Linux Audio

Check our new training course

Loading...
Note: File does not exist in v3.1.
  1/*
  2 * Copyright © 2011-2012 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 * Authors:
 24 *    Ben Widawsky <ben@bwidawsk.net>
 25 *
 26 */
 27
 28/*
 29 * This file implements HW context support. On gen5+ a HW context consists of an
 30 * opaque GPU object which is referenced at times of context saves and restores.
 31 * With RC6 enabled, the context is also referenced as the GPU enters and exists
 32 * from RC6 (GPU has it's own internal power context, except on gen5). Though
 33 * something like a context does exist for the media ring, the code only
 34 * supports contexts for the render ring.
 35 *
 36 * In software, there is a distinction between contexts created by the user,
 37 * and the default HW context. The default HW context is used by GPU clients
 38 * that do not request setup of their own hardware context. The default
 39 * context's state is never restored to help prevent programming errors. This
 40 * would happen if a client ran and piggy-backed off another clients GPU state.
 41 * The default context only exists to give the GPU some offset to load as the
 42 * current to invoke a save of the context we actually care about. In fact, the
 43 * code could likely be constructed, albeit in a more complicated fashion, to
 44 * never use the default context, though that limits the driver's ability to
 45 * swap out, and/or destroy other contexts.
 46 *
 47 * All other contexts are created as a request by the GPU client. These contexts
 48 * store GPU state, and thus allow GPU clients to not re-emit state (and
 49 * potentially query certain state) at any time. The kernel driver makes
 50 * certain that the appropriate commands are inserted.
 51 *
 52 * The context life cycle is semi-complicated in that context BOs may live
 53 * longer than the context itself because of the way the hardware, and object
 54 * tracking works. Below is a very crude representation of the state machine
 55 * describing the context life.
 56 *                                         refcount     pincount     active
 57 * S0: initial state                          0            0           0
 58 * S1: context created                        1            0           0
 59 * S2: context is currently running           2            1           X
 60 * S3: GPU referenced, but not current        2            0           1
 61 * S4: context is current, but destroyed      1            1           0
 62 * S5: like S3, but destroyed                 1            0           1
 63 *
 64 * The most common (but not all) transitions:
 65 * S0->S1: client creates a context
 66 * S1->S2: client submits execbuf with context
 67 * S2->S3: other clients submits execbuf with context
 68 * S3->S1: context object was retired
 69 * S3->S2: clients submits another execbuf
 70 * S2->S4: context destroy called with current context
 71 * S3->S5->S0: destroy path
 72 * S4->S5->S0: destroy path on current context
 73 *
 74 * There are two confusing terms used above:
 75 *  The "current context" means the context which is currently running on the
 76 *  GPU. The GPU has loaded its state already and has stored away the gtt
 77 *  offset of the BO. The GPU is not actively referencing the data at this
 78 *  offset, but it will on the next context switch. The only way to avoid this
 79 *  is to do a GPU reset.
 80 *
 81 *  An "active context' is one which was previously the "current context" and is
 82 *  on the active list waiting for the next context switch to occur. Until this
 83 *  happens, the object must remain at the same gtt offset. It is therefore
 84 *  possible to destroy a context, but it is still active.
 85 *
 86 */
 87
 88#include <drm/drmP.h>
 89#include <drm/i915_drm.h>
 90#include "i915_drv.h"
 91
 92/* This is a HW constraint. The value below is the largest known requirement
 93 * I've seen in a spec to date, and that was a workaround for a non-shipping
 94 * part. It should be safe to decrease this, but it's more future proof as is.
 95 */
 96#define GEN6_CONTEXT_ALIGN (64<<10)
 97#define GEN7_CONTEXT_ALIGN 4096
 98
 99static void do_ppgtt_cleanup(struct i915_hw_ppgtt *ppgtt)
100{
101	struct drm_device *dev = ppgtt->base.dev;
102	struct drm_i915_private *dev_priv = dev->dev_private;
103	struct i915_address_space *vm = &ppgtt->base;
104
105	if (ppgtt == dev_priv->mm.aliasing_ppgtt ||
106	    (list_empty(&vm->active_list) && list_empty(&vm->inactive_list))) {
107		ppgtt->base.cleanup(&ppgtt->base);
108		return;
109	}
110
111	/*
112	 * Make sure vmas are unbound before we take down the drm_mm
113	 *
114	 * FIXME: Proper refcounting should take care of this, this shouldn't be
115	 * needed at all.
116	 */
117	if (!list_empty(&vm->active_list)) {
118		struct i915_vma *vma;
119
120		list_for_each_entry(vma, &vm->active_list, mm_list)
121			if (WARN_ON(list_empty(&vma->vma_link) ||
122				    list_is_singular(&vma->vma_link)))
123				break;
124
125		i915_gem_evict_vm(&ppgtt->base, true);
126	} else {
127		i915_gem_retire_requests(dev);
128		i915_gem_evict_vm(&ppgtt->base, false);
129	}
130
131	ppgtt->base.cleanup(&ppgtt->base);
132}
133
134static void ppgtt_release(struct kref *kref)
135{
136	struct i915_hw_ppgtt *ppgtt =
137		container_of(kref, struct i915_hw_ppgtt, ref);
138
139	do_ppgtt_cleanup(ppgtt);
140	kfree(ppgtt);
141}
142
143static size_t get_context_alignment(struct drm_device *dev)
144{
145	if (IS_GEN6(dev))
146		return GEN6_CONTEXT_ALIGN;
147
148	return GEN7_CONTEXT_ALIGN;
149}
150
151static int get_context_size(struct drm_device *dev)
152{
153	struct drm_i915_private *dev_priv = dev->dev_private;
154	int ret;
155	u32 reg;
156
157	switch (INTEL_INFO(dev)->gen) {
158	case 6:
159		reg = I915_READ(CXT_SIZE);
160		ret = GEN6_CXT_TOTAL_SIZE(reg) * 64;
161		break;
162	case 7:
163		reg = I915_READ(GEN7_CXT_SIZE);
164		if (IS_HASWELL(dev))
165			ret = HSW_CXT_TOTAL_SIZE;
166		else
167			ret = GEN7_CXT_TOTAL_SIZE(reg) * 64;
168		break;
169	case 8:
170		ret = GEN8_CXT_TOTAL_SIZE;
171		break;
172	default:
173		BUG();
174	}
175
176	return ret;
177}
178
179void i915_gem_context_free(struct kref *ctx_ref)
180{
181	struct i915_hw_context *ctx = container_of(ctx_ref,
182						   typeof(*ctx), ref);
183	struct i915_hw_ppgtt *ppgtt = NULL;
184
185	if (ctx->obj) {
186		/* We refcount even the aliasing PPGTT to keep the code symmetric */
187		if (USES_PPGTT(ctx->obj->base.dev))
188			ppgtt = ctx_to_ppgtt(ctx);
189
190		/* XXX: Free up the object before tearing down the address space, in
191		 * case we're bound in the PPGTT */
192		drm_gem_object_unreference(&ctx->obj->base);
193	}
194
195	if (ppgtt)
196		kref_put(&ppgtt->ref, ppgtt_release);
197	list_del(&ctx->link);
198	kfree(ctx);
199}
200
201static struct i915_hw_ppgtt *
202create_vm_for_ctx(struct drm_device *dev, struct i915_hw_context *ctx)
203{
204	struct i915_hw_ppgtt *ppgtt;
205	int ret;
206
207	ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
208	if (!ppgtt)
209		return ERR_PTR(-ENOMEM);
210
211	ret = i915_gem_init_ppgtt(dev, ppgtt);
212	if (ret) {
213		kfree(ppgtt);
214		return ERR_PTR(ret);
215	}
216
217	ppgtt->ctx = ctx;
218	return ppgtt;
219}
220
221static struct i915_hw_context *
222__create_hw_context(struct drm_device *dev,
223		  struct drm_i915_file_private *file_priv)
224{
225	struct drm_i915_private *dev_priv = dev->dev_private;
226	struct i915_hw_context *ctx;
227	int ret;
228
229	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
230	if (ctx == NULL)
231		return ERR_PTR(-ENOMEM);
232
233	kref_init(&ctx->ref);
234	list_add_tail(&ctx->link, &dev_priv->context_list);
235
236	if (dev_priv->hw_context_size) {
237		ctx->obj = i915_gem_alloc_object(dev, dev_priv->hw_context_size);
238		if (ctx->obj == NULL) {
239			ret = -ENOMEM;
240			goto err_out;
241		}
242
243		if (INTEL_INFO(dev)->gen >= 7) {
244			ret = i915_gem_object_set_cache_level(ctx->obj,
245							      I915_CACHE_L3_LLC);
246			/* Failure shouldn't ever happen this early */
247			if (WARN_ON(ret))
248				goto err_out;
249		}
250	}
251
252	/* Default context will never have a file_priv */
253	if (file_priv != NULL) {
254		ret = idr_alloc(&file_priv->context_idr, ctx,
255				DEFAULT_CONTEXT_ID, 0, GFP_KERNEL);
256		if (ret < 0)
257			goto err_out;
258	} else
259		ret = DEFAULT_CONTEXT_ID;
260
261	ctx->file_priv = file_priv;
262	ctx->id = ret;
263	/* NB: Mark all slices as needing a remap so that when the context first
264	 * loads it will restore whatever remap state already exists. If there
265	 * is no remap info, it will be a NOP. */
266	ctx->remap_slice = (1 << NUM_L3_SLICES(dev)) - 1;
267
268	return ctx;
269
270err_out:
271	i915_gem_context_unreference(ctx);
272	return ERR_PTR(ret);
273}
274
275/**
276 * The default context needs to exist per ring that uses contexts. It stores the
277 * context state of the GPU for applications that don't utilize HW contexts, as
278 * well as an idle case.
279 */
280static struct i915_hw_context *
281i915_gem_create_context(struct drm_device *dev,
282			struct drm_i915_file_private *file_priv,
283			bool create_vm)
284{
285	const bool is_global_default_ctx = file_priv == NULL;
286	struct drm_i915_private *dev_priv = dev->dev_private;
287	struct i915_hw_context *ctx;
288	int ret = 0;
289
290	BUG_ON(!mutex_is_locked(&dev->struct_mutex));
291
292	ctx = __create_hw_context(dev, file_priv);
293	if (IS_ERR(ctx))
294		return ctx;
295
296	if (is_global_default_ctx && ctx->obj) {
297		/* We may need to do things with the shrinker which
298		 * require us to immediately switch back to the default
299		 * context. This can cause a problem as pinning the
300		 * default context also requires GTT space which may not
301		 * be available. To avoid this we always pin the default
302		 * context.
303		 */
304		ret = i915_gem_obj_ggtt_pin(ctx->obj,
305					    get_context_alignment(dev), 0);
306		if (ret) {
307			DRM_DEBUG_DRIVER("Couldn't pin %d\n", ret);
308			goto err_destroy;
309		}
310	}
311
312	if (create_vm) {
313		struct i915_hw_ppgtt *ppgtt = create_vm_for_ctx(dev, ctx);
314
315		if (IS_ERR_OR_NULL(ppgtt)) {
316			DRM_DEBUG_DRIVER("PPGTT setup failed (%ld)\n",
317					 PTR_ERR(ppgtt));
318			ret = PTR_ERR(ppgtt);
319			goto err_unpin;
320		} else
321			ctx->vm = &ppgtt->base;
322
323		/* This case is reserved for the global default context and
324		 * should only happen once. */
325		if (is_global_default_ctx) {
326			if (WARN_ON(dev_priv->mm.aliasing_ppgtt)) {
327				ret = -EEXIST;
328				goto err_unpin;
329			}
330
331			dev_priv->mm.aliasing_ppgtt = ppgtt;
332		}
333	} else if (USES_PPGTT(dev)) {
334		/* For platforms which only have aliasing PPGTT, we fake the
335		 * address space and refcounting. */
336		ctx->vm = &dev_priv->mm.aliasing_ppgtt->base;
337		kref_get(&dev_priv->mm.aliasing_ppgtt->ref);
338	} else
339		ctx->vm = &dev_priv->gtt.base;
340
341	return ctx;
342
343err_unpin:
344	if (is_global_default_ctx && ctx->obj)
345		i915_gem_object_ggtt_unpin(ctx->obj);
346err_destroy:
347	i915_gem_context_unreference(ctx);
348	return ERR_PTR(ret);
349}
350
351void i915_gem_context_reset(struct drm_device *dev)
352{
353	struct drm_i915_private *dev_priv = dev->dev_private;
354	int i;
355
356	/* Prevent the hardware from restoring the last context (which hung) on
357	 * the next switch */
358	for (i = 0; i < I915_NUM_RINGS; i++) {
359		struct intel_ring_buffer *ring = &dev_priv->ring[i];
360		struct i915_hw_context *dctx = ring->default_context;
361
362		/* Do a fake switch to the default context */
363		if (ring->last_context == dctx)
364			continue;
365
366		if (!ring->last_context)
367			continue;
368
369		if (dctx->obj && i == RCS) {
370			WARN_ON(i915_gem_obj_ggtt_pin(dctx->obj,
371						      get_context_alignment(dev), 0));
372			/* Fake a finish/inactive */
373			dctx->obj->base.write_domain = 0;
374			dctx->obj->active = 0;
375		}
376
377		i915_gem_context_unreference(ring->last_context);
378		i915_gem_context_reference(dctx);
379		ring->last_context = dctx;
380	}
381}
382
383int i915_gem_context_init(struct drm_device *dev)
384{
385	struct drm_i915_private *dev_priv = dev->dev_private;
386	struct i915_hw_context *ctx;
387	int i;
388
389	/* Init should only be called once per module load. Eventually the
390	 * restriction on the context_disabled check can be loosened. */
391	if (WARN_ON(dev_priv->ring[RCS].default_context))
392		return 0;
393
394	if (HAS_HW_CONTEXTS(dev)) {
395		dev_priv->hw_context_size = round_up(get_context_size(dev), 4096);
396		if (dev_priv->hw_context_size > (1<<20)) {
397			DRM_DEBUG_DRIVER("Disabling HW Contexts; invalid size %d\n",
398					 dev_priv->hw_context_size);
399			dev_priv->hw_context_size = 0;
400		}
401	}
402
403	ctx = i915_gem_create_context(dev, NULL, USES_PPGTT(dev));
404	if (IS_ERR(ctx)) {
405		DRM_ERROR("Failed to create default global context (error %ld)\n",
406			  PTR_ERR(ctx));
407		return PTR_ERR(ctx);
408	}
409
410	/* NB: RCS will hold a ref for all rings */
411	for (i = 0; i < I915_NUM_RINGS; i++)
412		dev_priv->ring[i].default_context = ctx;
413
414	DRM_DEBUG_DRIVER("%s context support initialized\n", dev_priv->hw_context_size ? "HW" : "fake");
415	return 0;
416}
417
418void i915_gem_context_fini(struct drm_device *dev)
419{
420	struct drm_i915_private *dev_priv = dev->dev_private;
421	struct i915_hw_context *dctx = dev_priv->ring[RCS].default_context;
422	int i;
423
424	if (dctx->obj) {
425		/* The only known way to stop the gpu from accessing the hw context is
426		 * to reset it. Do this as the very last operation to avoid confusing
427		 * other code, leading to spurious errors. */
428		intel_gpu_reset(dev);
429
430		/* When default context is created and switched to, base object refcount
431		 * will be 2 (+1 from object creation and +1 from do_switch()).
432		 * i915_gem_context_fini() will be called after gpu_idle() has switched
433		 * to default context. So we need to unreference the base object once
434		 * to offset the do_switch part, so that i915_gem_context_unreference()
435		 * can then free the base object correctly. */
436		WARN_ON(!dev_priv->ring[RCS].last_context);
437		if (dev_priv->ring[RCS].last_context == dctx) {
438			/* Fake switch to NULL context */
439			WARN_ON(dctx->obj->active);
440			i915_gem_object_ggtt_unpin(dctx->obj);
441			i915_gem_context_unreference(dctx);
442			dev_priv->ring[RCS].last_context = NULL;
443		}
444	}
445
446	for (i = 0; i < I915_NUM_RINGS; i++) {
447		struct intel_ring_buffer *ring = &dev_priv->ring[i];
448
449		if (ring->last_context)
450			i915_gem_context_unreference(ring->last_context);
451
452		ring->default_context = NULL;
453		ring->last_context = NULL;
454	}
455
456	i915_gem_object_ggtt_unpin(dctx->obj);
457	i915_gem_context_unreference(dctx);
458}
459
460int i915_gem_context_enable(struct drm_i915_private *dev_priv)
461{
462	struct intel_ring_buffer *ring;
463	int ret, i;
464
465	/* This is the only place the aliasing PPGTT gets enabled, which means
466	 * it has to happen before we bail on reset */
467	if (dev_priv->mm.aliasing_ppgtt) {
468		struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
469		ppgtt->enable(ppgtt);
470	}
471
472	/* FIXME: We should make this work, even in reset */
473	if (i915_reset_in_progress(&dev_priv->gpu_error))
474		return 0;
475
476	BUG_ON(!dev_priv->ring[RCS].default_context);
477
478	for_each_ring(ring, dev_priv, i) {
479		ret = i915_switch_context(ring, ring->default_context);
480		if (ret)
481			return ret;
482	}
483
484	return 0;
485}
486
487static int context_idr_cleanup(int id, void *p, void *data)
488{
489	struct i915_hw_context *ctx = p;
490
491	/* Ignore the default context because close will handle it */
492	if (i915_gem_context_is_default(ctx))
493		return 0;
494
495	i915_gem_context_unreference(ctx);
496	return 0;
497}
498
499int i915_gem_context_open(struct drm_device *dev, struct drm_file *file)
500{
501	struct drm_i915_file_private *file_priv = file->driver_priv;
502
503	idr_init(&file_priv->context_idr);
504
505	mutex_lock(&dev->struct_mutex);
506	file_priv->private_default_ctx =
507		i915_gem_create_context(dev, file_priv, USES_FULL_PPGTT(dev));
508	mutex_unlock(&dev->struct_mutex);
509
510	if (IS_ERR(file_priv->private_default_ctx)) {
511		idr_destroy(&file_priv->context_idr);
512		return PTR_ERR(file_priv->private_default_ctx);
513	}
514
515	return 0;
516}
517
518void i915_gem_context_close(struct drm_device *dev, struct drm_file *file)
519{
520	struct drm_i915_file_private *file_priv = file->driver_priv;
521
522	idr_for_each(&file_priv->context_idr, context_idr_cleanup, NULL);
523	idr_destroy(&file_priv->context_idr);
524
525	i915_gem_context_unreference(file_priv->private_default_ctx);
526}
527
528struct i915_hw_context *
529i915_gem_context_get(struct drm_i915_file_private *file_priv, u32 id)
530{
531	struct i915_hw_context *ctx;
532
533	ctx = (struct i915_hw_context *)idr_find(&file_priv->context_idr, id);
534	if (!ctx)
535		return ERR_PTR(-ENOENT);
536
537	return ctx;
538}
539
540static inline int
541mi_set_context(struct intel_ring_buffer *ring,
542	       struct i915_hw_context *new_context,
543	       u32 hw_flags)
544{
545	int ret;
546
547	/* w/a: If Flush TLB Invalidation Mode is enabled, driver must do a TLB
548	 * invalidation prior to MI_SET_CONTEXT. On GEN6 we don't set the value
549	 * explicitly, so we rely on the value at ring init, stored in
550	 * itlb_before_ctx_switch.
551	 */
552	if (IS_GEN6(ring->dev) && ring->itlb_before_ctx_switch) {
553		ret = ring->flush(ring, I915_GEM_GPU_DOMAINS, 0);
554		if (ret)
555			return ret;
556	}
557
558	ret = intel_ring_begin(ring, 6);
559	if (ret)
560		return ret;
561
562	/* WaProgramMiArbOnOffAroundMiSetContext:ivb,vlv,hsw */
563	if (IS_GEN7(ring->dev))
564		intel_ring_emit(ring, MI_ARB_ON_OFF | MI_ARB_DISABLE);
565	else
566		intel_ring_emit(ring, MI_NOOP);
567
568	intel_ring_emit(ring, MI_NOOP);
569	intel_ring_emit(ring, MI_SET_CONTEXT);
570	intel_ring_emit(ring, i915_gem_obj_ggtt_offset(new_context->obj) |
571			MI_MM_SPACE_GTT |
572			MI_SAVE_EXT_STATE_EN |
573			MI_RESTORE_EXT_STATE_EN |
574			hw_flags);
575	/*
576	 * w/a: MI_SET_CONTEXT must always be followed by MI_NOOP
577	 * WaMiSetContext_Hang:snb,ivb,vlv
578	 */
579	intel_ring_emit(ring, MI_NOOP);
580
581	if (IS_GEN7(ring->dev))
582		intel_ring_emit(ring, MI_ARB_ON_OFF | MI_ARB_ENABLE);
583	else
584		intel_ring_emit(ring, MI_NOOP);
585
586	intel_ring_advance(ring);
587
588	return ret;
589}
590
591static int do_switch(struct intel_ring_buffer *ring,
592		     struct i915_hw_context *to)
593{
594	struct drm_i915_private *dev_priv = ring->dev->dev_private;
595	struct i915_hw_context *from = ring->last_context;
596	struct i915_hw_ppgtt *ppgtt = ctx_to_ppgtt(to);
597	u32 hw_flags = 0;
598	int ret, i;
599
600	if (from != NULL && ring == &dev_priv->ring[RCS]) {
601		BUG_ON(from->obj == NULL);
602		BUG_ON(!i915_gem_obj_is_pinned(from->obj));
603	}
604
605	if (from == to && from->last_ring == ring && !to->remap_slice)
606		return 0;
607
608	/* Trying to pin first makes error handling easier. */
609	if (ring == &dev_priv->ring[RCS]) {
610		ret = i915_gem_obj_ggtt_pin(to->obj,
611					    get_context_alignment(ring->dev), 0);
612		if (ret)
613			return ret;
614	}
615
616	/*
617	 * Pin can switch back to the default context if we end up calling into
618	 * evict_everything - as a last ditch gtt defrag effort that also
619	 * switches to the default context. Hence we need to reload from here.
620	 */
621	from = ring->last_context;
622
623	if (USES_FULL_PPGTT(ring->dev)) {
624		ret = ppgtt->switch_mm(ppgtt, ring, false);
625		if (ret)
626			goto unpin_out;
627	}
628
629	if (ring != &dev_priv->ring[RCS]) {
630		if (from)
631			i915_gem_context_unreference(from);
632		goto done;
633	}
634
635	/*
636	 * Clear this page out of any CPU caches for coherent swap-in/out. Note
637	 * that thanks to write = false in this call and us not setting any gpu
638	 * write domains when putting a context object onto the active list
639	 * (when switching away from it), this won't block.
640	 *
641	 * XXX: We need a real interface to do this instead of trickery.
642	 */
643	ret = i915_gem_object_set_to_gtt_domain(to->obj, false);
644	if (ret)
645		goto unpin_out;
646
647	if (!to->obj->has_global_gtt_mapping) {
648		struct i915_vma *vma = i915_gem_obj_to_vma(to->obj,
649							   &dev_priv->gtt.base);
650		vma->bind_vma(vma, to->obj->cache_level, GLOBAL_BIND);
651	}
652
653	if (!to->is_initialized || i915_gem_context_is_default(to))
654		hw_flags |= MI_RESTORE_INHIBIT;
655
656	ret = mi_set_context(ring, to, hw_flags);
657	if (ret)
658		goto unpin_out;
659
660	for (i = 0; i < MAX_L3_SLICES; i++) {
661		if (!(to->remap_slice & (1<<i)))
662			continue;
663
664		ret = i915_gem_l3_remap(ring, i);
665		/* If it failed, try again next round */
666		if (ret)
667			DRM_DEBUG_DRIVER("L3 remapping failed\n");
668		else
669			to->remap_slice &= ~(1<<i);
670	}
671
672	/* The backing object for the context is done after switching to the
673	 * *next* context. Therefore we cannot retire the previous context until
674	 * the next context has already started running. In fact, the below code
675	 * is a bit suboptimal because the retiring can occur simply after the
676	 * MI_SET_CONTEXT instead of when the next seqno has completed.
677	 */
678	if (from != NULL) {
679		from->obj->base.read_domains = I915_GEM_DOMAIN_INSTRUCTION;
680		i915_vma_move_to_active(i915_gem_obj_to_ggtt(from->obj), ring);
681		/* As long as MI_SET_CONTEXT is serializing, ie. it flushes the
682		 * whole damn pipeline, we don't need to explicitly mark the
683		 * object dirty. The only exception is that the context must be
684		 * correct in case the object gets swapped out. Ideally we'd be
685		 * able to defer doing this until we know the object would be
686		 * swapped, but there is no way to do that yet.
687		 */
688		from->obj->dirty = 1;
689		BUG_ON(from->obj->ring != ring);
690
691		/* obj is kept alive until the next request by its active ref */
692		i915_gem_object_ggtt_unpin(from->obj);
693		i915_gem_context_unreference(from);
694	}
695
696	to->is_initialized = true;
697
698done:
699	i915_gem_context_reference(to);
700	ring->last_context = to;
701	to->last_ring = ring;
702
703	return 0;
704
705unpin_out:
706	if (ring->id == RCS)
707		i915_gem_object_ggtt_unpin(to->obj);
708	return ret;
709}
710
711/**
712 * i915_switch_context() - perform a GPU context switch.
713 * @ring: ring for which we'll execute the context switch
714 * @to: the context to switch to
715 *
716 * The context life cycle is simple. The context refcount is incremented and
717 * decremented by 1 and create and destroy. If the context is in use by the GPU,
718 * it will have a refoucnt > 1. This allows us to destroy the context abstract
719 * object while letting the normal object tracking destroy the backing BO.
720 */
721int i915_switch_context(struct intel_ring_buffer *ring,
722			struct i915_hw_context *to)
723{
724	struct drm_i915_private *dev_priv = ring->dev->dev_private;
725
726	WARN_ON(!mutex_is_locked(&dev_priv->dev->struct_mutex));
727
728	if (to->obj == NULL) { /* We have the fake context */
729		if (to != ring->last_context) {
730			i915_gem_context_reference(to);
731			if (ring->last_context)
732				i915_gem_context_unreference(ring->last_context);
733			ring->last_context = to;
734		}
735		return 0;
736	}
737
738	return do_switch(ring, to);
739}
740
741static bool hw_context_enabled(struct drm_device *dev)
742{
743	return to_i915(dev)->hw_context_size;
744}
745
746int i915_gem_context_create_ioctl(struct drm_device *dev, void *data,
747				  struct drm_file *file)
748{
749	struct drm_i915_gem_context_create *args = data;
750	struct drm_i915_file_private *file_priv = file->driver_priv;
751	struct i915_hw_context *ctx;
752	int ret;
753
754	if (!hw_context_enabled(dev))
755		return -ENODEV;
756
757	ret = i915_mutex_lock_interruptible(dev);
758	if (ret)
759		return ret;
760
761	ctx = i915_gem_create_context(dev, file_priv, USES_FULL_PPGTT(dev));
762	mutex_unlock(&dev->struct_mutex);
763	if (IS_ERR(ctx))
764		return PTR_ERR(ctx);
765
766	args->ctx_id = ctx->id;
767	DRM_DEBUG_DRIVER("HW context %d created\n", args->ctx_id);
768
769	return 0;
770}
771
772int i915_gem_context_destroy_ioctl(struct drm_device *dev, void *data,
773				   struct drm_file *file)
774{
775	struct drm_i915_gem_context_destroy *args = data;
776	struct drm_i915_file_private *file_priv = file->driver_priv;
777	struct i915_hw_context *ctx;
778	int ret;
779
780	if (args->ctx_id == DEFAULT_CONTEXT_ID)
781		return -ENOENT;
782
783	ret = i915_mutex_lock_interruptible(dev);
784	if (ret)
785		return ret;
786
787	ctx = i915_gem_context_get(file_priv, args->ctx_id);
788	if (IS_ERR(ctx)) {
789		mutex_unlock(&dev->struct_mutex);
790		return PTR_ERR(ctx);
791	}
792
793	idr_remove(&ctx->file_priv->context_idr, ctx->id);
794	i915_gem_context_unreference(ctx);
795	mutex_unlock(&dev->struct_mutex);
796
797	DRM_DEBUG_DRIVER("HW context %d destroyed\n", args->ctx_id);
798	return 0;
799}