Linux Audio

Check our new training course

Linux debugging, profiling, tracing and performance analysis training

Mar 24-27, 2025, special US time zones
Register
Loading...
Note: File does not exist in v4.6.
  1// SPDX-License-Identifier: MIT
  2/*
  3 * Copyright © 2019 Intel Corporation
  4 */
  5
  6#include "debugfs_gt.h"
  7#include "i915_drv.h"
  8#include "intel_context.h"
  9#include "intel_gt.h"
 10#include "intel_gt_buffer_pool.h"
 11#include "intel_gt_clock_utils.h"
 12#include "intel_gt_pm.h"
 13#include "intel_gt_requests.h"
 14#include "intel_mocs.h"
 15#include "intel_rc6.h"
 16#include "intel_renderstate.h"
 17#include "intel_rps.h"
 18#include "intel_uncore.h"
 19#include "intel_pm.h"
 20#include "shmem_utils.h"
 21
 22void intel_gt_init_early(struct intel_gt *gt, struct drm_i915_private *i915)
 23{
 24	gt->i915 = i915;
 25	gt->uncore = &i915->uncore;
 26
 27	spin_lock_init(&gt->irq_lock);
 28
 29	INIT_LIST_HEAD(&gt->closed_vma);
 30	spin_lock_init(&gt->closed_lock);
 31
 32	intel_gt_init_buffer_pool(gt);
 33	intel_gt_init_reset(gt);
 34	intel_gt_init_requests(gt);
 35	intel_gt_init_timelines(gt);
 36	intel_gt_pm_init_early(gt);
 37
 38	intel_rps_init_early(&gt->rps);
 39	intel_uc_init_early(&gt->uc);
 40}
 41
 42void intel_gt_init_hw_early(struct intel_gt *gt, struct i915_ggtt *ggtt)
 43{
 44	gt->ggtt = ggtt;
 45}
 46
 47int intel_gt_init_mmio(struct intel_gt *gt)
 48{
 49	intel_uc_init_mmio(&gt->uc);
 50	intel_sseu_info_init(gt);
 51
 52	return intel_engines_init_mmio(gt);
 53}
 54
 55static void init_unused_ring(struct intel_gt *gt, u32 base)
 56{
 57	struct intel_uncore *uncore = gt->uncore;
 58
 59	intel_uncore_write(uncore, RING_CTL(base), 0);
 60	intel_uncore_write(uncore, RING_HEAD(base), 0);
 61	intel_uncore_write(uncore, RING_TAIL(base), 0);
 62	intel_uncore_write(uncore, RING_START(base), 0);
 63}
 64
 65static void init_unused_rings(struct intel_gt *gt)
 66{
 67	struct drm_i915_private *i915 = gt->i915;
 68
 69	if (IS_I830(i915)) {
 70		init_unused_ring(gt, PRB1_BASE);
 71		init_unused_ring(gt, SRB0_BASE);
 72		init_unused_ring(gt, SRB1_BASE);
 73		init_unused_ring(gt, SRB2_BASE);
 74		init_unused_ring(gt, SRB3_BASE);
 75	} else if (IS_GEN(i915, 2)) {
 76		init_unused_ring(gt, SRB0_BASE);
 77		init_unused_ring(gt, SRB1_BASE);
 78	} else if (IS_GEN(i915, 3)) {
 79		init_unused_ring(gt, PRB1_BASE);
 80		init_unused_ring(gt, PRB2_BASE);
 81	}
 82}
 83
 84int intel_gt_init_hw(struct intel_gt *gt)
 85{
 86	struct drm_i915_private *i915 = gt->i915;
 87	struct intel_uncore *uncore = gt->uncore;
 88	int ret;
 89
 90	gt->last_init_time = ktime_get();
 91
 92	/* Double layer security blanket, see i915_gem_init() */
 93	intel_uncore_forcewake_get(uncore, FORCEWAKE_ALL);
 94
 95	if (HAS_EDRAM(i915) && INTEL_GEN(i915) < 9)
 96		intel_uncore_rmw(uncore, HSW_IDICR, 0, IDIHASHMSK(0xf));
 97
 98	if (IS_HASWELL(i915))
 99		intel_uncore_write(uncore,
100				   MI_PREDICATE_RESULT_2,
101				   IS_HSW_GT3(i915) ?
102				   LOWER_SLICE_ENABLED : LOWER_SLICE_DISABLED);
103
104	/* Apply the GT workarounds... */
105	intel_gt_apply_workarounds(gt);
106	/* ...and determine whether they are sticking. */
107	intel_gt_verify_workarounds(gt, "init");
108
109	intel_gt_init_swizzling(gt);
110
111	/*
112	 * At least 830 can leave some of the unused rings
113	 * "active" (ie. head != tail) after resume which
114	 * will prevent c3 entry. Makes sure all unused rings
115	 * are totally idle.
116	 */
117	init_unused_rings(gt);
118
119	ret = i915_ppgtt_init_hw(gt);
120	if (ret) {
121		DRM_ERROR("Enabling PPGTT failed (%d)\n", ret);
122		goto out;
123	}
124
125	/* We can't enable contexts until all firmware is loaded */
126	ret = intel_uc_init_hw(&gt->uc);
127	if (ret) {
128		i915_probe_error(i915, "Enabling uc failed (%d)\n", ret);
129		goto out;
130	}
131
132	intel_mocs_init(gt);
133
134out:
135	intel_uncore_forcewake_put(uncore, FORCEWAKE_ALL);
136	return ret;
137}
138
139static void rmw_set(struct intel_uncore *uncore, i915_reg_t reg, u32 set)
140{
141	intel_uncore_rmw(uncore, reg, 0, set);
142}
143
144static void rmw_clear(struct intel_uncore *uncore, i915_reg_t reg, u32 clr)
145{
146	intel_uncore_rmw(uncore, reg, clr, 0);
147}
148
149static void clear_register(struct intel_uncore *uncore, i915_reg_t reg)
150{
151	intel_uncore_rmw(uncore, reg, 0, 0);
152}
153
154static void gen8_clear_engine_error_register(struct intel_engine_cs *engine)
155{
156	GEN6_RING_FAULT_REG_RMW(engine, RING_FAULT_VALID, 0);
157	GEN6_RING_FAULT_REG_POSTING_READ(engine);
158}
159
160void
161intel_gt_clear_error_registers(struct intel_gt *gt,
162			       intel_engine_mask_t engine_mask)
163{
164	struct drm_i915_private *i915 = gt->i915;
165	struct intel_uncore *uncore = gt->uncore;
166	u32 eir;
167
168	if (!IS_GEN(i915, 2))
169		clear_register(uncore, PGTBL_ER);
170
171	if (INTEL_GEN(i915) < 4)
172		clear_register(uncore, IPEIR(RENDER_RING_BASE));
173	else
174		clear_register(uncore, IPEIR_I965);
175
176	clear_register(uncore, EIR);
177	eir = intel_uncore_read(uncore, EIR);
178	if (eir) {
179		/*
180		 * some errors might have become stuck,
181		 * mask them.
182		 */
183		DRM_DEBUG_DRIVER("EIR stuck: 0x%08x, masking\n", eir);
184		rmw_set(uncore, EMR, eir);
185		intel_uncore_write(uncore, GEN2_IIR,
186				   I915_MASTER_ERROR_INTERRUPT);
187	}
188
189	if (INTEL_GEN(i915) >= 12) {
190		rmw_clear(uncore, GEN12_RING_FAULT_REG, RING_FAULT_VALID);
191		intel_uncore_posting_read(uncore, GEN12_RING_FAULT_REG);
192	} else if (INTEL_GEN(i915) >= 8) {
193		rmw_clear(uncore, GEN8_RING_FAULT_REG, RING_FAULT_VALID);
194		intel_uncore_posting_read(uncore, GEN8_RING_FAULT_REG);
195	} else if (INTEL_GEN(i915) >= 6) {
196		struct intel_engine_cs *engine;
197		enum intel_engine_id id;
198
199		for_each_engine_masked(engine, gt, engine_mask, id)
200			gen8_clear_engine_error_register(engine);
201	}
202}
203
204static void gen6_check_faults(struct intel_gt *gt)
205{
206	struct intel_engine_cs *engine;
207	enum intel_engine_id id;
208	u32 fault;
209
210	for_each_engine(engine, gt, id) {
211		fault = GEN6_RING_FAULT_REG_READ(engine);
212		if (fault & RING_FAULT_VALID) {
213			drm_dbg(&engine->i915->drm, "Unexpected fault\n"
214				"\tAddr: 0x%08lx\n"
215				"\tAddress space: %s\n"
216				"\tSource ID: %d\n"
217				"\tType: %d\n",
218				fault & PAGE_MASK,
219				fault & RING_FAULT_GTTSEL_MASK ?
220				"GGTT" : "PPGTT",
221				RING_FAULT_SRCID(fault),
222				RING_FAULT_FAULT_TYPE(fault));
223		}
224	}
225}
226
227static void gen8_check_faults(struct intel_gt *gt)
228{
229	struct intel_uncore *uncore = gt->uncore;
230	i915_reg_t fault_reg, fault_data0_reg, fault_data1_reg;
231	u32 fault;
232
233	if (INTEL_GEN(gt->i915) >= 12) {
234		fault_reg = GEN12_RING_FAULT_REG;
235		fault_data0_reg = GEN12_FAULT_TLB_DATA0;
236		fault_data1_reg = GEN12_FAULT_TLB_DATA1;
237	} else {
238		fault_reg = GEN8_RING_FAULT_REG;
239		fault_data0_reg = GEN8_FAULT_TLB_DATA0;
240		fault_data1_reg = GEN8_FAULT_TLB_DATA1;
241	}
242
243	fault = intel_uncore_read(uncore, fault_reg);
244	if (fault & RING_FAULT_VALID) {
245		u32 fault_data0, fault_data1;
246		u64 fault_addr;
247
248		fault_data0 = intel_uncore_read(uncore, fault_data0_reg);
249		fault_data1 = intel_uncore_read(uncore, fault_data1_reg);
250
251		fault_addr = ((u64)(fault_data1 & FAULT_VA_HIGH_BITS) << 44) |
252			     ((u64)fault_data0 << 12);
253
254		drm_dbg(&uncore->i915->drm, "Unexpected fault\n"
255			"\tAddr: 0x%08x_%08x\n"
256			"\tAddress space: %s\n"
257			"\tEngine ID: %d\n"
258			"\tSource ID: %d\n"
259			"\tType: %d\n",
260			upper_32_bits(fault_addr), lower_32_bits(fault_addr),
261			fault_data1 & FAULT_GTT_SEL ? "GGTT" : "PPGTT",
262			GEN8_RING_FAULT_ENGINE_ID(fault),
263			RING_FAULT_SRCID(fault),
264			RING_FAULT_FAULT_TYPE(fault));
265	}
266}
267
268void intel_gt_check_and_clear_faults(struct intel_gt *gt)
269{
270	struct drm_i915_private *i915 = gt->i915;
271
272	/* From GEN8 onwards we only have one 'All Engine Fault Register' */
273	if (INTEL_GEN(i915) >= 8)
274		gen8_check_faults(gt);
275	else if (INTEL_GEN(i915) >= 6)
276		gen6_check_faults(gt);
277	else
278		return;
279
280	intel_gt_clear_error_registers(gt, ALL_ENGINES);
281}
282
283void intel_gt_flush_ggtt_writes(struct intel_gt *gt)
284{
285	struct intel_uncore *uncore = gt->uncore;
286	intel_wakeref_t wakeref;
287
288	/*
289	 * No actual flushing is required for the GTT write domain for reads
290	 * from the GTT domain. Writes to it "immediately" go to main memory
291	 * as far as we know, so there's no chipset flush. It also doesn't
292	 * land in the GPU render cache.
293	 *
294	 * However, we do have to enforce the order so that all writes through
295	 * the GTT land before any writes to the device, such as updates to
296	 * the GATT itself.
297	 *
298	 * We also have to wait a bit for the writes to land from the GTT.
299	 * An uncached read (i.e. mmio) seems to be ideal for the round-trip
300	 * timing. This issue has only been observed when switching quickly
301	 * between GTT writes and CPU reads from inside the kernel on recent hw,
302	 * and it appears to only affect discrete GTT blocks (i.e. on LLC
303	 * system agents we cannot reproduce this behaviour, until Cannonlake
304	 * that was!).
305	 */
306
307	wmb();
308
309	if (INTEL_INFO(gt->i915)->has_coherent_ggtt)
310		return;
311
312	intel_gt_chipset_flush(gt);
313
314	with_intel_runtime_pm_if_in_use(uncore->rpm, wakeref) {
315		unsigned long flags;
316
317		spin_lock_irqsave(&uncore->lock, flags);
318		intel_uncore_posting_read_fw(uncore,
319					     RING_HEAD(RENDER_RING_BASE));
320		spin_unlock_irqrestore(&uncore->lock, flags);
321	}
322}
323
324void intel_gt_chipset_flush(struct intel_gt *gt)
325{
326	wmb();
327	if (INTEL_GEN(gt->i915) < 6)
328		intel_gtt_chipset_flush();
329}
330
331void intel_gt_driver_register(struct intel_gt *gt)
332{
333	intel_rps_driver_register(&gt->rps);
334
335	debugfs_gt_register(gt);
336}
337
338static int intel_gt_init_scratch(struct intel_gt *gt, unsigned int size)
339{
340	struct drm_i915_private *i915 = gt->i915;
341	struct drm_i915_gem_object *obj;
342	struct i915_vma *vma;
343	int ret;
344
345	obj = i915_gem_object_create_stolen(i915, size);
346	if (IS_ERR(obj))
347		obj = i915_gem_object_create_internal(i915, size);
348	if (IS_ERR(obj)) {
349		DRM_ERROR("Failed to allocate scratch page\n");
350		return PTR_ERR(obj);
351	}
352
353	vma = i915_vma_instance(obj, &gt->ggtt->vm, NULL);
354	if (IS_ERR(vma)) {
355		ret = PTR_ERR(vma);
356		goto err_unref;
357	}
358
359	ret = i915_ggtt_pin(vma, 0, PIN_HIGH);
360	if (ret)
361		goto err_unref;
362
363	gt->scratch = i915_vma_make_unshrinkable(vma);
364
365	return 0;
366
367err_unref:
368	i915_gem_object_put(obj);
369	return ret;
370}
371
372static void intel_gt_fini_scratch(struct intel_gt *gt)
373{
374	i915_vma_unpin_and_release(&gt->scratch, 0);
375}
376
377static struct i915_address_space *kernel_vm(struct intel_gt *gt)
378{
379	if (INTEL_PPGTT(gt->i915) > INTEL_PPGTT_ALIASING)
380		return &i915_ppgtt_create(gt)->vm;
381	else
382		return i915_vm_get(&gt->ggtt->vm);
383}
384
385static int __engines_record_defaults(struct intel_gt *gt)
386{
387	struct i915_request *requests[I915_NUM_ENGINES] = {};
388	struct intel_engine_cs *engine;
389	enum intel_engine_id id;
390	int err = 0;
391
392	/*
393	 * As we reset the gpu during very early sanitisation, the current
394	 * register state on the GPU should reflect its defaults values.
395	 * We load a context onto the hw (with restore-inhibit), then switch
396	 * over to a second context to save that default register state. We
397	 * can then prime every new context with that state so they all start
398	 * from the same default HW values.
399	 */
400
401	for_each_engine(engine, gt, id) {
402		struct intel_renderstate so;
403		struct intel_context *ce;
404		struct i915_request *rq;
405
406		/* We must be able to switch to something! */
407		GEM_BUG_ON(!engine->kernel_context);
408
409		err = intel_renderstate_init(&so, engine);
410		if (err)
411			goto out;
412
413		ce = intel_context_create(engine);
414		if (IS_ERR(ce)) {
415			err = PTR_ERR(ce);
416			goto out;
417		}
418
419		rq = intel_context_create_request(ce);
420		if (IS_ERR(rq)) {
421			err = PTR_ERR(rq);
422			intel_context_put(ce);
423			goto out;
424		}
425
426		err = intel_engine_emit_ctx_wa(rq);
427		if (err)
428			goto err_rq;
429
430		err = intel_renderstate_emit(&so, rq);
431		if (err)
432			goto err_rq;
433
434err_rq:
435		requests[id] = i915_request_get(rq);
436		i915_request_add(rq);
437		intel_renderstate_fini(&so);
438		if (err)
439			goto out;
440	}
441
442	/* Flush the default context image to memory, and enable powersaving. */
443	if (intel_gt_wait_for_idle(gt, I915_GEM_IDLE_TIMEOUT) == -ETIME) {
444		err = -EIO;
445		goto out;
446	}
447
448	for (id = 0; id < ARRAY_SIZE(requests); id++) {
449		struct i915_request *rq;
450		struct file *state;
451
452		rq = requests[id];
453		if (!rq)
454			continue;
455
456		if (rq->fence.error) {
457			err = -EIO;
458			goto out;
459		}
460
461		GEM_BUG_ON(!test_bit(CONTEXT_ALLOC_BIT, &rq->context->flags));
462		if (!rq->context->state)
463			continue;
464
465		/* Keep a copy of the state's backing pages; free the obj */
466		state = shmem_create_from_object(rq->context->state->obj);
467		if (IS_ERR(state)) {
468			err = PTR_ERR(state);
469			goto out;
470		}
471		rq->engine->default_state = state;
472	}
473
474out:
475	/*
476	 * If we have to abandon now, we expect the engines to be idle
477	 * and ready to be torn-down. The quickest way we can accomplish
478	 * this is by declaring ourselves wedged.
479	 */
480	if (err)
481		intel_gt_set_wedged(gt);
482
483	for (id = 0; id < ARRAY_SIZE(requests); id++) {
484		struct intel_context *ce;
485		struct i915_request *rq;
486
487		rq = requests[id];
488		if (!rq)
489			continue;
490
491		ce = rq->context;
492		i915_request_put(rq);
493		intel_context_put(ce);
494	}
495	return err;
496}
497
498static int __engines_verify_workarounds(struct intel_gt *gt)
499{
500	struct intel_engine_cs *engine;
501	enum intel_engine_id id;
502	int err = 0;
503
504	if (!IS_ENABLED(CONFIG_DRM_I915_DEBUG_GEM))
505		return 0;
506
507	for_each_engine(engine, gt, id) {
508		if (intel_engine_verify_workarounds(engine, "load"))
509			err = -EIO;
510	}
511
512	/* Flush and restore the kernel context for safety */
513	if (intel_gt_wait_for_idle(gt, I915_GEM_IDLE_TIMEOUT) == -ETIME)
514		err = -EIO;
515
516	return err;
517}
518
519static void __intel_gt_disable(struct intel_gt *gt)
520{
521	intel_gt_set_wedged_on_fini(gt);
522
523	intel_gt_suspend_prepare(gt);
524	intel_gt_suspend_late(gt);
525
526	GEM_BUG_ON(intel_gt_pm_is_awake(gt));
527}
528
529int intel_gt_init(struct intel_gt *gt)
530{
531	int err;
532
533	err = i915_inject_probe_error(gt->i915, -ENODEV);
534	if (err)
535		return err;
536
537	/*
538	 * This is just a security blanket to placate dragons.
539	 * On some systems, we very sporadically observe that the first TLBs
540	 * used by the CS may be stale, despite us poking the TLB reset. If
541	 * we hold the forcewake during initialisation these problems
542	 * just magically go away.
543	 */
544	intel_uncore_forcewake_get(gt->uncore, FORCEWAKE_ALL);
545
546	intel_gt_init_clock_frequency(gt);
547
548	err = intel_gt_init_scratch(gt, IS_GEN(gt->i915, 2) ? SZ_256K : SZ_4K);
549	if (err)
550		goto out_fw;
551
552	intel_gt_pm_init(gt);
553
554	gt->vm = kernel_vm(gt);
555	if (!gt->vm) {
556		err = -ENOMEM;
557		goto err_pm;
558	}
559
560	err = intel_engines_init(gt);
561	if (err)
562		goto err_engines;
563
564	err = intel_uc_init(&gt->uc);
565	if (err)
566		goto err_engines;
567
568	err = intel_gt_resume(gt);
569	if (err)
570		goto err_uc_init;
571
572	err = __engines_record_defaults(gt);
573	if (err)
574		goto err_gt;
575
576	err = __engines_verify_workarounds(gt);
577	if (err)
578		goto err_gt;
579
580	err = i915_inject_probe_error(gt->i915, -EIO);
581	if (err)
582		goto err_gt;
583
584	goto out_fw;
585err_gt:
586	__intel_gt_disable(gt);
587	intel_uc_fini_hw(&gt->uc);
588err_uc_init:
589	intel_uc_fini(&gt->uc);
590err_engines:
591	intel_engines_release(gt);
592	i915_vm_put(fetch_and_zero(&gt->vm));
593err_pm:
594	intel_gt_pm_fini(gt);
595	intel_gt_fini_scratch(gt);
596out_fw:
597	if (err)
598		intel_gt_set_wedged_on_init(gt);
599	intel_uncore_forcewake_put(gt->uncore, FORCEWAKE_ALL);
600	return err;
601}
602
603void intel_gt_driver_remove(struct intel_gt *gt)
604{
605	__intel_gt_disable(gt);
606
607	intel_uc_driver_remove(&gt->uc);
608
609	intel_engines_release(gt);
610}
611
612void intel_gt_driver_unregister(struct intel_gt *gt)
613{
614	intel_rps_driver_unregister(&gt->rps);
615
616	/*
617	 * Upon unregistering the device to prevent any new users, cancel
618	 * all in-flight requests so that we can quickly unbind the active
619	 * resources.
620	 */
621	intel_gt_set_wedged(gt);
622}
623
624void intel_gt_driver_release(struct intel_gt *gt)
625{
626	struct i915_address_space *vm;
627	intel_wakeref_t wakeref;
628
629	/* Scrub all HW state upon release */
630	with_intel_runtime_pm(gt->uncore->rpm, wakeref)
631		__intel_gt_reset(gt, ALL_ENGINES);
632
633	vm = fetch_and_zero(&gt->vm);
634	if (vm) /* FIXME being called twice on error paths :( */
635		i915_vm_put(vm);
636
637	intel_gt_pm_fini(gt);
638	intel_gt_fini_scratch(gt);
639	intel_gt_fini_buffer_pool(gt);
640}
641
642void intel_gt_driver_late_release(struct intel_gt *gt)
643{
644	/* We need to wait for inflight RCU frees to release their grip */
645	rcu_barrier();
646
647	intel_uc_driver_late_release(&gt->uc);
648	intel_gt_fini_requests(gt);
649	intel_gt_fini_reset(gt);
650	intel_gt_fini_timelines(gt);
651	intel_engines_free(gt);
652}
653
654void intel_gt_info_print(const struct intel_gt_info *info,
655			 struct drm_printer *p)
656{
657	drm_printf(p, "available engines: %x\n", info->engine_mask);
658
659	intel_sseu_dump(&info->sseu, p);
660}