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 "gt/intel_gt.h"
32
33 #include "intel_engine.h"
34 #include "intel_engine_pm.h"
35 #include "intel_engine_pool.h"
36 #include "intel_engine_user.h"
37 #include "intel_context.h"
38 #include "intel_lrc.h"
39 #include "intel_reset.h"
40
41 /* Haswell does have the CXT_SIZE register however it does not appear to be
42 * valid. Now, docs explain in dwords what is in the context object. The full
43 * size is 70720 bytes, however, the power context and execlist context will
44 * never be saved (power context is stored elsewhere, and execlists don't work
45 * on HSW) - so the final size, including the extra state required for the
46 * Resource Streamer, is 66944 bytes, which rounds to 17 pages.
47 */
48 #define HSW_CXT_TOTAL_SIZE (17 * PAGE_SIZE)
49
50 #define DEFAULT_LR_CONTEXT_RENDER_SIZE (22 * PAGE_SIZE)
51 #define GEN8_LR_CONTEXT_RENDER_SIZE (20 * PAGE_SIZE)
52 #define GEN9_LR_CONTEXT_RENDER_SIZE (22 * PAGE_SIZE)
53 #define GEN10_LR_CONTEXT_RENDER_SIZE (18 * PAGE_SIZE)
54 #define GEN11_LR_CONTEXT_RENDER_SIZE (14 * PAGE_SIZE)
55
56 #define GEN8_LR_CONTEXT_OTHER_SIZE ( 2 * PAGE_SIZE)
57
58 #define MAX_MMIO_BASES 3
59 struct engine_info {
60 unsigned int hw_id;
61 u8 class;
62 u8 instance;
63 /* mmio bases table *must* be sorted in reverse gen order */
64 struct engine_mmio_base {
65 u32 gen : 8;
66 u32 base : 24;
67 } mmio_bases[MAX_MMIO_BASES];
68 };
69
70 static const struct engine_info intel_engines[] = {
71 [RCS0] = {
72 .hw_id = RCS0_HW,
73 .class = RENDER_CLASS,
74 .instance = 0,
75 .mmio_bases = {
76 { .gen = 1, .base = RENDER_RING_BASE }
77 },
78 },
79 [BCS0] = {
80 .hw_id = BCS0_HW,
81 .class = COPY_ENGINE_CLASS,
82 .instance = 0,
83 .mmio_bases = {
84 { .gen = 6, .base = BLT_RING_BASE }
85 },
86 },
87 [VCS0] = {
88 .hw_id = VCS0_HW,
89 .class = VIDEO_DECODE_CLASS,
90 .instance = 0,
91 .mmio_bases = {
92 { .gen = 11, .base = GEN11_BSD_RING_BASE },
93 { .gen = 6, .base = GEN6_BSD_RING_BASE },
94 { .gen = 4, .base = BSD_RING_BASE }
95 },
96 },
97 [VCS1] = {
98 .hw_id = VCS1_HW,
99 .class = VIDEO_DECODE_CLASS,
100 .instance = 1,
101 .mmio_bases = {
102 { .gen = 11, .base = GEN11_BSD2_RING_BASE },
103 { .gen = 8, .base = GEN8_BSD2_RING_BASE }
104 },
105 },
106 [VCS2] = {
107 .hw_id = VCS2_HW,
108 .class = VIDEO_DECODE_CLASS,
109 .instance = 2,
110 .mmio_bases = {
111 { .gen = 11, .base = GEN11_BSD3_RING_BASE }
112 },
113 },
114 [VCS3] = {
115 .hw_id = VCS3_HW,
116 .class = VIDEO_DECODE_CLASS,
117 .instance = 3,
118 .mmio_bases = {
119 { .gen = 11, .base = GEN11_BSD4_RING_BASE }
120 },
121 },
122 [VECS0] = {
123 .hw_id = VECS0_HW,
124 .class = VIDEO_ENHANCEMENT_CLASS,
125 .instance = 0,
126 .mmio_bases = {
127 { .gen = 11, .base = GEN11_VEBOX_RING_BASE },
128 { .gen = 7, .base = VEBOX_RING_BASE }
129 },
130 },
131 [VECS1] = {
132 .hw_id = VECS1_HW,
133 .class = VIDEO_ENHANCEMENT_CLASS,
134 .instance = 1,
135 .mmio_bases = {
136 { .gen = 11, .base = GEN11_VEBOX2_RING_BASE }
137 },
138 },
139 };
140
141 /**
142 * intel_engine_context_size() - return the size of the context for an engine
143 * @dev_priv: i915 device private
144 * @class: engine class
145 *
146 * Each engine class may require a different amount of space for a context
147 * image.
148 *
149 * Return: size (in bytes) of an engine class specific context image
150 *
151 * Note: this size includes the HWSP, which is part of the context image
152 * in LRC mode, but does not include the "shared data page" used with
153 * GuC submission. The caller should account for this if using the GuC.
154 */
intel_engine_context_size(struct drm_i915_private * dev_priv,u8 class)155 u32 intel_engine_context_size(struct drm_i915_private *dev_priv, u8 class)
156 {
157 u32 cxt_size;
158
159 BUILD_BUG_ON(I915_GTT_PAGE_SIZE != PAGE_SIZE);
160
161 switch (class) {
162 case RENDER_CLASS:
163 switch (INTEL_GEN(dev_priv)) {
164 default:
165 MISSING_CASE(INTEL_GEN(dev_priv));
166 return DEFAULT_LR_CONTEXT_RENDER_SIZE;
167 case 12:
168 case 11:
169 return GEN11_LR_CONTEXT_RENDER_SIZE;
170 case 10:
171 return GEN10_LR_CONTEXT_RENDER_SIZE;
172 case 9:
173 return GEN9_LR_CONTEXT_RENDER_SIZE;
174 case 8:
175 return GEN8_LR_CONTEXT_RENDER_SIZE;
176 case 7:
177 if (IS_HASWELL(dev_priv))
178 return HSW_CXT_TOTAL_SIZE;
179
180 cxt_size = I915_READ(GEN7_CXT_SIZE);
181 return round_up(GEN7_CXT_TOTAL_SIZE(cxt_size) * 64,
182 PAGE_SIZE);
183 case 6:
184 cxt_size = I915_READ(CXT_SIZE);
185 return round_up(GEN6_CXT_TOTAL_SIZE(cxt_size) * 64,
186 PAGE_SIZE);
187 case 5:
188 case 4:
189 /*
190 * There is a discrepancy here between the size reported
191 * by the register and the size of the context layout
192 * in the docs. Both are described as authorative!
193 *
194 * The discrepancy is on the order of a few cachelines,
195 * but the total is under one page (4k), which is our
196 * minimum allocation anyway so it should all come
197 * out in the wash.
198 */
199 cxt_size = I915_READ(CXT_SIZE) + 1;
200 DRM_DEBUG_DRIVER("gen%d CXT_SIZE = %d bytes [0x%08x]\n",
201 INTEL_GEN(dev_priv),
202 cxt_size * 64,
203 cxt_size - 1);
204 return round_up(cxt_size * 64, PAGE_SIZE);
205 case 3:
206 case 2:
207 /* For the special day when i810 gets merged. */
208 case 1:
209 return 0;
210 }
211 break;
212 default:
213 MISSING_CASE(class);
214 /* fall through */
215 case VIDEO_DECODE_CLASS:
216 case VIDEO_ENHANCEMENT_CLASS:
217 case COPY_ENGINE_CLASS:
218 if (INTEL_GEN(dev_priv) < 8)
219 return 0;
220 return GEN8_LR_CONTEXT_OTHER_SIZE;
221 }
222 }
223
__engine_mmio_base(struct drm_i915_private * i915,const struct engine_mmio_base * bases)224 static u32 __engine_mmio_base(struct drm_i915_private *i915,
225 const struct engine_mmio_base *bases)
226 {
227 int i;
228
229 for (i = 0; i < MAX_MMIO_BASES; i++)
230 if (INTEL_GEN(i915) >= bases[i].gen)
231 break;
232
233 GEM_BUG_ON(i == MAX_MMIO_BASES);
234 GEM_BUG_ON(!bases[i].base);
235
236 return bases[i].base;
237 }
238
__sprint_engine_name(struct intel_engine_cs * engine)239 static void __sprint_engine_name(struct intel_engine_cs *engine)
240 {
241 /*
242 * Before we know what the uABI name for this engine will be,
243 * we still would like to keep track of this engine in the debug logs.
244 * We throw in a ' here as a reminder that this isn't its final name.
245 */
246 GEM_WARN_ON(snprintf(engine->name, sizeof(engine->name), "%s'%u",
247 intel_engine_class_repr(engine->class),
248 engine->instance) >= sizeof(engine->name));
249 }
250
intel_engine_set_hwsp_writemask(struct intel_engine_cs * engine,u32 mask)251 void intel_engine_set_hwsp_writemask(struct intel_engine_cs *engine, u32 mask)
252 {
253 /*
254 * Though they added more rings on g4x/ilk, they did not add
255 * per-engine HWSTAM until gen6.
256 */
257 if (INTEL_GEN(engine->i915) < 6 && engine->class != RENDER_CLASS)
258 return;
259
260 if (INTEL_GEN(engine->i915) >= 3)
261 ENGINE_WRITE(engine, RING_HWSTAM, mask);
262 else
263 ENGINE_WRITE16(engine, RING_HWSTAM, mask);
264 }
265
intel_engine_sanitize_mmio(struct intel_engine_cs * engine)266 static void intel_engine_sanitize_mmio(struct intel_engine_cs *engine)
267 {
268 /* Mask off all writes into the unknown HWSP */
269 intel_engine_set_hwsp_writemask(engine, ~0u);
270 }
271
intel_engine_setup(struct intel_gt * gt,enum intel_engine_id id)272 static int intel_engine_setup(struct intel_gt *gt, enum intel_engine_id id)
273 {
274 const struct engine_info *info = &intel_engines[id];
275 struct intel_engine_cs *engine;
276
277 BUILD_BUG_ON(MAX_ENGINE_CLASS >= BIT(GEN11_ENGINE_CLASS_WIDTH));
278 BUILD_BUG_ON(MAX_ENGINE_INSTANCE >= BIT(GEN11_ENGINE_INSTANCE_WIDTH));
279
280 if (GEM_DEBUG_WARN_ON(info->class > MAX_ENGINE_CLASS))
281 return -EINVAL;
282
283 if (GEM_DEBUG_WARN_ON(info->instance > MAX_ENGINE_INSTANCE))
284 return -EINVAL;
285
286 if (GEM_DEBUG_WARN_ON(gt->engine_class[info->class][info->instance]))
287 return -EINVAL;
288
289 engine = kzalloc(sizeof(*engine), GFP_KERNEL);
290 if (!engine)
291 return -ENOMEM;
292
293 BUILD_BUG_ON(BITS_PER_TYPE(engine->mask) < I915_NUM_ENGINES);
294
295 engine->id = id;
296 engine->mask = BIT(id);
297 engine->i915 = gt->i915;
298 engine->gt = gt;
299 engine->uncore = gt->uncore;
300 engine->hw_id = engine->guc_id = info->hw_id;
301 engine->mmio_base = __engine_mmio_base(gt->i915, info->mmio_bases);
302
303 engine->class = info->class;
304 engine->instance = info->instance;
305 __sprint_engine_name(engine);
306
307 /*
308 * To be overridden by the backend on setup. However to facilitate
309 * cleanup on error during setup, we always provide the destroy vfunc.
310 */
311 engine->destroy = (typeof(engine->destroy))kfree;
312
313 engine->context_size = intel_engine_context_size(gt->i915,
314 engine->class);
315 if (WARN_ON(engine->context_size > BIT(20)))
316 engine->context_size = 0;
317 if (engine->context_size)
318 DRIVER_CAPS(gt->i915)->has_logical_contexts = true;
319
320 /* Nothing to do here, execute in order of dependencies */
321 engine->schedule = NULL;
322
323 seqlock_init(&engine->stats.lock);
324
325 ATOMIC_INIT_NOTIFIER_HEAD(&engine->context_status_notifier);
326
327 /* Scrub mmio state on takeover */
328 intel_engine_sanitize_mmio(engine);
329
330 gt->engine_class[info->class][info->instance] = engine;
331
332 intel_engine_add_user(engine);
333 gt->i915->engine[id] = engine;
334
335 return 0;
336 }
337
__setup_engine_capabilities(struct intel_engine_cs * engine)338 static void __setup_engine_capabilities(struct intel_engine_cs *engine)
339 {
340 struct drm_i915_private *i915 = engine->i915;
341
342 if (engine->class == VIDEO_DECODE_CLASS) {
343 /*
344 * HEVC support is present on first engine instance
345 * before Gen11 and on all instances afterwards.
346 */
347 if (INTEL_GEN(i915) >= 11 ||
348 (INTEL_GEN(i915) >= 9 && engine->instance == 0))
349 engine->uabi_capabilities |=
350 I915_VIDEO_CLASS_CAPABILITY_HEVC;
351
352 /*
353 * SFC block is present only on even logical engine
354 * instances.
355 */
356 if ((INTEL_GEN(i915) >= 11 &&
357 RUNTIME_INFO(i915)->vdbox_sfc_access & engine->mask) ||
358 (INTEL_GEN(i915) >= 9 && engine->instance == 0))
359 engine->uabi_capabilities |=
360 I915_VIDEO_AND_ENHANCE_CLASS_CAPABILITY_SFC;
361 } else if (engine->class == VIDEO_ENHANCEMENT_CLASS) {
362 if (INTEL_GEN(i915) >= 9)
363 engine->uabi_capabilities |=
364 I915_VIDEO_AND_ENHANCE_CLASS_CAPABILITY_SFC;
365 }
366 }
367
intel_setup_engine_capabilities(struct drm_i915_private * i915)368 static void intel_setup_engine_capabilities(struct drm_i915_private *i915)
369 {
370 struct intel_engine_cs *engine;
371 enum intel_engine_id id;
372
373 for_each_engine(engine, i915, id)
374 __setup_engine_capabilities(engine);
375 }
376
377 /**
378 * intel_engines_cleanup() - free the resources allocated for Command Streamers
379 * @i915: the i915 devic
380 */
intel_engines_cleanup(struct drm_i915_private * i915)381 void intel_engines_cleanup(struct drm_i915_private *i915)
382 {
383 struct intel_engine_cs *engine;
384 enum intel_engine_id id;
385
386 for_each_engine(engine, i915, id) {
387 engine->destroy(engine);
388 i915->engine[id] = NULL;
389 }
390 }
391
392 /**
393 * intel_engines_init_mmio() - allocate and prepare the Engine Command Streamers
394 * @i915: the i915 device
395 *
396 * Return: non-zero if the initialization failed.
397 */
intel_engines_init_mmio(struct drm_i915_private * i915)398 int intel_engines_init_mmio(struct drm_i915_private *i915)
399 {
400 struct intel_device_info *device_info = mkwrite_device_info(i915);
401 const unsigned int engine_mask = INTEL_INFO(i915)->engine_mask;
402 unsigned int mask = 0;
403 unsigned int i;
404 int err;
405
406 WARN_ON(engine_mask == 0);
407 WARN_ON(engine_mask &
408 GENMASK(BITS_PER_TYPE(mask) - 1, I915_NUM_ENGINES));
409
410 if (i915_inject_probe_failure(i915))
411 return -ENODEV;
412
413 for (i = 0; i < ARRAY_SIZE(intel_engines); i++) {
414 if (!HAS_ENGINE(i915, i))
415 continue;
416
417 err = intel_engine_setup(&i915->gt, i);
418 if (err)
419 goto cleanup;
420
421 mask |= BIT(i);
422 }
423
424 /*
425 * Catch failures to update intel_engines table when the new engines
426 * are added to the driver by a warning and disabling the forgotten
427 * engines.
428 */
429 if (WARN_ON(mask != engine_mask))
430 device_info->engine_mask = mask;
431
432 RUNTIME_INFO(i915)->num_engines = hweight32(mask);
433
434 intel_gt_check_and_clear_faults(&i915->gt);
435
436 intel_setup_engine_capabilities(i915);
437
438 return 0;
439
440 cleanup:
441 intel_engines_cleanup(i915);
442 return err;
443 }
444
445 /**
446 * intel_engines_init() - init the Engine Command Streamers
447 * @i915: i915 device private
448 *
449 * Return: non-zero if the initialization failed.
450 */
intel_engines_init(struct drm_i915_private * i915)451 int intel_engines_init(struct drm_i915_private *i915)
452 {
453 int (*init)(struct intel_engine_cs *engine);
454 struct intel_engine_cs *engine;
455 enum intel_engine_id id;
456 int err;
457
458 if (HAS_EXECLISTS(i915))
459 init = intel_execlists_submission_init;
460 else
461 init = intel_ring_submission_init;
462
463 for_each_engine(engine, i915, id) {
464 err = init(engine);
465 if (err)
466 goto cleanup;
467 }
468
469 return 0;
470
471 cleanup:
472 intel_engines_cleanup(i915);
473 return err;
474 }
475
intel_engine_init_execlists(struct intel_engine_cs * engine)476 void intel_engine_init_execlists(struct intel_engine_cs *engine)
477 {
478 struct intel_engine_execlists * const execlists = &engine->execlists;
479
480 execlists->port_mask = 1;
481 GEM_BUG_ON(!is_power_of_2(execlists_num_ports(execlists)));
482 GEM_BUG_ON(execlists_num_ports(execlists) > EXECLIST_MAX_PORTS);
483
484 memset(execlists->pending, 0, sizeof(execlists->pending));
485 execlists->active =
486 memset(execlists->inflight, 0, sizeof(execlists->inflight));
487
488 execlists->queue_priority_hint = INT_MIN;
489 execlists->queue = RB_ROOT_CACHED;
490 }
491
cleanup_status_page(struct intel_engine_cs * engine)492 static void cleanup_status_page(struct intel_engine_cs *engine)
493 {
494 struct i915_vma *vma;
495
496 /* Prevent writes into HWSP after returning the page to the system */
497 intel_engine_set_hwsp_writemask(engine, ~0u);
498
499 vma = fetch_and_zero(&engine->status_page.vma);
500 if (!vma)
501 return;
502
503 if (!HWS_NEEDS_PHYSICAL(engine->i915))
504 i915_vma_unpin(vma);
505
506 i915_gem_object_unpin_map(vma->obj);
507 i915_gem_object_put(vma->obj);
508 }
509
pin_ggtt_status_page(struct intel_engine_cs * engine,struct i915_vma * vma)510 static int pin_ggtt_status_page(struct intel_engine_cs *engine,
511 struct i915_vma *vma)
512 {
513 unsigned int flags;
514
515 flags = PIN_GLOBAL;
516 if (!HAS_LLC(engine->i915))
517 /*
518 * On g33, we cannot place HWS above 256MiB, so
519 * restrict its pinning to the low mappable arena.
520 * Though this restriction is not documented for
521 * gen4, gen5, or byt, they also behave similarly
522 * and hang if the HWS is placed at the top of the
523 * GTT. To generalise, it appears that all !llc
524 * platforms have issues with us placing the HWS
525 * above the mappable region (even though we never
526 * actually map it).
527 */
528 flags |= PIN_MAPPABLE;
529 else
530 flags |= PIN_HIGH;
531
532 return i915_vma_pin(vma, 0, 0, flags);
533 }
534
init_status_page(struct intel_engine_cs * engine)535 static int init_status_page(struct intel_engine_cs *engine)
536 {
537 struct drm_i915_gem_object *obj;
538 struct i915_vma *vma;
539 void *vaddr;
540 int ret;
541
542 /*
543 * Though the HWS register does support 36bit addresses, historically
544 * we have had hangs and corruption reported due to wild writes if
545 * the HWS is placed above 4G. We only allow objects to be allocated
546 * in GFP_DMA32 for i965, and no earlier physical address users had
547 * access to more than 4G.
548 */
549 obj = i915_gem_object_create_internal(engine->i915, PAGE_SIZE);
550 if (IS_ERR(obj)) {
551 DRM_ERROR("Failed to allocate status page\n");
552 return PTR_ERR(obj);
553 }
554
555 i915_gem_object_set_cache_coherency(obj, I915_CACHE_LLC);
556
557 vma = i915_vma_instance(obj, &engine->gt->ggtt->vm, NULL);
558 if (IS_ERR(vma)) {
559 ret = PTR_ERR(vma);
560 goto err;
561 }
562
563 vaddr = i915_gem_object_pin_map(obj, I915_MAP_WB);
564 if (IS_ERR(vaddr)) {
565 ret = PTR_ERR(vaddr);
566 goto err;
567 }
568
569 engine->status_page.addr = memset(vaddr, 0, PAGE_SIZE);
570 engine->status_page.vma = vma;
571
572 if (!HWS_NEEDS_PHYSICAL(engine->i915)) {
573 ret = pin_ggtt_status_page(engine, vma);
574 if (ret)
575 goto err_unpin;
576 }
577
578 return 0;
579
580 err_unpin:
581 i915_gem_object_unpin_map(obj);
582 err:
583 i915_gem_object_put(obj);
584 return ret;
585 }
586
intel_engine_setup_common(struct intel_engine_cs * engine)587 static int intel_engine_setup_common(struct intel_engine_cs *engine)
588 {
589 int err;
590
591 init_llist_head(&engine->barrier_tasks);
592
593 err = init_status_page(engine);
594 if (err)
595 return err;
596
597 intel_engine_init_active(engine, ENGINE_PHYSICAL);
598 intel_engine_init_breadcrumbs(engine);
599 intel_engine_init_execlists(engine);
600 intel_engine_init_hangcheck(engine);
601 intel_engine_init_cmd_parser(engine);
602 intel_engine_init__pm(engine);
603
604 intel_engine_pool_init(&engine->pool);
605
606 /* Use the whole device by default */
607 engine->sseu =
608 intel_sseu_from_device_info(&RUNTIME_INFO(engine->i915)->sseu);
609
610 intel_engine_init_workarounds(engine);
611 intel_engine_init_whitelist(engine);
612 intel_engine_init_ctx_wa(engine);
613
614 return 0;
615 }
616
617 /**
618 * intel_engines_setup- setup engine state not requiring hw access
619 * @i915: Device to setup.
620 *
621 * Initializes engine structure members shared between legacy and execlists
622 * submission modes which do not require hardware access.
623 *
624 * Typically done early in the submission mode specific engine setup stage.
625 */
intel_engines_setup(struct drm_i915_private * i915)626 int intel_engines_setup(struct drm_i915_private *i915)
627 {
628 int (*setup)(struct intel_engine_cs *engine);
629 struct intel_engine_cs *engine;
630 enum intel_engine_id id;
631 int err;
632
633 if (HAS_EXECLISTS(i915))
634 setup = intel_execlists_submission_setup;
635 else
636 setup = intel_ring_submission_setup;
637
638 for_each_engine(engine, i915, id) {
639 err = intel_engine_setup_common(engine);
640 if (err)
641 goto cleanup;
642
643 err = setup(engine);
644 if (err)
645 goto cleanup;
646
647 /* We expect the backend to take control over its state */
648 GEM_BUG_ON(engine->destroy == (typeof(engine->destroy))kfree);
649
650 GEM_BUG_ON(!engine->cops);
651 }
652
653 return 0;
654
655 cleanup:
656 intel_engines_cleanup(i915);
657 return err;
658 }
659
660 struct measure_breadcrumb {
661 struct i915_request rq;
662 struct intel_timeline timeline;
663 struct intel_ring ring;
664 u32 cs[1024];
665 };
666
measure_breadcrumb_dw(struct intel_engine_cs * engine)667 static int measure_breadcrumb_dw(struct intel_engine_cs *engine)
668 {
669 struct measure_breadcrumb *frame;
670 int dw = -ENOMEM;
671
672 GEM_BUG_ON(!engine->gt->scratch);
673
674 frame = kzalloc(sizeof(*frame), GFP_KERNEL);
675 if (!frame)
676 return -ENOMEM;
677
678 if (intel_timeline_init(&frame->timeline,
679 engine->gt,
680 engine->status_page.vma))
681 goto out_frame;
682
683 frame->ring.vaddr = frame->cs;
684 frame->ring.size = sizeof(frame->cs);
685 frame->ring.effective_size = frame->ring.size;
686 intel_ring_update_space(&frame->ring);
687
688 frame->rq.i915 = engine->i915;
689 frame->rq.engine = engine;
690 frame->rq.ring = &frame->ring;
691 frame->rq.timeline = &frame->timeline;
692
693 dw = intel_timeline_pin(&frame->timeline);
694 if (dw < 0)
695 goto out_timeline;
696
697 dw = engine->emit_fini_breadcrumb(&frame->rq, frame->cs) - frame->cs;
698 GEM_BUG_ON(dw & 1); /* RING_TAIL must be qword aligned */
699
700 intel_timeline_unpin(&frame->timeline);
701
702 out_timeline:
703 intel_timeline_fini(&frame->timeline);
704 out_frame:
705 kfree(frame);
706 return dw;
707 }
708
709 void
intel_engine_init_active(struct intel_engine_cs * engine,unsigned int subclass)710 intel_engine_init_active(struct intel_engine_cs *engine, unsigned int subclass)
711 {
712 INIT_LIST_HEAD(&engine->active.requests);
713
714 spin_lock_init(&engine->active.lock);
715 lockdep_set_subclass(&engine->active.lock, subclass);
716
717 /*
718 * Due to an interesting quirk in lockdep's internal debug tracking,
719 * after setting a subclass we must ensure the lock is used. Otherwise,
720 * nr_unused_locks is incremented once too often.
721 */
722 #ifdef CONFIG_DEBUG_LOCK_ALLOC
723 local_irq_disable();
724 lock_map_acquire(&engine->active.lock.dep_map);
725 lock_map_release(&engine->active.lock.dep_map);
726 local_irq_enable();
727 #endif
728 }
729
730 static struct intel_context *
create_kernel_context(struct intel_engine_cs * engine)731 create_kernel_context(struct intel_engine_cs *engine)
732 {
733 struct intel_context *ce;
734 int err;
735
736 ce = intel_context_create(engine->i915->kernel_context, engine);
737 if (IS_ERR(ce))
738 return ce;
739
740 ce->ring = __intel_context_ring_size(SZ_4K);
741
742 err = intel_context_pin(ce);
743 if (err) {
744 intel_context_put(ce);
745 return ERR_PTR(err);
746 }
747
748 return ce;
749 }
750
751 /**
752 * intel_engines_init_common - initialize cengine state which might require hw access
753 * @engine: Engine to initialize.
754 *
755 * Initializes @engine@ structure members shared between legacy and execlists
756 * submission modes which do require hardware access.
757 *
758 * Typcally done at later stages of submission mode specific engine setup.
759 *
760 * Returns zero on success or an error code on failure.
761 */
intel_engine_init_common(struct intel_engine_cs * engine)762 int intel_engine_init_common(struct intel_engine_cs *engine)
763 {
764 struct intel_context *ce;
765 int ret;
766
767 engine->set_default_submission(engine);
768
769 /*
770 * We may need to do things with the shrinker which
771 * require us to immediately switch back to the default
772 * context. This can cause a problem as pinning the
773 * default context also requires GTT space which may not
774 * be available. To avoid this we always pin the default
775 * context.
776 */
777 ce = create_kernel_context(engine);
778 if (IS_ERR(ce))
779 return PTR_ERR(ce);
780
781 engine->kernel_context = ce;
782
783 ret = measure_breadcrumb_dw(engine);
784 if (ret < 0)
785 goto err_unpin;
786
787 engine->emit_fini_breadcrumb_dw = ret;
788
789 return 0;
790
791 err_unpin:
792 intel_context_unpin(ce);
793 intel_context_put(ce);
794 return ret;
795 }
796
797 /**
798 * intel_engines_cleanup_common - cleans up the engine state created by
799 * the common initiailizers.
800 * @engine: Engine to cleanup.
801 *
802 * This cleans up everything created by the common helpers.
803 */
intel_engine_cleanup_common(struct intel_engine_cs * engine)804 void intel_engine_cleanup_common(struct intel_engine_cs *engine)
805 {
806 GEM_BUG_ON(!list_empty(&engine->active.requests));
807
808 cleanup_status_page(engine);
809
810 intel_engine_pool_fini(&engine->pool);
811 intel_engine_fini_breadcrumbs(engine);
812 intel_engine_cleanup_cmd_parser(engine);
813
814 if (engine->default_state)
815 i915_gem_object_put(engine->default_state);
816
817 intel_context_unpin(engine->kernel_context);
818 intel_context_put(engine->kernel_context);
819 GEM_BUG_ON(!llist_empty(&engine->barrier_tasks));
820
821 intel_wa_list_free(&engine->ctx_wa_list);
822 intel_wa_list_free(&engine->wa_list);
823 intel_wa_list_free(&engine->whitelist);
824 }
825
intel_engine_get_active_head(const struct intel_engine_cs * engine)826 u64 intel_engine_get_active_head(const struct intel_engine_cs *engine)
827 {
828 struct drm_i915_private *i915 = engine->i915;
829
830 u64 acthd;
831
832 if (INTEL_GEN(i915) >= 8)
833 acthd = ENGINE_READ64(engine, RING_ACTHD, RING_ACTHD_UDW);
834 else if (INTEL_GEN(i915) >= 4)
835 acthd = ENGINE_READ(engine, RING_ACTHD);
836 else
837 acthd = ENGINE_READ(engine, ACTHD);
838
839 return acthd;
840 }
841
intel_engine_get_last_batch_head(const struct intel_engine_cs * engine)842 u64 intel_engine_get_last_batch_head(const struct intel_engine_cs *engine)
843 {
844 u64 bbaddr;
845
846 if (INTEL_GEN(engine->i915) >= 8)
847 bbaddr = ENGINE_READ64(engine, RING_BBADDR, RING_BBADDR_UDW);
848 else
849 bbaddr = ENGINE_READ(engine, RING_BBADDR);
850
851 return bbaddr;
852 }
853
intel_engine_stop_cs(struct intel_engine_cs * engine)854 int intel_engine_stop_cs(struct intel_engine_cs *engine)
855 {
856 struct intel_uncore *uncore = engine->uncore;
857 const u32 base = engine->mmio_base;
858 const i915_reg_t mode = RING_MI_MODE(base);
859 int err;
860
861 if (INTEL_GEN(engine->i915) < 3)
862 return -ENODEV;
863
864 GEM_TRACE("%s\n", engine->name);
865
866 intel_uncore_write_fw(uncore, mode, _MASKED_BIT_ENABLE(STOP_RING));
867
868 err = 0;
869 if (__intel_wait_for_register_fw(uncore,
870 mode, MODE_IDLE, MODE_IDLE,
871 1000, 0,
872 NULL)) {
873 GEM_TRACE("%s: timed out on STOP_RING -> IDLE\n", engine->name);
874 err = -ETIMEDOUT;
875 }
876
877 /* A final mmio read to let GPU writes be hopefully flushed to memory */
878 intel_uncore_posting_read_fw(uncore, mode);
879
880 return err;
881 }
882
intel_engine_cancel_stop_cs(struct intel_engine_cs * engine)883 void intel_engine_cancel_stop_cs(struct intel_engine_cs *engine)
884 {
885 GEM_TRACE("%s\n", engine->name);
886
887 ENGINE_WRITE_FW(engine, RING_MI_MODE, _MASKED_BIT_DISABLE(STOP_RING));
888 }
889
i915_cache_level_str(struct drm_i915_private * i915,int type)890 const char *i915_cache_level_str(struct drm_i915_private *i915, int type)
891 {
892 switch (type) {
893 case I915_CACHE_NONE: return " uncached";
894 case I915_CACHE_LLC: return HAS_LLC(i915) ? " LLC" : " snooped";
895 case I915_CACHE_L3_LLC: return " L3+LLC";
896 case I915_CACHE_WT: return " WT";
897 default: return "";
898 }
899 }
900
901 static u32
read_subslice_reg(struct intel_engine_cs * engine,int slice,int subslice,i915_reg_t reg)902 read_subslice_reg(struct intel_engine_cs *engine, int slice, int subslice,
903 i915_reg_t reg)
904 {
905 struct drm_i915_private *i915 = engine->i915;
906 struct intel_uncore *uncore = engine->uncore;
907 u32 mcr_mask, mcr_ss, mcr, old_mcr, val;
908 enum forcewake_domains fw_domains;
909
910 if (INTEL_GEN(i915) >= 11) {
911 mcr_mask = GEN11_MCR_SLICE_MASK | GEN11_MCR_SUBSLICE_MASK;
912 mcr_ss = GEN11_MCR_SLICE(slice) | GEN11_MCR_SUBSLICE(subslice);
913 } else {
914 mcr_mask = GEN8_MCR_SLICE_MASK | GEN8_MCR_SUBSLICE_MASK;
915 mcr_ss = GEN8_MCR_SLICE(slice) | GEN8_MCR_SUBSLICE(subslice);
916 }
917
918 fw_domains = intel_uncore_forcewake_for_reg(uncore, reg,
919 FW_REG_READ);
920 fw_domains |= intel_uncore_forcewake_for_reg(uncore,
921 GEN8_MCR_SELECTOR,
922 FW_REG_READ | FW_REG_WRITE);
923
924 spin_lock_irq(&uncore->lock);
925 intel_uncore_forcewake_get__locked(uncore, fw_domains);
926
927 old_mcr = mcr = intel_uncore_read_fw(uncore, GEN8_MCR_SELECTOR);
928
929 mcr &= ~mcr_mask;
930 mcr |= mcr_ss;
931 intel_uncore_write_fw(uncore, GEN8_MCR_SELECTOR, mcr);
932
933 val = intel_uncore_read_fw(uncore, reg);
934
935 mcr &= ~mcr_mask;
936 mcr |= old_mcr & mcr_mask;
937
938 intel_uncore_write_fw(uncore, GEN8_MCR_SELECTOR, mcr);
939
940 intel_uncore_forcewake_put__locked(uncore, fw_domains);
941 spin_unlock_irq(&uncore->lock);
942
943 return val;
944 }
945
946 /* NB: please notice the memset */
intel_engine_get_instdone(struct intel_engine_cs * engine,struct intel_instdone * instdone)947 void intel_engine_get_instdone(struct intel_engine_cs *engine,
948 struct intel_instdone *instdone)
949 {
950 struct drm_i915_private *i915 = engine->i915;
951 struct intel_uncore *uncore = engine->uncore;
952 u32 mmio_base = engine->mmio_base;
953 int slice;
954 int subslice;
955
956 memset(instdone, 0, sizeof(*instdone));
957
958 switch (INTEL_GEN(i915)) {
959 default:
960 instdone->instdone =
961 intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
962
963 if (engine->id != RCS0)
964 break;
965
966 instdone->slice_common =
967 intel_uncore_read(uncore, GEN7_SC_INSTDONE);
968 for_each_instdone_slice_subslice(i915, slice, subslice) {
969 instdone->sampler[slice][subslice] =
970 read_subslice_reg(engine, slice, subslice,
971 GEN7_SAMPLER_INSTDONE);
972 instdone->row[slice][subslice] =
973 read_subslice_reg(engine, slice, subslice,
974 GEN7_ROW_INSTDONE);
975 }
976 break;
977 case 7:
978 instdone->instdone =
979 intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
980
981 if (engine->id != RCS0)
982 break;
983
984 instdone->slice_common =
985 intel_uncore_read(uncore, GEN7_SC_INSTDONE);
986 instdone->sampler[0][0] =
987 intel_uncore_read(uncore, GEN7_SAMPLER_INSTDONE);
988 instdone->row[0][0] =
989 intel_uncore_read(uncore, GEN7_ROW_INSTDONE);
990
991 break;
992 case 6:
993 case 5:
994 case 4:
995 instdone->instdone =
996 intel_uncore_read(uncore, RING_INSTDONE(mmio_base));
997 if (engine->id == RCS0)
998 /* HACK: Using the wrong struct member */
999 instdone->slice_common =
1000 intel_uncore_read(uncore, GEN4_INSTDONE1);
1001 break;
1002 case 3:
1003 case 2:
1004 instdone->instdone = intel_uncore_read(uncore, GEN2_INSTDONE);
1005 break;
1006 }
1007 }
1008
ring_is_idle(struct intel_engine_cs * engine)1009 static bool ring_is_idle(struct intel_engine_cs *engine)
1010 {
1011 bool idle = true;
1012
1013 if (I915_SELFTEST_ONLY(!engine->mmio_base))
1014 return true;
1015
1016 if (!intel_engine_pm_get_if_awake(engine))
1017 return true;
1018
1019 /* First check that no commands are left in the ring */
1020 if ((ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR) !=
1021 (ENGINE_READ(engine, RING_TAIL) & TAIL_ADDR))
1022 idle = false;
1023
1024 /* No bit for gen2, so assume the CS parser is idle */
1025 if (INTEL_GEN(engine->i915) > 2 &&
1026 !(ENGINE_READ(engine, RING_MI_MODE) & MODE_IDLE))
1027 idle = false;
1028
1029 intel_engine_pm_put(engine);
1030
1031 return idle;
1032 }
1033
1034 /**
1035 * intel_engine_is_idle() - Report if the engine has finished process all work
1036 * @engine: the intel_engine_cs
1037 *
1038 * Return true if there are no requests pending, nothing left to be submitted
1039 * to hardware, and that the engine is idle.
1040 */
intel_engine_is_idle(struct intel_engine_cs * engine)1041 bool intel_engine_is_idle(struct intel_engine_cs *engine)
1042 {
1043 /* More white lies, if wedged, hw state is inconsistent */
1044 if (intel_gt_is_wedged(engine->gt))
1045 return true;
1046
1047 if (!intel_engine_pm_is_awake(engine))
1048 return true;
1049
1050 /* Waiting to drain ELSP? */
1051 if (execlists_active(&engine->execlists)) {
1052 struct tasklet_struct *t = &engine->execlists.tasklet;
1053
1054 synchronize_hardirq(engine->i915->drm.pdev->irq);
1055
1056 local_bh_disable();
1057 if (tasklet_trylock(t)) {
1058 /* Must wait for any GPU reset in progress. */
1059 if (__tasklet_is_enabled(t))
1060 t->func(t->data);
1061 tasklet_unlock(t);
1062 }
1063 local_bh_enable();
1064
1065 /* Otherwise flush the tasklet if it was on another cpu */
1066 tasklet_unlock_wait(t);
1067
1068 if (execlists_active(&engine->execlists))
1069 return false;
1070 }
1071
1072 /* ELSP is empty, but there are ready requests? E.g. after reset */
1073 if (!RB_EMPTY_ROOT(&engine->execlists.queue.rb_root))
1074 return false;
1075
1076 /* Ring stopped? */
1077 return ring_is_idle(engine);
1078 }
1079
intel_engines_are_idle(struct intel_gt * gt)1080 bool intel_engines_are_idle(struct intel_gt *gt)
1081 {
1082 struct intel_engine_cs *engine;
1083 enum intel_engine_id id;
1084
1085 /*
1086 * If the driver is wedged, HW state may be very inconsistent and
1087 * report that it is still busy, even though we have stopped using it.
1088 */
1089 if (intel_gt_is_wedged(gt))
1090 return true;
1091
1092 /* Already parked (and passed an idleness test); must still be idle */
1093 if (!READ_ONCE(gt->awake))
1094 return true;
1095
1096 for_each_engine(engine, gt->i915, id) {
1097 if (!intel_engine_is_idle(engine))
1098 return false;
1099 }
1100
1101 return true;
1102 }
1103
intel_engines_reset_default_submission(struct intel_gt * gt)1104 void intel_engines_reset_default_submission(struct intel_gt *gt)
1105 {
1106 struct intel_engine_cs *engine;
1107 enum intel_engine_id id;
1108
1109 for_each_engine(engine, gt->i915, id)
1110 engine->set_default_submission(engine);
1111 }
1112
intel_engine_can_store_dword(struct intel_engine_cs * engine)1113 bool intel_engine_can_store_dword(struct intel_engine_cs *engine)
1114 {
1115 switch (INTEL_GEN(engine->i915)) {
1116 case 2:
1117 return false; /* uses physical not virtual addresses */
1118 case 3:
1119 /* maybe only uses physical not virtual addresses */
1120 return !(IS_I915G(engine->i915) || IS_I915GM(engine->i915));
1121 case 6:
1122 return engine->class != VIDEO_DECODE_CLASS; /* b0rked */
1123 default:
1124 return true;
1125 }
1126 }
1127
print_sched_attr(struct drm_i915_private * i915,const struct i915_sched_attr * attr,char * buf,int x,int len)1128 static int print_sched_attr(struct drm_i915_private *i915,
1129 const struct i915_sched_attr *attr,
1130 char *buf, int x, int len)
1131 {
1132 if (attr->priority == I915_PRIORITY_INVALID)
1133 return x;
1134
1135 x += snprintf(buf + x, len - x,
1136 " prio=%d", attr->priority);
1137
1138 return x;
1139 }
1140
print_request(struct drm_printer * m,struct i915_request * rq,const char * prefix)1141 static void print_request(struct drm_printer *m,
1142 struct i915_request *rq,
1143 const char *prefix)
1144 {
1145 const char *name = rq->fence.ops->get_timeline_name(&rq->fence);
1146 char buf[80] = "";
1147 int x = 0;
1148
1149 x = print_sched_attr(rq->i915, &rq->sched.attr, buf, x, sizeof(buf));
1150
1151 drm_printf(m, "%s %llx:%llx%s%s %s @ %dms: %s\n",
1152 prefix,
1153 rq->fence.context, rq->fence.seqno,
1154 i915_request_completed(rq) ? "!" :
1155 i915_request_started(rq) ? "*" :
1156 "",
1157 test_bit(DMA_FENCE_FLAG_SIGNALED_BIT,
1158 &rq->fence.flags) ? "+" :
1159 test_bit(DMA_FENCE_FLAG_ENABLE_SIGNAL_BIT,
1160 &rq->fence.flags) ? "-" :
1161 "",
1162 buf,
1163 jiffies_to_msecs(jiffies - rq->emitted_jiffies),
1164 name);
1165 }
1166
hexdump(struct drm_printer * m,const void * buf,size_t len)1167 static void hexdump(struct drm_printer *m, const void *buf, size_t len)
1168 {
1169 const size_t rowsize = 8 * sizeof(u32);
1170 const void *prev = NULL;
1171 bool skip = false;
1172 size_t pos;
1173
1174 for (pos = 0; pos < len; pos += rowsize) {
1175 char line[128];
1176
1177 if (prev && !memcmp(prev, buf + pos, rowsize)) {
1178 if (!skip) {
1179 drm_printf(m, "*\n");
1180 skip = true;
1181 }
1182 continue;
1183 }
1184
1185 WARN_ON_ONCE(hex_dump_to_buffer(buf + pos, len - pos,
1186 rowsize, sizeof(u32),
1187 line, sizeof(line),
1188 false) >= sizeof(line));
1189 drm_printf(m, "[%04zx] %s\n", pos, line);
1190
1191 prev = buf + pos;
1192 skip = false;
1193 }
1194 }
1195
intel_engine_print_registers(struct intel_engine_cs * engine,struct drm_printer * m)1196 static void intel_engine_print_registers(struct intel_engine_cs *engine,
1197 struct drm_printer *m)
1198 {
1199 struct drm_i915_private *dev_priv = engine->i915;
1200 struct intel_engine_execlists * const execlists = &engine->execlists;
1201 u64 addr;
1202
1203 if (engine->id == RENDER_CLASS && IS_GEN_RANGE(dev_priv, 4, 7))
1204 drm_printf(m, "\tCCID: 0x%08x\n", ENGINE_READ(engine, CCID));
1205 drm_printf(m, "\tRING_START: 0x%08x\n",
1206 ENGINE_READ(engine, RING_START));
1207 drm_printf(m, "\tRING_HEAD: 0x%08x\n",
1208 ENGINE_READ(engine, RING_HEAD) & HEAD_ADDR);
1209 drm_printf(m, "\tRING_TAIL: 0x%08x\n",
1210 ENGINE_READ(engine, RING_TAIL) & TAIL_ADDR);
1211 drm_printf(m, "\tRING_CTL: 0x%08x%s\n",
1212 ENGINE_READ(engine, RING_CTL),
1213 ENGINE_READ(engine, RING_CTL) & (RING_WAIT | RING_WAIT_SEMAPHORE) ? " [waiting]" : "");
1214 if (INTEL_GEN(engine->i915) > 2) {
1215 drm_printf(m, "\tRING_MODE: 0x%08x%s\n",
1216 ENGINE_READ(engine, RING_MI_MODE),
1217 ENGINE_READ(engine, RING_MI_MODE) & (MODE_IDLE) ? " [idle]" : "");
1218 }
1219
1220 if (INTEL_GEN(dev_priv) >= 6) {
1221 drm_printf(m, "\tRING_IMR: %08x\n",
1222 ENGINE_READ(engine, RING_IMR));
1223 }
1224
1225 addr = intel_engine_get_active_head(engine);
1226 drm_printf(m, "\tACTHD: 0x%08x_%08x\n",
1227 upper_32_bits(addr), lower_32_bits(addr));
1228 addr = intel_engine_get_last_batch_head(engine);
1229 drm_printf(m, "\tBBADDR: 0x%08x_%08x\n",
1230 upper_32_bits(addr), lower_32_bits(addr));
1231 if (INTEL_GEN(dev_priv) >= 8)
1232 addr = ENGINE_READ64(engine, RING_DMA_FADD, RING_DMA_FADD_UDW);
1233 else if (INTEL_GEN(dev_priv) >= 4)
1234 addr = ENGINE_READ(engine, RING_DMA_FADD);
1235 else
1236 addr = ENGINE_READ(engine, DMA_FADD_I8XX);
1237 drm_printf(m, "\tDMA_FADDR: 0x%08x_%08x\n",
1238 upper_32_bits(addr), lower_32_bits(addr));
1239 if (INTEL_GEN(dev_priv) >= 4) {
1240 drm_printf(m, "\tIPEIR: 0x%08x\n",
1241 ENGINE_READ(engine, RING_IPEIR));
1242 drm_printf(m, "\tIPEHR: 0x%08x\n",
1243 ENGINE_READ(engine, RING_IPEHR));
1244 } else {
1245 drm_printf(m, "\tIPEIR: 0x%08x\n", ENGINE_READ(engine, IPEIR));
1246 drm_printf(m, "\tIPEHR: 0x%08x\n", ENGINE_READ(engine, IPEHR));
1247 }
1248
1249 if (HAS_EXECLISTS(dev_priv)) {
1250 struct i915_request * const *port, *rq;
1251 const u32 *hws =
1252 &engine->status_page.addr[I915_HWS_CSB_BUF0_INDEX];
1253 const u8 num_entries = execlists->csb_size;
1254 unsigned int idx;
1255 u8 read, write;
1256
1257 drm_printf(m, "\tExeclist status: 0x%08x %08x, entries %u\n",
1258 ENGINE_READ(engine, RING_EXECLIST_STATUS_LO),
1259 ENGINE_READ(engine, RING_EXECLIST_STATUS_HI),
1260 num_entries);
1261
1262 read = execlists->csb_head;
1263 write = READ_ONCE(*execlists->csb_write);
1264
1265 drm_printf(m, "\tExeclist CSB read %d, write %d, tasklet queued? %s (%s)\n",
1266 read, write,
1267 yesno(test_bit(TASKLET_STATE_SCHED,
1268 &engine->execlists.tasklet.state)),
1269 enableddisabled(!atomic_read(&engine->execlists.tasklet.count)));
1270 if (read >= num_entries)
1271 read = 0;
1272 if (write >= num_entries)
1273 write = 0;
1274 if (read > write)
1275 write += num_entries;
1276 while (read < write) {
1277 idx = ++read % num_entries;
1278 drm_printf(m, "\tExeclist CSB[%d]: 0x%08x, context: %d\n",
1279 idx, hws[idx * 2], hws[idx * 2 + 1]);
1280 }
1281
1282 execlists_active_lock_bh(execlists);
1283 for (port = execlists->active; (rq = *port); port++) {
1284 char hdr[80];
1285 int len;
1286
1287 len = snprintf(hdr, sizeof(hdr),
1288 "\t\tActive[%d: ",
1289 (int)(port - execlists->active));
1290 if (!i915_request_signaled(rq))
1291 len += snprintf(hdr + len, sizeof(hdr) - len,
1292 "ring:{start:%08x, hwsp:%08x, seqno:%08x}, ",
1293 i915_ggtt_offset(rq->ring->vma),
1294 rq->timeline->hwsp_offset,
1295 hwsp_seqno(rq));
1296 snprintf(hdr + len, sizeof(hdr) - len, "rq: ");
1297 print_request(m, rq, hdr);
1298 }
1299 for (port = execlists->pending; (rq = *port); port++) {
1300 char hdr[80];
1301
1302 snprintf(hdr, sizeof(hdr),
1303 "\t\tPending[%d] ring:{start:%08x, hwsp:%08x, seqno:%08x}, rq: ",
1304 (int)(port - execlists->pending),
1305 i915_ggtt_offset(rq->ring->vma),
1306 rq->timeline->hwsp_offset,
1307 hwsp_seqno(rq));
1308 print_request(m, rq, hdr);
1309 }
1310 execlists_active_unlock_bh(execlists);
1311 } else if (INTEL_GEN(dev_priv) > 6) {
1312 drm_printf(m, "\tPP_DIR_BASE: 0x%08x\n",
1313 ENGINE_READ(engine, RING_PP_DIR_BASE));
1314 drm_printf(m, "\tPP_DIR_BASE_READ: 0x%08x\n",
1315 ENGINE_READ(engine, RING_PP_DIR_BASE_READ));
1316 drm_printf(m, "\tPP_DIR_DCLV: 0x%08x\n",
1317 ENGINE_READ(engine, RING_PP_DIR_DCLV));
1318 }
1319 }
1320
print_request_ring(struct drm_printer * m,struct i915_request * rq)1321 static void print_request_ring(struct drm_printer *m, struct i915_request *rq)
1322 {
1323 void *ring;
1324 int size;
1325
1326 drm_printf(m,
1327 "[head %04x, postfix %04x, tail %04x, batch 0x%08x_%08x]:\n",
1328 rq->head, rq->postfix, rq->tail,
1329 rq->batch ? upper_32_bits(rq->batch->node.start) : ~0u,
1330 rq->batch ? lower_32_bits(rq->batch->node.start) : ~0u);
1331
1332 size = rq->tail - rq->head;
1333 if (rq->tail < rq->head)
1334 size += rq->ring->size;
1335
1336 ring = kmalloc(size, GFP_ATOMIC);
1337 if (ring) {
1338 const void *vaddr = rq->ring->vaddr;
1339 unsigned int head = rq->head;
1340 unsigned int len = 0;
1341
1342 if (rq->tail < head) {
1343 len = rq->ring->size - head;
1344 memcpy(ring, vaddr + head, len);
1345 head = 0;
1346 }
1347 memcpy(ring + len, vaddr + head, size - len);
1348
1349 hexdump(m, ring, size);
1350 kfree(ring);
1351 }
1352 }
1353
intel_engine_dump(struct intel_engine_cs * engine,struct drm_printer * m,const char * header,...)1354 void intel_engine_dump(struct intel_engine_cs *engine,
1355 struct drm_printer *m,
1356 const char *header, ...)
1357 {
1358 struct i915_gpu_error * const error = &engine->i915->gpu_error;
1359 struct i915_request *rq;
1360 intel_wakeref_t wakeref;
1361 unsigned long flags;
1362
1363 if (header) {
1364 va_list ap;
1365
1366 va_start(ap, header);
1367 drm_vprintf(m, header, &ap);
1368 va_end(ap);
1369 }
1370
1371 if (intel_gt_is_wedged(engine->gt))
1372 drm_printf(m, "*** WEDGED ***\n");
1373
1374 drm_printf(m, "\tAwake? %d\n", atomic_read(&engine->wakeref.count));
1375 drm_printf(m, "\tHangcheck: %d ms ago\n",
1376 jiffies_to_msecs(jiffies - engine->hangcheck.action_timestamp));
1377 drm_printf(m, "\tReset count: %d (global %d)\n",
1378 i915_reset_engine_count(error, engine),
1379 i915_reset_count(error));
1380
1381 drm_printf(m, "\tRequests:\n");
1382
1383 spin_lock_irqsave(&engine->active.lock, flags);
1384 rq = intel_engine_find_active_request(engine);
1385 if (rq) {
1386 print_request(m, rq, "\t\tactive ");
1387
1388 drm_printf(m, "\t\tring->start: 0x%08x\n",
1389 i915_ggtt_offset(rq->ring->vma));
1390 drm_printf(m, "\t\tring->head: 0x%08x\n",
1391 rq->ring->head);
1392 drm_printf(m, "\t\tring->tail: 0x%08x\n",
1393 rq->ring->tail);
1394 drm_printf(m, "\t\tring->emit: 0x%08x\n",
1395 rq->ring->emit);
1396 drm_printf(m, "\t\tring->space: 0x%08x\n",
1397 rq->ring->space);
1398 drm_printf(m, "\t\tring->hwsp: 0x%08x\n",
1399 rq->timeline->hwsp_offset);
1400
1401 print_request_ring(m, rq);
1402 }
1403 spin_unlock_irqrestore(&engine->active.lock, flags);
1404
1405 drm_printf(m, "\tMMIO base: 0x%08x\n", engine->mmio_base);
1406 wakeref = intel_runtime_pm_get_if_in_use(&engine->i915->runtime_pm);
1407 if (wakeref) {
1408 intel_engine_print_registers(engine, m);
1409 intel_runtime_pm_put(&engine->i915->runtime_pm, wakeref);
1410 } else {
1411 drm_printf(m, "\tDevice is asleep; skipping register dump\n");
1412 }
1413
1414 intel_execlists_show_requests(engine, m, print_request, 8);
1415
1416 drm_printf(m, "HWSP:\n");
1417 hexdump(m, engine->status_page.addr, PAGE_SIZE);
1418
1419 drm_printf(m, "Idle? %s\n", yesno(intel_engine_is_idle(engine)));
1420
1421 intel_engine_print_breadcrumbs(engine, m);
1422 }
1423
1424 /**
1425 * intel_enable_engine_stats() - Enable engine busy tracking on engine
1426 * @engine: engine to enable stats collection
1427 *
1428 * Start collecting the engine busyness data for @engine.
1429 *
1430 * Returns 0 on success or a negative error code.
1431 */
intel_enable_engine_stats(struct intel_engine_cs * engine)1432 int intel_enable_engine_stats(struct intel_engine_cs *engine)
1433 {
1434 struct intel_engine_execlists *execlists = &engine->execlists;
1435 unsigned long flags;
1436 int err = 0;
1437
1438 if (!intel_engine_supports_stats(engine))
1439 return -ENODEV;
1440
1441 execlists_active_lock_bh(execlists);
1442 write_seqlock_irqsave(&engine->stats.lock, flags);
1443
1444 if (unlikely(engine->stats.enabled == ~0)) {
1445 err = -EBUSY;
1446 goto unlock;
1447 }
1448
1449 if (engine->stats.enabled++ == 0) {
1450 struct i915_request * const *port;
1451 struct i915_request *rq;
1452
1453 engine->stats.enabled_at = ktime_get();
1454
1455 /* XXX submission method oblivious? */
1456 for (port = execlists->active; (rq = *port); port++)
1457 engine->stats.active++;
1458
1459 for (port = execlists->pending; (rq = *port); port++) {
1460 /* Exclude any contexts already counted in active */
1461 if (!intel_context_inflight_count(rq->hw_context))
1462 engine->stats.active++;
1463 }
1464
1465 if (engine->stats.active)
1466 engine->stats.start = engine->stats.enabled_at;
1467 }
1468
1469 unlock:
1470 write_sequnlock_irqrestore(&engine->stats.lock, flags);
1471 execlists_active_unlock_bh(execlists);
1472
1473 return err;
1474 }
1475
__intel_engine_get_busy_time(struct intel_engine_cs * engine)1476 static ktime_t __intel_engine_get_busy_time(struct intel_engine_cs *engine)
1477 {
1478 ktime_t total = engine->stats.total;
1479
1480 /*
1481 * If the engine is executing something at the moment
1482 * add it to the total.
1483 */
1484 if (engine->stats.active)
1485 total = ktime_add(total,
1486 ktime_sub(ktime_get(), engine->stats.start));
1487
1488 return total;
1489 }
1490
1491 /**
1492 * intel_engine_get_busy_time() - Return current accumulated engine busyness
1493 * @engine: engine to report on
1494 *
1495 * Returns accumulated time @engine was busy since engine stats were enabled.
1496 */
intel_engine_get_busy_time(struct intel_engine_cs * engine)1497 ktime_t intel_engine_get_busy_time(struct intel_engine_cs *engine)
1498 {
1499 unsigned int seq;
1500 ktime_t total;
1501
1502 do {
1503 seq = read_seqbegin(&engine->stats.lock);
1504 total = __intel_engine_get_busy_time(engine);
1505 } while (read_seqretry(&engine->stats.lock, seq));
1506
1507 return total;
1508 }
1509
1510 /**
1511 * intel_disable_engine_stats() - Disable engine busy tracking on engine
1512 * @engine: engine to disable stats collection
1513 *
1514 * Stops collecting the engine busyness data for @engine.
1515 */
intel_disable_engine_stats(struct intel_engine_cs * engine)1516 void intel_disable_engine_stats(struct intel_engine_cs *engine)
1517 {
1518 unsigned long flags;
1519
1520 if (!intel_engine_supports_stats(engine))
1521 return;
1522
1523 write_seqlock_irqsave(&engine->stats.lock, flags);
1524 WARN_ON_ONCE(engine->stats.enabled == 0);
1525 if (--engine->stats.enabled == 0) {
1526 engine->stats.total = __intel_engine_get_busy_time(engine);
1527 engine->stats.active = 0;
1528 }
1529 write_sequnlock_irqrestore(&engine->stats.lock, flags);
1530 }
1531
match_ring(struct i915_request * rq)1532 static bool match_ring(struct i915_request *rq)
1533 {
1534 u32 ring = ENGINE_READ(rq->engine, RING_START);
1535
1536 return ring == i915_ggtt_offset(rq->ring->vma);
1537 }
1538
1539 struct i915_request *
intel_engine_find_active_request(struct intel_engine_cs * engine)1540 intel_engine_find_active_request(struct intel_engine_cs *engine)
1541 {
1542 struct i915_request *request, *active = NULL;
1543
1544 /*
1545 * We are called by the error capture, reset and to dump engine
1546 * state at random points in time. In particular, note that neither is
1547 * crucially ordered with an interrupt. After a hang, the GPU is dead
1548 * and we assume that no more writes can happen (we waited long enough
1549 * for all writes that were in transaction to be flushed) - adding an
1550 * extra delay for a recent interrupt is pointless. Hence, we do
1551 * not need an engine->irq_seqno_barrier() before the seqno reads.
1552 * At all other times, we must assume the GPU is still running, but
1553 * we only care about the snapshot of this moment.
1554 */
1555 lockdep_assert_held(&engine->active.lock);
1556 list_for_each_entry(request, &engine->active.requests, sched.link) {
1557 if (i915_request_completed(request))
1558 continue;
1559
1560 if (!i915_request_started(request))
1561 continue;
1562
1563 /* More than one preemptible request may match! */
1564 if (!match_ring(request))
1565 continue;
1566
1567 active = request;
1568 break;
1569 }
1570
1571 return active;
1572 }
1573
1574 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
1575 #include "mock_engine.c"
1576 #include "selftest_engine.c"
1577 #include "selftest_engine_cs.c"
1578 #endif
1579
