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_breadcrumbs.h"
32 #include "intel_context.h"
33 #include "intel_engine.h"
34 #include "intel_engine_pm.h"
35 #include "intel_engine_user.h"
36 #include "intel_gt.h"
37 #include "intel_gt_requests.h"
38 #include "intel_gt_pm.h"
39 #include "intel_lrc.h"
40 #include "intel_reset.h"
41 #include "intel_ring.h"
42 
43 /* Haswell does have the CXT_SIZE register however it does not appear to be
44  * valid. Now, docs explain in dwords what is in the context object. The full
45  * size is 70720 bytes, however, the power context and execlist context will
46  * never be saved (power context is stored elsewhere, and execlists don't work
47  * on HSW) - so the final size, including the extra state required for the
48  * Resource Streamer, is 66944 bytes, which rounds to 17 pages.
49  */
50 #define HSW_CXT_TOTAL_SIZE		(17 * PAGE_SIZE)
51 
52 #define DEFAULT_LR_CONTEXT_RENDER_SIZE	(22 * PAGE_SIZE)
53 #define GEN8_LR_CONTEXT_RENDER_SIZE	(20 * PAGE_SIZE)
54 #define GEN9_LR_CONTEXT_RENDER_SIZE	(22 * PAGE_SIZE)
55 #define GEN10_LR_CONTEXT_RENDER_SIZE	(18 * PAGE_SIZE)
56 #define GEN11_LR_CONTEXT_RENDER_SIZE	(14 * PAGE_SIZE)
57 
58 #define GEN8_LR_CONTEXT_OTHER_SIZE	( 2 * PAGE_SIZE)
59 
60 #define MAX_MMIO_BASES 3
61 struct engine_info {
62 	unsigned int hw_id;
63 	u8 class;
64 	u8 instance;
65 	/* mmio bases table *must* be sorted in reverse gen order */
66 	struct engine_mmio_base {
67 		u32 gen : 8;
68 		u32 base : 24;
69 	} mmio_bases[MAX_MMIO_BASES];
70 };
71 
72 static const struct engine_info intel_engines[] = {
73 	[RCS0] = {
74 		.hw_id = RCS0_HW,
75 		.class = RENDER_CLASS,
76 		.instance = 0,
77 		.mmio_bases = {
78 			{ .gen = 1, .base = RENDER_RING_BASE }
79 		},
80 	},
81 	[BCS0] = {
82 		.hw_id = BCS0_HW,
83 		.class = COPY_ENGINE_CLASS,
84 		.instance = 0,
85 		.mmio_bases = {
86 			{ .gen = 6, .base = BLT_RING_BASE }
87 		},
88 	},
89 	[VCS0] = {
90 		.hw_id = VCS0_HW,
91 		.class = VIDEO_DECODE_CLASS,
92 		.instance = 0,
93 		.mmio_bases = {
94 			{ .gen = 11, .base = GEN11_BSD_RING_BASE },
95 			{ .gen = 6, .base = GEN6_BSD_RING_BASE },
96 			{ .gen = 4, .base = BSD_RING_BASE }
97 		},
98 	},
99 	[VCS1] = {
100 		.hw_id = VCS1_HW,
101 		.class = VIDEO_DECODE_CLASS,
102 		.instance = 1,
103 		.mmio_bases = {
104 			{ .gen = 11, .base = GEN11_BSD2_RING_BASE },
105 			{ .gen = 8, .base = GEN8_BSD2_RING_BASE }
106 		},
107 	},
108 	[VCS2] = {
109 		.hw_id = VCS2_HW,
110 		.class = VIDEO_DECODE_CLASS,
111 		.instance = 2,
112 		.mmio_bases = {
113 			{ .gen = 11, .base = GEN11_BSD3_RING_BASE }
114 		},
115 	},
116 	[VCS3] = {
117 		.hw_id = VCS3_HW,
118 		.class = VIDEO_DECODE_CLASS,
119 		.instance = 3,
120 		.mmio_bases = {
121 			{ .gen = 11, .base = GEN11_BSD4_RING_BASE }
122 		},
123 	},
124 	[VECS0] = {
125 		.hw_id = VECS0_HW,
126 		.class = VIDEO_ENHANCEMENT_CLASS,
127 		.instance = 0,
128 		.mmio_bases = {
129 			{ .gen = 11, .base = GEN11_VEBOX_RING_BASE },
130 			{ .gen = 7, .base = VEBOX_RING_BASE }
131 		},
132 	},
133 	[VECS1] = {
134 		.hw_id = VECS1_HW,
135 		.class = VIDEO_ENHANCEMENT_CLASS,
136 		.instance = 1,
137 		.mmio_bases = {
138 			{ .gen = 11, .base = GEN11_VEBOX2_RING_BASE }
139 		},
140 	},
141 };
142 
143 /**
144  * intel_engine_context_size() - return the size of the context for an engine
145  * @gt: the gt
146  * @class: engine class
147  *
148  * Each engine class may require a different amount of space for a context
149  * image.
150  *
151  * Return: size (in bytes) of an engine class specific context image
152  *
153  * Note: this size includes the HWSP, which is part of the context image
154  * in LRC mode, but does not include the "shared data page" used with
155  * GuC submission. The caller should account for this if using the GuC.
156  */
157 u32 intel_engine_context_size(struct intel_gt *gt, u8 class)
158 {
159 	struct intel_uncore *uncore = gt->uncore;
160 	u32 cxt_size;
161 
162 	BUILD_BUG_ON(I915_GTT_PAGE_SIZE != PAGE_SIZE);
163 
164 	switch (class) {
165 	case RENDER_CLASS:
166 		switch (INTEL_GEN(gt->i915)) {
167 		default:
168 			MISSING_CASE(INTEL_GEN(gt->i915));
169 			return DEFAULT_LR_CONTEXT_RENDER_SIZE;
170 		case 12:
171 		case 11:
172 			return GEN11_LR_CONTEXT_RENDER_SIZE;
173 		case 10:
174 			return GEN10_LR_CONTEXT_RENDER_SIZE;
175 		case 9:
176 			return GEN9_LR_CONTEXT_RENDER_SIZE;
177 		case 8:
178 			return GEN8_LR_CONTEXT_RENDER_SIZE;
179 		case 7:
180 			if (IS_HASWELL(gt->i915))
181 				return HSW_CXT_TOTAL_SIZE;
182 
183 			cxt_size = intel_uncore_read(uncore, GEN7_CXT_SIZE);
184 			return round_up(GEN7_CXT_TOTAL_SIZE(cxt_size) * 64,
185 					PAGE_SIZE);
186 		case 6:
187 			cxt_size = intel_uncore_read(uncore, CXT_SIZE);
188 			return round_up(GEN6_CXT_TOTAL_SIZE(cxt_size) * 64,
189 					PAGE_SIZE);
190 		case 5:
191 		case 4:
192 			/*
193 			 * There is a discrepancy here between the size reported
194 			 * by the register and the size of the context layout
195 			 * in the docs. Both are described as authorative!
196 			 *
197 			 * The discrepancy is on the order of a few cachelines,
198 			 * but the total is under one page (4k), which is our
199 			 * minimum allocation anyway so it should all come
200 			 * out in the wash.
201 			 */
202 			cxt_size = intel_uncore_read(uncore, CXT_SIZE) + 1;
203 			drm_dbg(&gt->i915->drm,
204 				"gen%d CXT_SIZE = %d bytes [0x%08x]\n",
205 				INTEL_GEN(gt->i915), cxt_size * 64,
206 				cxt_size - 1);
207 			return round_up(cxt_size * 64, PAGE_SIZE);
208 		case 3:
209 		case 2:
210 		/* For the special day when i810 gets merged. */
211 		case 1:
212 			return 0;
213 		}
214 		break;
215 	default:
216 		MISSING_CASE(class);
217 		fallthrough;
218 	case VIDEO_DECODE_CLASS:
219 	case VIDEO_ENHANCEMENT_CLASS:
220 	case COPY_ENGINE_CLASS:
221 		if (INTEL_GEN(gt->i915) < 8)
222 			return 0;
223 		return GEN8_LR_CONTEXT_OTHER_SIZE;
224 	}
225 }
226 
227 static u32 __engine_mmio_base(struct drm_i915_private *i915,
228 			      const struct engine_mmio_base *bases)
229 {
230 	int i;
231 
232 	for (i = 0; i < MAX_MMIO_BASES; i++)
233 		if (INTEL_GEN(i915) >= bases[i].gen)
234 			break;
235 
236 	GEM_BUG_ON(i == MAX_MMIO_BASES);
237 	GEM_BUG_ON(!bases[i].base);
238 
239 	return bases[i].base;
240 }
241 
242 static void __sprint_engine_name(struct intel_engine_cs *engine)
243 {
244 	/*
245 	 * Before we know what the uABI name for this engine will be,
246 	 * we still would like to keep track of this engine in the debug logs.
247 	 * We throw in a ' here as a reminder that this isn't its final name.
248 	 */
249 	GEM_WARN_ON(snprintf(engine->name, sizeof(engine->name), "%s'%u",
250 			     intel_engine_class_repr(engine->class),
251 			     engine->instance) >= sizeof(engine->name));
252 }
253 
254 void intel_engine_set_hwsp_writemask(struct intel_engine_cs *engine, u32 mask)
255 {
256 	/*
257 	 * Though they added more rings on g4x/ilk, they did not add
258 	 * per-engine HWSTAM until gen6.
259 	 */
260 	if (INTEL_GEN(engine->i915) < 6 && engine->class != RENDER_CLASS)
261 		return;
262 
263 	if (INTEL_GEN(engine->i915) >= 3)
264 		ENGINE_WRITE(engine, RING_HWSTAM, mask);
265 	else
266 		ENGINE_WRITE16(engine, RING_HWSTAM, mask);
267 }
268 
269 static void intel_engine_sanitize_mmio(struct intel_engine_cs *engine)
270 {
271 	/* Mask off all writes into the unknown HWSP */
272 	intel_engine_set_hwsp_writemask(engine, ~0u);
273 }
274 
275 static int intel_engine_setup(struct intel_gt *gt, enum intel_engine_id id)
276 {
277 	const struct engine_info *info = &intel_engines[id];
278 	struct drm_i915_private *i915 = gt->i915;
279 	struct intel_engine_cs *engine;
280 
281 	BUILD_BUG_ON(MAX_ENGINE_CLASS >= BIT(GEN11_ENGINE_CLASS_WIDTH));
282 	BUILD_BUG_ON(MAX_ENGINE_INSTANCE >= BIT(GEN11_ENGINE_INSTANCE_WIDTH));
283 
284 	if (GEM_DEBUG_WARN_ON(id >= ARRAY_SIZE(gt->engine)))
285 		return -EINVAL;
286 
287 	if (GEM_DEBUG_WARN_ON(info->class > MAX_ENGINE_CLASS))
288 		return -EINVAL;
289 
290 	if (GEM_DEBUG_WARN_ON(info->instance > MAX_ENGINE_INSTANCE))
291 		return -EINVAL;
292 
293 	if (GEM_DEBUG_WARN_ON(gt->engine_class[info->class][info->instance]))
294 		return -EINVAL;
295 
296 	engine = kzalloc(sizeof(*engine), GFP_KERNEL);
297 	if (!engine)
298 		return -ENOMEM;
299 
300 	BUILD_BUG_ON(BITS_PER_TYPE(engine->mask) < I915_NUM_ENGINES);
301 
302 	engine->id = id;
303 	engine->legacy_idx = INVALID_ENGINE;
304 	engine->mask = BIT(id);
305 	engine->i915 = i915;
306 	engine->gt = gt;
307 	engine->uncore = gt->uncore;
308 	engine->hw_id = engine->guc_id = info->hw_id;
309 	engine->mmio_base = __engine_mmio_base(i915, info->mmio_bases);
310 
311 	engine->class = info->class;
312 	engine->instance = info->instance;
313 	__sprint_engine_name(engine);
314 
315 	engine->props.heartbeat_interval_ms =
316 		CONFIG_DRM_I915_HEARTBEAT_INTERVAL;
317 	engine->props.max_busywait_duration_ns =
318 		CONFIG_DRM_I915_MAX_REQUEST_BUSYWAIT;
319 	engine->props.preempt_timeout_ms =
320 		CONFIG_DRM_I915_PREEMPT_TIMEOUT;
321 	engine->props.stop_timeout_ms =
322 		CONFIG_DRM_I915_STOP_TIMEOUT;
323 	engine->props.timeslice_duration_ms =
324 		CONFIG_DRM_I915_TIMESLICE_DURATION;
325 
326 	/* Override to uninterruptible for OpenCL workloads. */
327 	if (INTEL_GEN(i915) == 12 && engine->class == RENDER_CLASS)
328 		engine->props.preempt_timeout_ms = 0;
329 
330 	engine->defaults = engine->props; /* never to change again */
331 
332 	engine->context_size = intel_engine_context_size(gt, engine->class);
333 	if (WARN_ON(engine->context_size > BIT(20)))
334 		engine->context_size = 0;
335 	if (engine->context_size)
336 		DRIVER_CAPS(i915)->has_logical_contexts = true;
337 
338 	/* Nothing to do here, execute in order of dependencies */
339 	engine->schedule = NULL;
340 
341 	ewma__engine_latency_init(&engine->latency);
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 	gt->engine_class[info->class][info->instance] = engine;
350 	gt->engine[id] = engine;
351 
352 	return 0;
353 }
354 
355 static void __setup_engine_capabilities(struct intel_engine_cs *engine)
356 {
357 	struct drm_i915_private *i915 = engine->i915;
358 
359 	if (engine->class == VIDEO_DECODE_CLASS) {
360 		/*
361 		 * HEVC support is present on first engine instance
362 		 * before Gen11 and on all instances afterwards.
363 		 */
364 		if (INTEL_GEN(i915) >= 11 ||
365 		    (INTEL_GEN(i915) >= 9 && engine->instance == 0))
366 			engine->uabi_capabilities |=
367 				I915_VIDEO_CLASS_CAPABILITY_HEVC;
368 
369 		/*
370 		 * SFC block is present only on even logical engine
371 		 * instances.
372 		 */
373 		if ((INTEL_GEN(i915) >= 11 &&
374 		     engine->gt->info.vdbox_sfc_access & engine->mask) ||
375 		    (INTEL_GEN(i915) >= 9 && engine->instance == 0))
376 			engine->uabi_capabilities |=
377 				I915_VIDEO_AND_ENHANCE_CLASS_CAPABILITY_SFC;
378 	} else if (engine->class == VIDEO_ENHANCEMENT_CLASS) {
379 		if (INTEL_GEN(i915) >= 9)
380 			engine->uabi_capabilities |=
381 				I915_VIDEO_AND_ENHANCE_CLASS_CAPABILITY_SFC;
382 	}
383 }
384 
385 static void intel_setup_engine_capabilities(struct intel_gt *gt)
386 {
387 	struct intel_engine_cs *engine;
388 	enum intel_engine_id id;
389 
390 	for_each_engine(engine, gt, id)
391 		__setup_engine_capabilities(engine);
392 }
393 
394 /**
395  * intel_engines_release() - free the resources allocated for Command Streamers
396  * @gt: pointer to struct intel_gt
397  */
398 void intel_engines_release(struct intel_gt *gt)
399 {
400 	struct intel_engine_cs *engine;
401 	enum intel_engine_id id;
402 
403 	/*
404 	 * Before we release the resources held by engine, we must be certain
405 	 * that the HW is no longer accessing them -- having the GPU scribble
406 	 * to or read from a page being used for something else causes no end
407 	 * of fun.
408 	 *
409 	 * The GPU should be reset by this point, but assume the worst just
410 	 * in case we aborted before completely initialising the engines.
411 	 */
412 	GEM_BUG_ON(intel_gt_pm_is_awake(gt));
413 	if (!INTEL_INFO(gt->i915)->gpu_reset_clobbers_display)
414 		__intel_gt_reset(gt, ALL_ENGINES);
415 
416 	/* Decouple the backend; but keep the layout for late GPU resets */
417 	for_each_engine(engine, gt, id) {
418 		if (!engine->release)
419 			continue;
420 
421 		intel_wakeref_wait_for_idle(&engine->wakeref);
422 		GEM_BUG_ON(intel_engine_pm_is_awake(engine));
423 
424 		engine->release(engine);
425 		engine->release = NULL;
426 
427 		memset(&engine->reset, 0, sizeof(engine->reset));
428 	}
429 }
430 
431 void intel_engine_free_request_pool(struct intel_engine_cs *engine)
432 {
433 	if (!engine->request_pool)
434 		return;
435 
436 	kmem_cache_free(i915_request_slab_cache(), engine->request_pool);
437 }
438 
439 void intel_engines_free(struct intel_gt *gt)
440 {
441 	struct intel_engine_cs *engine;
442 	enum intel_engine_id id;
443 
444 	/* Free the requests! dma-resv keeps fences around for an eternity */
445 	rcu_barrier();
446 
447 	for_each_engine(engine, gt, id) {
448 		intel_engine_free_request_pool(engine);
449 		kfree(engine);
450 		gt->engine[id] = NULL;
451 	}
452 }
453 
454 /*
455  * Determine which engines are fused off in our particular hardware.
456  * Note that we have a catch-22 situation where we need to be able to access
457  * the blitter forcewake domain to read the engine fuses, but at the same time
458  * we need to know which engines are available on the system to know which
459  * forcewake domains are present. We solve this by intializing the forcewake
460  * domains based on the full engine mask in the platform capabilities before
461  * calling this function and pruning the domains for fused-off engines
462  * afterwards.
463  */
464 static intel_engine_mask_t init_engine_mask(struct intel_gt *gt)
465 {
466 	struct drm_i915_private *i915 = gt->i915;
467 	struct intel_gt_info *info = &gt->info;
468 	struct intel_uncore *uncore = gt->uncore;
469 	unsigned int logical_vdbox = 0;
470 	unsigned int i;
471 	u32 media_fuse;
472 	u16 vdbox_mask;
473 	u16 vebox_mask;
474 
475 	info->engine_mask = INTEL_INFO(i915)->platform_engine_mask;
476 
477 	if (INTEL_GEN(i915) < 11)
478 		return info->engine_mask;
479 
480 	media_fuse = ~intel_uncore_read(uncore, GEN11_GT_VEBOX_VDBOX_DISABLE);
481 
482 	vdbox_mask = media_fuse & GEN11_GT_VDBOX_DISABLE_MASK;
483 	vebox_mask = (media_fuse & GEN11_GT_VEBOX_DISABLE_MASK) >>
484 		      GEN11_GT_VEBOX_DISABLE_SHIFT;
485 
486 	for (i = 0; i < I915_MAX_VCS; i++) {
487 		if (!HAS_ENGINE(gt, _VCS(i))) {
488 			vdbox_mask &= ~BIT(i);
489 			continue;
490 		}
491 
492 		if (!(BIT(i) & vdbox_mask)) {
493 			info->engine_mask &= ~BIT(_VCS(i));
494 			drm_dbg(&i915->drm, "vcs%u fused off\n", i);
495 			continue;
496 		}
497 
498 		/*
499 		 * In Gen11, only even numbered logical VDBOXes are
500 		 * hooked up to an SFC (Scaler & Format Converter) unit.
501 		 * In TGL each VDBOX has access to an SFC.
502 		 */
503 		if (INTEL_GEN(i915) >= 12 || logical_vdbox++ % 2 == 0)
504 			gt->info.vdbox_sfc_access |= BIT(i);
505 	}
506 	drm_dbg(&i915->drm, "vdbox enable: %04x, instances: %04lx\n",
507 		vdbox_mask, VDBOX_MASK(gt));
508 	GEM_BUG_ON(vdbox_mask != VDBOX_MASK(gt));
509 
510 	for (i = 0; i < I915_MAX_VECS; i++) {
511 		if (!HAS_ENGINE(gt, _VECS(i))) {
512 			vebox_mask &= ~BIT(i);
513 			continue;
514 		}
515 
516 		if (!(BIT(i) & vebox_mask)) {
517 			info->engine_mask &= ~BIT(_VECS(i));
518 			drm_dbg(&i915->drm, "vecs%u fused off\n", i);
519 		}
520 	}
521 	drm_dbg(&i915->drm, "vebox enable: %04x, instances: %04lx\n",
522 		vebox_mask, VEBOX_MASK(gt));
523 	GEM_BUG_ON(vebox_mask != VEBOX_MASK(gt));
524 
525 	return info->engine_mask;
526 }
527 
528 /**
529  * intel_engines_init_mmio() - allocate and prepare the Engine Command Streamers
530  * @gt: pointer to struct intel_gt
531  *
532  * Return: non-zero if the initialization failed.
533  */
534 int intel_engines_init_mmio(struct intel_gt *gt)
535 {
536 	struct drm_i915_private *i915 = gt->i915;
537 	const unsigned int engine_mask = init_engine_mask(gt);
538 	unsigned int mask = 0;
539 	unsigned int i;
540 	int err;
541 
542 	drm_WARN_ON(&i915->drm, engine_mask == 0);
543 	drm_WARN_ON(&i915->drm, engine_mask &
544 		    GENMASK(BITS_PER_TYPE(mask) - 1, I915_NUM_ENGINES));
545 
546 	if (i915_inject_probe_failure(i915))
547 		return -ENODEV;
548 
549 	for (i = 0; i < ARRAY_SIZE(intel_engines); i++) {
550 		if (!HAS_ENGINE(gt, i))
551 			continue;
552 
553 		err = intel_engine_setup(gt, i);
554 		if (err)
555 			goto cleanup;
556 
557 		mask |= BIT(i);
558 	}
559 
560 	/*
561 	 * Catch failures to update intel_engines table when the new engines
562 	 * are added to the driver by a warning and disabling the forgotten
563 	 * engines.
564 	 */
565 	if (drm_WARN_ON(&i915->drm, mask != engine_mask))
566 		gt->info.engine_mask = mask;
567 
568 	gt->info.num_engines = hweight32(mask);
569 
570 	intel_gt_check_and_clear_faults(gt);
571 
572 	intel_setup_engine_capabilities(gt);
573 
574 	intel_uncore_prune_engine_fw_domains(gt->uncore, gt);
575 
576 	return 0;
577 
578 cleanup:
579 	intel_engines_free(gt);
580 	return err;
581 }
582 
583 void intel_engine_init_execlists(struct intel_engine_cs *engine)
584 {
585 	struct intel_engine_execlists * const execlists = &engine->execlists;
586 
587 	execlists->port_mask = 1;
588 	GEM_BUG_ON(!is_power_of_2(execlists_num_ports(execlists)));
589 	GEM_BUG_ON(execlists_num_ports(execlists) > EXECLIST_MAX_PORTS);
590 
591 	memset(execlists->pending, 0, sizeof(execlists->pending));
592 	execlists->active =
593 		memset(execlists->inflight, 0, sizeof(execlists->inflight));
594 
595 	execlists->queue_priority_hint = INT_MIN;
596 	execlists->queue = RB_ROOT_CACHED;
597 }
598 
599 static void cleanup_status_page(struct intel_engine_cs *engine)
600 {
601 	struct i915_vma *vma;
602 
603 	/* Prevent writes into HWSP after returning the page to the system */
604 	intel_engine_set_hwsp_writemask(engine, ~0u);
605 
606 	vma = fetch_and_zero(&engine->status_page.vma);
607 	if (!vma)
608 		return;
609 
610 	if (!HWS_NEEDS_PHYSICAL(engine->i915))
611 		i915_vma_unpin(vma);
612 
613 	i915_gem_object_unpin_map(vma->obj);
614 	i915_gem_object_put(vma->obj);
615 }
616 
617 static int pin_ggtt_status_page(struct intel_engine_cs *engine,
618 				struct i915_vma *vma)
619 {
620 	unsigned int flags;
621 
622 	if (!HAS_LLC(engine->i915) && i915_ggtt_has_aperture(engine->gt->ggtt))
623 		/*
624 		 * On g33, we cannot place HWS above 256MiB, so
625 		 * restrict its pinning to the low mappable arena.
626 		 * Though this restriction is not documented for
627 		 * gen4, gen5, or byt, they also behave similarly
628 		 * and hang if the HWS is placed at the top of the
629 		 * GTT. To generalise, it appears that all !llc
630 		 * platforms have issues with us placing the HWS
631 		 * above the mappable region (even though we never
632 		 * actually map it).
633 		 */
634 		flags = PIN_MAPPABLE;
635 	else
636 		flags = PIN_HIGH;
637 
638 	return i915_ggtt_pin(vma, NULL, 0, flags);
639 }
640 
641 static int init_status_page(struct intel_engine_cs *engine)
642 {
643 	struct drm_i915_gem_object *obj;
644 	struct i915_vma *vma;
645 	void *vaddr;
646 	int ret;
647 
648 	/*
649 	 * Though the HWS register does support 36bit addresses, historically
650 	 * we have had hangs and corruption reported due to wild writes if
651 	 * the HWS is placed above 4G. We only allow objects to be allocated
652 	 * in GFP_DMA32 for i965, and no earlier physical address users had
653 	 * access to more than 4G.
654 	 */
655 	obj = i915_gem_object_create_internal(engine->i915, PAGE_SIZE);
656 	if (IS_ERR(obj)) {
657 		drm_err(&engine->i915->drm,
658 			"Failed to allocate status page\n");
659 		return PTR_ERR(obj);
660 	}
661 
662 	i915_gem_object_set_cache_coherency(obj, I915_CACHE_LLC);
663 
664 	vma = i915_vma_instance(obj, &engine->gt->ggtt->vm, NULL);
665 	if (IS_ERR(vma)) {
666 		ret = PTR_ERR(vma);
667 		goto err;
668 	}
669 
670 	vaddr = i915_gem_object_pin_map(obj, I915_MAP_WB);
671 	if (IS_ERR(vaddr)) {
672 		ret = PTR_ERR(vaddr);
673 		goto err;
674 	}
675 
676 	engine->status_page.addr = memset(vaddr, 0, PAGE_SIZE);
677 	engine->status_page.vma = vma;
678 
679 	if (!HWS_NEEDS_PHYSICAL(engine->i915)) {
680 		ret = pin_ggtt_status_page(engine, vma);
681 		if (ret)
682 			goto err_unpin;
683 	}
684 
685 	return 0;
686 
687 err_unpin:
688 	i915_gem_object_unpin_map(obj);
689 err:
690 	i915_gem_object_put(obj);
691 	return ret;
692 }
693 
694 static int engine_setup_common(struct intel_engine_cs *engine)
695 {
696 	int err;
697 
698 	init_llist_head(&engine->barrier_tasks);
699 
700 	err = init_status_page(engine);
701 	if (err)
702 		return err;
703 
704 	engine->breadcrumbs = intel_breadcrumbs_create(engine);
705 	if (!engine->breadcrumbs) {
706 		err = -ENOMEM;
707 		goto err_status;
708 	}
709 
710 	intel_engine_init_active(engine, ENGINE_PHYSICAL);
711 	intel_engine_init_execlists(engine);
712 	intel_engine_init_cmd_parser(engine);
713 	intel_engine_init__pm(engine);
714 	intel_engine_init_retire(engine);
715 
716 	/* Use the whole device by default */
717 	engine->sseu =
718 		intel_sseu_from_device_info(&engine->gt->info.sseu);
719 
720 	intel_engine_init_workarounds(engine);
721 	intel_engine_init_whitelist(engine);
722 	intel_engine_init_ctx_wa(engine);
723 
724 	return 0;
725 
726 err_status:
727 	cleanup_status_page(engine);
728 	return err;
729 }
730 
731 struct measure_breadcrumb {
732 	struct i915_request rq;
733 	struct intel_ring ring;
734 	u32 cs[2048];
735 };
736 
737 static int measure_breadcrumb_dw(struct intel_context *ce)
738 {
739 	struct intel_engine_cs *engine = ce->engine;
740 	struct measure_breadcrumb *frame;
741 	int dw;
742 
743 	GEM_BUG_ON(!engine->gt->scratch);
744 
745 	frame = kzalloc(sizeof(*frame), GFP_KERNEL);
746 	if (!frame)
747 		return -ENOMEM;
748 
749 	frame->rq.engine = engine;
750 	frame->rq.context = ce;
751 	rcu_assign_pointer(frame->rq.timeline, ce->timeline);
752 
753 	frame->ring.vaddr = frame->cs;
754 	frame->ring.size = sizeof(frame->cs);
755 	frame->ring.wrap =
756 		BITS_PER_TYPE(frame->ring.size) - ilog2(frame->ring.size);
757 	frame->ring.effective_size = frame->ring.size;
758 	intel_ring_update_space(&frame->ring);
759 	frame->rq.ring = &frame->ring;
760 
761 	mutex_lock(&ce->timeline->mutex);
762 	spin_lock_irq(&engine->active.lock);
763 
764 	dw = engine->emit_fini_breadcrumb(&frame->rq, frame->cs) - frame->cs;
765 
766 	spin_unlock_irq(&engine->active.lock);
767 	mutex_unlock(&ce->timeline->mutex);
768 
769 	GEM_BUG_ON(dw & 1); /* RING_TAIL must be qword aligned */
770 
771 	kfree(frame);
772 	return dw;
773 }
774 
775 void
776 intel_engine_init_active(struct intel_engine_cs *engine, unsigned int subclass)
777 {
778 	INIT_LIST_HEAD(&engine->active.requests);
779 	INIT_LIST_HEAD(&engine->active.hold);
780 
781 	spin_lock_init(&engine->active.lock);
782 	lockdep_set_subclass(&engine->active.lock, subclass);
783 
784 	/*
785 	 * Due to an interesting quirk in lockdep's internal debug tracking,
786 	 * after setting a subclass we must ensure the lock is used. Otherwise,
787 	 * nr_unused_locks is incremented once too often.
788 	 */
789 #ifdef CONFIG_DEBUG_LOCK_ALLOC
790 	local_irq_disable();
791 	lock_map_acquire(&engine->active.lock.dep_map);
792 	lock_map_release(&engine->active.lock.dep_map);
793 	local_irq_enable();
794 #endif
795 }
796 
797 static struct intel_context *
798 create_pinned_context(struct intel_engine_cs *engine,
799 		      unsigned int hwsp,
800 		      struct lock_class_key *key,
801 		      const char *name)
802 {
803 	struct intel_context *ce;
804 	int err;
805 
806 	ce = intel_context_create(engine);
807 	if (IS_ERR(ce))
808 		return ce;
809 
810 	__set_bit(CONTEXT_BARRIER_BIT, &ce->flags);
811 	ce->timeline = page_pack_bits(NULL, hwsp);
812 
813 	err = intel_context_pin(ce); /* perma-pin so it is always available */
814 	if (err) {
815 		intel_context_put(ce);
816 		return ERR_PTR(err);
817 	}
818 
819 	/*
820 	 * Give our perma-pinned kernel timelines a separate lockdep class,
821 	 * so that we can use them from within the normal user timelines
822 	 * should we need to inject GPU operations during their request
823 	 * construction.
824 	 */
825 	lockdep_set_class_and_name(&ce->timeline->mutex, key, name);
826 
827 	return ce;
828 }
829 
830 static struct intel_context *
831 create_kernel_context(struct intel_engine_cs *engine)
832 {
833 	static struct lock_class_key kernel;
834 
835 	return create_pinned_context(engine, I915_GEM_HWS_SEQNO_ADDR,
836 				     &kernel, "kernel_context");
837 }
838 
839 /**
840  * intel_engines_init_common - initialize cengine state which might require hw access
841  * @engine: Engine to initialize.
842  *
843  * Initializes @engine@ structure members shared between legacy and execlists
844  * submission modes which do require hardware access.
845  *
846  * Typcally done at later stages of submission mode specific engine setup.
847  *
848  * Returns zero on success or an error code on failure.
849  */
850 static int engine_init_common(struct intel_engine_cs *engine)
851 {
852 	struct intel_context *ce;
853 	int ret;
854 
855 	engine->set_default_submission(engine);
856 
857 	/*
858 	 * We may need to do things with the shrinker which
859 	 * require us to immediately switch back to the default
860 	 * context. This can cause a problem as pinning the
861 	 * default context also requires GTT space which may not
862 	 * be available. To avoid this we always pin the default
863 	 * context.
864 	 */
865 	ce = create_kernel_context(engine);
866 	if (IS_ERR(ce))
867 		return PTR_ERR(ce);
868 
869 	ret = measure_breadcrumb_dw(ce);
870 	if (ret < 0)
871 		goto err_context;
872 
873 	engine->emit_fini_breadcrumb_dw = ret;
874 	engine->kernel_context = ce;
875 
876 	return 0;
877 
878 err_context:
879 	intel_context_put(ce);
880 	return ret;
881 }
882 
883 int intel_engines_init(struct intel_gt *gt)
884 {
885 	int (*setup)(struct intel_engine_cs *engine);
886 	struct intel_engine_cs *engine;
887 	enum intel_engine_id id;
888 	int err;
889 
890 	if (HAS_EXECLISTS(gt->i915))
891 		setup = intel_execlists_submission_setup;
892 	else
893 		setup = intel_ring_submission_setup;
894 
895 	for_each_engine(engine, gt, id) {
896 		err = engine_setup_common(engine);
897 		if (err)
898 			return err;
899 
900 		err = setup(engine);
901 		if (err)
902 			return err;
903 
904 		err = engine_init_common(engine);
905 		if (err)
906 			return err;
907 
908 		intel_engine_add_user(engine);
909 	}
910 
911 	return 0;
912 }
913 
914 /**
915  * intel_engines_cleanup_common - cleans up the engine state created by
916  *                                the common initiailizers.
917  * @engine: Engine to cleanup.
918  *
919  * This cleans up everything created by the common helpers.
920  */
921 void intel_engine_cleanup_common(struct intel_engine_cs *engine)
922 {
923 	GEM_BUG_ON(!list_empty(&engine->active.requests));
924 	tasklet_kill(&engine->execlists.tasklet); /* flush the callback */
925 
926 	cleanup_status_page(engine);
927 	intel_breadcrumbs_free(engine->breadcrumbs);
928 
929 	intel_engine_fini_retire(engine);
930 	intel_engine_cleanup_cmd_parser(engine);
931 
932 	if (engine->default_state)
933 		fput(engine->default_state);
934 
935 	if (engine->kernel_context) {
936 		intel_context_unpin(engine->kernel_context);
937 		intel_context_put(engine->kernel_context);
938 	}
939 	GEM_BUG_ON(!llist_empty(&engine->barrier_tasks));
940 
941 	intel_wa_list_free(&engine->ctx_wa_list);
942 	intel_wa_list_free(&engine->wa_list);
943 	intel_wa_list_free(&engine->whitelist);
944 }
945 
946 /**
947  * intel_engine_resume - re-initializes the HW state of the engine
948  * @engine: Engine to resume.
949  *
950  * Returns zero on success or an error code on failure.
951  */
952 int intel_engine_resume(struct intel_engine_cs *engine)
953 {
954 	intel_engine_apply_workarounds(engine);
955 	intel_engine_apply_whitelist(engine);
956 
957 	return engine->resume(engine);
958 }
959 
960 u64 intel_engine_get_active_head(const struct intel_engine_cs *engine)
961 {
962 	struct drm_i915_private *i915 = engine->i915;
963 
964 	u64 acthd;
965 
966 	if (INTEL_GEN(i915) >= 8)
967 		acthd = ENGINE_READ64(engine, RING_ACTHD, RING_ACTHD_UDW);
968 	else if (INTEL_GEN(i915) >= 4)
969 		acthd = ENGINE_READ(engine, RING_ACTHD);
970 	else
971 		acthd = ENGINE_READ(engine, ACTHD);
972 
973 	return acthd;
974 }
975 
976 u64 intel_engine_get_last_batch_head(const struct intel_engine_cs *engine)
977 {
978 	u64 bbaddr;
979 
980 	if (INTEL_GEN(engine->i915) >= 8)
981 		bbaddr = ENGINE_READ64(engine, RING_BBADDR, RING_BBADDR_UDW);
982 	else
983 		bbaddr = ENGINE_READ(engine, RING_BBADDR);
984 
985 	return bbaddr;
986 }
987 
988 static unsigned long stop_timeout(const struct intel_engine_cs *engine)
989 {
990 	if (in_atomic() || irqs_disabled()) /* inside atomic preempt-reset? */
991 		return 0;
992 
993 	/*
994 	 * If we are doing a normal GPU reset, we can take our time and allow
995 	 * the engine to quiesce. We've stopped submission to the engine, and
996 	 * if we wait long enough an innocent context should complete and
997 	 * leave the engine idle. So they should not be caught unaware by
998 	 * the forthcoming GPU reset (which usually follows the stop_cs)!
999 	 */
1000 	return READ_ONCE(engine->props.stop_timeout_ms);
1001 }
1002 
1003 int intel_engine_stop_cs(struct intel_engine_cs *engine)
1004 {
1005 	struct intel_uncore *uncore = engine->uncore;
1006 	const u32 base = engine->mmio_base;
1007 	const i915_reg_t mode = RING_MI_MODE(base);
1008 	int err;
1009 
1010 	if (INTEL_GEN(engine->i915) < 3)
1011 		return -ENODEV;
1012 
1013 	ENGINE_TRACE(engine, "\n");
1014 
1015 	intel_uncore_write_fw(uncore, mode, _MASKED_BIT_ENABLE(STOP_RING));
1016 
1017 	err = 0;
1018 	if (__intel_wait_for_register_fw(uncore,
1019 					 mode, MODE_IDLE, MODE_IDLE,
1020 					 1000, stop_timeout(engine),
1021 					 NULL)) {
1022 		ENGINE_TRACE(engine, "timed out on STOP_RING -> IDLE\n");
1023 		err = -ETIMEDOUT;
1024 	}
1025 
1026 	/* A final mmio read to let GPU writes be hopefully flushed to memory */
1027 	intel_uncore_posting_read_fw(uncore, mode);
1028 
1029 	return err;
1030 }
1031 
1032 void intel_engine_cancel_stop_cs(struct intel_engine_cs *engine)
1033 {
1034 	ENGINE_TRACE(engine, "\n");
1035 
1036 	ENGINE_WRITE_FW(engine, RING_MI_MODE, _MASKED_BIT_DISABLE(STOP_RING));
1037 }
1038 
1039 const char *i915_cache_level_str(struct drm_i915_private *i915, int type)
1040 {
1041 	switch (type) {
1042 	case I915_CACHE_NONE: return " uncached";
1043 	case I915_CACHE_LLC: return HAS_LLC(i915) ? " LLC" : " snooped";
1044 	case I915_CACHE_L3_LLC: return " L3+LLC";
1045 	case I915_CACHE_WT: return " WT";
1046 	default: return "";
1047 	}
1048 }
1049 
1050 static u32
1051 read_subslice_reg(const struct intel_engine_cs *engine,
1052 		  int slice, int subslice, i915_reg_t reg)
1053 {
1054 	struct drm_i915_private *i915 = engine->i915;
1055 	struct intel_uncore *uncore = engine->uncore;
1056 	u32 mcr_mask, mcr_ss, mcr, old_mcr, val;
1057 	enum forcewake_domains fw_domains;
1058 
1059 	if (INTEL_GEN(i915) >= 11) {
1060 		mcr_mask = GEN11_MCR_SLICE_MASK | GEN11_MCR_SUBSLICE_MASK;
1061 		mcr_ss = GEN11_MCR_SLICE(slice) | GEN11_MCR_SUBSLICE(subslice);
1062 	} else {
1063 		mcr_mask = GEN8_MCR_SLICE_MASK | GEN8_MCR_SUBSLICE_MASK;
1064 		mcr_ss = GEN8_MCR_SLICE(slice) | GEN8_MCR_SUBSLICE(subslice);
1065 	}
1066 
1067 	fw_domains = intel_uncore_forcewake_for_reg(uncore, reg,
1068 						    FW_REG_READ);
1069 	fw_domains |= intel_uncore_forcewake_for_reg(uncore,
1070 						     GEN8_MCR_SELECTOR,
1071 						     FW_REG_READ | FW_REG_WRITE);
1072 
1073 	spin_lock_irq(&uncore->lock);
1074 	intel_uncore_forcewake_get__locked(uncore, fw_domains);
1075 
1076 	old_mcr = mcr = intel_uncore_read_fw(uncore, GEN8_MCR_SELECTOR);
1077 
1078 	mcr &= ~mcr_mask;
1079 	mcr |= mcr_ss;
1080 	intel_uncore_write_fw(uncore, GEN8_MCR_SELECTOR, mcr);
1081 
1082 	val = intel_uncore_read_fw(uncore, reg);
1083 
1084 	mcr &= ~mcr_mask;
1085 	mcr |= old_mcr & mcr_mask;
1086 
1087 	intel_uncore_write_fw(uncore, GEN8_MCR_SELECTOR, mcr);
1088 
1089 	intel_uncore_forcewake_put__locked(uncore, fw_domains);
1090 	spin_unlock_irq(&uncore->lock);
1091 
1092 	return val;
1093 }
1094 
1095 /* NB: please notice the memset */
1096 void intel_engine_get_instdone(const struct intel_engine_cs *engine,
1097 			       struct intel_instdone *instdone)
1098 {
1099 	struct drm_i915_private *i915 = engine->i915;
1100 	const struct sseu_dev_info *sseu = &engine->gt->info.sseu;
1101 	struct intel_uncore *uncore = engine->uncore;
1102 	u32 mmio_base = engine->mmio_base;
1103 	int slice;
1104 	int subslice;
1105 
1106 	memset(instdone, 0, sizeof(*instdone));
1107 
1108 	switch (INTEL_GEN(i915)) {
1109 	default:
1110 		instdone->instdone =
1111 			intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
1112 
1113 		if (engine->id != RCS0)
1114 			break;
1115 
1116 		instdone->slice_common =
1117 			intel_uncore_read(uncore, GEN7_SC_INSTDONE);
1118 		if (INTEL_GEN(i915) >= 12) {
1119 			instdone->slice_common_extra[0] =
1120 				intel_uncore_read(uncore, GEN12_SC_INSTDONE_EXTRA);
1121 			instdone->slice_common_extra[1] =
1122 				intel_uncore_read(uncore, GEN12_SC_INSTDONE_EXTRA2);
1123 		}
1124 		for_each_instdone_slice_subslice(i915, sseu, slice, subslice) {
1125 			instdone->sampler[slice][subslice] =
1126 				read_subslice_reg(engine, slice, subslice,
1127 						  GEN7_SAMPLER_INSTDONE);
1128 			instdone->row[slice][subslice] =
1129 				read_subslice_reg(engine, slice, subslice,
1130 						  GEN7_ROW_INSTDONE);
1131 		}
1132 		break;
1133 	case 7:
1134 		instdone->instdone =
1135 			intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
1136 
1137 		if (engine->id != RCS0)
1138 			break;
1139 
1140 		instdone->slice_common =
1141 			intel_uncore_read(uncore, GEN7_SC_INSTDONE);
1142 		instdone->sampler[0][0] =
1143 			intel_uncore_read(uncore, GEN7_SAMPLER_INSTDONE);
1144 		instdone->row[0][0] =
1145 			intel_uncore_read(uncore, GEN7_ROW_INSTDONE);
1146 
1147 		break;
1148 	case 6:
1149 	case 5:
1150 	case 4:
1151 		instdone->instdone =
1152 			intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
1153 		if (engine->id == RCS0)
1154 			/* HACK: Using the wrong struct member */
1155 			instdone->slice_common =
1156 				intel_uncore_read(uncore, GEN4_INSTDONE1);
1157 		break;
1158 	case 3:
1159 	case 2:
1160 		instdone->instdone = intel_uncore_read(uncore, GEN2_INSTDONE);
1161 		break;
1162 	}
1163 }
1164 
1165 static bool ring_is_idle(struct intel_engine_cs *engine)
1166 {
1167 	bool idle = true;
1168 
1169 	if (I915_SELFTEST_ONLY(!engine->mmio_base))
1170 		return true;
1171 
1172 	if (!intel_engine_pm_get_if_awake(engine))
1173 		return true;
1174 
1175 	/* First check that no commands are left in the ring */
1176 	if ((ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR) !=
1177 	    (ENGINE_READ(engine, RING_TAIL) & TAIL_ADDR))
1178 		idle = false;
1179 
1180 	/* No bit for gen2, so assume the CS parser is idle */
1181 	if (INTEL_GEN(engine->i915) > 2 &&
1182 	    !(ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE))
1183 		idle = false;
1184 
1185 	intel_engine_pm_put(engine);
1186 
1187 	return idle;
1188 }
1189 
1190 void intel_engine_flush_submission(struct intel_engine_cs *engine)
1191 {
1192 	struct tasklet_struct *t = &engine->execlists.tasklet;
1193 
1194 	if (!t->func)
1195 		return;
1196 
1197 	/* Synchronise and wait for the tasklet on another CPU */
1198 	tasklet_kill(t);
1199 
1200 	/* Having cancelled the tasklet, ensure that is run */
1201 	local_bh_disable();
1202 	if (tasklet_trylock(t)) {
1203 		/* Must wait for any GPU reset in progress. */
1204 		if (__tasklet_is_enabled(t))
1205 			t->func(t->data);
1206 		tasklet_unlock(t);
1207 	}
1208 	local_bh_enable();
1209 }
1210 
1211 /**
1212  * intel_engine_is_idle() - Report if the engine has finished process all work
1213  * @engine: the intel_engine_cs
1214  *
1215  * Return true if there are no requests pending, nothing left to be submitted
1216  * to hardware, and that the engine is idle.
1217  */
1218 bool intel_engine_is_idle(struct intel_engine_cs *engine)
1219 {
1220 	/* More white lies, if wedged, hw state is inconsistent */
1221 	if (intel_gt_is_wedged(engine->gt))
1222 		return true;
1223 
1224 	if (!intel_engine_pm_is_awake(engine))
1225 		return true;
1226 
1227 	/* Waiting to drain ELSP? */
1228 	if (execlists_active(&engine->execlists)) {
1229 		synchronize_hardirq(engine->i915->drm.pdev->irq);
1230 
1231 		intel_engine_flush_submission(engine);
1232 
1233 		if (execlists_active(&engine->execlists))
1234 			return false;
1235 	}
1236 
1237 	/* ELSP is empty, but there are ready requests? E.g. after reset */
1238 	if (!RB_EMPTY_ROOT(&engine->execlists.queue.rb_root))
1239 		return false;
1240 
1241 	/* Ring stopped? */
1242 	return ring_is_idle(engine);
1243 }
1244 
1245 bool intel_engines_are_idle(struct intel_gt *gt)
1246 {
1247 	struct intel_engine_cs *engine;
1248 	enum intel_engine_id id;
1249 
1250 	/*
1251 	 * If the driver is wedged, HW state may be very inconsistent and
1252 	 * report that it is still busy, even though we have stopped using it.
1253 	 */
1254 	if (intel_gt_is_wedged(gt))
1255 		return true;
1256 
1257 	/* Already parked (and passed an idleness test); must still be idle */
1258 	if (!READ_ONCE(gt->awake))
1259 		return true;
1260 
1261 	for_each_engine(engine, gt, id) {
1262 		if (!intel_engine_is_idle(engine))
1263 			return false;
1264 	}
1265 
1266 	return true;
1267 }
1268 
1269 void intel_engines_reset_default_submission(struct intel_gt *gt)
1270 {
1271 	struct intel_engine_cs *engine;
1272 	enum intel_engine_id id;
1273 
1274 	for_each_engine(engine, gt, id)
1275 		engine->set_default_submission(engine);
1276 }
1277 
1278 bool intel_engine_can_store_dword(struct intel_engine_cs *engine)
1279 {
1280 	switch (INTEL_GEN(engine->i915)) {
1281 	case 2:
1282 		return false; /* uses physical not virtual addresses */
1283 	case 3:
1284 		/* maybe only uses physical not virtual addresses */
1285 		return !(IS_I915G(engine->i915) || IS_I915GM(engine->i915));
1286 	case 4:
1287 		return !IS_I965G(engine->i915); /* who knows! */
1288 	case 6:
1289 		return engine->class != VIDEO_DECODE_CLASS; /* b0rked */
1290 	default:
1291 		return true;
1292 	}
1293 }
1294 
1295 static int print_sched_attr(const struct i915_sched_attr *attr,
1296 			    char *buf, int x, int len)
1297 {
1298 	if (attr->priority == I915_PRIORITY_INVALID)
1299 		return x;
1300 
1301 	x += snprintf(buf + x, len - x,
1302 		      " prio=%d", attr->priority);
1303 
1304 	return x;
1305 }
1306 
1307 static void print_request(struct drm_printer *m,
1308 			  struct i915_request *rq,
1309 			  const char *prefix)
1310 {
1311 	const char *name = rq->fence.ops->get_timeline_name(&rq->fence);
1312 	char buf[80] = "";
1313 	int x = 0;
1314 
1315 	x = print_sched_attr(&rq->sched.attr, buf, x, sizeof(buf));
1316 
1317 	drm_printf(m, "%s %llx:%llx%s%s %s @ %dms: %s\n",
1318 		   prefix,
1319 		   rq->fence.context, rq->fence.seqno,
1320 		   i915_request_completed(rq) ? "!" :
1321 		   i915_request_started(rq) ? "*" :
1322 		   "",
1323 		   test_bit(DMA_FENCE_FLAG_SIGNALED_BIT,
1324 			    &rq->fence.flags) ? "+" :
1325 		   test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT,
1326 			    &rq->fence.flags) ? "-" :
1327 		   "",
1328 		   buf,
1329 		   jiffies_to_msecs(jiffies - rq->emitted_jiffies),
1330 		   name);
1331 }
1332 
1333 static struct intel_timeline *get_timeline(struct i915_request *rq)
1334 {
1335 	struct intel_timeline *tl;
1336 
1337 	/*
1338 	 * Even though we are holding the engine->active.lock here, there
1339 	 * is no control over the submission queue per-se and we are
1340 	 * inspecting the active state at a random point in time, with an
1341 	 * unknown queue. Play safe and make sure the timeline remains valid.
1342 	 * (Only being used for pretty printing, one extra kref shouldn't
1343 	 * cause a camel stampede!)
1344 	 */
1345 	rcu_read_lock();
1346 	tl = rcu_dereference(rq->timeline);
1347 	if (!kref_get_unless_zero(&tl->kref))
1348 		tl = NULL;
1349 	rcu_read_unlock();
1350 
1351 	return tl;
1352 }
1353 
1354 static int print_ring(char *buf, int sz, struct i915_request *rq)
1355 {
1356 	int len = 0;
1357 
1358 	if (!i915_request_signaled(rq)) {
1359 		struct intel_timeline *tl = get_timeline(rq);
1360 
1361 		len = scnprintf(buf, sz,
1362 				"ring:{start:%08x, hwsp:%08x, seqno:%08x, runtime:%llums}, ",
1363 				i915_ggtt_offset(rq->ring->vma),
1364 				tl ? tl->hwsp_offset : 0,
1365 				hwsp_seqno(rq),
1366 				DIV_ROUND_CLOSEST_ULL(intel_context_get_total_runtime_ns(rq->context),
1367 						      1000 * 1000));
1368 
1369 		if (tl)
1370 			intel_timeline_put(tl);
1371 	}
1372 
1373 	return len;
1374 }
1375 
1376 static void hexdump(struct drm_printer *m, const void *buf, size_t len)
1377 {
1378 	const size_t rowsize = 8 * sizeof(u32);
1379 	const void *prev = NULL;
1380 	bool skip = false;
1381 	size_t pos;
1382 
1383 	for (pos = 0; pos < len; pos += rowsize) {
1384 		char line[128];
1385 
1386 		if (prev && !memcmp(prev, buf + pos, rowsize)) {
1387 			if (!skip) {
1388 				drm_printf(m, "*\n");
1389 				skip = true;
1390 			}
1391 			continue;
1392 		}
1393 
1394 		WARN_ON_ONCE(hex_dump_to_buffer(buf + pos, len - pos,
1395 						rowsize, sizeof(u32),
1396 						line, sizeof(line),
1397 						false) >= sizeof(line));
1398 		drm_printf(m, "[%04zx] %s\n", pos, line);
1399 
1400 		prev = buf + pos;
1401 		skip = false;
1402 	}
1403 }
1404 
1405 static const char *repr_timer(const struct timer_list *t)
1406 {
1407 	if (!READ_ONCE(t->expires))
1408 		return "inactive";
1409 
1410 	if (timer_pending(t))
1411 		return "active";
1412 
1413 	return "expired";
1414 }
1415 
1416 static void intel_engine_print_registers(struct intel_engine_cs *engine,
1417 					 struct drm_printer *m)
1418 {
1419 	struct drm_i915_private *dev_priv = engine->i915;
1420 	struct intel_engine_execlists * const execlists = &engine->execlists;
1421 	u64 addr;
1422 
1423 	if (engine->id == RENDER_CLASS && IS_GEN_RANGE(dev_priv, 4, 7))
1424 		drm_printf(m, "\tCCID: 0x%08x\n", ENGINE_READ(engine, CCID));
1425 	if (HAS_EXECLISTS(dev_priv)) {
1426 		drm_printf(m, "\tEL_STAT_HI: 0x%08x\n",
1427 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_HI));
1428 		drm_printf(m, "\tEL_STAT_LO: 0x%08x\n",
1429 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_LO));
1430 	}
1431 	drm_printf(m, "\tRING_START: 0x%08x\n",
1432 		   ENGINE_READ(engine, RING_START));
1433 	drm_printf(m, "\tRING_HEAD:  0x%08x\n",
1434 		   ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR);
1435 	drm_printf(m, "\tRING_TAIL:  0x%08x\n",
1436 		   ENGINE_READ(engine, RING_TAIL) & TAIL_ADDR);
1437 	drm_printf(m, "\tRING_CTL:   0x%08x%s\n",
1438 		   ENGINE_READ(engine, RING_CTL),
1439 		   ENGINE_READ(engine, RING_CTL) & (RING_WAIT | RING_WAIT_SEMAPHORE) ? " [waiting]" : "");
1440 	if (INTEL_GEN(engine->i915) > 2) {
1441 		drm_printf(m, "\tRING_MODE:  0x%08x%s\n",
1442 			   ENGINE_READ(engine, RING_MI_MODE),
1443 			   ENGINE_READ(engine, RING_MI_MODE) & (MODE_IDLE) ? " [idle]" : "");
1444 	}
1445 
1446 	if (INTEL_GEN(dev_priv) >= 6) {
1447 		drm_printf(m, "\tRING_IMR:   0x%08x\n",
1448 			   ENGINE_READ(engine, RING_IMR));
1449 		drm_printf(m, "\tRING_ESR:   0x%08x\n",
1450 			   ENGINE_READ(engine, RING_ESR));
1451 		drm_printf(m, "\tRING_EMR:   0x%08x\n",
1452 			   ENGINE_READ(engine, RING_EMR));
1453 		drm_printf(m, "\tRING_EIR:   0x%08x\n",
1454 			   ENGINE_READ(engine, RING_EIR));
1455 	}
1456 
1457 	addr = intel_engine_get_active_head(engine);
1458 	drm_printf(m, "\tACTHD:  0x%08x_%08x\n",
1459 		   upper_32_bits(addr), lower_32_bits(addr));
1460 	addr = intel_engine_get_last_batch_head(engine);
1461 	drm_printf(m, "\tBBADDR: 0x%08x_%08x\n",
1462 		   upper_32_bits(addr), lower_32_bits(addr));
1463 	if (INTEL_GEN(dev_priv) >= 8)
1464 		addr = ENGINE_READ64(engine, RING_DMA_FADD, RING_DMA_FADD_UDW);
1465 	else if (INTEL_GEN(dev_priv) >= 4)
1466 		addr = ENGINE_READ(engine, RING_DMA_FADD);
1467 	else
1468 		addr = ENGINE_READ(engine, DMA_FADD_I8XX);
1469 	drm_printf(m, "\tDMA_FADDR: 0x%08x_%08x\n",
1470 		   upper_32_bits(addr), lower_32_bits(addr));
1471 	if (INTEL_GEN(dev_priv) >= 4) {
1472 		drm_printf(m, "\tIPEIR: 0x%08x\n",
1473 			   ENGINE_READ(engine, RING_IPEIR));
1474 		drm_printf(m, "\tIPEHR: 0x%08x\n",
1475 			   ENGINE_READ(engine, RING_IPEHR));
1476 	} else {
1477 		drm_printf(m, "\tIPEIR: 0x%08x\n", ENGINE_READ(engine, IPEIR));
1478 		drm_printf(m, "\tIPEHR: 0x%08x\n", ENGINE_READ(engine, IPEHR));
1479 	}
1480 
1481 	if (HAS_EXECLISTS(dev_priv)) {
1482 		struct i915_request * const *port, *rq;
1483 		const u32 *hws =
1484 			&engine->status_page.addr[I915_HWS_CSB_BUF0_INDEX];
1485 		const u8 num_entries = execlists->csb_size;
1486 		unsigned int idx;
1487 		u8 read, write;
1488 
1489 		drm_printf(m, "\tExeclist tasklet queued? %s (%s), preempt? %s, timeslice? %s\n",
1490 			   yesno(test_bit(TASKLET_STATE_SCHED,
1491 					  &engine->execlists.tasklet.state)),
1492 			   enableddisabled(!atomic_read(&engine->execlists.tasklet.count)),
1493 			   repr_timer(&engine->execlists.preempt),
1494 			   repr_timer(&engine->execlists.timer));
1495 
1496 		read = execlists->csb_head;
1497 		write = READ_ONCE(*execlists->csb_write);
1498 
1499 		drm_printf(m, "\tExeclist status: 0x%08x %08x; CSB read:%d, write:%d, entries:%d\n",
1500 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_LO),
1501 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_HI),
1502 			   read, write, num_entries);
1503 
1504 		if (read >= num_entries)
1505 			read = 0;
1506 		if (write >= num_entries)
1507 			write = 0;
1508 		if (read > write)
1509 			write += num_entries;
1510 		while (read < write) {
1511 			idx = ++read % num_entries;
1512 			drm_printf(m, "\tExeclist CSB[%d]: 0x%08x, context: %d\n",
1513 				   idx, hws[idx * 2], hws[idx * 2 + 1]);
1514 		}
1515 
1516 		execlists_active_lock_bh(execlists);
1517 		rcu_read_lock();
1518 		for (port = execlists->active; (rq = *port); port++) {
1519 			char hdr[160];
1520 			int len;
1521 
1522 			len = scnprintf(hdr, sizeof(hdr),
1523 					"\t\tActive[%d]:  ccid:%08x%s%s, ",
1524 					(int)(port - execlists->active),
1525 					rq->context->lrc.ccid,
1526 					intel_context_is_closed(rq->context) ? "!" : "",
1527 					intel_context_is_banned(rq->context) ? "*" : "");
1528 			len += print_ring(hdr + len, sizeof(hdr) - len, rq);
1529 			scnprintf(hdr + len, sizeof(hdr) - len, "rq: ");
1530 			print_request(m, rq, hdr);
1531 		}
1532 		for (port = execlists->pending; (rq = *port); port++) {
1533 			char hdr[160];
1534 			int len;
1535 
1536 			len = scnprintf(hdr, sizeof(hdr),
1537 					"\t\tPending[%d]: ccid:%08x%s%s, ",
1538 					(int)(port - execlists->pending),
1539 					rq->context->lrc.ccid,
1540 					intel_context_is_closed(rq->context) ? "!" : "",
1541 					intel_context_is_banned(rq->context) ? "*" : "");
1542 			len += print_ring(hdr + len, sizeof(hdr) - len, rq);
1543 			scnprintf(hdr + len, sizeof(hdr) - len, "rq: ");
1544 			print_request(m, rq, hdr);
1545 		}
1546 		rcu_read_unlock();
1547 		execlists_active_unlock_bh(execlists);
1548 	} else if (INTEL_GEN(dev_priv) > 6) {
1549 		drm_printf(m, "\tPP_DIR_BASE: 0x%08x\n",
1550 			   ENGINE_READ(engine, RING_PP_DIR_BASE));
1551 		drm_printf(m, "\tPP_DIR_BASE_READ: 0x%08x\n",
1552 			   ENGINE_READ(engine, RING_PP_DIR_BASE_READ));
1553 		drm_printf(m, "\tPP_DIR_DCLV: 0x%08x\n",
1554 			   ENGINE_READ(engine, RING_PP_DIR_DCLV));
1555 	}
1556 }
1557 
1558 static void print_request_ring(struct drm_printer *m, struct i915_request *rq)
1559 {
1560 	void *ring;
1561 	int size;
1562 
1563 	drm_printf(m,
1564 		   "[head %04x, postfix %04x, tail %04x, batch 0x%08x_%08x]:\n",
1565 		   rq->head, rq->postfix, rq->tail,
1566 		   rq->batch ? upper_32_bits(rq->batch->node.start) : ~0u,
1567 		   rq->batch ? lower_32_bits(rq->batch->node.start) : ~0u);
1568 
1569 	size = rq->tail - rq->head;
1570 	if (rq->tail < rq->head)
1571 		size += rq->ring->size;
1572 
1573 	ring = kmalloc(size, GFP_ATOMIC);
1574 	if (ring) {
1575 		const void *vaddr = rq->ring->vaddr;
1576 		unsigned int head = rq->head;
1577 		unsigned int len = 0;
1578 
1579 		if (rq->tail < head) {
1580 			len = rq->ring->size - head;
1581 			memcpy(ring, vaddr + head, len);
1582 			head = 0;
1583 		}
1584 		memcpy(ring + len, vaddr + head, size - len);
1585 
1586 		hexdump(m, ring, size);
1587 		kfree(ring);
1588 	}
1589 }
1590 
1591 static unsigned long list_count(struct list_head *list)
1592 {
1593 	struct list_head *pos;
1594 	unsigned long count = 0;
1595 
1596 	list_for_each(pos, list)
1597 		count++;
1598 
1599 	return count;
1600 }
1601 
1602 void intel_engine_dump(struct intel_engine_cs *engine,
1603 		       struct drm_printer *m,
1604 		       const char *header, ...)
1605 {
1606 	struct i915_gpu_error * const error = &engine->i915->gpu_error;
1607 	struct i915_request *rq;
1608 	intel_wakeref_t wakeref;
1609 	unsigned long flags;
1610 	ktime_t dummy;
1611 
1612 	if (header) {
1613 		va_list ap;
1614 
1615 		va_start(ap, header);
1616 		drm_vprintf(m, header, &ap);
1617 		va_end(ap);
1618 	}
1619 
1620 	if (intel_gt_is_wedged(engine->gt))
1621 		drm_printf(m, "*** WEDGED ***\n");
1622 
1623 	drm_printf(m, "\tAwake? %d\n", atomic_read(&engine->wakeref.count));
1624 	drm_printf(m, "\tBarriers?: %s\n",
1625 		   yesno(!llist_empty(&engine->barrier_tasks)));
1626 	drm_printf(m, "\tLatency: %luus\n",
1627 		   ewma__engine_latency_read(&engine->latency));
1628 	if (intel_engine_supports_stats(engine))
1629 		drm_printf(m, "\tRuntime: %llums\n",
1630 			   ktime_to_ms(intel_engine_get_busy_time(engine,
1631 								  &dummy)));
1632 	drm_printf(m, "\tForcewake: %x domains, %d active\n",
1633 		   engine->fw_domain, atomic_read(&engine->fw_active));
1634 
1635 	rcu_read_lock();
1636 	rq = READ_ONCE(engine->heartbeat.systole);
1637 	if (rq)
1638 		drm_printf(m, "\tHeartbeat: %d ms ago\n",
1639 			   jiffies_to_msecs(jiffies - rq->emitted_jiffies));
1640 	rcu_read_unlock();
1641 	drm_printf(m, "\tReset count: %d (global %d)\n",
1642 		   i915_reset_engine_count(error, engine),
1643 		   i915_reset_count(error));
1644 
1645 	drm_printf(m, "\tRequests:\n");
1646 
1647 	spin_lock_irqsave(&engine->active.lock, flags);
1648 	rq = intel_engine_find_active_request(engine);
1649 	if (rq) {
1650 		struct intel_timeline *tl = get_timeline(rq);
1651 
1652 		print_request(m, rq, "\t\tactive ");
1653 
1654 		drm_printf(m, "\t\tring->start:  0x%08x\n",
1655 			   i915_ggtt_offset(rq->ring->vma));
1656 		drm_printf(m, "\t\tring->head:   0x%08x\n",
1657 			   rq->ring->head);
1658 		drm_printf(m, "\t\tring->tail:   0x%08x\n",
1659 			   rq->ring->tail);
1660 		drm_printf(m, "\t\tring->emit:   0x%08x\n",
1661 			   rq->ring->emit);
1662 		drm_printf(m, "\t\tring->space:  0x%08x\n",
1663 			   rq->ring->space);
1664 
1665 		if (tl) {
1666 			drm_printf(m, "\t\tring->hwsp:   0x%08x\n",
1667 				   tl->hwsp_offset);
1668 			intel_timeline_put(tl);
1669 		}
1670 
1671 		print_request_ring(m, rq);
1672 
1673 		if (rq->context->lrc_reg_state) {
1674 			drm_printf(m, "Logical Ring Context:\n");
1675 			hexdump(m, rq->context->lrc_reg_state, PAGE_SIZE);
1676 		}
1677 	}
1678 	drm_printf(m, "\tOn hold?: %lu\n", list_count(&engine->active.hold));
1679 	spin_unlock_irqrestore(&engine->active.lock, flags);
1680 
1681 	drm_printf(m, "\tMMIO base:  0x%08x\n", engine->mmio_base);
1682 	wakeref = intel_runtime_pm_get_if_in_use(engine->uncore->rpm);
1683 	if (wakeref) {
1684 		intel_engine_print_registers(engine, m);
1685 		intel_runtime_pm_put(engine->uncore->rpm, wakeref);
1686 	} else {
1687 		drm_printf(m, "\tDevice is asleep; skipping register dump\n");
1688 	}
1689 
1690 	intel_execlists_show_requests(engine, m, print_request, 8);
1691 
1692 	drm_printf(m, "HWSP:\n");
1693 	hexdump(m, engine->status_page.addr, PAGE_SIZE);
1694 
1695 	drm_printf(m, "Idle? %s\n", yesno(intel_engine_is_idle(engine)));
1696 
1697 	intel_engine_print_breadcrumbs(engine, m);
1698 }
1699 
1700 static ktime_t __intel_engine_get_busy_time(struct intel_engine_cs *engine,
1701 					    ktime_t *now)
1702 {
1703 	ktime_t total = engine->stats.total;
1704 
1705 	/*
1706 	 * If the engine is executing something at the moment
1707 	 * add it to the total.
1708 	 */
1709 	*now = ktime_get();
1710 	if (atomic_read(&engine->stats.active))
1711 		total = ktime_add(total, ktime_sub(*now, engine->stats.start));
1712 
1713 	return total;
1714 }
1715 
1716 /**
1717  * intel_engine_get_busy_time() - Return current accumulated engine busyness
1718  * @engine: engine to report on
1719  * @now: monotonic timestamp of sampling
1720  *
1721  * Returns accumulated time @engine was busy since engine stats were enabled.
1722  */
1723 ktime_t intel_engine_get_busy_time(struct intel_engine_cs *engine, ktime_t *now)
1724 {
1725 	unsigned int seq;
1726 	ktime_t total;
1727 
1728 	do {
1729 		seq = read_seqbegin(&engine->stats.lock);
1730 		total = __intel_engine_get_busy_time(engine, now);
1731 	} while (read_seqretry(&engine->stats.lock, seq));
1732 
1733 	return total;
1734 }
1735 
1736 static bool match_ring(struct i915_request *rq)
1737 {
1738 	u32 ring = ENGINE_READ(rq->engine, RING_START);
1739 
1740 	return ring == i915_ggtt_offset(rq->ring->vma);
1741 }
1742 
1743 struct i915_request *
1744 intel_engine_find_active_request(struct intel_engine_cs *engine)
1745 {
1746 	struct i915_request *request, *active = NULL;
1747 
1748 	/*
1749 	 * We are called by the error capture, reset and to dump engine
1750 	 * state at random points in time. In particular, note that neither is
1751 	 * crucially ordered with an interrupt. After a hang, the GPU is dead
1752 	 * and we assume that no more writes can happen (we waited long enough
1753 	 * for all writes that were in transaction to be flushed) - adding an
1754 	 * extra delay for a recent interrupt is pointless. Hence, we do
1755 	 * not need an engine->irq_seqno_barrier() before the seqno reads.
1756 	 * At all other times, we must assume the GPU is still running, but
1757 	 * we only care about the snapshot of this moment.
1758 	 */
1759 	lockdep_assert_held(&engine->active.lock);
1760 
1761 	rcu_read_lock();
1762 	request = execlists_active(&engine->execlists);
1763 	if (request) {
1764 		struct intel_timeline *tl = request->context->timeline;
1765 
1766 		list_for_each_entry_from_reverse(request, &tl->requests, link) {
1767 			if (i915_request_completed(request))
1768 				break;
1769 
1770 			active = request;
1771 		}
1772 	}
1773 	rcu_read_unlock();
1774 	if (active)
1775 		return active;
1776 
1777 	list_for_each_entry(request, &engine->active.requests, sched.link) {
1778 		if (i915_request_completed(request))
1779 			continue;
1780 
1781 		if (!i915_request_started(request))
1782 			continue;
1783 
1784 		/* More than one preemptible request may match! */
1785 		if (!match_ring(request))
1786 			continue;
1787 
1788 		active = request;
1789 		break;
1790 	}
1791 
1792 	return active;
1793 }
1794 
1795 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1796 #include "mock_engine.c"
1797 #include "selftest_engine.c"
1798 #include "selftest_engine_cs.c"
1799 #endif
1800