512 	execlists->queue = RB_ROOT_CACHED;
513 }
514 
515 static void cleanup_status_page(struct intel_engine_cs *engine)
516 {
517 	struct i915_vma *vma;
518 
519 	/* Prevent writes into HWSP after returning the page to the system */
520 	intel_engine_set_hwsp_writemask(engine, ~0u);
521 
522 	vma = fetch_and_zero(&engine->status_page.vma);
523 	if (!vma)
524 		return;
525 
526 	if (!HWS_NEEDS_PHYSICAL(engine->i915))
527 		i915_vma_unpin(vma);
528 
529 	i915_gem_object_unpin_map(vma->obj);
530 	i915_gem_object_put(vma->obj);
531 }
532 
533 static int pin_ggtt_status_page(struct intel_engine_cs *engine,
534 				struct i915_vma *vma)
535 {
536 	unsigned int flags;
537 
538 	flags = PIN_GLOBAL;
539 	if (!HAS_LLC(engine->i915))
540 		/*
541 		 * On g33, we cannot place HWS above 256MiB, so
542 		 * restrict its pinning to the low mappable arena.
543 		 * Though this restriction is not documented for
544 		 * gen4, gen5, or byt, they also behave similarly
545 		 * and hang if the HWS is placed at the top of the
546 		 * GTT. To generalise, it appears that all !llc
547 		 * platforms have issues with us placing the HWS
548 		 * above the mappable region (even though we never
549 		 * actually map it).
550 		 */
551 		flags |= PIN_MAPPABLE;
552 	else
553 		flags |= PIN_HIGH;
554 
555 	return i915_vma_pin(vma, 0, 0, flags);
556 }
557 
558 static int init_status_page(struct intel_engine_cs *engine)
559 {
560 	struct drm_i915_gem_object *obj;
561 	struct i915_vma *vma;
562 	void *vaddr;
563 	int ret;
564 
565 	/*
566 	 * Though the HWS register does support 36bit addresses, historically
567 	 * we have had hangs and corruption reported due to wild writes if
568 	 * the HWS is placed above 4G. We only allow objects to be allocated
569 	 * in GFP_DMA32 for i965, and no earlier physical address users had
570 	 * access to more than 4G.
571 	 */
572 	obj = i915_gem_object_create_internal(engine->i915, PAGE_SIZE);
573 	if (IS_ERR(obj)) {
574 		DRM_ERROR("Failed to allocate status page\n");
575 		return PTR_ERR(obj);
576 	}
577 
578 	i915_gem_object_set_cache_coherency(obj, I915_CACHE_LLC);
579 
580 	vma = i915_vma_instance(obj, &engine->i915->ggtt.vm, NULL);
581 	if (IS_ERR(vma)) {
582 		ret = PTR_ERR(vma);
583 		goto err;
584 	}
585 
586 	vaddr = i915_gem_object_pin_map(obj, I915_MAP_WB);
587 	if (IS_ERR(vaddr)) {
588 		ret = PTR_ERR(vaddr);
589 		goto err;
590 	}
591 
592 	engine->status_page.addr = memset(vaddr, 0, PAGE_SIZE);
593 	engine->status_page.vma = vma;
594 
595 	if (!HWS_NEEDS_PHYSICAL(engine->i915)) {
596 		ret = pin_ggtt_status_page(engine, vma);
597 		if (ret)
598 			goto err_unpin;
599 	}
600 
601 	return 0;
602 
603 err_unpin:
604 	i915_gem_object_unpin_map(obj);
605 err:
606 	i915_gem_object_put(obj);
607 	return ret;
608 }
609 
610 static int intel_engine_setup_common(struct intel_engine_cs *engine)
611 {
612 	int err;
613 
614 	init_llist_head(&engine->barrier_tasks);
615 
616 	err = init_status_page(engine);
617 	if (err)
618 		return err;
619 
620 	intel_engine_init_active(engine, ENGINE_PHYSICAL);
621 	intel_engine_init_breadcrumbs(engine);
622 	intel_engine_init_execlists(engine);
623 	intel_engine_init_hangcheck(engine);
624 	intel_engine_init_batch_pool(engine);
625 	intel_engine_init_cmd_parser(engine);
626 	intel_engine_init__pm(engine);
627 
628 	/* Use the whole device by default */
629 	engine->sseu =
630 		intel_sseu_from_device_info(&RUNTIME_INFO(engine->i915)->sseu);
631 
632 	return 0;
633 }
634 
635 /**
636  * intel_engines_setup- setup engine state not requiring hw access
637  * @i915: Device to setup.
638  *
639  * Initializes engine structure members shared between legacy and execlists
640  * submission modes which do not require hardware access.
641  *
642  * Typically done early in the submission mode specific engine setup stage.
643  */
644 int intel_engines_setup(struct drm_i915_private *i915)
645 {
646 	int (*setup)(struct intel_engine_cs *engine);
647 	struct intel_engine_cs *engine;
648 	enum intel_engine_id id;
649 	int err;
650 
651 	if (HAS_EXECLISTS(i915))
652 		setup = intel_execlists_submission_setup;
653 	else
654 		setup = intel_ring_submission_setup;
655 
656 	for_each_engine(engine, i915, id) {
657 		err = intel_engine_setup_common(engine);
658 		if (err)
659 			goto cleanup;
660 
661 		err = setup(engine);
662 		if (err)
663 			goto cleanup;
664 
665 		/* We expect the backend to take control over its state */
666 		GEM_BUG_ON(engine->destroy == (typeof(engine->destroy))kfree);
667 
668 		GEM_BUG_ON(!engine->cops);
669 	}
670 
671 	return 0;
672 
673 cleanup:
674 	intel_engines_cleanup(i915);
675 	return err;
676 }
677 
678 void intel_engines_set_scheduler_caps(struct drm_i915_private *i915)
679 {
680 	static const struct {
681 		u8 engine;
682 		u8 sched;
683 	} map[] = {
684 #define MAP(x, y) { ilog2(I915_ENGINE_HAS_##x), ilog2(I915_SCHEDULER_CAP_##y) }
685 		MAP(PREEMPTION, PREEMPTION),
686 		MAP(SEMAPHORES, SEMAPHORES),
687 #undef MAP
688 	};
689 	struct intel_engine_cs *engine;
690 	enum intel_engine_id id;
691 	u32 enabled, disabled;
692 
693 	enabled = 0;
694 	disabled = 0;
695 	for_each_engine(engine, i915, id) { /* all engines must agree! */
696 		int i;
697 
698 		if (engine->schedule)
699 			enabled |= (I915_SCHEDULER_CAP_ENABLED |
700 				    I915_SCHEDULER_CAP_PRIORITY);
701 		else
702 			disabled |= (I915_SCHEDULER_CAP_ENABLED |
703 				     I915_SCHEDULER_CAP_PRIORITY);
704 
705 		for (i = 0; i < ARRAY_SIZE(map); i++) {
706 			if (engine->flags & BIT(map[i].engine))
707 				enabled |= BIT(map[i].sched);
708 			else
709 				disabled |= BIT(map[i].sched);
710 		}
711 	}
712 
713 	i915->caps.scheduler = enabled & ~disabled;
714 	if (!(i915->caps.scheduler & I915_SCHEDULER_CAP_ENABLED))
715 		i915->caps.scheduler = 0;
716 }
717 
718 struct measure_breadcrumb {
719 	struct i915_request rq;
720 	struct i915_timeline timeline;
721 	struct intel_ring ring;
722 	u32 cs[1024];
723 };
724 
725 static int measure_breadcrumb_dw(struct intel_engine_cs *engine)
726 {
727 	struct measure_breadcrumb *frame;
728 	int dw = -ENOMEM;
729 
730 	GEM_BUG_ON(!engine->i915->gt.scratch);
731 
732 	frame = kzalloc(sizeof(*frame), GFP_KERNEL);
733 	if (!frame)
734 		return -ENOMEM;
735 
736 	if (i915_timeline_init(engine->i915,
737 			       &frame->timeline,
738 			       engine->status_page.vma))
739 		goto out_frame;
740 
741 	INIT_LIST_HEAD(&frame->ring.request_list);
742 	frame->ring.timeline = &frame->timeline;
743 	frame->ring.vaddr = frame->cs;
744 	frame->ring.size = sizeof(frame->cs);
745 	frame->ring.effective_size = frame->ring.size;
746 	intel_ring_update_space(&frame->ring);
747 
748 	frame->rq.i915 = engine->i915;
749 	frame->rq.engine = engine;
750 	frame->rq.ring = &frame->ring;
751 	frame->rq.timeline = &frame->timeline;
752 
753 	dw = i915_timeline_pin(&frame->timeline);
754 	if (dw < 0)
755 		goto out_timeline;
756 
757 	dw = engine->emit_fini_breadcrumb(&frame->rq, frame->cs) - frame->cs;
758 	GEM_BUG_ON(dw & 1); /* RING_TAIL must be qword aligned */
759 
760 	i915_timeline_unpin(&frame->timeline);
761 
762 out_timeline:
763 	i915_timeline_fini(&frame->timeline);
764 out_frame:
765 	kfree(frame);
766 	return dw;
767 }
768 
769 static int pin_context(struct i915_gem_context *ctx,
770 		       struct intel_engine_cs *engine,
771 		       struct intel_context **out)
772 {
773 	struct intel_context *ce;
774 	int err;
775 
776 	ce = i915_gem_context_get_engine(ctx, engine->id);
777 	if (IS_ERR(ce))
778 		return PTR_ERR(ce);
779 
780 	err = intel_context_pin(ce);
781 	intel_context_put(ce);
782 	if (err)
783 		return err;
784 
785 	*out = ce;
786 	return 0;
787 }
788 
789 void
790 intel_engine_init_active(struct intel_engine_cs *engine, unsigned int subclass)
791 {
792 	INIT_LIST_HEAD(&engine->active.requests);
793 
794 	spin_lock_init(&engine->active.lock);
795 	lockdep_set_subclass(&engine->active.lock, subclass);
796 
797 	/*
798 	 * Due to an interesting quirk in lockdep's internal debug tracking,
799 	 * after setting a subclass we must ensure the lock is used. Otherwise,
800 	 * nr_unused_locks is incremented once too often.
801 	 */
802 #ifdef CONFIG_DEBUG_LOCK_ALLOC
803 	local_irq_disable();
804 	lock_map_acquire(&engine->active.lock.dep_map);
805 	lock_map_release(&engine->active.lock.dep_map);
806 	local_irq_enable();
807 #endif
808 }
809 
810 /**
811  * intel_engines_init_common - initialize cengine state which might require hw access
812  * @engine: Engine to initialize.
813  *
814  * Initializes @engine@ structure members shared between legacy and execlists
815  * submission modes which do require hardware access.
816  *
817  * Typcally done at later stages of submission mode specific engine setup.
818  *
819  * Returns zero on success or an error code on failure.
820  */
821 int intel_engine_init_common(struct intel_engine_cs *engine)
822 {
823 	struct drm_i915_private *i915 = engine->i915;
824 	int ret;
825 
826 	/* We may need to do things with the shrinker which
827 	 * require us to immediately switch back to the default
828 	 * context. This can cause a problem as pinning the
829 	 * default context also requires GTT space which may not
830 	 * be available. To avoid this we always pin the default
831 	 * context.
832 	 */
833 	ret = pin_context(i915->kernel_context, engine,
834 			  &engine->kernel_context);
835 	if (ret)
836 		return ret;
837 
838 	/*
839 	 * Similarly the preempt context must always be available so that
840 	 * we can interrupt the engine at any time. However, as preemption
841 	 * is optional, we allow it to fail.
842 	 */
843 	if (i915->preempt_context)
844 		pin_context(i915->preempt_context, engine,
845 			    &engine->preempt_context);
846 
847 	ret = measure_breadcrumb_dw(engine);
848 	if (ret < 0)
849 		goto err_unpin;
850 
851 	engine->emit_fini_breadcrumb_dw = ret;
852 
853 	engine->set_default_submission(engine);
854 
855 	return 0;
856 
857 err_unpin:
858 	if (engine->preempt_context)
859 		intel_context_unpin(engine->preempt_context);
860 	intel_context_unpin(engine->kernel_context);
861 	return ret;
862 }
863 
864 /**
865  * intel_engines_cleanup_common - cleans up the engine state created by
866  *                                the common initiailizers.
867  * @engine: Engine to cleanup.
868  *
869  * This cleans up everything created by the common helpers.
870  */
871 void intel_engine_cleanup_common(struct intel_engine_cs *engine)
872 {
873 	GEM_BUG_ON(!list_empty(&engine->active.requests));
874 
875 	cleanup_status_page(engine);
876 
877 	intel_engine_fini_breadcrumbs(engine);
878 	intel_engine_cleanup_cmd_parser(engine);
879 	i915_gem_batch_pool_fini(&engine->batch_pool);
880 
881 	if (engine->default_state)
882 		i915_gem_object_put(engine->default_state);
883 
884 	if (engine->preempt_context)
885 		intel_context_unpin(engine->preempt_context);
886 	intel_context_unpin(engine->kernel_context);
887 	GEM_BUG_ON(!llist_empty(&engine->barrier_tasks));
888 
889 	intel_wa_list_free(&engine->ctx_wa_list);
890 	intel_wa_list_free(&engine->wa_list);
891 	intel_wa_list_free(&engine->whitelist);
892 }
893 
894 u64 intel_engine_get_active_head(const struct intel_engine_cs *engine)
895 {
896 	struct drm_i915_private *i915 = engine->i915;
897 
898 	u64 acthd;
899 
900 	if (INTEL_GEN(i915) >= 8)
901 		acthd = ENGINE_READ64(engine, RING_ACTHD, RING_ACTHD_UDW);
902 	else if (INTEL_GEN(i915) >= 4)
903 		acthd = ENGINE_READ(engine, RING_ACTHD);
904 	else
905 		acthd = ENGINE_READ(engine, ACTHD);
906 
907 	return acthd;
908 }
909 
910 u64 intel_engine_get_last_batch_head(const struct intel_engine_cs *engine)
911 {
912 	u64 bbaddr;
913 
914 	if (INTEL_GEN(engine->i915) >= 8)
915 		bbaddr = ENGINE_READ64(engine, RING_BBADDR, RING_BBADDR_UDW);
916 	else
917 		bbaddr = ENGINE_READ(engine, RING_BBADDR);
918 
919 	return bbaddr;
920 }
921 
922 int intel_engine_stop_cs(struct intel_engine_cs *engine)
923 {
924 	struct intel_uncore *uncore = engine->uncore;
925 	const u32 base = engine->mmio_base;
926 	const i915_reg_t mode = RING_MI_MODE(base);
927 	int err;
928 
929 	if (INTEL_GEN(engine->i915) < 3)
930 		return -ENODEV;
931 
932 	GEM_TRACE("%s\n", engine->name);
933 
934 	intel_uncore_write_fw(uncore, mode, _MASKED_BIT_ENABLE(STOP_RING));
935 
936 	err = 0;
937 	if (__intel_wait_for_register_fw(uncore,
938 					 mode, MODE_IDLE, MODE_IDLE,
939 					 1000, 0,
940 					 NULL)) {
941 		GEM_TRACE("%s: timed out on STOP_RING -> IDLE\n", engine->name);
942 		err = -ETIMEDOUT;
943 	}
944 
945 	/* A final mmio read to let GPU writes be hopefully flushed to memory */
946 	intel_uncore_posting_read_fw(uncore, mode);
947 
948 	return err;
949 }
950 
951 void intel_engine_cancel_stop_cs(struct intel_engine_cs *engine)
952 {
953 	GEM_TRACE("%s\n", engine->name);
954 
955 	ENGINE_WRITE_FW(engine, RING_MI_MODE, _MASKED_BIT_DISABLE(STOP_RING));
956 }
957 
958 const char *i915_cache_level_str(struct drm_i915_private *i915, int type)
959 {
960 	switch (type) {
961 	case I915_CACHE_NONE: return " uncached";
962 	case I915_CACHE_LLC: return HAS_LLC(i915) ? " LLC" : " snooped";
963 	case I915_CACHE_L3_LLC: return " L3+LLC";
964 	case I915_CACHE_WT: return " WT";
965 	default: return "";
966 	}
967 }
968 
969 u32 intel_calculate_mcr_s_ss_select(struct drm_i915_private *dev_priv)
970 {
971 	const struct sseu_dev_info *sseu = &RUNTIME_INFO(dev_priv)->sseu;
972 	unsigned int slice = fls(sseu->slice_mask) - 1;
973 	unsigned int subslice;
974 	u32 mcr_s_ss_select;
975 
976 	GEM_BUG_ON(slice >= ARRAY_SIZE(sseu->subslice_mask));
977 	subslice = fls(sseu->subslice_mask[slice]);
978 	GEM_BUG_ON(!subslice);
979 	subslice--;
980 
981 	if (IS_GEN(dev_priv, 10))
982 		mcr_s_ss_select = GEN8_MCR_SLICE(slice) |
983 				  GEN8_MCR_SUBSLICE(subslice);
984 	else if (INTEL_GEN(dev_priv) >= 11)
985 		mcr_s_ss_select = GEN11_MCR_SLICE(slice) |
986 				  GEN11_MCR_SUBSLICE(subslice);
987 	else
988 		mcr_s_ss_select = 0;
989 
990 	return mcr_s_ss_select;
991 }
992 
993 static u32
994 read_subslice_reg(struct intel_engine_cs *engine, int slice, int subslice,
995 		  i915_reg_t reg)
996 {
997 	struct drm_i915_private *i915 = engine->i915;
998 	struct intel_uncore *uncore = engine->uncore;
999 	u32 mcr_slice_subslice_mask;
1000 	u32 mcr_slice_subslice_select;
1001 	u32 default_mcr_s_ss_select;
1002 	u32 mcr;
1003 	u32 ret;
1004 	enum forcewake_domains fw_domains;
1005 
1006 	if (INTEL_GEN(i915) >= 11) {
1007 		mcr_slice_subslice_mask = GEN11_MCR_SLICE_MASK |
1008 					  GEN11_MCR_SUBSLICE_MASK;
1009 		mcr_slice_subslice_select = GEN11_MCR_SLICE(slice) |
1010 					    GEN11_MCR_SUBSLICE(subslice);
1011 	} else {
1012 		mcr_slice_subslice_mask = GEN8_MCR_SLICE_MASK |
1013 					  GEN8_MCR_SUBSLICE_MASK;
1014 		mcr_slice_subslice_select = GEN8_MCR_SLICE(slice) |
1015 					    GEN8_MCR_SUBSLICE(subslice);
1016 	}
1017 
1018 	default_mcr_s_ss_select = intel_calculate_mcr_s_ss_select(i915);
1019 
1020 	fw_domains = intel_uncore_forcewake_for_reg(uncore, reg,
1021 						    FW_REG_READ);
1022 	fw_domains |= intel_uncore_forcewake_for_reg(uncore,
1023 						     GEN8_MCR_SELECTOR,
1024 						     FW_REG_READ | FW_REG_WRITE);
1025 
1026 	spin_lock_irq(&uncore->lock);
1027 	intel_uncore_forcewake_get__locked(uncore, fw_domains);
1028 
1029 	mcr = intel_uncore_read_fw(uncore, GEN8_MCR_SELECTOR);
1030 
1031 	WARN_ON_ONCE((mcr & mcr_slice_subslice_mask) !=
1032 		     default_mcr_s_ss_select);
1033 
1034 	mcr &= ~mcr_slice_subslice_mask;
1035 	mcr |= mcr_slice_subslice_select;
1036 	intel_uncore_write_fw(uncore, GEN8_MCR_SELECTOR, mcr);
1037 
1038 	ret = intel_uncore_read_fw(uncore, reg);
1039 
1040 	mcr &= ~mcr_slice_subslice_mask;
1041 	mcr |= default_mcr_s_ss_select;
1042 
1043 	intel_uncore_write_fw(uncore, GEN8_MCR_SELECTOR, mcr);
1044 
1045 	intel_uncore_forcewake_put__locked(uncore, fw_domains);
1046 	spin_unlock_irq(&uncore->lock);
1047 
1048 	return ret;
1049 }
1050 
1051 /* NB: please notice the memset */
1052 void intel_engine_get_instdone(struct intel_engine_cs *engine,
1053 			       struct intel_instdone *instdone)
1054 {
1055 	struct drm_i915_private *i915 = engine->i915;
1056 	struct intel_uncore *uncore = engine->uncore;
1057 	u32 mmio_base = engine->mmio_base;
1058 	int slice;
1059 	int subslice;
1060 
1061 	memset(instdone, 0, sizeof(*instdone));
1062 
1063 	switch (INTEL_GEN(i915)) {
1064 	default:
1065 		instdone->instdone =
1066 			intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
1067 
1068 		if (engine->id != RCS0)
1069 			break;
1070 
1071 		instdone->slice_common =
1072 			intel_uncore_read(uncore, GEN7_SC_INSTDONE);
1073 		for_each_instdone_slice_subslice(i915, slice, subslice) {
1074 			instdone->sampler[slice][subslice] =
1075 				read_subslice_reg(engine, slice, subslice,
1076 						  GEN7_SAMPLER_INSTDONE);
1077 			instdone->row[slice][subslice] =
1078 				read_subslice_reg(engine, slice, subslice,
1079 						  GEN7_ROW_INSTDONE);
1080 		}
1081 		break;
1082 	case 7:
1083 		instdone->instdone =
1084 			intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
1085 
1086 		if (engine->id != RCS0)
1087 			break;
1088 
1089 		instdone->slice_common =
1090 			intel_uncore_read(uncore, GEN7_SC_INSTDONE);
1091 		instdone->sampler[0][0] =
1092 			intel_uncore_read(uncore, GEN7_SAMPLER_INSTDONE);
1093 		instdone->row[0][0] =
1094 			intel_uncore_read(uncore, GEN7_ROW_INSTDONE);
1095 
1096 		break;
1097 	case 6:
1098 	case 5:
1099 	case 4:
1100 		instdone->instdone =
1101 			intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
1102 		if (engine->id == RCS0)
1103 			/* HACK: Using the wrong struct member */
1104 			instdone->slice_common =
1105 				intel_uncore_read(uncore, GEN4_INSTDONE1);
1106 		break;
1107 	case 3:
1108 	case 2:
1109 		instdone->instdone = intel_uncore_read(uncore, GEN2_INSTDONE);
1110 		break;
1111 	}
1112 }
1113 
1114 static bool ring_is_idle(struct intel_engine_cs *engine)
1115 {
1116 	struct drm_i915_private *dev_priv = engine->i915;
1117 	intel_wakeref_t wakeref;
1118 	bool idle = true;
1119 
1120 	if (I915_SELFTEST_ONLY(!engine->mmio_base))
1121 		return true;
1122 
1123 	/* If the whole device is asleep, the engine must be idle */
1124 	wakeref = intel_runtime_pm_get_if_in_use(&dev_priv->runtime_pm);
1125 	if (!wakeref)
1126 		return true;
1127 
1128 	/* First check that no commands are left in the ring */
1129 	if ((ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR) !=
1130 	    (ENGINE_READ(engine, RING_TAIL) & TAIL_ADDR))
1131 		idle = false;
1132 
1133 	/* No bit for gen2, so assume the CS parser is idle */
1134 	if (INTEL_GEN(dev_priv) > 2 &&
1135 	    !(ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE))
1136 		idle = false;
1137 
1138 	intel_runtime_pm_put(&dev_priv->runtime_pm, wakeref);
1139 
1140 	return idle;
1141 }
1142 
1143 /**
1144  * intel_engine_is_idle() - Report if the engine has finished process all work
1145  * @engine: the intel_engine_cs
1146  *
1147  * Return true if there are no requests pending, nothing left to be submitted
1148  * to hardware, and that the engine is idle.
1149  */
1150 bool intel_engine_is_idle(struct intel_engine_cs *engine)
1151 {
1152 	/* More white lies, if wedged, hw state is inconsistent */
1153 	if (i915_reset_failed(engine->i915))
1154 		return true;
1155 
1156 	if (!intel_wakeref_active(&engine->wakeref))
1157 		return true;
1158 
1159 	/* Waiting to drain ELSP? */
1160 	if (READ_ONCE(engine->execlists.active)) {
1161 		struct tasklet_struct *t = &engine->execlists.tasklet;
1162 
1163 		synchronize_hardirq(engine->i915->drm.irq);
1164 
1165 		local_bh_disable();
1166 		if (tasklet_trylock(t)) {
1167 			/* Must wait for any GPU reset in progress. */
1168 			if (__tasklet_is_enabled(t))
1169 				t->func(t->data);
1170 			tasklet_unlock(t);
1171 		}
1172 		local_bh_enable();
1173 
1174 		/* Otherwise flush the tasklet if it was on another cpu */
1175 		tasklet_unlock_wait(t);
1176 
1177 		if (READ_ONCE(engine->execlists.active))
1178 			return false;
1179 	}
1180 
1181 	/* ELSP is empty, but there are ready requests? E.g. after reset */
1182 	if (!RB_EMPTY_ROOT(&engine->execlists.queue.rb_root))
1183 		return false;
1184 
1185 	/* Ring stopped? */
1186 	return ring_is_idle(engine);
1187 }
1188 
1189 bool intel_engines_are_idle(struct drm_i915_private *i915)
1190 {
1191 	struct intel_engine_cs *engine;
1192 	enum intel_engine_id id;
1193 
1194 	/*
1195 	 * If the driver is wedged, HW state may be very inconsistent and
1196 	 * report that it is still busy, even though we have stopped using it.
1197 	 */
1198 	if (i915_reset_failed(i915))
1199 		return true;
1200 
1201 	/* Already parked (and passed an idleness test); must still be idle */
1202 	if (!READ_ONCE(i915->gt.awake))
1203 		return true;
1204 
1205 	for_each_engine(engine, i915, id) {
1206 		if (!intel_engine_is_idle(engine))
1207 			return false;
1208 	}
1209 
1210 	return true;
1211 }
1212 
1213 void intel_engines_reset_default_submission(struct drm_i915_private *i915)
1214 {
1215 	struct intel_engine_cs *engine;
1216 	enum intel_engine_id id;
1217 
1218 	for_each_engine(engine, i915, id)
1219 		engine->set_default_submission(engine);
1220 }
1221 
1222 bool intel_engine_can_store_dword(struct intel_engine_cs *engine)
1223 {
1224 	switch (INTEL_GEN(engine->i915)) {
1225 	case 2:
1226 		return false; /* uses physical not virtual addresses */
1227 	case 3:
1228 		/* maybe only uses physical not virtual addresses */
1229 		return !(IS_I915G(engine->i915) || IS_I915GM(engine->i915));
1230 	case 6:
1231 		return engine->class != VIDEO_DECODE_CLASS; /* b0rked */
1232 	default:
1233 		return true;
1234 	}
1235 }
1236 
1237 unsigned int intel_engines_has_context_isolation(struct drm_i915_private *i915)
1238 {
1239 	struct intel_engine_cs *engine;
1240 	enum intel_engine_id id;
1241 	unsigned int which;
1242 
1243 	which = 0;
1244 	for_each_engine(engine, i915, id)
1245 		if (engine->default_state)
1246 			which |= BIT(engine->uabi_class);
1247 
1248 	return which;
1249 }
1250 
1251 static int print_sched_attr(struct drm_i915_private *i915,
1252 			    const struct i915_sched_attr *attr,
1253 			    char *buf, int x, int len)
1254 {
1255 	if (attr->priority == I915_PRIORITY_INVALID)
1256 		return x;
1257 
1258 	x += snprintf(buf + x, len - x,
1259 		      " prio=%d", attr->priority);
1260 
1261 	return x;
1262 }
1263 
1264 static void print_request(struct drm_printer *m,
1265 			  struct i915_request *rq,
1266 			  const char *prefix)
1267 {
1268 	const char *name = rq->fence.ops->get_timeline_name(&rq->fence);
1269 	char buf[80] = "";
1270 	int x = 0;
1271 
1272 	x = print_sched_attr(rq->i915, &rq->sched.attr, buf, x, sizeof(buf));
1273 
1274 	drm_printf(m, "%s %llx:%llx%s%s %s @ %dms: %s\n",
1275 		   prefix,
1276 		   rq->fence.context, rq->fence.seqno,
1277 		   i915_request_completed(rq) ? "!" :
1278 		   i915_request_started(rq) ? "*" :
1279 		   "",
1280 		   test_bit(DMA_FENCE_FLAG_SIGNALED_BIT,
1281 			    &rq->fence.flags) ? "+" :
1282 		   test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT,
1283 			    &rq->fence.flags) ? "-" :
1284 		   "",
1285 		   buf,
1286 		   jiffies_to_msecs(jiffies - rq->emitted_jiffies),
1287 		   name);
1288 }
1289 
1290 static void hexdump(struct drm_printer *m, const void *buf, size_t len)
1291 {
1292 	const size_t rowsize = 8 * sizeof(u32);
1293 	const void *prev = NULL;
1294 	bool skip = false;
1295 	size_t pos;
1296 
1297 	for (pos = 0; pos < len; pos += rowsize) {
1298 		char line[128];
1299 
1300 		if (prev && !memcmp(prev, buf + pos, rowsize)) {
1301 			if (!skip) {
1302 				drm_printf(m, "*\n");
1303 				skip = true;
1304 			}
1305 			continue;
1306 		}
1307 
1308 		WARN_ON_ONCE(hex_dump_to_buffer(buf + pos, len - pos,
1309 						rowsize, sizeof(u32),
1310 						line, sizeof(line),
1311 						false) >= sizeof(line));
1312 		drm_printf(m, "[%04zx] %s\n", pos, line);
1313 
1314 		prev = buf + pos;
1315 		skip = false;
1316 	}
1317 }
1318 
1319 static void intel_engine_print_registers(struct intel_engine_cs *engine,
1320 					 struct drm_printer *m)
1321 {
1322 	struct drm_i915_private *dev_priv = engine->i915;
1323 	const struct intel_engine_execlists * const execlists =
1324 		&engine->execlists;
1325 	unsigned long flags;
1326 	u64 addr;
1327 
1328 	if (engine->id == RCS0 && IS_GEN_RANGE(dev_priv, 4, 7))
1329 		drm_printf(m, "\tCCID: 0x%08x\n", ENGINE_READ(engine, CCID));
1330 	drm_printf(m, "\tRING_START: 0x%08x\n",
1331 		   ENGINE_READ(engine, RING_START));
1332 	drm_printf(m, "\tRING_HEAD:  0x%08x\n",
1333 		   ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR);
1334 	drm_printf(m, "\tRING_TAIL:  0x%08x\n",
1335 		   ENGINE_READ(engine, RING_TAIL) & TAIL_ADDR);
1336 	drm_printf(m, "\tRING_CTL:   0x%08x%s\n",
1337 		   ENGINE_READ(engine, RING_CTL),
1338 		   ENGINE_READ(engine, RING_CTL) & (RING_WAIT | RING_WAIT_SEMAPHORE) ? " [waiting]" : "");
1339 	if (INTEL_GEN(engine->i915) > 2) {
1340 		drm_printf(m, "\tRING_MODE:  0x%08x%s\n",
1341 			   ENGINE_READ(engine, RING_MI_MODE),
1342 			   ENGINE_READ(engine, RING_MI_MODE) & (MODE_IDLE) ? " [idle]" : "");
1343 	}
1344 
1345 	if (INTEL_GEN(dev_priv) >= 6) {
1346 		drm_printf(m, "\tRING_IMR: %08x\n",
1347 			   ENGINE_READ(engine, RING_IMR));
1348 	}
1349 
1350 	addr = intel_engine_get_active_head(engine);
1351 	drm_printf(m, "\tACTHD:  0x%08x_%08x\n",
1352 		   upper_32_bits(addr), lower_32_bits(addr));
1353 	addr = intel_engine_get_last_batch_head(engine);
1354 	drm_printf(m, "\tBBADDR: 0x%08x_%08x\n",
1355 		   upper_32_bits(addr), lower_32_bits(addr));
1356 	if (INTEL_GEN(dev_priv) >= 8)
1357 		addr = ENGINE_READ64(engine, RING_DMA_FADD, RING_DMA_FADD_UDW);
1358 	else if (INTEL_GEN(dev_priv) >= 4)
1359 		addr = ENGINE_READ(engine, RING_DMA_FADD);
1360 	else
1361 		addr = ENGINE_READ(engine, DMA_FADD_I8XX);
1362 	drm_printf(m, "\tDMA_FADDR: 0x%08x_%08x\n",
1363 		   upper_32_bits(addr), lower_32_bits(addr));
1364 	if (INTEL_GEN(dev_priv) >= 4) {
1365 		drm_printf(m, "\tIPEIR: 0x%08x\n",
1366 			   ENGINE_READ(engine, RING_IPEIR));
1367 		drm_printf(m, "\tIPEHR: 0x%08x\n",
1368 			   ENGINE_READ(engine, RING_IPEHR));
1369 	} else {
1370 		drm_printf(m, "\tIPEIR: 0x%08x\n", ENGINE_READ(engine, IPEIR));
1371 		drm_printf(m, "\tIPEHR: 0x%08x\n", ENGINE_READ(engine, IPEHR));
1372 	}
1373 
1374 	if (HAS_EXECLISTS(dev_priv)) {
1375 		const u32 *hws =
1376 			&engine->status_page.addr[I915_HWS_CSB_BUF0_INDEX];
1377 		const u8 num_entries = execlists->csb_size;
1378 		unsigned int idx;
1379 		u8 read, write;
1380 
1381 		drm_printf(m, "\tExeclist status: 0x%08x %08x, entries %u\n",
1382 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_LO),
1383 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_HI),
1384 			   num_entries);
1385 
1386 		read = execlists->csb_head;
1387 		write = READ_ONCE(*execlists->csb_write);
1388 
1389 		drm_printf(m, "\tExeclist CSB read %d, write %d, tasklet queued? %s (%s)\n",
1390 			   read, write,
1391 			   yesno(test_bit(TASKLET_STATE_SCHED,
1392 					  &engine->execlists.tasklet.state)),
1393 			   enableddisabled(!atomic_read(&engine->execlists.tasklet.count)));
1394 		if (read >= num_entries)
1395 			read = 0;
1396 		if (write >= num_entries)
1397 			write = 0;
1398 		if (read > write)
1399 			write += num_entries;
1400 		while (read < write) {
1401 			idx = ++read % num_entries;
1402 			drm_printf(m, "\tExeclist CSB[%d]: 0x%08x, context: %d\n",
1403 				   idx, hws[idx * 2], hws[idx * 2 + 1]);
1404 		}
1405 
1406 		spin_lock_irqsave(&engine->active.lock, flags);
1407 		for (idx = 0; idx < execlists_num_ports(execlists); idx++) {
1408 			struct i915_request *rq;
1409 			unsigned int count;
1410 			char hdr[80];
1411 
1412 			rq = port_unpack(&execlists->port[idx], &count);
1413 			if (!rq) {
1414 				drm_printf(m, "\t\tELSP[%d] idle\n", idx);
1415 			} else if (!i915_request_signaled(rq)) {
1416 				snprintf(hdr, sizeof(hdr),
1417 					 "\t\tELSP[%d] count=%d, ring:{start:%08x, hwsp:%08x, seqno:%08x}, rq: ",
1418 					 idx, count,
1419 					 i915_ggtt_offset(rq->ring->vma),
1420 					 rq->timeline->hwsp_offset,
1421 					 hwsp_seqno(rq));
1422 				print_request(m, rq, hdr);
1423 			} else {
1424 				print_request(m, rq, "\t\tELSP[%d] rq: ");
1425 			}
1426 		}
1427 		drm_printf(m, "\t\tHW active? 0x%x\n", execlists->active);
1428 		spin_unlock_irqrestore(&engine->active.lock, flags);
1429 	} else if (INTEL_GEN(dev_priv) > 6) {
1430 		drm_printf(m, "\tPP_DIR_BASE: 0x%08x\n",
1431 			   ENGINE_READ(engine, RING_PP_DIR_BASE));
1432 		drm_printf(m, "\tPP_DIR_BASE_READ: 0x%08x\n",
1433 			   ENGINE_READ(engine, RING_PP_DIR_BASE_READ));
1434 		drm_printf(m, "\tPP_DIR_DCLV: 0x%08x\n",
1435 			   ENGINE_READ(engine, RING_PP_DIR_DCLV));
1436 	}
1437 }
1438 
1439 static void print_request_ring(struct drm_printer *m, struct i915_request *rq)
1440 {
1441 	void *ring;
1442 	int size;
1443 
1444 	drm_printf(m,
1445 		   "[head %04x, postfix %04x, tail %04x, batch 0x%08x_%08x]:\n",
1446 		   rq->head, rq->postfix, rq->tail,
1447 		   rq->batch ? upper_32_bits(rq->batch->node.start) : ~0u,
1448 		   rq->batch ? lower_32_bits(rq->batch->node.start) : ~0u);
1449 
1450 	size = rq->tail - rq->head;
1451 	if (rq->tail < rq->head)
1452 		size += rq->ring->size;
1453 
1454 	ring = kmalloc(size, GFP_ATOMIC);
1455 	if (ring) {
1456 		const void *vaddr = rq->ring->vaddr;
1457 		unsigned int head = rq->head;
1458 		unsigned int len = 0;
1459 
1460 		if (rq->tail < head) {
1461 			len = rq->ring->size - head;
1462 			memcpy(ring, vaddr + head, len);
1463 			head = 0;
1464 		}
1465 		memcpy(ring + len, vaddr + head, size - len);
1466 
1467 		hexdump(m, ring, size);
1468 		kfree(ring);
1469 	}
1470 }
1471 
1472 void intel_engine_dump(struct intel_engine_cs *engine,
1473 		       struct drm_printer *m,
1474 		       const char *header, ...)
1475 {
1476 	struct i915_gpu_error * const error = &engine->i915->gpu_error;
1477 	struct i915_request *rq;
1478 	intel_wakeref_t wakeref;
1479 	unsigned long flags;
1480 
1481 	if (header) {
1482 		va_list ap;
1483 
1484 		va_start(ap, header);
1485 		drm_vprintf(m, header, &ap);
1486 		va_end(ap);
1487 	}
1488 
1489 	if (i915_reset_failed(engine->i915))
1490 		drm_printf(m, "*** WEDGED ***\n");
1491 
1492 	drm_printf(m, "\tAwake? %d\n", atomic_read(&engine->wakeref.count));
1493 	drm_printf(m, "\tHangcheck: %d ms ago\n",
1494 		   jiffies_to_msecs(jiffies - engine->hangcheck.action_timestamp));
1495 	drm_printf(m, "\tReset count: %d (global %d)\n",
1496 		   i915_reset_engine_count(error, engine),
1497 		   i915_reset_count(error));
1498 
1499 	drm_printf(m, "\tRequests:\n");
1500 
1501 	spin_lock_irqsave(&engine->active.lock, flags);
1502 	rq = intel_engine_find_active_request(engine);
1503 	if (rq) {
1504 		print_request(m, rq, "\t\tactive ");
1505 
1506 		drm_printf(m, "\t\tring->start:  0x%08x\n",
1507 			   i915_ggtt_offset(rq->ring->vma));
1508 		drm_printf(m, "\t\tring->head:   0x%08x\n",
1509 			   rq->ring->head);
1510 		drm_printf(m, "\t\tring->tail:   0x%08x\n",
1511 			   rq->ring->tail);
1512 		drm_printf(m, "\t\tring->emit:   0x%08x\n",
1513 			   rq->ring->emit);
1514 		drm_printf(m, "\t\tring->space:  0x%08x\n",
1515 			   rq->ring->space);
1516 		drm_printf(m, "\t\tring->hwsp:   0x%08x\n",
1517 			   rq->timeline->hwsp_offset);
1518 
1519 		print_request_ring(m, rq);
1520 	}
1521 	spin_unlock_irqrestore(&engine->active.lock, flags);
1522 
1523 	wakeref = intel_runtime_pm_get_if_in_use(&engine->i915->runtime_pm);
1524 	if (wakeref) {
1525 		intel_engine_print_registers(engine, m);
1526 		intel_runtime_pm_put(&engine->i915->runtime_pm, wakeref);
1527 	} else {
1528 		drm_printf(m, "\tDevice is asleep; skipping register dump\n");
1529 	}
1530 
1531 	intel_execlists_show_requests(engine, m, print_request, 8);
1532 
1533 	drm_printf(m, "HWSP:\n");
1534 	hexdump(m, engine->status_page.addr, PAGE_SIZE);
1535 
1536 	drm_printf(m, "Idle? %s\n", yesno(intel_engine_is_idle(engine)));
1537 
1538 	intel_engine_print_breadcrumbs(engine, m);
1539 }
1540 
1541 static u8 user_class_map[] = {
1542 	[I915_ENGINE_CLASS_RENDER] = RENDER_CLASS,
1543 	[I915_ENGINE_CLASS_COPY] = COPY_ENGINE_CLASS,
1544 	[I915_ENGINE_CLASS_VIDEO] = VIDEO_DECODE_CLASS,
1545 	[I915_ENGINE_CLASS_VIDEO_ENHANCE] = VIDEO_ENHANCEMENT_CLASS,
1546 };
1547 
1548 struct intel_engine_cs *
1549 intel_engine_lookup_user(struct drm_i915_private *i915, u8 class, u8 instance)
1550 {
1551 	if (class >= ARRAY_SIZE(user_class_map))
1552 		return NULL;
1553 
1554 	class = user_class_map[class];
1555 
1556 	GEM_BUG_ON(class > MAX_ENGINE_CLASS);
1557 
1558 	if (instance > MAX_ENGINE_INSTANCE)
1559 		return NULL;
1560 
1561 	return i915->engine_class[class][instance];
1562 }
1563 
1564 /**
1565  * intel_enable_engine_stats() - Enable engine busy tracking on engine
1566  * @engine: engine to enable stats collection
1567  *
1568  * Start collecting the engine busyness data for @engine.
1569  *
1570  * Returns 0 on success or a negative error code.
1571  */
1572 int intel_enable_engine_stats(struct intel_engine_cs *engine)
1573 {
1574 	struct intel_engine_execlists *execlists = &engine->execlists;
1575 	unsigned long flags;
1576 	int err = 0;
1577 
1578 	if (!intel_engine_supports_stats(engine))
1579 		return -ENODEV;
1580 
1581 	spin_lock_irqsave(&engine->active.lock, flags);
1582 	write_seqlock(&engine->stats.lock);
1583 
1584 	if (unlikely(engine->stats.enabled == ~0)) {
1585 		err = -EBUSY;
1586 		goto unlock;
1587 	}
1588 
1589 	if (engine->stats.enabled++ == 0) {
1590 		const struct execlist_port *port = execlists->port;
1591 		unsigned int num_ports = execlists_num_ports(execlists);
1592 
1593 		engine->stats.enabled_at = ktime_get();
1594 
1595 		/* XXX submission method oblivious? */
1596 		while (num_ports-- && port_isset(port)) {
1597 			engine->stats.active++;
1598 			port++;
1599 		}
1600 
1601 		if (engine->stats.active)
1602 			engine->stats.start = engine->stats.enabled_at;
1603 	}
1604 
1605 unlock:
1606 	write_sequnlock(&engine->stats.lock);
1607 	spin_unlock_irqrestore(&engine->active.lock, flags);
1608 
1609 	return err;
1610 }
1611 
1612 static ktime_t __intel_engine_get_busy_time(struct intel_engine_cs *engine)
1613 {
1614 	ktime_t total = engine->stats.total;
1615 
1616 	/*
1617 	 * If the engine is executing something at the moment
1618 	 * add it to the total.
1619 	 */
1620 	if (engine->stats.active)
1621 		total = ktime_add(total,
1622 				  ktime_sub(ktime_get(), engine->stats.start));
1623 
1624 	return total;
1625 }
1626 
1627 /**
1628  * intel_engine_get_busy_time() - Return current accumulated engine busyness
1629  * @engine: engine to report on
1630  *
1631  * Returns accumulated time @engine was busy since engine stats were enabled.
1632  */
1633 ktime_t intel_engine_get_busy_time(struct intel_engine_cs *engine)
1634 {
1635 	unsigned int seq;
1636 	ktime_t total;
1637 
1638 	do {
1639 		seq = read_seqbegin(&engine->stats.lock);
1640 		total = __intel_engine_get_busy_time(engine);
1641 	} while (read_seqretry(&engine->stats.lock, seq));
1642 
1643 	return total;
1644 }
1645 
1646 /**
1647  * intel_disable_engine_stats() - Disable engine busy tracking on engine
1648  * @engine: engine to disable stats collection
1649  *
1650  * Stops collecting the engine busyness data for @engine.
1651  */
1652 void intel_disable_engine_stats(struct intel_engine_cs *engine)
1653 {
1654 	unsigned long flags;
1655 
1656 	if (!intel_engine_supports_stats(engine))
1657 		return;
1658 
1659 	write_seqlock_irqsave(&engine->stats.lock, flags);
1660 	WARN_ON_ONCE(engine->stats.enabled == 0);
1661 	if (--engine->stats.enabled == 0) {
1662 		engine->stats.total = __intel_engine_get_busy_time(engine);
1663 		engine->stats.active = 0;
1664 	}
1665 	write_sequnlock_irqrestore(&engine->stats.lock, flags);
1666 }
1667 
1668 static bool match_ring(struct i915_request *rq)
1669 {
1670 	u32 ring = ENGINE_READ(rq->engine, RING_START);
1671 
1672 	return ring == i915_ggtt_offset(rq->ring->vma);
1673 }
1674 
1675 struct i915_request *
1676 intel_engine_find_active_request(struct intel_engine_cs *engine)
1677 {
1678 	struct i915_request *request, *active = NULL;
1679 
1680 	/*
1681 	 * We are called by the error capture, reset and to dump engine
1682 	 * state at random points in time. In particular, note that neither is
1683 	 * crucially ordered with an interrupt. After a hang, the GPU is dead
1684 	 * and we assume that no more writes can happen (we waited long enough
1685 	 * for all writes that were in transaction to be flushed) - adding an
1686 	 * extra delay for a recent interrupt is pointless. Hence, we do
1687 	 * not need an engine->irq_seqno_barrier() before the seqno reads.
1688 	 * At all other times, we must assume the GPU is still running, but
1689 	 * we only care about the snapshot of this moment.
1690 	 */
1691 	lockdep_assert_held(&engine->active.lock);
1692 	list_for_each_entry(request, &engine->active.requests, sched.link) {
1693 		if (i915_request_completed(request))
1694 			continue;
1695 
1696 		if (!i915_request_started(request))
1697 			continue;
1698 
1699 		/* More than one preemptible request may match! */
1700 		if (!match_ring(request))
1701 			continue;
1702 
1703 		active = request;
1704 		break;
1705 	}
1706 
1707 	return active;
1708 }
1709 
1710 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1711 #include "selftest_engine_cs.c"
1712 #endif
1713