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_context.h"
32 #include "intel_engine.h"
33 #include "intel_engine_pm.h"
34 #include "intel_engine_user.h"
35 #include "intel_gt.h"
36 #include "intel_gt_requests.h"
37 #include "intel_gt_pm.h"
38 #include "intel_lrc.h"
39 #include "intel_reset.h"
40 #include "intel_ring.h"
41 
42 /* Haswell does have the CXT_SIZE register however it does not appear to be
43  * valid. Now, docs explain in dwords what is in the context object. The full
44  * size is 70720 bytes, however, the power context and execlist context will
45  * never be saved (power context is stored elsewhere, and execlists don't work
46  * on HSW) - so the final size, including the extra state required for the
47  * Resource Streamer, is 66944 bytes, which rounds to 17 pages.
48  */
49 #define HSW_CXT_TOTAL_SIZE		(17 * PAGE_SIZE)
50 
51 #define DEFAULT_LR_CONTEXT_RENDER_SIZE	(22 * PAGE_SIZE)
52 #define GEN8_LR_CONTEXT_RENDER_SIZE	(20 * PAGE_SIZE)
53 #define GEN9_LR_CONTEXT_RENDER_SIZE	(22 * PAGE_SIZE)
54 #define GEN10_LR_CONTEXT_RENDER_SIZE	(18 * PAGE_SIZE)
55 #define GEN11_LR_CONTEXT_RENDER_SIZE	(14 * PAGE_SIZE)
56 
57 #define GEN8_LR_CONTEXT_OTHER_SIZE	( 2 * PAGE_SIZE)
58 
59 #define MAX_MMIO_BASES 3
60 struct engine_info {
61 	unsigned int hw_id;
62 	u8 class;
63 	u8 instance;
64 	/* mmio bases table *must* be sorted in reverse gen order */
65 	struct engine_mmio_base {
66 		u32 gen : 8;
67 		u32 base : 24;
68 	} mmio_bases[MAX_MMIO_BASES];
69 };
70 
71 static const struct engine_info intel_engines[] = {
72 	[RCS0] = {
73 		.hw_id = RCS0_HW,
74 		.class = RENDER_CLASS,
75 		.instance = 0,
76 		.mmio_bases = {
77 			{ .gen = 1, .base = RENDER_RING_BASE }
78 		},
79 	},
80 	[BCS0] = {
81 		.hw_id = BCS0_HW,
82 		.class = COPY_ENGINE_CLASS,
83 		.instance = 0,
84 		.mmio_bases = {
85 			{ .gen = 6, .base = BLT_RING_BASE }
86 		},
87 	},
88 	[VCS0] = {
89 		.hw_id = VCS0_HW,
90 		.class = VIDEO_DECODE_CLASS,
91 		.instance = 0,
92 		.mmio_bases = {
93 			{ .gen = 11, .base = GEN11_BSD_RING_BASE },
94 			{ .gen = 6, .base = GEN6_BSD_RING_BASE },
95 			{ .gen = 4, .base = BSD_RING_BASE }
96 		},
97 	},
98 	[VCS1] = {
99 		.hw_id = VCS1_HW,
100 		.class = VIDEO_DECODE_CLASS,
101 		.instance = 1,
102 		.mmio_bases = {
103 			{ .gen = 11, .base = GEN11_BSD2_RING_BASE },
104 			{ .gen = 8, .base = GEN8_BSD2_RING_BASE }
105 		},
106 	},
107 	[VCS2] = {
108 		.hw_id = VCS2_HW,
109 		.class = VIDEO_DECODE_CLASS,
110 		.instance = 2,
111 		.mmio_bases = {
112 			{ .gen = 11, .base = GEN11_BSD3_RING_BASE }
113 		},
114 	},
115 	[VCS3] = {
116 		.hw_id = VCS3_HW,
117 		.class = VIDEO_DECODE_CLASS,
118 		.instance = 3,
119 		.mmio_bases = {
120 			{ .gen = 11, .base = GEN11_BSD4_RING_BASE }
121 		},
122 	},
123 	[VECS0] = {
124 		.hw_id = VECS0_HW,
125 		.class = VIDEO_ENHANCEMENT_CLASS,
126 		.instance = 0,
127 		.mmio_bases = {
128 			{ .gen = 11, .base = GEN11_VEBOX_RING_BASE },
129 			{ .gen = 7, .base = VEBOX_RING_BASE }
130 		},
131 	},
132 	[VECS1] = {
133 		.hw_id = VECS1_HW,
134 		.class = VIDEO_ENHANCEMENT_CLASS,
135 		.instance = 1,
136 		.mmio_bases = {
137 			{ .gen = 11, .base = GEN11_VEBOX2_RING_BASE }
138 		},
139 	},
140 };
141 
142 /**
143  * intel_engine_context_size() - return the size of the context for an engine
144  * @gt: the gt
145  * @class: engine class
146  *
147  * Each engine class may require a different amount of space for a context
148  * image.
149  *
150  * Return: size (in bytes) of an engine class specific context image
151  *
152  * Note: this size includes the HWSP, which is part of the context image
153  * in LRC mode, but does not include the "shared data page" used with
154  * GuC submission. The caller should account for this if using the GuC.
155  */
156 u32 intel_engine_context_size(struct intel_gt *gt, u8 class)
157 {
158 	struct intel_uncore *uncore = gt->uncore;
159 	u32 cxt_size;
160 
161 	BUILD_BUG_ON(I915_GTT_PAGE_SIZE != PAGE_SIZE);
162 
163 	switch (class) {
164 	case RENDER_CLASS:
165 		switch (INTEL_GEN(gt->i915)) {
166 		default:
167 			MISSING_CASE(INTEL_GEN(gt->i915));
168 			return DEFAULT_LR_CONTEXT_RENDER_SIZE;
169 		case 12:
170 		case 11:
171 			return GEN11_LR_CONTEXT_RENDER_SIZE;
172 		case 10:
173 			return GEN10_LR_CONTEXT_RENDER_SIZE;
174 		case 9:
175 			return GEN9_LR_CONTEXT_RENDER_SIZE;
176 		case 8:
177 			return GEN8_LR_CONTEXT_RENDER_SIZE;
178 		case 7:
179 			if (IS_HASWELL(gt->i915))
180 				return HSW_CXT_TOTAL_SIZE;
181 
182 			cxt_size = intel_uncore_read(uncore, GEN7_CXT_SIZE);
183 			return round_up(GEN7_CXT_TOTAL_SIZE(cxt_size) * 64,
184 					PAGE_SIZE);
185 		case 6:
186 			cxt_size = intel_uncore_read(uncore, CXT_SIZE);
187 			return round_up(GEN6_CXT_TOTAL_SIZE(cxt_size) * 64,
188 					PAGE_SIZE);
189 		case 5:
190 		case 4:
191 			/*
192 			 * There is a discrepancy here between the size reported
193 			 * by the register and the size of the context layout
194 			 * in the docs. Both are described as authorative!
195 			 *
196 			 * The discrepancy is on the order of a few cachelines,
197 			 * but the total is under one page (4k), which is our
198 			 * minimum allocation anyway so it should all come
199 			 * out in the wash.
200 			 */
201 			cxt_size = intel_uncore_read(uncore, CXT_SIZE) + 1;
202 			drm_dbg(&gt->i915->drm,
203 				"gen%d CXT_SIZE = %d bytes [0x%08x]\n",
204 				INTEL_GEN(gt->i915), cxt_size * 64,
205 				cxt_size - 1);
206 			return round_up(cxt_size * 64, PAGE_SIZE);
207 		case 3:
208 		case 2:
209 		/* For the special day when i810 gets merged. */
210 		case 1:
211 			return 0;
212 		}
213 		break;
214 	default:
215 		MISSING_CASE(class);
216 		fallthrough;
217 	case VIDEO_DECODE_CLASS:
218 	case VIDEO_ENHANCEMENT_CLASS:
219 	case COPY_ENGINE_CLASS:
220 		if (INTEL_GEN(gt->i915) < 8)
221 			return 0;
222 		return GEN8_LR_CONTEXT_OTHER_SIZE;
223 	}
224 }
225 
226 static u32 __engine_mmio_base(struct drm_i915_private *i915,
227 			      const struct engine_mmio_base *bases)
228 {
229 	int i;
230 
231 	for (i = 0; i < MAX_MMIO_BASES; i++)
232 		if (INTEL_GEN(i915) >= bases[i].gen)
233 			break;
234 
235 	GEM_BUG_ON(i == MAX_MMIO_BASES);
236 	GEM_BUG_ON(!bases[i].base);
237 
238 	return bases[i].base;
239 }
240 
241 static void __sprint_engine_name(struct intel_engine_cs *engine)
242 {
243 	/*
244 	 * Before we know what the uABI name for this engine will be,
245 	 * we still would like to keep track of this engine in the debug logs.
246 	 * We throw in a ' here as a reminder that this isn't its final name.
247 	 */
248 	GEM_WARN_ON(snprintf(engine->name, sizeof(engine->name), "%s'%u",
249 			     intel_engine_class_repr(engine->class),
250 			     engine->instance) >= sizeof(engine->name));
251 }
252 
253 void intel_engine_set_hwsp_writemask(struct intel_engine_cs *engine, u32 mask)
254 {
255 	/*
256 	 * Though they added more rings on g4x/ilk, they did not add
257 	 * per-engine HWSTAM until gen6.
258 	 */
259 	if (INTEL_GEN(engine->i915) < 6 && engine->class != RENDER_CLASS)
260 		return;
261 
262 	if (INTEL_GEN(engine->i915) >= 3)
263 		ENGINE_WRITE(engine, RING_HWSTAM, mask);
264 	else
265 		ENGINE_WRITE16(engine, RING_HWSTAM, mask);
266 }
267 
268 static void intel_engine_sanitize_mmio(struct intel_engine_cs *engine)
269 {
270 	/* Mask off all writes into the unknown HWSP */
271 	intel_engine_set_hwsp_writemask(engine, ~0u);
272 }
273 
274 static int intel_engine_setup(struct intel_gt *gt, enum intel_engine_id id)
275 {
276 	const struct engine_info *info = &intel_engines[id];
277 	struct drm_i915_private *i915 = gt->i915;
278 	struct intel_engine_cs *engine;
279 
280 	BUILD_BUG_ON(MAX_ENGINE_CLASS >= BIT(GEN11_ENGINE_CLASS_WIDTH));
281 	BUILD_BUG_ON(MAX_ENGINE_INSTANCE >= BIT(GEN11_ENGINE_INSTANCE_WIDTH));
282 
283 	if (GEM_DEBUG_WARN_ON(id >= ARRAY_SIZE(gt->engine)))
284 		return -EINVAL;
285 
286 	if (GEM_DEBUG_WARN_ON(info->class > MAX_ENGINE_CLASS))
287 		return -EINVAL;
288 
289 	if (GEM_DEBUG_WARN_ON(info->instance > MAX_ENGINE_INSTANCE))
290 		return -EINVAL;
291 
292 	if (GEM_DEBUG_WARN_ON(gt->engine_class[info->class][info->instance]))
293 		return -EINVAL;
294 
295 	engine = kzalloc(sizeof(*engine), GFP_KERNEL);
296 	if (!engine)
297 		return -ENOMEM;
298 
299 	BUILD_BUG_ON(BITS_PER_TYPE(engine->mask) < I915_NUM_ENGINES);
300 
301 	engine->id = id;
302 	engine->legacy_idx = INVALID_ENGINE;
303 	engine->mask = BIT(id);
304 	engine->i915 = i915;
305 	engine->gt = gt;
306 	engine->uncore = gt->uncore;
307 	engine->hw_id = engine->guc_id = info->hw_id;
308 	engine->mmio_base = __engine_mmio_base(i915, info->mmio_bases);
309 
310 	engine->class = info->class;
311 	engine->instance = info->instance;
312 	__sprint_engine_name(engine);
313 
314 	engine->props.heartbeat_interval_ms =
315 		CONFIG_DRM_I915_HEARTBEAT_INTERVAL;
316 	engine->props.max_busywait_duration_ns =
317 		CONFIG_DRM_I915_MAX_REQUEST_BUSYWAIT;
318 	engine->props.preempt_timeout_ms =
319 		CONFIG_DRM_I915_PREEMPT_TIMEOUT;
320 	engine->props.stop_timeout_ms =
321 		CONFIG_DRM_I915_STOP_TIMEOUT;
322 	engine->props.timeslice_duration_ms =
323 		CONFIG_DRM_I915_TIMESLICE_DURATION;
324 
325 	/* Override to uninterruptible for OpenCL workloads. */
326 	if (INTEL_GEN(i915) == 12 && engine->class == RENDER_CLASS)
327 		engine->props.preempt_timeout_ms = 0;
328 
329 	engine->defaults = engine->props; /* never to change again */
330 
331 	engine->context_size = intel_engine_context_size(gt, engine->class);
332 	if (WARN_ON(engine->context_size > BIT(20)))
333 		engine->context_size = 0;
334 	if (engine->context_size)
335 		DRIVER_CAPS(i915)->has_logical_contexts = true;
336 
337 	/* Nothing to do here, execute in order of dependencies */
338 	engine->schedule = NULL;
339 
340 	ewma__engine_latency_init(&engine->latency);
341 	seqlock_init(&engine->stats.lock);
342 
343 	ATOMIC_INIT_NOTIFIER_HEAD(&engine->context_status_notifier);
344 
345 	/* Scrub mmio state on takeover */
346 	intel_engine_sanitize_mmio(engine);
347 
348 	gt->engine_class[info->class][info->instance] = engine;
349 	gt->engine[id] = engine;
350 
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 		     engine->gt->info.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 intel_gt *gt)
385 {
386 	struct intel_engine_cs *engine;
387 	enum intel_engine_id id;
388 
389 	for_each_engine(engine, gt, id)
390 		__setup_engine_capabilities(engine);
391 }
392 
393 /**
394  * intel_engines_release() - free the resources allocated for Command Streamers
395  * @gt: pointer to struct intel_gt
396  */
397 void intel_engines_release(struct intel_gt *gt)
398 {
399 	struct intel_engine_cs *engine;
400 	enum intel_engine_id id;
401 
402 	/*
403 	 * Before we release the resources held by engine, we must be certain
404 	 * that the HW is no longer accessing them -- having the GPU scribble
405 	 * to or read from a page being used for something else causes no end
406 	 * of fun.
407 	 *
408 	 * The GPU should be reset by this point, but assume the worst just
409 	 * in case we aborted before completely initialising the engines.
410 	 */
411 	GEM_BUG_ON(intel_gt_pm_is_awake(gt));
412 	if (!INTEL_INFO(gt->i915)->gpu_reset_clobbers_display)
413 		__intel_gt_reset(gt, ALL_ENGINES);
414 
415 	/* Decouple the backend; but keep the layout for late GPU resets */
416 	for_each_engine(engine, gt, id) {
417 		if (!engine->release)
418 			continue;
419 
420 		intel_wakeref_wait_for_idle(&engine->wakeref);
421 		GEM_BUG_ON(intel_engine_pm_is_awake(engine));
422 
423 		engine->release(engine);
424 		engine->release = NULL;
425 
426 		memset(&engine->reset, 0, sizeof(engine->reset));
427 	}
428 }
429 
430 void intel_engine_free_request_pool(struct intel_engine_cs *engine)
431 {
432 	if (!engine->request_pool)
433 		return;
434 
435 	kmem_cache_free(i915_request_slab_cache(), engine->request_pool);
436 }
437 
438 void intel_engines_free(struct intel_gt *gt)
439 {
440 	struct intel_engine_cs *engine;
441 	enum intel_engine_id id;
442 
443 	/* Free the requests! dma-resv keeps fences around for an eternity */
444 	rcu_barrier();
445 
446 	for_each_engine(engine, gt, id) {
447 		intel_engine_free_request_pool(engine);
448 		kfree(engine);
449 		gt->engine[id] = NULL;
450 	}
451 }
452 
453 /*
454  * Determine which engines are fused off in our particular hardware.
455  * Note that we have a catch-22 situation where we need to be able to access
456  * the blitter forcewake domain to read the engine fuses, but at the same time
457  * we need to know which engines are available on the system to know which
458  * forcewake domains are present. We solve this by intializing the forcewake
459  * domains based on the full engine mask in the platform capabilities before
460  * calling this function and pruning the domains for fused-off engines
461  * afterwards.
462  */
463 static intel_engine_mask_t init_engine_mask(struct intel_gt *gt)
464 {
465 	struct drm_i915_private *i915 = gt->i915;
466 	struct intel_gt_info *info = &gt->info;
467 	struct intel_uncore *uncore = gt->uncore;
468 	unsigned int logical_vdbox = 0;
469 	unsigned int i;
470 	u32 media_fuse;
471 	u16 vdbox_mask;
472 	u16 vebox_mask;
473 
474 	info->engine_mask = INTEL_INFO(i915)->platform_engine_mask;
475 
476 	if (INTEL_GEN(i915) < 11)
477 		return info->engine_mask;
478 
479 	media_fuse = ~intel_uncore_read(uncore, GEN11_GT_VEBOX_VDBOX_DISABLE);
480 
481 	vdbox_mask = media_fuse & GEN11_GT_VDBOX_DISABLE_MASK;
482 	vebox_mask = (media_fuse & GEN11_GT_VEBOX_DISABLE_MASK) >>
483 		      GEN11_GT_VEBOX_DISABLE_SHIFT;
484 
485 	for (i = 0; i < I915_MAX_VCS; i++) {
486 		if (!HAS_ENGINE(gt, _VCS(i))) {
487 			vdbox_mask &= ~BIT(i);
488 			continue;
489 		}
490 
491 		if (!(BIT(i) & vdbox_mask)) {
492 			info->engine_mask &= ~BIT(_VCS(i));
493 			drm_dbg(&i915->drm, "vcs%u fused off\n", i);
494 			continue;
495 		}
496 
497 		/*
498 		 * In Gen11, only even numbered logical VDBOXes are
499 		 * hooked up to an SFC (Scaler & Format Converter) unit.
500 		 * In TGL each VDBOX has access to an SFC.
501 		 */
502 		if (INTEL_GEN(i915) >= 12 || logical_vdbox++ % 2 == 0)
503 			gt->info.vdbox_sfc_access |= BIT(i);
504 	}
505 	drm_dbg(&i915->drm, "vdbox enable: %04x, instances: %04lx\n",
506 		vdbox_mask, VDBOX_MASK(gt));
507 	GEM_BUG_ON(vdbox_mask != VDBOX_MASK(gt));
508 
509 	for (i = 0; i < I915_MAX_VECS; i++) {
510 		if (!HAS_ENGINE(gt, _VECS(i))) {
511 			vebox_mask &= ~BIT(i);
512 			continue;
513 		}
514 
515 		if (!(BIT(i) & vebox_mask)) {
516 			info->engine_mask &= ~BIT(_VECS(i));
517 			drm_dbg(&i915->drm, "vecs%u fused off\n", i);
518 		}
519 	}
520 	drm_dbg(&i915->drm, "vebox enable: %04x, instances: %04lx\n",
521 		vebox_mask, VEBOX_MASK(gt));
522 	GEM_BUG_ON(vebox_mask != VEBOX_MASK(gt));
523 
524 	return info->engine_mask;
525 }
526 
527 /**
528  * intel_engines_init_mmio() - allocate and prepare the Engine Command Streamers
529  * @gt: pointer to struct intel_gt
530  *
531  * Return: non-zero if the initialization failed.
532  */
533 int intel_engines_init_mmio(struct intel_gt *gt)
534 {
535 	struct drm_i915_private *i915 = gt->i915;
536 	const unsigned int engine_mask = init_engine_mask(gt);
537 	unsigned int mask = 0;
538 	unsigned int i;
539 	int err;
540 
541 	drm_WARN_ON(&i915->drm, engine_mask == 0);
542 	drm_WARN_ON(&i915->drm, engine_mask &
543 		    GENMASK(BITS_PER_TYPE(mask) - 1, I915_NUM_ENGINES));
544 
545 	if (i915_inject_probe_failure(i915))
546 		return -ENODEV;
547 
548 	for (i = 0; i < ARRAY_SIZE(intel_engines); i++) {
549 		if (!HAS_ENGINE(gt, i))
550 			continue;
551 
552 		err = intel_engine_setup(gt, i);
553 		if (err)
554 			goto cleanup;
555 
556 		mask |= BIT(i);
557 	}
558 
559 	/*
560 	 * Catch failures to update intel_engines table when the new engines
561 	 * are added to the driver by a warning and disabling the forgotten
562 	 * engines.
563 	 */
564 	if (drm_WARN_ON(&i915->drm, mask != engine_mask))
565 		gt->info.engine_mask = mask;
566 
567 	gt->info.num_engines = hweight32(mask);
568 
569 	intel_gt_check_and_clear_faults(gt);
570 
571 	intel_setup_engine_capabilities(gt);
572 
573 	intel_uncore_prune_engine_fw_domains(gt->uncore, gt);
574 
575 	return 0;
576 
577 cleanup:
578 	intel_engines_free(gt);
579 	return err;
580 }
581 
582 void intel_engine_init_execlists(struct intel_engine_cs *engine)
583 {
584 	struct intel_engine_execlists * const execlists = &engine->execlists;
585 
586 	execlists->port_mask = 1;
587 	GEM_BUG_ON(!is_power_of_2(execlists_num_ports(execlists)));
588 	GEM_BUG_ON(execlists_num_ports(execlists) > EXECLIST_MAX_PORTS);
589 
590 	memset(execlists->pending, 0, sizeof(execlists->pending));
591 	execlists->active =
592 		memset(execlists->inflight, 0, sizeof(execlists->inflight));
593 
594 	execlists->queue_priority_hint = INT_MIN;
595 	execlists->queue = RB_ROOT_CACHED;
596 }
597 
598 static void cleanup_status_page(struct intel_engine_cs *engine)
599 {
600 	struct i915_vma *vma;
601 
602 	/* Prevent writes into HWSP after returning the page to the system */
603 	intel_engine_set_hwsp_writemask(engine, ~0u);
604 
605 	vma = fetch_and_zero(&engine->status_page.vma);
606 	if (!vma)
607 		return;
608 
609 	if (!HWS_NEEDS_PHYSICAL(engine->i915))
610 		i915_vma_unpin(vma);
611 
612 	i915_gem_object_unpin_map(vma->obj);
613 	i915_gem_object_put(vma->obj);
614 }
615 
616 static int pin_ggtt_status_page(struct intel_engine_cs *engine,
617 				struct i915_vma *vma)
618 {
619 	unsigned int flags;
620 
621 	if (!HAS_LLC(engine->i915) && i915_ggtt_has_aperture(engine->gt->ggtt))
622 		/*
623 		 * On g33, we cannot place HWS above 256MiB, so
624 		 * restrict its pinning to the low mappable arena.
625 		 * Though this restriction is not documented for
626 		 * gen4, gen5, or byt, they also behave similarly
627 		 * and hang if the HWS is placed at the top of the
628 		 * GTT. To generalise, it appears that all !llc
629 		 * platforms have issues with us placing the HWS
630 		 * above the mappable region (even though we never
631 		 * actually map it).
632 		 */
633 		flags = PIN_MAPPABLE;
634 	else
635 		flags = PIN_HIGH;
636 
637 	return i915_ggtt_pin(vma, 0, flags);
638 }
639 
640 static int init_status_page(struct intel_engine_cs *engine)
641 {
642 	struct drm_i915_gem_object *obj;
643 	struct i915_vma *vma;
644 	void *vaddr;
645 	int ret;
646 
647 	/*
648 	 * Though the HWS register does support 36bit addresses, historically
649 	 * we have had hangs and corruption reported due to wild writes if
650 	 * the HWS is placed above 4G. We only allow objects to be allocated
651 	 * in GFP_DMA32 for i965, and no earlier physical address users had
652 	 * access to more than 4G.
653 	 */
654 	obj = i915_gem_object_create_internal(engine->i915, PAGE_SIZE);
655 	if (IS_ERR(obj)) {
656 		drm_err(&engine->i915->drm,
657 			"Failed to allocate status page\n");
658 		return PTR_ERR(obj);
659 	}
660 
661 	i915_gem_object_set_cache_coherency(obj, I915_CACHE_LLC);
662 
663 	vma = i915_vma_instance(obj, &engine->gt->ggtt->vm, NULL);
664 	if (IS_ERR(vma)) {
665 		ret = PTR_ERR(vma);
666 		goto err;
667 	}
668 
669 	vaddr = i915_gem_object_pin_map(obj, I915_MAP_WB);
670 	if (IS_ERR(vaddr)) {
671 		ret = PTR_ERR(vaddr);
672 		goto err;
673 	}
674 
675 	engine->status_page.addr = memset(vaddr, 0, PAGE_SIZE);
676 	engine->status_page.vma = vma;
677 
678 	if (!HWS_NEEDS_PHYSICAL(engine->i915)) {
679 		ret = pin_ggtt_status_page(engine, vma);
680 		if (ret)
681 			goto err_unpin;
682 	}
683 
684 	return 0;
685 
686 err_unpin:
687 	i915_gem_object_unpin_map(obj);
688 err:
689 	i915_gem_object_put(obj);
690 	return ret;
691 }
692 
693 static int engine_setup_common(struct intel_engine_cs *engine)
694 {
695 	int err;
696 
697 	init_llist_head(&engine->barrier_tasks);
698 
699 	err = init_status_page(engine);
700 	if (err)
701 		return err;
702 
703 	intel_engine_init_active(engine, ENGINE_PHYSICAL);
704 	intel_engine_init_breadcrumbs(engine);
705 	intel_engine_init_execlists(engine);
706 	intel_engine_init_cmd_parser(engine);
707 	intel_engine_init__pm(engine);
708 	intel_engine_init_retire(engine);
709 
710 	/* Use the whole device by default */
711 	engine->sseu =
712 		intel_sseu_from_device_info(&engine->gt->info.sseu);
713 
714 	intel_engine_init_workarounds(engine);
715 	intel_engine_init_whitelist(engine);
716 	intel_engine_init_ctx_wa(engine);
717 
718 	return 0;
719 }
720 
721 struct measure_breadcrumb {
722 	struct i915_request rq;
723 	struct intel_ring ring;
724 	u32 cs[2048];
725 };
726 
727 static int measure_breadcrumb_dw(struct intel_context *ce)
728 {
729 	struct intel_engine_cs *engine = ce->engine;
730 	struct measure_breadcrumb *frame;
731 	int dw;
732 
733 	GEM_BUG_ON(!engine->gt->scratch);
734 
735 	frame = kzalloc(sizeof(*frame), GFP_KERNEL);
736 	if (!frame)
737 		return -ENOMEM;
738 
739 	frame->rq.engine = engine;
740 	frame->rq.context = ce;
741 	rcu_assign_pointer(frame->rq.timeline, ce->timeline);
742 
743 	frame->ring.vaddr = frame->cs;
744 	frame->ring.size = sizeof(frame->cs);
745 	frame->ring.wrap =
746 		BITS_PER_TYPE(frame->ring.size) - ilog2(frame->ring.size);
747 	frame->ring.effective_size = frame->ring.size;
748 	intel_ring_update_space(&frame->ring);
749 	frame->rq.ring = &frame->ring;
750 
751 	mutex_lock(&ce->timeline->mutex);
752 	spin_lock_irq(&engine->active.lock);
753 
754 	dw = engine->emit_fini_breadcrumb(&frame->rq, frame->cs) - frame->cs;
755 
756 	spin_unlock_irq(&engine->active.lock);
757 	mutex_unlock(&ce->timeline->mutex);
758 
759 	GEM_BUG_ON(dw & 1); /* RING_TAIL must be qword aligned */
760 
761 	kfree(frame);
762 	return dw;
763 }
764 
765 void
766 intel_engine_init_active(struct intel_engine_cs *engine, unsigned int subclass)
767 {
768 	INIT_LIST_HEAD(&engine->active.requests);
769 	INIT_LIST_HEAD(&engine->active.hold);
770 
771 	spin_lock_init(&engine->active.lock);
772 	lockdep_set_subclass(&engine->active.lock, subclass);
773 
774 	/*
775 	 * Due to an interesting quirk in lockdep's internal debug tracking,
776 	 * after setting a subclass we must ensure the lock is used. Otherwise,
777 	 * nr_unused_locks is incremented once too often.
778 	 */
779 #ifdef CONFIG_DEBUG_LOCK_ALLOC
780 	local_irq_disable();
781 	lock_map_acquire(&engine->active.lock.dep_map);
782 	lock_map_release(&engine->active.lock.dep_map);
783 	local_irq_enable();
784 #endif
785 }
786 
787 static struct intel_context *
788 create_kernel_context(struct intel_engine_cs *engine)
789 {
790 	static struct lock_class_key kernel;
791 	struct intel_context *ce;
792 	int err;
793 
794 	ce = intel_context_create(engine);
795 	if (IS_ERR(ce))
796 		return ce;
797 
798 	__set_bit(CONTEXT_BARRIER_BIT, &ce->flags);
799 
800 	err = intel_context_pin(ce); /* perma-pin so it is always available */
801 	if (err) {
802 		intel_context_put(ce);
803 		return ERR_PTR(err);
804 	}
805 
806 	/*
807 	 * Give our perma-pinned kernel timelines a separate lockdep class,
808 	 * so that we can use them from within the normal user timelines
809 	 * should we need to inject GPU operations during their request
810 	 * construction.
811 	 */
812 	lockdep_set_class(&ce->timeline->mutex, &kernel);
813 
814 	return ce;
815 }
816 
817 /**
818  * intel_engines_init_common - initialize cengine state which might require hw access
819  * @engine: Engine to initialize.
820  *
821  * Initializes @engine@ structure members shared between legacy and execlists
822  * submission modes which do require hardware access.
823  *
824  * Typcally done at later stages of submission mode specific engine setup.
825  *
826  * Returns zero on success or an error code on failure.
827  */
828 static int engine_init_common(struct intel_engine_cs *engine)
829 {
830 	struct intel_context *ce;
831 	int ret;
832 
833 	engine->set_default_submission(engine);
834 
835 	/*
836 	 * We may need to do things with the shrinker which
837 	 * require us to immediately switch back to the default
838 	 * context. This can cause a problem as pinning the
839 	 * default context also requires GTT space which may not
840 	 * be available. To avoid this we always pin the default
841 	 * context.
842 	 */
843 	ce = create_kernel_context(engine);
844 	if (IS_ERR(ce))
845 		return PTR_ERR(ce);
846 
847 	ret = measure_breadcrumb_dw(ce);
848 	if (ret < 0)
849 		goto err_context;
850 
851 	engine->emit_fini_breadcrumb_dw = ret;
852 	engine->kernel_context = ce;
853 
854 	return 0;
855 
856 err_context:
857 	intel_context_put(ce);
858 	return ret;
859 }
860 
861 int intel_engines_init(struct intel_gt *gt)
862 {
863 	int (*setup)(struct intel_engine_cs *engine);
864 	struct intel_engine_cs *engine;
865 	enum intel_engine_id id;
866 	int err;
867 
868 	if (HAS_EXECLISTS(gt->i915))
869 		setup = intel_execlists_submission_setup;
870 	else
871 		setup = intel_ring_submission_setup;
872 
873 	for_each_engine(engine, gt, id) {
874 		err = engine_setup_common(engine);
875 		if (err)
876 			return err;
877 
878 		err = setup(engine);
879 		if (err)
880 			return err;
881 
882 		err = engine_init_common(engine);
883 		if (err)
884 			return err;
885 
886 		intel_engine_add_user(engine);
887 	}
888 
889 	return 0;
890 }
891 
892 /**
893  * intel_engines_cleanup_common - cleans up the engine state created by
894  *                                the common initiailizers.
895  * @engine: Engine to cleanup.
896  *
897  * This cleans up everything created by the common helpers.
898  */
899 void intel_engine_cleanup_common(struct intel_engine_cs *engine)
900 {
901 	GEM_BUG_ON(!list_empty(&engine->active.requests));
902 	tasklet_kill(&engine->execlists.tasklet); /* flush the callback */
903 
904 	cleanup_status_page(engine);
905 
906 	intel_engine_fini_retire(engine);
907 	intel_engine_fini_breadcrumbs(engine);
908 	intel_engine_cleanup_cmd_parser(engine);
909 
910 	if (engine->default_state)
911 		fput(engine->default_state);
912 
913 	if (engine->kernel_context) {
914 		intel_context_unpin(engine->kernel_context);
915 		intel_context_put(engine->kernel_context);
916 	}
917 	GEM_BUG_ON(!llist_empty(&engine->barrier_tasks));
918 
919 	intel_wa_list_free(&engine->ctx_wa_list);
920 	intel_wa_list_free(&engine->wa_list);
921 	intel_wa_list_free(&engine->whitelist);
922 }
923 
924 /**
925  * intel_engine_resume - re-initializes the HW state of the engine
926  * @engine: Engine to resume.
927  *
928  * Returns zero on success or an error code on failure.
929  */
930 int intel_engine_resume(struct intel_engine_cs *engine)
931 {
932 	intel_engine_apply_workarounds(engine);
933 	intel_engine_apply_whitelist(engine);
934 
935 	return engine->resume(engine);
936 }
937 
938 u64 intel_engine_get_active_head(const struct intel_engine_cs *engine)
939 {
940 	struct drm_i915_private *i915 = engine->i915;
941 
942 	u64 acthd;
943 
944 	if (INTEL_GEN(i915) >= 8)
945 		acthd = ENGINE_READ64(engine, RING_ACTHD, RING_ACTHD_UDW);
946 	else if (INTEL_GEN(i915) >= 4)
947 		acthd = ENGINE_READ(engine, RING_ACTHD);
948 	else
949 		acthd = ENGINE_READ(engine, ACTHD);
950 
951 	return acthd;
952 }
953 
954 u64 intel_engine_get_last_batch_head(const struct intel_engine_cs *engine)
955 {
956 	u64 bbaddr;
957 
958 	if (INTEL_GEN(engine->i915) >= 8)
959 		bbaddr = ENGINE_READ64(engine, RING_BBADDR, RING_BBADDR_UDW);
960 	else
961 		bbaddr = ENGINE_READ(engine, RING_BBADDR);
962 
963 	return bbaddr;
964 }
965 
966 static unsigned long stop_timeout(const struct intel_engine_cs *engine)
967 {
968 	if (in_atomic() || irqs_disabled()) /* inside atomic preempt-reset? */
969 		return 0;
970 
971 	/*
972 	 * If we are doing a normal GPU reset, we can take our time and allow
973 	 * the engine to quiesce. We've stopped submission to the engine, and
974 	 * if we wait long enough an innocent context should complete and
975 	 * leave the engine idle. So they should not be caught unaware by
976 	 * the forthcoming GPU reset (which usually follows the stop_cs)!
977 	 */
978 	return READ_ONCE(engine->props.stop_timeout_ms);
979 }
980 
981 int intel_engine_stop_cs(struct intel_engine_cs *engine)
982 {
983 	struct intel_uncore *uncore = engine->uncore;
984 	const u32 base = engine->mmio_base;
985 	const i915_reg_t mode = RING_MI_MODE(base);
986 	int err;
987 
988 	if (INTEL_GEN(engine->i915) < 3)
989 		return -ENODEV;
990 
991 	ENGINE_TRACE(engine, "\n");
992 
993 	intel_uncore_write_fw(uncore, mode, _MASKED_BIT_ENABLE(STOP_RING));
994 
995 	err = 0;
996 	if (__intel_wait_for_register_fw(uncore,
997 					 mode, MODE_IDLE, MODE_IDLE,
998 					 1000, stop_timeout(engine),
999 					 NULL)) {
1000 		ENGINE_TRACE(engine, "timed out on STOP_RING -> IDLE\n");
1001 		err = -ETIMEDOUT;
1002 	}
1003 
1004 	/* A final mmio read to let GPU writes be hopefully flushed to memory */
1005 	intel_uncore_posting_read_fw(uncore, mode);
1006 
1007 	return err;
1008 }
1009 
1010 void intel_engine_cancel_stop_cs(struct intel_engine_cs *engine)
1011 {
1012 	ENGINE_TRACE(engine, "\n");
1013 
1014 	ENGINE_WRITE_FW(engine, RING_MI_MODE, _MASKED_BIT_DISABLE(STOP_RING));
1015 }
1016 
1017 const char *i915_cache_level_str(struct drm_i915_private *i915, int type)
1018 {
1019 	switch (type) {
1020 	case I915_CACHE_NONE: return " uncached";
1021 	case I915_CACHE_LLC: return HAS_LLC(i915) ? " LLC" : " snooped";
1022 	case I915_CACHE_L3_LLC: return " L3+LLC";
1023 	case I915_CACHE_WT: return " WT";
1024 	default: return "";
1025 	}
1026 }
1027 
1028 static u32
1029 read_subslice_reg(const struct intel_engine_cs *engine,
1030 		  int slice, int subslice, i915_reg_t reg)
1031 {
1032 	struct drm_i915_private *i915 = engine->i915;
1033 	struct intel_uncore *uncore = engine->uncore;
1034 	u32 mcr_mask, mcr_ss, mcr, old_mcr, val;
1035 	enum forcewake_domains fw_domains;
1036 
1037 	if (INTEL_GEN(i915) >= 11) {
1038 		mcr_mask = GEN11_MCR_SLICE_MASK | GEN11_MCR_SUBSLICE_MASK;
1039 		mcr_ss = GEN11_MCR_SLICE(slice) | GEN11_MCR_SUBSLICE(subslice);
1040 	} else {
1041 		mcr_mask = GEN8_MCR_SLICE_MASK | GEN8_MCR_SUBSLICE_MASK;
1042 		mcr_ss = GEN8_MCR_SLICE(slice) | GEN8_MCR_SUBSLICE(subslice);
1043 	}
1044 
1045 	fw_domains = intel_uncore_forcewake_for_reg(uncore, reg,
1046 						    FW_REG_READ);
1047 	fw_domains |= intel_uncore_forcewake_for_reg(uncore,
1048 						     GEN8_MCR_SELECTOR,
1049 						     FW_REG_READ | FW_REG_WRITE);
1050 
1051 	spin_lock_irq(&uncore->lock);
1052 	intel_uncore_forcewake_get__locked(uncore, fw_domains);
1053 
1054 	old_mcr = mcr = intel_uncore_read_fw(uncore, GEN8_MCR_SELECTOR);
1055 
1056 	mcr &= ~mcr_mask;
1057 	mcr |= mcr_ss;
1058 	intel_uncore_write_fw(uncore, GEN8_MCR_SELECTOR, mcr);
1059 
1060 	val = intel_uncore_read_fw(uncore, reg);
1061 
1062 	mcr &= ~mcr_mask;
1063 	mcr |= old_mcr & mcr_mask;
1064 
1065 	intel_uncore_write_fw(uncore, GEN8_MCR_SELECTOR, mcr);
1066 
1067 	intel_uncore_forcewake_put__locked(uncore, fw_domains);
1068 	spin_unlock_irq(&uncore->lock);
1069 
1070 	return val;
1071 }
1072 
1073 /* NB: please notice the memset */
1074 void intel_engine_get_instdone(const struct intel_engine_cs *engine,
1075 			       struct intel_instdone *instdone)
1076 {
1077 	struct drm_i915_private *i915 = engine->i915;
1078 	const struct sseu_dev_info *sseu = &engine->gt->info.sseu;
1079 	struct intel_uncore *uncore = engine->uncore;
1080 	u32 mmio_base = engine->mmio_base;
1081 	int slice;
1082 	int subslice;
1083 
1084 	memset(instdone, 0, sizeof(*instdone));
1085 
1086 	switch (INTEL_GEN(i915)) {
1087 	default:
1088 		instdone->instdone =
1089 			intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
1090 
1091 		if (engine->id != RCS0)
1092 			break;
1093 
1094 		instdone->slice_common =
1095 			intel_uncore_read(uncore, GEN7_SC_INSTDONE);
1096 		if (INTEL_GEN(i915) >= 12) {
1097 			instdone->slice_common_extra[0] =
1098 				intel_uncore_read(uncore, GEN12_SC_INSTDONE_EXTRA);
1099 			instdone->slice_common_extra[1] =
1100 				intel_uncore_read(uncore, GEN12_SC_INSTDONE_EXTRA2);
1101 		}
1102 		for_each_instdone_slice_subslice(i915, sseu, slice, subslice) {
1103 			instdone->sampler[slice][subslice] =
1104 				read_subslice_reg(engine, slice, subslice,
1105 						  GEN7_SAMPLER_INSTDONE);
1106 			instdone->row[slice][subslice] =
1107 				read_subslice_reg(engine, slice, subslice,
1108 						  GEN7_ROW_INSTDONE);
1109 		}
1110 		break;
1111 	case 7:
1112 		instdone->instdone =
1113 			intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
1114 
1115 		if (engine->id != RCS0)
1116 			break;
1117 
1118 		instdone->slice_common =
1119 			intel_uncore_read(uncore, GEN7_SC_INSTDONE);
1120 		instdone->sampler[0][0] =
1121 			intel_uncore_read(uncore, GEN7_SAMPLER_INSTDONE);
1122 		instdone->row[0][0] =
1123 			intel_uncore_read(uncore, GEN7_ROW_INSTDONE);
1124 
1125 		break;
1126 	case 6:
1127 	case 5:
1128 	case 4:
1129 		instdone->instdone =
1130 			intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
1131 		if (engine->id == RCS0)
1132 			/* HACK: Using the wrong struct member */
1133 			instdone->slice_common =
1134 				intel_uncore_read(uncore, GEN4_INSTDONE1);
1135 		break;
1136 	case 3:
1137 	case 2:
1138 		instdone->instdone = intel_uncore_read(uncore, GEN2_INSTDONE);
1139 		break;
1140 	}
1141 }
1142 
1143 static bool ring_is_idle(struct intel_engine_cs *engine)
1144 {
1145 	bool idle = true;
1146 
1147 	if (I915_SELFTEST_ONLY(!engine->mmio_base))
1148 		return true;
1149 
1150 	if (!intel_engine_pm_get_if_awake(engine))
1151 		return true;
1152 
1153 	/* First check that no commands are left in the ring */
1154 	if ((ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR) !=
1155 	    (ENGINE_READ(engine, RING_TAIL) & TAIL_ADDR))
1156 		idle = false;
1157 
1158 	/* No bit for gen2, so assume the CS parser is idle */
1159 	if (INTEL_GEN(engine->i915) > 2 &&
1160 	    !(ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE))
1161 		idle = false;
1162 
1163 	intel_engine_pm_put(engine);
1164 
1165 	return idle;
1166 }
1167 
1168 void intel_engine_flush_submission(struct intel_engine_cs *engine)
1169 {
1170 	struct tasklet_struct *t = &engine->execlists.tasklet;
1171 
1172 	if (!t->func)
1173 		return;
1174 
1175 	/* Synchronise and wait for the tasklet on another CPU */
1176 	tasklet_kill(t);
1177 
1178 	/* Having cancelled the tasklet, ensure that is run */
1179 	local_bh_disable();
1180 	if (tasklet_trylock(t)) {
1181 		/* Must wait for any GPU reset in progress. */
1182 		if (__tasklet_is_enabled(t))
1183 			t->func(t->data);
1184 		tasklet_unlock(t);
1185 	}
1186 	local_bh_enable();
1187 }
1188 
1189 /**
1190  * intel_engine_is_idle() - Report if the engine has finished process all work
1191  * @engine: the intel_engine_cs
1192  *
1193  * Return true if there are no requests pending, nothing left to be submitted
1194  * to hardware, and that the engine is idle.
1195  */
1196 bool intel_engine_is_idle(struct intel_engine_cs *engine)
1197 {
1198 	/* More white lies, if wedged, hw state is inconsistent */
1199 	if (intel_gt_is_wedged(engine->gt))
1200 		return true;
1201 
1202 	if (!intel_engine_pm_is_awake(engine))
1203 		return true;
1204 
1205 	/* Waiting to drain ELSP? */
1206 	if (execlists_active(&engine->execlists)) {
1207 		synchronize_hardirq(engine->i915->drm.pdev->irq);
1208 
1209 		intel_engine_flush_submission(engine);
1210 
1211 		if (execlists_active(&engine->execlists))
1212 			return false;
1213 	}
1214 
1215 	/* ELSP is empty, but there are ready requests? E.g. after reset */
1216 	if (!RB_EMPTY_ROOT(&engine->execlists.queue.rb_root))
1217 		return false;
1218 
1219 	/* Ring stopped? */
1220 	return ring_is_idle(engine);
1221 }
1222 
1223 bool intel_engines_are_idle(struct intel_gt *gt)
1224 {
1225 	struct intel_engine_cs *engine;
1226 	enum intel_engine_id id;
1227 
1228 	/*
1229 	 * If the driver is wedged, HW state may be very inconsistent and
1230 	 * report that it is still busy, even though we have stopped using it.
1231 	 */
1232 	if (intel_gt_is_wedged(gt))
1233 		return true;
1234 
1235 	/* Already parked (and passed an idleness test); must still be idle */
1236 	if (!READ_ONCE(gt->awake))
1237 		return true;
1238 
1239 	for_each_engine(engine, gt, id) {
1240 		if (!intel_engine_is_idle(engine))
1241 			return false;
1242 	}
1243 
1244 	return true;
1245 }
1246 
1247 void intel_engines_reset_default_submission(struct intel_gt *gt)
1248 {
1249 	struct intel_engine_cs *engine;
1250 	enum intel_engine_id id;
1251 
1252 	for_each_engine(engine, gt, id)
1253 		engine->set_default_submission(engine);
1254 }
1255 
1256 bool intel_engine_can_store_dword(struct intel_engine_cs *engine)
1257 {
1258 	switch (INTEL_GEN(engine->i915)) {
1259 	case 2:
1260 		return false; /* uses physical not virtual addresses */
1261 	case 3:
1262 		/* maybe only uses physical not virtual addresses */
1263 		return !(IS_I915G(engine->i915) || IS_I915GM(engine->i915));
1264 	case 4:
1265 		return !IS_I965G(engine->i915); /* who knows! */
1266 	case 6:
1267 		return engine->class != VIDEO_DECODE_CLASS; /* b0rked */
1268 	default:
1269 		return true;
1270 	}
1271 }
1272 
1273 static int print_sched_attr(const struct i915_sched_attr *attr,
1274 			    char *buf, int x, int len)
1275 {
1276 	if (attr->priority == I915_PRIORITY_INVALID)
1277 		return x;
1278 
1279 	x += snprintf(buf + x, len - x,
1280 		      " prio=%d", attr->priority);
1281 
1282 	return x;
1283 }
1284 
1285 static void print_request(struct drm_printer *m,
1286 			  struct i915_request *rq,
1287 			  const char *prefix)
1288 {
1289 	const char *name = rq->fence.ops->get_timeline_name(&rq->fence);
1290 	char buf[80] = "";
1291 	int x = 0;
1292 
1293 	x = print_sched_attr(&rq->sched.attr, buf, x, sizeof(buf));
1294 
1295 	drm_printf(m, "%s %llx:%llx%s%s %s @ %dms: %s\n",
1296 		   prefix,
1297 		   rq->fence.context, rq->fence.seqno,
1298 		   i915_request_completed(rq) ? "!" :
1299 		   i915_request_started(rq) ? "*" :
1300 		   "",
1301 		   test_bit(DMA_FENCE_FLAG_SIGNALED_BIT,
1302 			    &rq->fence.flags) ? "+" :
1303 		   test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT,
1304 			    &rq->fence.flags) ? "-" :
1305 		   "",
1306 		   buf,
1307 		   jiffies_to_msecs(jiffies - rq->emitted_jiffies),
1308 		   name);
1309 }
1310 
1311 static struct intel_timeline *get_timeline(struct i915_request *rq)
1312 {
1313 	struct intel_timeline *tl;
1314 
1315 	/*
1316 	 * Even though we are holding the engine->active.lock here, there
1317 	 * is no control over the submission queue per-se and we are
1318 	 * inspecting the active state at a random point in time, with an
1319 	 * unknown queue. Play safe and make sure the timeline remains valid.
1320 	 * (Only being used for pretty printing, one extra kref shouldn't
1321 	 * cause a camel stampede!)
1322 	 */
1323 	rcu_read_lock();
1324 	tl = rcu_dereference(rq->timeline);
1325 	if (!kref_get_unless_zero(&tl->kref))
1326 		tl = NULL;
1327 	rcu_read_unlock();
1328 
1329 	return tl;
1330 }
1331 
1332 static int print_ring(char *buf, int sz, struct i915_request *rq)
1333 {
1334 	int len = 0;
1335 
1336 	if (!i915_request_signaled(rq)) {
1337 		struct intel_timeline *tl = get_timeline(rq);
1338 
1339 		len = scnprintf(buf, sz,
1340 				"ring:{start:%08x, hwsp:%08x, seqno:%08x, runtime:%llums}, ",
1341 				i915_ggtt_offset(rq->ring->vma),
1342 				tl ? tl->hwsp_offset : 0,
1343 				hwsp_seqno(rq),
1344 				DIV_ROUND_CLOSEST_ULL(intel_context_get_total_runtime_ns(rq->context),
1345 						      1000 * 1000));
1346 
1347 		if (tl)
1348 			intel_timeline_put(tl);
1349 	}
1350 
1351 	return len;
1352 }
1353 
1354 static void hexdump(struct drm_printer *m, const void *buf, size_t len)
1355 {
1356 	const size_t rowsize = 8 * sizeof(u32);
1357 	const void *prev = NULL;
1358 	bool skip = false;
1359 	size_t pos;
1360 
1361 	for (pos = 0; pos < len; pos += rowsize) {
1362 		char line[128];
1363 
1364 		if (prev && !memcmp(prev, buf + pos, rowsize)) {
1365 			if (!skip) {
1366 				drm_printf(m, "*\n");
1367 				skip = true;
1368 			}
1369 			continue;
1370 		}
1371 
1372 		WARN_ON_ONCE(hex_dump_to_buffer(buf + pos, len - pos,
1373 						rowsize, sizeof(u32),
1374 						line, sizeof(line),
1375 						false) >= sizeof(line));
1376 		drm_printf(m, "[%04zx] %s\n", pos, line);
1377 
1378 		prev = buf + pos;
1379 		skip = false;
1380 	}
1381 }
1382 
1383 static const char *repr_timer(const struct timer_list *t)
1384 {
1385 	if (!READ_ONCE(t->expires))
1386 		return "inactive";
1387 
1388 	if (timer_pending(t))
1389 		return "active";
1390 
1391 	return "expired";
1392 }
1393 
1394 static void intel_engine_print_registers(struct intel_engine_cs *engine,
1395 					 struct drm_printer *m)
1396 {
1397 	struct drm_i915_private *dev_priv = engine->i915;
1398 	struct intel_engine_execlists * const execlists = &engine->execlists;
1399 	u64 addr;
1400 
1401 	if (engine->id == RENDER_CLASS && IS_GEN_RANGE(dev_priv, 4, 7))
1402 		drm_printf(m, "\tCCID: 0x%08x\n", ENGINE_READ(engine, CCID));
1403 	if (HAS_EXECLISTS(dev_priv)) {
1404 		drm_printf(m, "\tEL_STAT_HI: 0x%08x\n",
1405 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_HI));
1406 		drm_printf(m, "\tEL_STAT_LO: 0x%08x\n",
1407 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_LO));
1408 	}
1409 	drm_printf(m, "\tRING_START: 0x%08x\n",
1410 		   ENGINE_READ(engine, RING_START));
1411 	drm_printf(m, "\tRING_HEAD:  0x%08x\n",
1412 		   ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR);
1413 	drm_printf(m, "\tRING_TAIL:  0x%08x\n",
1414 		   ENGINE_READ(engine, RING_TAIL) & TAIL_ADDR);
1415 	drm_printf(m, "\tRING_CTL:   0x%08x%s\n",
1416 		   ENGINE_READ(engine, RING_CTL),
1417 		   ENGINE_READ(engine, RING_CTL) & (RING_WAIT | RING_WAIT_SEMAPHORE) ? " [waiting]" : "");
1418 	if (INTEL_GEN(engine->i915) > 2) {
1419 		drm_printf(m, "\tRING_MODE:  0x%08x%s\n",
1420 			   ENGINE_READ(engine, RING_MI_MODE),
1421 			   ENGINE_READ(engine, RING_MI_MODE) & (MODE_IDLE) ? " [idle]" : "");
1422 	}
1423 
1424 	if (INTEL_GEN(dev_priv) >= 6) {
1425 		drm_printf(m, "\tRING_IMR:   0x%08x\n",
1426 			   ENGINE_READ(engine, RING_IMR));
1427 		drm_printf(m, "\tRING_ESR:   0x%08x\n",
1428 			   ENGINE_READ(engine, RING_ESR));
1429 		drm_printf(m, "\tRING_EMR:   0x%08x\n",
1430 			   ENGINE_READ(engine, RING_EMR));
1431 		drm_printf(m, "\tRING_EIR:   0x%08x\n",
1432 			   ENGINE_READ(engine, RING_EIR));
1433 	}
1434 
1435 	addr = intel_engine_get_active_head(engine);
1436 	drm_printf(m, "\tACTHD:  0x%08x_%08x\n",
1437 		   upper_32_bits(addr), lower_32_bits(addr));
1438 	addr = intel_engine_get_last_batch_head(engine);
1439 	drm_printf(m, "\tBBADDR: 0x%08x_%08x\n",
1440 		   upper_32_bits(addr), lower_32_bits(addr));
1441 	if (INTEL_GEN(dev_priv) >= 8)
1442 		addr = ENGINE_READ64(engine, RING_DMA_FADD, RING_DMA_FADD_UDW);
1443 	else if (INTEL_GEN(dev_priv) >= 4)
1444 		addr = ENGINE_READ(engine, RING_DMA_FADD);
1445 	else
1446 		addr = ENGINE_READ(engine, DMA_FADD_I8XX);
1447 	drm_printf(m, "\tDMA_FADDR: 0x%08x_%08x\n",
1448 		   upper_32_bits(addr), lower_32_bits(addr));
1449 	if (INTEL_GEN(dev_priv) >= 4) {
1450 		drm_printf(m, "\tIPEIR: 0x%08x\n",
1451 			   ENGINE_READ(engine, RING_IPEIR));
1452 		drm_printf(m, "\tIPEHR: 0x%08x\n",
1453 			   ENGINE_READ(engine, RING_IPEHR));
1454 	} else {
1455 		drm_printf(m, "\tIPEIR: 0x%08x\n", ENGINE_READ(engine, IPEIR));
1456 		drm_printf(m, "\tIPEHR: 0x%08x\n", ENGINE_READ(engine, IPEHR));
1457 	}
1458 
1459 	if (HAS_EXECLISTS(dev_priv)) {
1460 		struct i915_request * const *port, *rq;
1461 		const u32 *hws =
1462 			&engine->status_page.addr[I915_HWS_CSB_BUF0_INDEX];
1463 		const u8 num_entries = execlists->csb_size;
1464 		unsigned int idx;
1465 		u8 read, write;
1466 
1467 		drm_printf(m, "\tExeclist tasklet queued? %s (%s), preempt? %s, timeslice? %s\n",
1468 			   yesno(test_bit(TASKLET_STATE_SCHED,
1469 					  &engine->execlists.tasklet.state)),
1470 			   enableddisabled(!atomic_read(&engine->execlists.tasklet.count)),
1471 			   repr_timer(&engine->execlists.preempt),
1472 			   repr_timer(&engine->execlists.timer));
1473 
1474 		read = execlists->csb_head;
1475 		write = READ_ONCE(*execlists->csb_write);
1476 
1477 		drm_printf(m, "\tExeclist status: 0x%08x %08x; CSB read:%d, write:%d, entries:%d\n",
1478 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_LO),
1479 			   ENGINE_READ(engine, RING_EXECLIST_STATUS_HI),
1480 			   read, write, num_entries);
1481 
1482 		if (read >= num_entries)
1483 			read = 0;
1484 		if (write >= num_entries)
1485 			write = 0;
1486 		if (read > write)
1487 			write += num_entries;
1488 		while (read < write) {
1489 			idx = ++read % num_entries;
1490 			drm_printf(m, "\tExeclist CSB[%d]: 0x%08x, context: %d\n",
1491 				   idx, hws[idx * 2], hws[idx * 2 + 1]);
1492 		}
1493 
1494 		execlists_active_lock_bh(execlists);
1495 		rcu_read_lock();
1496 		for (port = execlists->active; (rq = *port); port++) {
1497 			char hdr[160];
1498 			int len;
1499 
1500 			len = scnprintf(hdr, sizeof(hdr),
1501 					"\t\tActive[%d]:  ccid:%08x%s%s, ",
1502 					(int)(port - execlists->active),
1503 					rq->context->lrc.ccid,
1504 					intel_context_is_closed(rq->context) ? "!" : "",
1505 					intel_context_is_banned(rq->context) ? "*" : "");
1506 			len += print_ring(hdr + len, sizeof(hdr) - len, rq);
1507 			scnprintf(hdr + len, sizeof(hdr) - len, "rq: ");
1508 			print_request(m, rq, hdr);
1509 		}
1510 		for (port = execlists->pending; (rq = *port); port++) {
1511 			char hdr[160];
1512 			int len;
1513 
1514 			len = scnprintf(hdr, sizeof(hdr),
1515 					"\t\tPending[%d]: ccid:%08x%s%s, ",
1516 					(int)(port - execlists->pending),
1517 					rq->context->lrc.ccid,
1518 					intel_context_is_closed(rq->context) ? "!" : "",
1519 					intel_context_is_banned(rq->context) ? "*" : "");
1520 			len += print_ring(hdr + len, sizeof(hdr) - len, rq);
1521 			scnprintf(hdr + len, sizeof(hdr) - len, "rq: ");
1522 			print_request(m, rq, hdr);
1523 		}
1524 		rcu_read_unlock();
1525 		execlists_active_unlock_bh(execlists);
1526 	} else if (INTEL_GEN(dev_priv) > 6) {
1527 		drm_printf(m, "\tPP_DIR_BASE: 0x%08x\n",
1528 			   ENGINE_READ(engine, RING_PP_DIR_BASE));
1529 		drm_printf(m, "\tPP_DIR_BASE_READ: 0x%08x\n",
1530 			   ENGINE_READ(engine, RING_PP_DIR_BASE_READ));
1531 		drm_printf(m, "\tPP_DIR_DCLV: 0x%08x\n",
1532 			   ENGINE_READ(engine, RING_PP_DIR_DCLV));
1533 	}
1534 }
1535 
1536 static void print_request_ring(struct drm_printer *m, struct i915_request *rq)
1537 {
1538 	void *ring;
1539 	int size;
1540 
1541 	drm_printf(m,
1542 		   "[head %04x, postfix %04x, tail %04x, batch 0x%08x_%08x]:\n",
1543 		   rq->head, rq->postfix, rq->tail,
1544 		   rq->batch ? upper_32_bits(rq->batch->node.start) : ~0u,
1545 		   rq->batch ? lower_32_bits(rq->batch->node.start) : ~0u);
1546 
1547 	size = rq->tail - rq->head;
1548 	if (rq->tail < rq->head)
1549 		size += rq->ring->size;
1550 
1551 	ring = kmalloc(size, GFP_ATOMIC);
1552 	if (ring) {
1553 		const void *vaddr = rq->ring->vaddr;
1554 		unsigned int head = rq->head;
1555 		unsigned int len = 0;
1556 
1557 		if (rq->tail < head) {
1558 			len = rq->ring->size - head;
1559 			memcpy(ring, vaddr + head, len);
1560 			head = 0;
1561 		}
1562 		memcpy(ring + len, vaddr + head, size - len);
1563 
1564 		hexdump(m, ring, size);
1565 		kfree(ring);
1566 	}
1567 }
1568 
1569 static unsigned long list_count(struct list_head *list)
1570 {
1571 	struct list_head *pos;
1572 	unsigned long count = 0;
1573 
1574 	list_for_each(pos, list)
1575 		count++;
1576 
1577 	return count;
1578 }
1579 
1580 void intel_engine_dump(struct intel_engine_cs *engine,
1581 		       struct drm_printer *m,
1582 		       const char *header, ...)
1583 {
1584 	struct i915_gpu_error * const error = &engine->i915->gpu_error;
1585 	struct i915_request *rq;
1586 	intel_wakeref_t wakeref;
1587 	unsigned long flags;
1588 	ktime_t dummy;
1589 
1590 	if (header) {
1591 		va_list ap;
1592 
1593 		va_start(ap, header);
1594 		drm_vprintf(m, header, &ap);
1595 		va_end(ap);
1596 	}
1597 
1598 	if (intel_gt_is_wedged(engine->gt))
1599 		drm_printf(m, "*** WEDGED ***\n");
1600 
1601 	drm_printf(m, "\tAwake? %d\n", atomic_read(&engine->wakeref.count));
1602 	drm_printf(m, "\tBarriers?: %s\n",
1603 		   yesno(!llist_empty(&engine->barrier_tasks)));
1604 	drm_printf(m, "\tLatency: %luus\n",
1605 		   ewma__engine_latency_read(&engine->latency));
1606 	if (intel_engine_supports_stats(engine))
1607 		drm_printf(m, "\tRuntime: %llums\n",
1608 			   ktime_to_ms(intel_engine_get_busy_time(engine,
1609 								  &dummy)));
1610 	drm_printf(m, "\tForcewake: %x domains, %d active\n",
1611 		   engine->fw_domain, atomic_read(&engine->fw_active));
1612 
1613 	rcu_read_lock();
1614 	rq = READ_ONCE(engine->heartbeat.systole);
1615 	if (rq)
1616 		drm_printf(m, "\tHeartbeat: %d ms ago\n",
1617 			   jiffies_to_msecs(jiffies - rq->emitted_jiffies));
1618 	rcu_read_unlock();
1619 	drm_printf(m, "\tReset count: %d (global %d)\n",
1620 		   i915_reset_engine_count(error, engine),
1621 		   i915_reset_count(error));
1622 
1623 	drm_printf(m, "\tRequests:\n");
1624 
1625 	spin_lock_irqsave(&engine->active.lock, flags);
1626 	rq = intel_engine_find_active_request(engine);
1627 	if (rq) {
1628 		struct intel_timeline *tl = get_timeline(rq);
1629 
1630 		print_request(m, rq, "\t\tactive ");
1631 
1632 		drm_printf(m, "\t\tring->start:  0x%08x\n",
1633 			   i915_ggtt_offset(rq->ring->vma));
1634 		drm_printf(m, "\t\tring->head:   0x%08x\n",
1635 			   rq->ring->head);
1636 		drm_printf(m, "\t\tring->tail:   0x%08x\n",
1637 			   rq->ring->tail);
1638 		drm_printf(m, "\t\tring->emit:   0x%08x\n",
1639 			   rq->ring->emit);
1640 		drm_printf(m, "\t\tring->space:  0x%08x\n",
1641 			   rq->ring->space);
1642 
1643 		if (tl) {
1644 			drm_printf(m, "\t\tring->hwsp:   0x%08x\n",
1645 				   tl->hwsp_offset);
1646 			intel_timeline_put(tl);
1647 		}
1648 
1649 		print_request_ring(m, rq);
1650 
1651 		if (rq->context->lrc_reg_state) {
1652 			drm_printf(m, "Logical Ring Context:\n");
1653 			hexdump(m, rq->context->lrc_reg_state, PAGE_SIZE);
1654 		}
1655 	}
1656 	drm_printf(m, "\tOn hold?: %lu\n", list_count(&engine->active.hold));
1657 	spin_unlock_irqrestore(&engine->active.lock, flags);
1658 
1659 	drm_printf(m, "\tMMIO base:  0x%08x\n", engine->mmio_base);
1660 	wakeref = intel_runtime_pm_get_if_in_use(engine->uncore->rpm);
1661 	if (wakeref) {
1662 		intel_engine_print_registers(engine, m);
1663 		intel_runtime_pm_put(engine->uncore->rpm, wakeref);
1664 	} else {
1665 		drm_printf(m, "\tDevice is asleep; skipping register dump\n");
1666 	}
1667 
1668 	intel_execlists_show_requests(engine, m, print_request, 8);
1669 
1670 	drm_printf(m, "HWSP:\n");
1671 	hexdump(m, engine->status_page.addr, PAGE_SIZE);
1672 
1673 	drm_printf(m, "Idle? %s\n", yesno(intel_engine_is_idle(engine)));
1674 
1675 	intel_engine_print_breadcrumbs(engine, m);
1676 }
1677 
1678 static ktime_t __intel_engine_get_busy_time(struct intel_engine_cs *engine,
1679 					    ktime_t *now)
1680 {
1681 	ktime_t total = engine->stats.total;
1682 
1683 	/*
1684 	 * If the engine is executing something at the moment
1685 	 * add it to the total.
1686 	 */
1687 	*now = ktime_get();
1688 	if (atomic_read(&engine->stats.active))
1689 		total = ktime_add(total, ktime_sub(*now, engine->stats.start));
1690 
1691 	return total;
1692 }
1693 
1694 /**
1695  * intel_engine_get_busy_time() - Return current accumulated engine busyness
1696  * @engine: engine to report on
1697  * @now: monotonic timestamp of sampling
1698  *
1699  * Returns accumulated time @engine was busy since engine stats were enabled.
1700  */
1701 ktime_t intel_engine_get_busy_time(struct intel_engine_cs *engine, ktime_t *now)
1702 {
1703 	unsigned int seq;
1704 	ktime_t total;
1705 
1706 	do {
1707 		seq = read_seqbegin(&engine->stats.lock);
1708 		total = __intel_engine_get_busy_time(engine, now);
1709 	} while (read_seqretry(&engine->stats.lock, seq));
1710 
1711 	return total;
1712 }
1713 
1714 static bool match_ring(struct i915_request *rq)
1715 {
1716 	u32 ring = ENGINE_READ(rq->engine, RING_START);
1717 
1718 	return ring == i915_ggtt_offset(rq->ring->vma);
1719 }
1720 
1721 struct i915_request *
1722 intel_engine_find_active_request(struct intel_engine_cs *engine)
1723 {
1724 	struct i915_request *request, *active = NULL;
1725 
1726 	/*
1727 	 * We are called by the error capture, reset and to dump engine
1728 	 * state at random points in time. In particular, note that neither is
1729 	 * crucially ordered with an interrupt. After a hang, the GPU is dead
1730 	 * and we assume that no more writes can happen (we waited long enough
1731 	 * for all writes that were in transaction to be flushed) - adding an
1732 	 * extra delay for a recent interrupt is pointless. Hence, we do
1733 	 * not need an engine->irq_seqno_barrier() before the seqno reads.
1734 	 * At all other times, we must assume the GPU is still running, but
1735 	 * we only care about the snapshot of this moment.
1736 	 */
1737 	lockdep_assert_held(&engine->active.lock);
1738 
1739 	rcu_read_lock();
1740 	request = execlists_active(&engine->execlists);
1741 	if (request) {
1742 		struct intel_timeline *tl = request->context->timeline;
1743 
1744 		list_for_each_entry_from_reverse(request, &tl->requests, link) {
1745 			if (i915_request_completed(request))
1746 				break;
1747 
1748 			active = request;
1749 		}
1750 	}
1751 	rcu_read_unlock();
1752 	if (active)
1753 		return active;
1754 
1755 	list_for_each_entry(request, &engine->active.requests, sched.link) {
1756 		if (i915_request_completed(request))
1757 			continue;
1758 
1759 		if (!i915_request_started(request))
1760 			continue;
1761 
1762 		/* More than one preemptible request may match! */
1763 		if (!match_ring(request))
1764 			continue;
1765 
1766 		active = request;
1767 		break;
1768 	}
1769 
1770 	return active;
1771 }
1772 
1773 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1774 #include "mock_engine.c"
1775 #include "selftest_engine.c"
1776 #include "selftest_engine_cs.c"
1777 #endif
1778