Linux Audio

Check our new training course

Loading...
v4.6
  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#include "i915_trace.h"
 92
 93/* This is a HW constraint. The value below is the largest known requirement
 94 * I've seen in a spec to date, and that was a workaround for a non-shipping
 95 * part. It should be safe to decrease this, but it's more future proof as is.
 96 */
 97#define GEN6_CONTEXT_ALIGN (64<<10)
 98#define GEN7_CONTEXT_ALIGN 4096
 99
100static size_t get_context_alignment(struct drm_device *dev)
101{
102	if (IS_GEN6(dev))
103		return GEN6_CONTEXT_ALIGN;
 
 
 
 
 
 
104
105	return GEN7_CONTEXT_ALIGN;
 
 
 
 
 
 
 
106}
107
108static int get_context_size(struct drm_device *dev)
109{
110	struct drm_i915_private *dev_priv = dev->dev_private;
111	int ret;
112	u32 reg;
113
114	switch (INTEL_INFO(dev)->gen) {
115	case 6:
116		reg = I915_READ(CXT_SIZE);
117		ret = GEN6_CXT_TOTAL_SIZE(reg) * 64;
118		break;
119	case 7:
120		reg = I915_READ(GEN7_CXT_SIZE);
121		if (IS_HASWELL(dev))
122			ret = HSW_CXT_TOTAL_SIZE;
123		else
124			ret = GEN7_CXT_TOTAL_SIZE(reg) * 64;
125		break;
126	case 8:
127		ret = GEN8_CXT_TOTAL_SIZE;
128		break;
129	default:
130		BUG();
131	}
132
133	return ret;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
134}
135
136static void i915_gem_context_clean(struct intel_context *ctx)
137{
138	struct i915_hw_ppgtt *ppgtt = ctx->ppgtt;
139	struct i915_vma *vma, *next;
140
141	if (!ppgtt)
 
 
 
142		return;
143
144	list_for_each_entry_safe(vma, next, &ppgtt->base.inactive_list,
145				 vm_link) {
146		if (WARN_ON(__i915_vma_unbind_no_wait(vma)))
147			break;
148	}
149}
150
151void i915_gem_context_free(struct kref *ctx_ref)
152{
153	struct intel_context *ctx = container_of(ctx_ref, typeof(*ctx), ref);
 
 
 
 
 
 
 
 
 
 
 
154
155	trace_i915_context_free(ctx);
 
 
 
156
157	if (i915.enable_execlists)
158		intel_lr_context_free(ctx);
 
159
160	/*
161	 * This context is going away and we need to remove all VMAs still
162	 * around. This is to handle imported shared objects for which
163	 * destructor did not run when their handles were closed.
164	 */
165	i915_gem_context_clean(ctx);
166
167	i915_ppgtt_put(ctx->ppgtt);
168
169	if (ctx->legacy_hw_ctx.rcs_state)
170		drm_gem_object_unreference(&ctx->legacy_hw_ctx.rcs_state->base);
171	list_del(&ctx->link);
172	kfree(ctx);
173}
174
175struct drm_i915_gem_object *
176i915_gem_alloc_context_obj(struct drm_device *dev, size_t size)
177{
178	struct drm_i915_gem_object *obj;
179	int ret;
 
180
181	obj = i915_gem_alloc_object(dev, size);
182	if (obj == NULL)
183		return ERR_PTR(-ENOMEM);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
184
185	/*
186	 * Try to make the context utilize L3 as well as LLC.
187	 *
188	 * On VLV we don't have L3 controls in the PTEs so we
189	 * shouldn't touch the cache level, especially as that
190	 * would make the object snooped which might have a
191	 * negative performance impact.
192	 *
193	 * Snooping is required on non-llc platforms in execlist
194	 * mode, but since all GGTT accesses use PAT entry 0 we
195	 * get snooping anyway regardless of cache_level.
196	 *
197	 * This is only applicable for Ivy Bridge devices since
198	 * later platforms don't have L3 control bits in the PTE.
 
 
 
 
 
 
 
 
 
199	 */
200	if (IS_IVYBRIDGE(dev)) {
201		ret = i915_gem_object_set_cache_level(obj, I915_CACHE_L3_LLC);
202		/* Failure shouldn't ever happen this early */
203		if (WARN_ON(ret)) {
204			drm_gem_object_unreference(&obj->base);
205			return ERR_PTR(ret);
206		}
207	}
208
209	return obj;
210}
211
212static struct intel_context *
213__create_hw_context(struct drm_device *dev,
214		    struct drm_i915_file_private *file_priv)
215{
216	struct drm_i915_private *dev_priv = dev->dev_private;
217	struct intel_context *ctx;
218	int ret;
219
220	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
221	if (ctx == NULL)
222		return ERR_PTR(-ENOMEM);
223
 
 
 
 
 
 
224	kref_init(&ctx->ref);
225	list_add_tail(&ctx->link, &dev_priv->context_list);
226	ctx->i915 = dev_priv;
 
227
228	if (dev_priv->hw_context_size) {
229		struct drm_i915_gem_object *obj =
230				i915_gem_alloc_context_obj(dev, dev_priv->hw_context_size);
231		if (IS_ERR(obj)) {
232			ret = PTR_ERR(obj);
233			goto err_out;
234		}
235		ctx->legacy_hw_ctx.rcs_state = obj;
236	}
237
238	/* Default context will never have a file_priv */
239	if (file_priv != NULL) {
 
240		ret = idr_alloc(&file_priv->context_idr, ctx,
241				DEFAULT_CONTEXT_HANDLE, 0, GFP_KERNEL);
242		if (ret < 0)
243			goto err_out;
244	} else
245		ret = DEFAULT_CONTEXT_HANDLE;
246
247	ctx->file_priv = file_priv;
248	ctx->user_handle = ret;
 
 
 
 
 
 
 
 
 
 
 
249	/* NB: Mark all slices as needing a remap so that when the context first
250	 * loads it will restore whatever remap state already exists. If there
251	 * is no remap info, it will be a NOP. */
252	ctx->remap_slice = (1 << NUM_L3_SLICES(dev)) - 1;
253
254	ctx->hang_stats.ban_period_seconds = DRM_I915_CTX_BAN_PERIOD;
 
 
 
 
 
 
 
 
 
 
 
 
255
256	return ctx;
257
258err_out:
259	i915_gem_context_unreference(ctx);
 
 
 
260	return ERR_PTR(ret);
261}
262
263/**
264 * The default context needs to exist per ring that uses contexts. It stores the
265 * context state of the GPU for applications that don't utilize HW contexts, as
266 * well as an idle case.
267 */
268static struct intel_context *
269i915_gem_create_context(struct drm_device *dev,
 
 
270			struct drm_i915_file_private *file_priv)
271{
272	const bool is_global_default_ctx = file_priv == NULL;
273	struct intel_context *ctx;
274	int ret = 0;
275
276	BUG_ON(!mutex_is_locked(&dev->struct_mutex));
 
277
278	ctx = __create_hw_context(dev, file_priv);
279	if (IS_ERR(ctx))
280		return ctx;
281
282	if (is_global_default_ctx && ctx->legacy_hw_ctx.rcs_state) {
283		/* We may need to do things with the shrinker which
284		 * require us to immediately switch back to the default
285		 * context. This can cause a problem as pinning the
286		 * default context also requires GTT space which may not
287		 * be available. To avoid this we always pin the default
288		 * context.
289		 */
290		ret = i915_gem_obj_ggtt_pin(ctx->legacy_hw_ctx.rcs_state,
291					    get_context_alignment(dev), 0);
292		if (ret) {
293			DRM_DEBUG_DRIVER("Couldn't pin %d\n", ret);
294			goto err_destroy;
295		}
296	}
297
298	if (USES_FULL_PPGTT(dev)) {
299		struct i915_hw_ppgtt *ppgtt = i915_ppgtt_create(dev, file_priv);
300
301		if (IS_ERR_OR_NULL(ppgtt)) {
302			DRM_DEBUG_DRIVER("PPGTT setup failed (%ld)\n",
303					 PTR_ERR(ppgtt));
304			ret = PTR_ERR(ppgtt);
305			goto err_unpin;
306		}
307
308		ctx->ppgtt = ppgtt;
 
309	}
310
311	trace_i915_context_create(ctx);
312
313	return ctx;
314
315err_unpin:
316	if (is_global_default_ctx && ctx->legacy_hw_ctx.rcs_state)
317		i915_gem_object_ggtt_unpin(ctx->legacy_hw_ctx.rcs_state);
318err_destroy:
319	idr_remove(&file_priv->context_idr, ctx->user_handle);
320	i915_gem_context_unreference(ctx);
321	return ERR_PTR(ret);
322}
323
324static void i915_gem_context_unpin(struct intel_context *ctx,
325				   struct intel_engine_cs *engine)
326{
327	if (i915.enable_execlists) {
328		intel_lr_context_unpin(ctx, engine);
329	} else {
330		if (engine->id == RCS && ctx->legacy_hw_ctx.rcs_state)
331			i915_gem_object_ggtt_unpin(ctx->legacy_hw_ctx.rcs_state);
332		i915_gem_context_unreference(ctx);
333	}
334}
335
336void i915_gem_context_reset(struct drm_device *dev)
337{
338	struct drm_i915_private *dev_priv = dev->dev_private;
339	int i;
340
341	if (i915.enable_execlists) {
342		struct intel_context *ctx;
343
344		list_for_each_entry(ctx, &dev_priv->context_list, link)
345			intel_lr_context_reset(dev, ctx);
346	}
347
348	for (i = 0; i < I915_NUM_RINGS; i++) {
349		struct intel_engine_cs *ring = &dev_priv->ring[i];
 
350
351		if (ring->last_context) {
352			i915_gem_context_unpin(ring->last_context, ring);
353			ring->last_context = NULL;
354		}
355	}
 
356
357	/* Force the GPU state to be reinitialised on enabling */
358	dev_priv->kernel_context->legacy_hw_ctx.initialized = false;
 
 
359}
360
361int i915_gem_context_init(struct drm_device *dev)
 
362{
363	struct drm_i915_private *dev_priv = dev->dev_private;
364	struct intel_context *ctx;
365
366	/* Init should only be called once per module load. Eventually the
367	 * restriction on the context_disabled check can be loosened. */
368	if (WARN_ON(dev_priv->kernel_context))
369		return 0;
370
371	if (intel_vgpu_active(dev) && HAS_LOGICAL_RING_CONTEXTS(dev)) {
372		if (!i915.enable_execlists) {
373			DRM_INFO("Only EXECLIST mode is supported in vgpu.\n");
374			return -EINVAL;
375		}
376	}
377
378	if (i915.enable_execlists) {
379		/* NB: intentionally left blank. We will allocate our own
380		 * backing objects as we need them, thank you very much */
381		dev_priv->hw_context_size = 0;
382	} else if (HAS_HW_CONTEXTS(dev)) {
383		dev_priv->hw_context_size = round_up(get_context_size(dev), 4096);
384		if (dev_priv->hw_context_size > (1<<20)) {
385			DRM_DEBUG_DRIVER("Disabling HW Contexts; invalid size %d\n",
386					 dev_priv->hw_context_size);
387			dev_priv->hw_context_size = 0;
388		}
389	}
390
391	ctx = i915_gem_create_context(dev, NULL);
392	if (IS_ERR(ctx)) {
393		DRM_ERROR("Failed to create default global context (error %ld)\n",
394			  PTR_ERR(ctx));
395		return PTR_ERR(ctx);
396	}
397
398	dev_priv->kernel_context = ctx;
399
400	DRM_DEBUG_DRIVER("%s context support initialized\n",
401			i915.enable_execlists ? "LR" :
402			dev_priv->hw_context_size ? "HW" : "fake");
403	return 0;
404}
405
406void i915_gem_context_fini(struct drm_device *dev)
 
407{
408	struct drm_i915_private *dev_priv = dev->dev_private;
409	struct intel_context *dctx = dev_priv->kernel_context;
410	int i;
411
412	if (dctx->legacy_hw_ctx.rcs_state) {
413		/* The only known way to stop the gpu from accessing the hw context is
414		 * to reset it. Do this as the very last operation to avoid confusing
415		 * other code, leading to spurious errors. */
416		intel_gpu_reset(dev);
417
418		/* When default context is created and switched to, base object refcount
419		 * will be 2 (+1 from object creation and +1 from do_switch()).
420		 * i915_gem_context_fini() will be called after gpu_idle() has switched
421		 * to default context. So we need to unreference the base object once
422		 * to offset the do_switch part, so that i915_gem_context_unreference()
423		 * can then free the base object correctly. */
424		WARN_ON(!dev_priv->ring[RCS].last_context);
425
426		i915_gem_object_ggtt_unpin(dctx->legacy_hw_ctx.rcs_state);
427	}
428
429	for (i = I915_NUM_RINGS; --i >= 0;) {
430		struct intel_engine_cs *ring = &dev_priv->ring[i];
431
432		if (ring->last_context) {
433			i915_gem_context_unpin(ring->last_context, ring);
434			ring->last_context = NULL;
435		}
436	}
437
438	i915_gem_context_unreference(dctx);
439	dev_priv->kernel_context = NULL;
440}
441
442int i915_gem_context_enable(struct drm_i915_gem_request *req)
443{
444	struct intel_engine_cs *ring = req->ring;
445	int ret;
446
447	if (i915.enable_execlists) {
448		if (ring->init_context == NULL)
449			return 0;
450
451		ret = ring->init_context(req);
452	} else
453		ret = i915_switch_context(req);
454
455	if (ret) {
456		DRM_ERROR("ring init context: %d\n", ret);
457		return ret;
458	}
459
460	return 0;
461}
462
463static int context_idr_cleanup(int id, void *p, void *data)
464{
465	struct intel_context *ctx = p;
466
467	i915_gem_context_unreference(ctx);
468	return 0;
469}
470
471int i915_gem_context_open(struct drm_device *dev, struct drm_file *file)
472{
473	struct drm_i915_file_private *file_priv = file->driver_priv;
474	struct intel_context *ctx;
475
476	idr_init(&file_priv->context_idr);
477
478	mutex_lock(&dev->struct_mutex);
479	ctx = i915_gem_create_context(dev, file_priv);
480	mutex_unlock(&dev->struct_mutex);
481
 
 
482	if (IS_ERR(ctx)) {
483		idr_destroy(&file_priv->context_idr);
484		return PTR_ERR(ctx);
485	}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
486
 
 
 
487	return 0;
488}
489
490void i915_gem_context_close(struct drm_device *dev, struct drm_file *file)
491{
492	struct drm_i915_file_private *file_priv = file->driver_priv;
 
493
494	idr_for_each(&file_priv->context_idr, context_idr_cleanup, NULL);
495	idr_destroy(&file_priv->context_idr);
496}
497
498struct intel_context *
499i915_gem_context_get(struct drm_i915_file_private *file_priv, u32 id)
500{
501	struct intel_context *ctx;
502
503	ctx = (struct intel_context *)idr_find(&file_priv->context_idr, id);
504	if (!ctx)
505		return ERR_PTR(-ENOENT);
506
507	return ctx;
 
 
508}
509
510static inline int
511mi_set_context(struct drm_i915_gem_request *req, u32 hw_flags)
512{
513	struct intel_engine_cs *ring = req->ring;
514	u32 flags = hw_flags | MI_MM_SPACE_GTT;
515	const int num_rings =
516		/* Use an extended w/a on ivb+ if signalling from other rings */
517		i915_semaphore_is_enabled(ring->dev) ?
518		hweight32(INTEL_INFO(ring->dev)->ring_mask) - 1 :
519		0;
520	int len, i, ret;
521
522	/* w/a: If Flush TLB Invalidation Mode is enabled, driver must do a TLB
523	 * invalidation prior to MI_SET_CONTEXT. On GEN6 we don't set the value
524	 * explicitly, so we rely on the value at ring init, stored in
525	 * itlb_before_ctx_switch.
526	 */
527	if (IS_GEN6(ring->dev)) {
528		ret = ring->flush(req, I915_GEM_GPU_DOMAINS, 0);
529		if (ret)
530			return ret;
531	}
532
533	/* These flags are for resource streamer on HSW+ */
534	if (IS_HASWELL(ring->dev) || INTEL_INFO(ring->dev)->gen >= 8)
535		flags |= (HSW_MI_RS_SAVE_STATE_EN | HSW_MI_RS_RESTORE_STATE_EN);
536	else if (INTEL_INFO(ring->dev)->gen < 8)
537		flags |= (MI_SAVE_EXT_STATE_EN | MI_RESTORE_EXT_STATE_EN);
538
539
540	len = 4;
541	if (INTEL_INFO(ring->dev)->gen >= 7)
542		len += 2 + (num_rings ? 4*num_rings + 2 : 0);
543
544	ret = intel_ring_begin(req, len);
545	if (ret)
546		return ret;
547
548	/* WaProgramMiArbOnOffAroundMiSetContext:ivb,vlv,hsw,bdw,chv */
549	if (INTEL_INFO(ring->dev)->gen >= 7) {
550		intel_ring_emit(ring, MI_ARB_ON_OFF | MI_ARB_DISABLE);
551		if (num_rings) {
552			struct intel_engine_cs *signaller;
553
554			intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(num_rings));
555			for_each_ring(signaller, to_i915(ring->dev), i) {
556				if (signaller == ring)
557					continue;
558
559				intel_ring_emit_reg(ring, RING_PSMI_CTL(signaller->mmio_base));
560				intel_ring_emit(ring, _MASKED_BIT_ENABLE(GEN6_PSMI_SLEEP_MSG_DISABLE));
561			}
562		}
563	}
564
565	intel_ring_emit(ring, MI_NOOP);
566	intel_ring_emit(ring, MI_SET_CONTEXT);
567	intel_ring_emit(ring, i915_gem_obj_ggtt_offset(req->ctx->legacy_hw_ctx.rcs_state) |
568			flags);
569	/*
570	 * w/a: MI_SET_CONTEXT must always be followed by MI_NOOP
571	 * WaMiSetContext_Hang:snb,ivb,vlv
572	 */
573	intel_ring_emit(ring, MI_NOOP);
574
575	if (INTEL_INFO(ring->dev)->gen >= 7) {
576		if (num_rings) {
577			struct intel_engine_cs *signaller;
578
579			intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(num_rings));
580			for_each_ring(signaller, to_i915(ring->dev), i) {
581				if (signaller == ring)
582					continue;
583
584				intel_ring_emit_reg(ring, RING_PSMI_CTL(signaller->mmio_base));
585				intel_ring_emit(ring, _MASKED_BIT_DISABLE(GEN6_PSMI_SLEEP_MSG_DISABLE));
586			}
587		}
588		intel_ring_emit(ring, MI_ARB_ON_OFF | MI_ARB_ENABLE);
589	}
590
591	intel_ring_advance(ring);
592
593	return ret;
594}
595
596static inline bool should_skip_switch(struct intel_engine_cs *ring,
597				      struct intel_context *from,
598				      struct intel_context *to)
599{
600	if (to->remap_slice)
601		return false;
602
603	if (to->ppgtt && from == to &&
604	    !(intel_ring_flag(ring) & to->ppgtt->pd_dirty_rings))
605		return true;
606
607	return false;
 
608}
609
610static bool
611needs_pd_load_pre(struct intel_engine_cs *ring, struct intel_context *to)
612{
613	struct drm_i915_private *dev_priv = ring->dev->dev_private;
 
614
615	if (!to->ppgtt)
616		return false;
617
618	if (INTEL_INFO(ring->dev)->gen < 8)
619		return true;
 
 
 
 
 
620
621	if (ring != &dev_priv->ring[RCS])
622		return true;
623
624	return false;
625}
626
627static bool
628needs_pd_load_post(struct intel_engine_cs *ring, struct intel_context *to,
629		u32 hw_flags)
630{
631	struct drm_i915_private *dev_priv = ring->dev->dev_private;
632
633	if (!to->ppgtt)
634		return false;
635
636	if (!IS_GEN8(ring->dev))
637		return false;
638
639	if (ring != &dev_priv->ring[RCS])
640		return false;
641
642	if (hw_flags & MI_RESTORE_INHIBIT)
643		return true;
644
645	return false;
 
646}
647
648static int do_switch(struct drm_i915_gem_request *req)
649{
650	struct intel_context *to = req->ctx;
651	struct intel_engine_cs *ring = req->ring;
652	struct drm_i915_private *dev_priv = ring->dev->dev_private;
653	struct intel_context *from = ring->last_context;
654	u32 hw_flags = 0;
655	bool uninitialized = false;
656	int ret, i;
657
658	if (from != NULL && ring == &dev_priv->ring[RCS]) {
659		BUG_ON(from->legacy_hw_ctx.rcs_state == NULL);
660		BUG_ON(!i915_gem_obj_is_pinned(from->legacy_hw_ctx.rcs_state));
661	}
662
663	if (should_skip_switch(ring, from, to))
664		return 0;
665
666	/* Trying to pin first makes error handling easier. */
667	if (ring == &dev_priv->ring[RCS]) {
668		ret = i915_gem_obj_ggtt_pin(to->legacy_hw_ctx.rcs_state,
669					    get_context_alignment(ring->dev), 0);
670		if (ret)
671			return ret;
672	}
673
674	/*
675	 * Pin can switch back to the default context if we end up calling into
676	 * evict_everything - as a last ditch gtt defrag effort that also
677	 * switches to the default context. Hence we need to reload from here.
678	 */
679	from = ring->last_context;
680
681	if (needs_pd_load_pre(ring, to)) {
682		/* Older GENs and non render rings still want the load first,
683		 * "PP_DCLV followed by PP_DIR_BASE register through Load
684		 * Register Immediate commands in Ring Buffer before submitting
685		 * a context."*/
686		trace_switch_mm(ring, to);
687		ret = to->ppgtt->switch_mm(to->ppgtt, req);
688		if (ret)
689			goto unpin_out;
690
691		/* Doing a PD load always reloads the page dirs */
692		to->ppgtt->pd_dirty_rings &= ~intel_ring_flag(ring);
693	}
694
695	if (ring != &dev_priv->ring[RCS]) {
696		if (from)
697			i915_gem_context_unreference(from);
698		goto done;
699	}
700
701	/*
702	 * Clear this page out of any CPU caches for coherent swap-in/out. Note
703	 * that thanks to write = false in this call and us not setting any gpu
704	 * write domains when putting a context object onto the active list
705	 * (when switching away from it), this won't block.
706	 *
707	 * XXX: We need a real interface to do this instead of trickery.
708	 */
709	ret = i915_gem_object_set_to_gtt_domain(to->legacy_hw_ctx.rcs_state, false);
710	if (ret)
711		goto unpin_out;
712
713	if (!to->legacy_hw_ctx.initialized || i915_gem_context_is_default(to)) {
714		hw_flags |= MI_RESTORE_INHIBIT;
715		/* NB: If we inhibit the restore, the context is not allowed to
716		 * die because future work may end up depending on valid address
717		 * space. This means we must enforce that a page table load
718		 * occur when this occurs. */
719	} else if (to->ppgtt &&
720		   (intel_ring_flag(ring) & to->ppgtt->pd_dirty_rings)) {
721		hw_flags |= MI_FORCE_RESTORE;
722		to->ppgtt->pd_dirty_rings &= ~intel_ring_flag(ring);
723	}
724
725	/* We should never emit switch_mm more than once */
726	WARN_ON(needs_pd_load_pre(ring, to) &&
727		needs_pd_load_post(ring, to, hw_flags));
728
729	ret = mi_set_context(req, hw_flags);
730	if (ret)
731		goto unpin_out;
732
733	/* GEN8 does *not* require an explicit reload if the PDPs have been
734	 * setup, and we do not wish to move them.
735	 */
736	if (needs_pd_load_post(ring, to, hw_flags)) {
737		trace_switch_mm(ring, to);
738		ret = to->ppgtt->switch_mm(to->ppgtt, req);
739		/* The hardware context switch is emitted, but we haven't
740		 * actually changed the state - so it's probably safe to bail
741		 * here. Still, let the user know something dangerous has
742		 * happened.
743		 */
744		if (ret) {
745			DRM_ERROR("Failed to change address space on context switch\n");
746			goto unpin_out;
747		}
748	}
749
750	for (i = 0; i < MAX_L3_SLICES; i++) {
751		if (!(to->remap_slice & (1<<i)))
752			continue;
753
754		ret = i915_gem_l3_remap(req, i);
755		/* If it failed, try again next round */
756		if (ret)
757			DRM_DEBUG_DRIVER("L3 remapping failed\n");
758		else
759			to->remap_slice &= ~(1<<i);
760	}
761
762	/* The backing object for the context is done after switching to the
763	 * *next* context. Therefore we cannot retire the previous context until
764	 * the next context has already started running. In fact, the below code
765	 * is a bit suboptimal because the retiring can occur simply after the
766	 * MI_SET_CONTEXT instead of when the next seqno has completed.
767	 */
768	if (from != NULL) {
769		from->legacy_hw_ctx.rcs_state->base.read_domains = I915_GEM_DOMAIN_INSTRUCTION;
770		i915_vma_move_to_active(i915_gem_obj_to_ggtt(from->legacy_hw_ctx.rcs_state), req);
771		/* As long as MI_SET_CONTEXT is serializing, ie. it flushes the
772		 * whole damn pipeline, we don't need to explicitly mark the
773		 * object dirty. The only exception is that the context must be
774		 * correct in case the object gets swapped out. Ideally we'd be
775		 * able to defer doing this until we know the object would be
776		 * swapped, but there is no way to do that yet.
 
777		 */
778		from->legacy_hw_ctx.rcs_state->dirty = 1;
779
780		/* obj is kept alive until the next request by its active ref */
781		i915_gem_object_ggtt_unpin(from->legacy_hw_ctx.rcs_state);
782		i915_gem_context_unreference(from);
783	}
784
785	uninitialized = !to->legacy_hw_ctx.initialized;
786	to->legacy_hw_ctx.initialized = true;
787
788done:
789	i915_gem_context_reference(to);
790	ring->last_context = to;
791
792	if (uninitialized) {
793		if (ring->init_context) {
794			ret = ring->init_context(req);
795			if (ret)
796				DRM_ERROR("ring init context: %d\n", ret);
797		}
798	}
799
800	return 0;
801
802unpin_out:
803	if (ring->id == RCS)
804		i915_gem_object_ggtt_unpin(to->legacy_hw_ctx.rcs_state);
805	return ret;
806}
807
808/**
809 * i915_switch_context() - perform a GPU context switch.
810 * @req: request for which we'll execute the context switch
811 *
812 * The context life cycle is simple. The context refcount is incremented and
813 * decremented by 1 and create and destroy. If the context is in use by the GPU,
814 * it will have a refcount > 1. This allows us to destroy the context abstract
815 * object while letting the normal object tracking destroy the backing BO.
816 *
817 * This function should not be used in execlists mode.  Instead the context is
818 * switched by writing to the ELSP and requests keep a reference to their
819 * context.
820 */
821int i915_switch_context(struct drm_i915_gem_request *req)
822{
823	struct intel_engine_cs *ring = req->ring;
824	struct drm_i915_private *dev_priv = ring->dev->dev_private;
825
826	WARN_ON(i915.enable_execlists);
827	WARN_ON(!mutex_is_locked(&dev_priv->dev->struct_mutex));
828
829	if (req->ctx->legacy_hw_ctx.rcs_state == NULL) { /* We have the fake context */
830		if (req->ctx != ring->last_context) {
831			i915_gem_context_reference(req->ctx);
832			if (ring->last_context)
833				i915_gem_context_unreference(ring->last_context);
834			ring->last_context = req->ctx;
835		}
836		return 0;
837	}
838
839	return do_switch(req);
840}
841
842static bool contexts_enabled(struct drm_device *dev)
843{
844	return i915.enable_execlists || to_i915(dev)->hw_context_size;
845}
846
847int i915_gem_context_create_ioctl(struct drm_device *dev, void *data,
848				  struct drm_file *file)
849{
 
850	struct drm_i915_gem_context_create *args = data;
851	struct drm_i915_file_private *file_priv = file->driver_priv;
852	struct intel_context *ctx;
853	int ret;
854
855	if (!contexts_enabled(dev))
856		return -ENODEV;
857
858	if (args->pad != 0)
859		return -EINVAL;
860
 
 
 
 
 
 
 
 
861	ret = i915_mutex_lock_interruptible(dev);
862	if (ret)
863		return ret;
864
865	ctx = i915_gem_create_context(dev, file_priv);
866	mutex_unlock(&dev->struct_mutex);
867	if (IS_ERR(ctx))
868		return PTR_ERR(ctx);
869
 
 
870	args->ctx_id = ctx->user_handle;
871	DRM_DEBUG_DRIVER("HW context %d created\n", args->ctx_id);
872
873	return 0;
874}
875
876int i915_gem_context_destroy_ioctl(struct drm_device *dev, void *data,
877				   struct drm_file *file)
878{
879	struct drm_i915_gem_context_destroy *args = data;
880	struct drm_i915_file_private *file_priv = file->driver_priv;
881	struct intel_context *ctx;
882	int ret;
883
884	if (args->pad != 0)
885		return -EINVAL;
886
887	if (args->ctx_id == DEFAULT_CONTEXT_HANDLE)
888		return -ENOENT;
889
890	ret = i915_mutex_lock_interruptible(dev);
891	if (ret)
892		return ret;
893
894	ctx = i915_gem_context_get(file_priv, args->ctx_id);
895	if (IS_ERR(ctx)) {
896		mutex_unlock(&dev->struct_mutex);
897		return PTR_ERR(ctx);
898	}
899
900	idr_remove(&ctx->file_priv->context_idr, ctx->user_handle);
901	i915_gem_context_unreference(ctx);
902	mutex_unlock(&dev->struct_mutex);
903
904	DRM_DEBUG_DRIVER("HW context %d destroyed\n", args->ctx_id);
 
905	return 0;
906}
907
908int i915_gem_context_getparam_ioctl(struct drm_device *dev, void *data,
909				    struct drm_file *file)
910{
911	struct drm_i915_file_private *file_priv = file->driver_priv;
912	struct drm_i915_gem_context_param *args = data;
913	struct intel_context *ctx;
914	int ret;
915
916	ret = i915_mutex_lock_interruptible(dev);
917	if (ret)
918		return ret;
919
920	ctx = i915_gem_context_get(file_priv, args->ctx_id);
921	if (IS_ERR(ctx)) {
922		mutex_unlock(&dev->struct_mutex);
923		return PTR_ERR(ctx);
924	}
925
926	args->size = 0;
927	switch (args->param) {
928	case I915_CONTEXT_PARAM_BAN_PERIOD:
929		args->value = ctx->hang_stats.ban_period_seconds;
930		break;
931	case I915_CONTEXT_PARAM_NO_ZEROMAP:
932		args->value = ctx->flags & CONTEXT_NO_ZEROMAP;
933		break;
934	case I915_CONTEXT_PARAM_GTT_SIZE:
935		if (ctx->ppgtt)
936			args->value = ctx->ppgtt->base.total;
937		else if (to_i915(dev)->mm.aliasing_ppgtt)
938			args->value = to_i915(dev)->mm.aliasing_ppgtt->base.total;
939		else
940			args->value = to_i915(dev)->gtt.base.total;
 
 
 
 
 
 
 
 
 
941		break;
942	default:
943		ret = -EINVAL;
944		break;
945	}
946	mutex_unlock(&dev->struct_mutex);
947
 
948	return ret;
949}
950
951int i915_gem_context_setparam_ioctl(struct drm_device *dev, void *data,
952				    struct drm_file *file)
953{
954	struct drm_i915_file_private *file_priv = file->driver_priv;
955	struct drm_i915_gem_context_param *args = data;
956	struct intel_context *ctx;
957	int ret;
958
 
 
 
 
959	ret = i915_mutex_lock_interruptible(dev);
960	if (ret)
961		return ret;
962
963	ctx = i915_gem_context_get(file_priv, args->ctx_id);
964	if (IS_ERR(ctx)) {
965		mutex_unlock(&dev->struct_mutex);
966		return PTR_ERR(ctx);
967	}
968
969	switch (args->param) {
970	case I915_CONTEXT_PARAM_BAN_PERIOD:
971		if (args->size)
972			ret = -EINVAL;
973		else if (args->value < ctx->hang_stats.ban_period_seconds &&
974			 !capable(CAP_SYS_ADMIN))
975			ret = -EPERM;
976		else
977			ctx->hang_stats.ban_period_seconds = args->value;
978		break;
979	case I915_CONTEXT_PARAM_NO_ZEROMAP:
980		if (args->size) {
981			ret = -EINVAL;
982		} else {
983			ctx->flags &= ~CONTEXT_NO_ZEROMAP;
984			ctx->flags |= args->value ? CONTEXT_NO_ZEROMAP : 0;
985		}
986		break;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
987	default:
988		ret = -EINVAL;
989		break;
990	}
991	mutex_unlock(&dev->struct_mutex);
992
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
993	return ret;
994}
v4.17
  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 <linux/log2.h>
 89#include <drm/drmP.h>
 90#include <drm/i915_drm.h>
 91#include "i915_drv.h"
 92#include "i915_trace.h"
 93
 94#define ALL_L3_SLICES(dev) (1 << NUM_L3_SLICES(dev)) - 1
 
 
 
 
 
 95
 96static void lut_close(struct i915_gem_context *ctx)
 97{
 98	struct i915_lut_handle *lut, *ln;
 99	struct radix_tree_iter iter;
100	void __rcu **slot;
101
102	list_for_each_entry_safe(lut, ln, &ctx->handles_list, ctx_link) {
103		list_del(&lut->obj_link);
104		kmem_cache_free(ctx->i915->luts, lut);
105	}
106
107	rcu_read_lock();
108	radix_tree_for_each_slot(slot, &ctx->handles_vma, &iter, 0) {
109		struct i915_vma *vma = rcu_dereference_raw(*slot);
110
111		radix_tree_iter_delete(&ctx->handles_vma, &iter, slot);
112		__i915_gem_object_release_unless_active(vma->obj);
113	}
114	rcu_read_unlock();
115}
116
117static void i915_gem_context_free(struct i915_gem_context *ctx)
118{
119	int i;
 
 
120
121	lockdep_assert_held(&ctx->i915->drm.struct_mutex);
122	GEM_BUG_ON(!i915_gem_context_is_closed(ctx));
123
124	i915_ppgtt_put(ctx->ppgtt);
125
126	for (i = 0; i < I915_NUM_ENGINES; i++) {
127		struct intel_context *ce = &ctx->engine[i];
128
129		if (!ce->state)
130			continue;
131
132		WARN_ON(ce->pin_count);
133		if (ce->ring)
134			intel_ring_free(ce->ring);
135
136		__i915_gem_object_release_unless_active(ce->state->obj);
 
137	}
138
139	kfree(ctx->name);
140	put_pid(ctx->pid);
141
142	list_del(&ctx->link);
143
144	ida_simple_remove(&ctx->i915->contexts.hw_ida, ctx->hw_id);
145	kfree_rcu(ctx, rcu);
146}
147
148static void contexts_free(struct drm_i915_private *i915)
149{
150	struct llist_node *freed = llist_del_all(&i915->contexts.free_list);
151	struct i915_gem_context *ctx, *cn;
152
153	lockdep_assert_held(&i915->drm.struct_mutex);
154
155	llist_for_each_entry_safe(ctx, cn, freed, free_link)
156		i915_gem_context_free(ctx);
157}
158
159static void contexts_free_first(struct drm_i915_private *i915)
160{
161	struct i915_gem_context *ctx;
162	struct llist_node *freed;
163
164	lockdep_assert_held(&i915->drm.struct_mutex);
165
166	freed = llist_del_first(&i915->contexts.free_list);
167	if (!freed)
168		return;
169
170	ctx = container_of(freed, typeof(*ctx), free_link);
171	i915_gem_context_free(ctx);
 
 
 
172}
173
174static void contexts_free_worker(struct work_struct *work)
175{
176	struct drm_i915_private *i915 =
177		container_of(work, typeof(*i915), contexts.free_work);
178
179	mutex_lock(&i915->drm.struct_mutex);
180	contexts_free(i915);
181	mutex_unlock(&i915->drm.struct_mutex);
182}
183
184void i915_gem_context_release(struct kref *ref)
185{
186	struct i915_gem_context *ctx = container_of(ref, typeof(*ctx), ref);
187	struct drm_i915_private *i915 = ctx->i915;
188
189	trace_i915_context_free(ctx);
190	if (llist_add(&ctx->free_link, &i915->contexts.free_list))
191		queue_work(i915->wq, &i915->contexts.free_work);
192}
193
194static void context_close(struct i915_gem_context *ctx)
195{
196	i915_gem_context_set_closed(ctx);
197
198	/*
199	 * The LUT uses the VMA as a backpointer to unref the object,
200	 * so we need to clear the LUT before we close all the VMA (inside
201	 * the ppgtt).
202	 */
203	lut_close(ctx);
204	if (ctx->ppgtt)
205		i915_ppgtt_close(&ctx->ppgtt->base);
206
207	ctx->file_priv = ERR_PTR(-EBADF);
208	i915_gem_context_put(ctx);
 
 
209}
210
211static int assign_hw_id(struct drm_i915_private *dev_priv, unsigned *out)
 
212{
 
213	int ret;
214	unsigned int max;
215
216	if (INTEL_GEN(dev_priv) >= 11)
217		max = GEN11_MAX_CONTEXT_HW_ID;
218	else
219		max = MAX_CONTEXT_HW_ID;
220
221	ret = ida_simple_get(&dev_priv->contexts.hw_ida,
222			     0, max, GFP_KERNEL);
223	if (ret < 0) {
224		/* Contexts are only released when no longer active.
225		 * Flush any pending retires to hopefully release some
226		 * stale contexts and try again.
227		 */
228		i915_retire_requests(dev_priv);
229		ret = ida_simple_get(&dev_priv->contexts.hw_ida,
230				     0, max, GFP_KERNEL);
231		if (ret < 0)
232			return ret;
233	}
234
235	*out = ret;
236	return 0;
237}
238
239static u32 default_desc_template(const struct drm_i915_private *i915,
240				 const struct i915_hw_ppgtt *ppgtt)
241{
242	u32 address_mode;
243	u32 desc;
244
245	desc = GEN8_CTX_VALID | GEN8_CTX_PRIVILEGE;
246
247	address_mode = INTEL_LEGACY_32B_CONTEXT;
248	if (ppgtt && i915_vm_is_48bit(&ppgtt->base))
249		address_mode = INTEL_LEGACY_64B_CONTEXT;
250	desc |= address_mode << GEN8_CTX_ADDRESSING_MODE_SHIFT;
251
252	if (IS_GEN8(i915))
253		desc |= GEN8_CTX_L3LLC_COHERENT;
254
255	/* TODO: WaDisableLiteRestore when we start using semaphore
256	 * signalling between Command Streamers
257	 * ring->ctx_desc_template |= GEN8_CTX_FORCE_RESTORE;
258	 */
 
 
 
 
 
 
 
 
259
260	return desc;
261}
262
263static struct i915_gem_context *
264__create_hw_context(struct drm_i915_private *dev_priv,
265		    struct drm_i915_file_private *file_priv)
266{
267	struct i915_gem_context *ctx;
 
268	int ret;
269
270	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
271	if (ctx == NULL)
272		return ERR_PTR(-ENOMEM);
273
274	ret = assign_hw_id(dev_priv, &ctx->hw_id);
275	if (ret) {
276		kfree(ctx);
277		return ERR_PTR(ret);
278	}
279
280	kref_init(&ctx->ref);
281	list_add_tail(&ctx->link, &dev_priv->contexts.list);
282	ctx->i915 = dev_priv;
283	ctx->priority = I915_PRIORITY_NORMAL;
284
285	INIT_RADIX_TREE(&ctx->handles_vma, GFP_KERNEL);
286	INIT_LIST_HEAD(&ctx->handles_list);
 
 
 
 
 
 
 
287
288	/* Default context will never have a file_priv */
289	ret = DEFAULT_CONTEXT_HANDLE;
290	if (file_priv) {
291		ret = idr_alloc(&file_priv->context_idr, ctx,
292				DEFAULT_CONTEXT_HANDLE, 0, GFP_KERNEL);
293		if (ret < 0)
294			goto err_lut;
295	}
296	ctx->user_handle = ret;
297
298	ctx->file_priv = file_priv;
299	if (file_priv) {
300		ctx->pid = get_task_pid(current, PIDTYPE_PID);
301		ctx->name = kasprintf(GFP_KERNEL, "%s[%d]/%x",
302				      current->comm,
303				      pid_nr(ctx->pid),
304				      ctx->user_handle);
305		if (!ctx->name) {
306			ret = -ENOMEM;
307			goto err_pid;
308		}
309	}
310
311	/* NB: Mark all slices as needing a remap so that when the context first
312	 * loads it will restore whatever remap state already exists. If there
313	 * is no remap info, it will be a NOP. */
314	ctx->remap_slice = ALL_L3_SLICES(dev_priv);
315
316	i915_gem_context_set_bannable(ctx);
317	ctx->ring_size = 4 * PAGE_SIZE;
318	ctx->desc_template =
319		default_desc_template(dev_priv, dev_priv->mm.aliasing_ppgtt);
320
321	/* GuC requires the ring to be placed above GUC_WOPCM_TOP. If GuC is not
322	 * present or not in use we still need a small bias as ring wraparound
323	 * at offset 0 sometimes hangs. No idea why.
324	 */
325	if (USES_GUC(dev_priv))
326		ctx->ggtt_offset_bias = GUC_WOPCM_TOP;
327	else
328		ctx->ggtt_offset_bias = I915_GTT_PAGE_SIZE;
329
330	return ctx;
331
332err_pid:
333	put_pid(ctx->pid);
334	idr_remove(&file_priv->context_idr, ctx->user_handle);
335err_lut:
336	context_close(ctx);
337	return ERR_PTR(ret);
338}
339
340static void __destroy_hw_context(struct i915_gem_context *ctx,
341				 struct drm_i915_file_private *file_priv)
342{
343	idr_remove(&file_priv->context_idr, ctx->user_handle);
344	context_close(ctx);
345}
346
347static struct i915_gem_context *
348i915_gem_create_context(struct drm_i915_private *dev_priv,
349			struct drm_i915_file_private *file_priv)
350{
351	struct i915_gem_context *ctx;
352
353	lockdep_assert_held(&dev_priv->drm.struct_mutex);
354
355	/* Reap the most stale context */
356	contexts_free_first(dev_priv);
357
358	ctx = __create_hw_context(dev_priv, file_priv);
359	if (IS_ERR(ctx))
360		return ctx;
361
362	if (USES_FULL_PPGTT(dev_priv)) {
363		struct i915_hw_ppgtt *ppgtt;
 
 
 
 
 
 
 
 
 
 
 
 
 
364
365		ppgtt = i915_ppgtt_create(dev_priv, file_priv, ctx->name);
366		if (IS_ERR(ppgtt)) {
 
 
367			DRM_DEBUG_DRIVER("PPGTT setup failed (%ld)\n",
368					 PTR_ERR(ppgtt));
369			__destroy_hw_context(ctx, file_priv);
370			return ERR_CAST(ppgtt);
371		}
372
373		ctx->ppgtt = ppgtt;
374		ctx->desc_template = default_desc_template(dev_priv, ppgtt);
375	}
376
377	trace_i915_context_create(ctx);
378
379	return ctx;
 
 
 
 
 
 
 
 
380}
381
382/**
383 * i915_gem_context_create_gvt - create a GVT GEM context
384 * @dev: drm device *
385 *
386 * This function is used to create a GVT specific GEM context.
387 *
388 * Returns:
389 * pointer to i915_gem_context on success, error pointer if failed
390 *
391 */
392struct i915_gem_context *
393i915_gem_context_create_gvt(struct drm_device *dev)
 
394{
395	struct i915_gem_context *ctx;
396	int ret;
397
398	if (!IS_ENABLED(CONFIG_DRM_I915_GVT))
399		return ERR_PTR(-ENODEV);
400
401	ret = i915_mutex_lock_interruptible(dev);
402	if (ret)
403		return ERR_PTR(ret);
404
405	ctx = __create_hw_context(to_i915(dev), NULL);
406	if (IS_ERR(ctx))
407		goto out;
408
409	ctx->file_priv = ERR_PTR(-EBADF);
410	i915_gem_context_set_closed(ctx); /* not user accessible */
411	i915_gem_context_clear_bannable(ctx);
412	i915_gem_context_set_force_single_submission(ctx);
413	if (!USES_GUC_SUBMISSION(to_i915(dev)))
414		ctx->ring_size = 512 * PAGE_SIZE; /* Max ring buffer size */
415
416	GEM_BUG_ON(i915_gem_context_is_kernel(ctx));
417out:
418	mutex_unlock(&dev->struct_mutex);
419	return ctx;
420}
421
422struct i915_gem_context *
423i915_gem_context_create_kernel(struct drm_i915_private *i915, int prio)
424{
425	struct i915_gem_context *ctx;
 
 
 
 
 
 
 
 
 
 
 
 
 
426
427	ctx = i915_gem_create_context(i915, NULL);
428	if (IS_ERR(ctx))
429		return ctx;
 
 
 
 
 
 
 
 
 
430
431	i915_gem_context_clear_bannable(ctx);
432	ctx->priority = prio;
433	ctx->ring_size = PAGE_SIZE;
 
 
 
434
435	GEM_BUG_ON(!i915_gem_context_is_kernel(ctx));
436
437	return ctx;
 
 
 
438}
439
440static void
441destroy_kernel_context(struct i915_gem_context **ctxp)
442{
443	struct i915_gem_context *ctx;
 
 
444
445	/* Keep the context ref so that we can free it immediately ourselves */
446	ctx = i915_gem_context_get(fetch_and_zero(ctxp));
447	GEM_BUG_ON(!i915_gem_context_is_kernel(ctx));
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
448
449	context_close(ctx);
450	i915_gem_context_free(ctx);
451}
452
453static bool needs_preempt_context(struct drm_i915_private *i915)
454{
455	return HAS_LOGICAL_RING_PREEMPTION(i915);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
456}
457
458int i915_gem_contexts_init(struct drm_i915_private *dev_priv)
459{
460	struct i915_gem_context *ctx;
461
462	/* Reassure ourselves we are only called once */
463	GEM_BUG_ON(dev_priv->kernel_context);
464	GEM_BUG_ON(dev_priv->preempt_context);
465
466	INIT_LIST_HEAD(&dev_priv->contexts.list);
467	INIT_WORK(&dev_priv->contexts.free_work, contexts_free_worker);
468	init_llist_head(&dev_priv->contexts.free_list);
 
469
470	/* Using the simple ida interface, the max is limited by sizeof(int) */
471	BUILD_BUG_ON(MAX_CONTEXT_HW_ID > INT_MAX);
472	BUILD_BUG_ON(GEN11_MAX_CONTEXT_HW_ID > INT_MAX);
473	ida_init(&dev_priv->contexts.hw_ida);
 
474
475	/* lowest priority; idle task */
476	ctx = i915_gem_context_create_kernel(dev_priv, I915_PRIORITY_MIN);
477	if (IS_ERR(ctx)) {
478		DRM_ERROR("Failed to create default global context\n");
479		return PTR_ERR(ctx);
480	}
481	/*
482	 * For easy recognisablity, we want the kernel context to be 0 and then
483	 * all user contexts will have non-zero hw_id.
484	 */
485	GEM_BUG_ON(ctx->hw_id);
486	dev_priv->kernel_context = ctx;
487
488	/* highest priority; preempting task */
489	if (needs_preempt_context(dev_priv)) {
490		ctx = i915_gem_context_create_kernel(dev_priv, INT_MAX);
491		if (!IS_ERR(ctx))
492			dev_priv->preempt_context = ctx;
493		else
494			DRM_ERROR("Failed to create preempt context; disabling preemption\n");
495	}
496
497	DRM_DEBUG_DRIVER("%s context support initialized\n",
498			 dev_priv->engine[RCS]->context_size ? "logical" :
499			 "fake");
500	return 0;
501}
502
503void i915_gem_contexts_lost(struct drm_i915_private *dev_priv)
504{
505	struct intel_engine_cs *engine;
506	enum intel_engine_id id;
507
508	lockdep_assert_held(&dev_priv->drm.struct_mutex);
 
 
509
510	for_each_engine(engine, dev_priv, id) {
511		engine->legacy_active_context = NULL;
512		engine->legacy_active_ppgtt = NULL;
 
513
514		if (!engine->last_retired_context)
515			continue;
 
516
517		engine->context_unpin(engine, engine->last_retired_context);
518		engine->last_retired_context = NULL;
519	}
520}
521
522void i915_gem_contexts_fini(struct drm_i915_private *i915)
 
523{
524	lockdep_assert_held(&i915->drm.struct_mutex);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
525
526	if (i915->preempt_context)
527		destroy_kernel_context(&i915->preempt_context);
528	destroy_kernel_context(&i915->kernel_context);
529
530	/* Must free all deferred contexts (via flush_workqueue) first */
531	ida_destroy(&i915->contexts.hw_ida);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
532}
533
534static int context_idr_cleanup(int id, void *p, void *data)
 
 
535{
536	struct i915_gem_context *ctx = p;
 
 
 
 
 
537
538	context_close(ctx);
539	return 0;
540}
541
542int i915_gem_context_open(struct drm_i915_private *i915,
543			  struct drm_file *file)
544{
545	struct drm_i915_file_private *file_priv = file->driver_priv;
546	struct i915_gem_context *ctx;
547
548	idr_init(&file_priv->context_idr);
 
549
550	mutex_lock(&i915->drm.struct_mutex);
551	ctx = i915_gem_create_context(i915, file_priv);
552	mutex_unlock(&i915->drm.struct_mutex);
553	if (IS_ERR(ctx)) {
554		idr_destroy(&file_priv->context_idr);
555		return PTR_ERR(ctx);
556	}
557
558	GEM_BUG_ON(i915_gem_context_is_kernel(ctx));
 
559
560	return 0;
561}
562
563void i915_gem_context_close(struct drm_file *file)
 
 
564{
565	struct drm_i915_file_private *file_priv = file->driver_priv;
 
 
 
 
 
 
 
 
 
566
567	lockdep_assert_held(&file_priv->dev_priv->drm.struct_mutex);
 
568
569	idr_for_each(&file_priv->context_idr, context_idr_cleanup, NULL);
570	idr_destroy(&file_priv->context_idr);
571}
572
573static bool engine_has_idle_kernel_context(struct intel_engine_cs *engine)
574{
575	struct i915_gem_timeline *timeline;
 
 
 
 
 
 
576
577	list_for_each_entry(timeline, &engine->i915->gt.timelines, link) {
578		struct intel_timeline *tl;
 
 
579
580		if (timeline == &engine->i915->gt.global_timeline)
581			continue;
582
583		tl = &timeline->engine[engine->id];
584		if (i915_gem_active_peek(&tl->last_request,
585					 &engine->i915->drm.struct_mutex))
586			return false;
 
 
587	}
588
589	return intel_engine_has_kernel_context(engine);
590}
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
591
592int i915_gem_switch_to_kernel_context(struct drm_i915_private *dev_priv)
593{
594	struct intel_engine_cs *engine;
595	struct i915_gem_timeline *timeline;
596	enum intel_engine_id id;
 
 
 
 
 
 
597
598	lockdep_assert_held(&dev_priv->drm.struct_mutex);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
599
600	i915_retire_requests(dev_priv);
 
 
601
602	for_each_engine(engine, dev_priv, id) {
603		struct i915_request *rq;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
604
605		if (engine_has_idle_kernel_context(engine))
 
606			continue;
607
608		rq = i915_request_alloc(engine, dev_priv->kernel_context);
609		if (IS_ERR(rq))
610			return PTR_ERR(rq);
611
612		/* Queue this switch after all other activity */
613		list_for_each_entry(timeline, &dev_priv->gt.timelines, link) {
614			struct i915_request *prev;
615			struct intel_timeline *tl;
616
617			tl = &timeline->engine[engine->id];
618			prev = i915_gem_active_raw(&tl->last_request,
619						   &dev_priv->drm.struct_mutex);
620			if (prev)
621				i915_sw_fence_await_sw_fence_gfp(&rq->submit,
622								 &prev->submit,
623								 I915_FENCE_GFP);
624		}
625
626		/*
627		 * Force a flush after the switch to ensure that all rendering
628		 * and operations prior to switching to the kernel context hits
629		 * memory. This should be guaranteed by the previous request,
630		 * but an extra layer of paranoia before we declare the system
631		 * idle (on suspend etc) is advisable!
632		 */
633		__i915_request_add(rq, true);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
634	}
635
636	return 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
637}
638
639static bool client_is_banned(struct drm_i915_file_private *file_priv)
640{
641	return atomic_read(&file_priv->context_bans) > I915_MAX_CLIENT_CONTEXT_BANS;
642}
643
644int i915_gem_context_create_ioctl(struct drm_device *dev, void *data,
645				  struct drm_file *file)
646{
647	struct drm_i915_private *dev_priv = to_i915(dev);
648	struct drm_i915_gem_context_create *args = data;
649	struct drm_i915_file_private *file_priv = file->driver_priv;
650	struct i915_gem_context *ctx;
651	int ret;
652
653	if (!dev_priv->engine[RCS]->context_size)
654		return -ENODEV;
655
656	if (args->pad != 0)
657		return -EINVAL;
658
659	if (client_is_banned(file_priv)) {
660		DRM_DEBUG("client %s[%d] banned from creating ctx\n",
661			  current->comm,
662			  pid_nr(get_task_pid(current, PIDTYPE_PID)));
663
664		return -EIO;
665	}
666
667	ret = i915_mutex_lock_interruptible(dev);
668	if (ret)
669		return ret;
670
671	ctx = i915_gem_create_context(dev_priv, file_priv);
672	mutex_unlock(&dev->struct_mutex);
673	if (IS_ERR(ctx))
674		return PTR_ERR(ctx);
675
676	GEM_BUG_ON(i915_gem_context_is_kernel(ctx));
677
678	args->ctx_id = ctx->user_handle;
679	DRM_DEBUG("HW context %d created\n", args->ctx_id);
680
681	return 0;
682}
683
684int i915_gem_context_destroy_ioctl(struct drm_device *dev, void *data,
685				   struct drm_file *file)
686{
687	struct drm_i915_gem_context_destroy *args = data;
688	struct drm_i915_file_private *file_priv = file->driver_priv;
689	struct i915_gem_context *ctx;
690	int ret;
691
692	if (args->pad != 0)
693		return -EINVAL;
694
695	if (args->ctx_id == DEFAULT_CONTEXT_HANDLE)
696		return -ENOENT;
697
698	ctx = i915_gem_context_lookup(file_priv, args->ctx_id);
699	if (!ctx)
700		return -ENOENT;
701
702	ret = mutex_lock_interruptible(&dev->struct_mutex);
703	if (ret)
704		goto out;
 
 
705
706	__destroy_hw_context(ctx, file_priv);
 
707	mutex_unlock(&dev->struct_mutex);
708
709out:
710	i915_gem_context_put(ctx);
711	return 0;
712}
713
714int i915_gem_context_getparam_ioctl(struct drm_device *dev, void *data,
715				    struct drm_file *file)
716{
717	struct drm_i915_file_private *file_priv = file->driver_priv;
718	struct drm_i915_gem_context_param *args = data;
719	struct i915_gem_context *ctx;
720	int ret = 0;
 
 
 
 
721
722	ctx = i915_gem_context_lookup(file_priv, args->ctx_id);
723	if (!ctx)
724		return -ENOENT;
 
 
725
726	args->size = 0;
727	switch (args->param) {
728	case I915_CONTEXT_PARAM_BAN_PERIOD:
729		ret = -EINVAL;
730		break;
731	case I915_CONTEXT_PARAM_NO_ZEROMAP:
732		args->value = ctx->flags & CONTEXT_NO_ZEROMAP;
733		break;
734	case I915_CONTEXT_PARAM_GTT_SIZE:
735		if (ctx->ppgtt)
736			args->value = ctx->ppgtt->base.total;
737		else if (to_i915(dev)->mm.aliasing_ppgtt)
738			args->value = to_i915(dev)->mm.aliasing_ppgtt->base.total;
739		else
740			args->value = to_i915(dev)->ggtt.base.total;
741		break;
742	case I915_CONTEXT_PARAM_NO_ERROR_CAPTURE:
743		args->value = i915_gem_context_no_error_capture(ctx);
744		break;
745	case I915_CONTEXT_PARAM_BANNABLE:
746		args->value = i915_gem_context_is_bannable(ctx);
747		break;
748	case I915_CONTEXT_PARAM_PRIORITY:
749		args->value = ctx->priority;
750		break;
751	default:
752		ret = -EINVAL;
753		break;
754	}
 
755
756	i915_gem_context_put(ctx);
757	return ret;
758}
759
760int i915_gem_context_setparam_ioctl(struct drm_device *dev, void *data,
761				    struct drm_file *file)
762{
763	struct drm_i915_file_private *file_priv = file->driver_priv;
764	struct drm_i915_gem_context_param *args = data;
765	struct i915_gem_context *ctx;
766	int ret;
767
768	ctx = i915_gem_context_lookup(file_priv, args->ctx_id);
769	if (!ctx)
770		return -ENOENT;
771
772	ret = i915_mutex_lock_interruptible(dev);
773	if (ret)
774		goto out;
 
 
 
 
 
 
775
776	switch (args->param) {
777	case I915_CONTEXT_PARAM_BAN_PERIOD:
778		ret = -EINVAL;
 
 
 
 
 
 
779		break;
780	case I915_CONTEXT_PARAM_NO_ZEROMAP:
781		if (args->size) {
782			ret = -EINVAL;
783		} else {
784			ctx->flags &= ~CONTEXT_NO_ZEROMAP;
785			ctx->flags |= args->value ? CONTEXT_NO_ZEROMAP : 0;
786		}
787		break;
788	case I915_CONTEXT_PARAM_NO_ERROR_CAPTURE:
789		if (args->size)
790			ret = -EINVAL;
791		else if (args->value)
792			i915_gem_context_set_no_error_capture(ctx);
793		else
794			i915_gem_context_clear_no_error_capture(ctx);
795		break;
796	case I915_CONTEXT_PARAM_BANNABLE:
797		if (args->size)
798			ret = -EINVAL;
799		else if (!capable(CAP_SYS_ADMIN) && !args->value)
800			ret = -EPERM;
801		else if (args->value)
802			i915_gem_context_set_bannable(ctx);
803		else
804			i915_gem_context_clear_bannable(ctx);
805		break;
806
807	case I915_CONTEXT_PARAM_PRIORITY:
808		{
809			s64 priority = args->value;
810
811			if (args->size)
812				ret = -EINVAL;
813			else if (!(to_i915(dev)->caps.scheduler & I915_SCHEDULER_CAP_PRIORITY))
814				ret = -ENODEV;
815			else if (priority > I915_CONTEXT_MAX_USER_PRIORITY ||
816				 priority < I915_CONTEXT_MIN_USER_PRIORITY)
817				ret = -EINVAL;
818			else if (priority > I915_CONTEXT_DEFAULT_PRIORITY &&
819				 !capable(CAP_SYS_NICE))
820				ret = -EPERM;
821			else
822				ctx->priority = priority;
823		}
824		break;
825
826	default:
827		ret = -EINVAL;
828		break;
829	}
830	mutex_unlock(&dev->struct_mutex);
831
832out:
833	i915_gem_context_put(ctx);
834	return ret;
835}
836
837int i915_gem_context_reset_stats_ioctl(struct drm_device *dev,
838				       void *data, struct drm_file *file)
839{
840	struct drm_i915_private *dev_priv = to_i915(dev);
841	struct drm_i915_reset_stats *args = data;
842	struct i915_gem_context *ctx;
843	int ret;
844
845	if (args->flags || args->pad)
846		return -EINVAL;
847
848	ret = -ENOENT;
849	rcu_read_lock();
850	ctx = __i915_gem_context_lookup_rcu(file->driver_priv, args->ctx_id);
851	if (!ctx)
852		goto out;
853
854	/*
855	 * We opt for unserialised reads here. This may result in tearing
856	 * in the extremely unlikely event of a GPU hang on this context
857	 * as we are querying them. If we need that extra layer of protection,
858	 * we should wrap the hangstats with a seqlock.
859	 */
860
861	if (capable(CAP_SYS_ADMIN))
862		args->reset_count = i915_reset_count(&dev_priv->gpu_error);
863	else
864		args->reset_count = 0;
865
866	args->batch_active = atomic_read(&ctx->guilty_count);
867	args->batch_pending = atomic_read(&ctx->active_count);
868
869	ret = 0;
870out:
871	rcu_read_unlock();
872	return ret;
873}
874
875#if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
876#include "selftests/mock_context.c"
877#include "selftests/i915_gem_context.c"
878#endif