2 * Copyright © 2011-2012 Intel Corporation
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:
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
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
24 * Ben Widawsky <ben@bwidawsk.net>
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.
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.
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.
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
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
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.
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.
89 #include <drm/i915_drm.h>
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.
96 #define GEN6_CONTEXT_ALIGN (64<<10)
97 #define GEN7_CONTEXT_ALIGN 4096
99 static size_t get_context_alignment(struct drm_device *dev)
102 return GEN6_CONTEXT_ALIGN;
104 return GEN7_CONTEXT_ALIGN;
107 static int get_context_size(struct drm_device *dev)
109 struct drm_i915_private *dev_priv = dev->dev_private;
113 switch (INTEL_INFO(dev)->gen) {
115 reg = I915_READ(CXT_SIZE);
116 ret = GEN6_CXT_TOTAL_SIZE(reg) * 64;
119 reg = I915_READ(GEN7_CXT_SIZE);
121 ret = HSW_CXT_TOTAL_SIZE;
123 ret = GEN7_CXT_TOTAL_SIZE(reg) * 64;
126 ret = GEN8_CXT_TOTAL_SIZE;
135 void i915_gem_context_free(struct kref *ctx_ref)
137 struct intel_context *ctx = container_of(ctx_ref,
140 if (i915.enable_execlists)
141 intel_lr_context_free(ctx);
143 i915_ppgtt_put(ctx->ppgtt);
145 if (ctx->legacy_hw_ctx.rcs_state)
146 drm_gem_object_unreference(&ctx->legacy_hw_ctx.rcs_state->base);
147 list_del(&ctx->link);
151 struct drm_i915_gem_object *
152 i915_gem_alloc_context_obj(struct drm_device *dev, size_t size)
154 struct drm_i915_gem_object *obj;
157 obj = i915_gem_alloc_object(dev, size);
159 return ERR_PTR(-ENOMEM);
162 * Try to make the context utilize L3 as well as LLC.
164 * On VLV we don't have L3 controls in the PTEs so we
165 * shouldn't touch the cache level, especially as that
166 * would make the object snooped which might have a
167 * negative performance impact.
169 if (INTEL_INFO(dev)->gen >= 7 && !IS_VALLEYVIEW(dev)) {
170 ret = i915_gem_object_set_cache_level(obj, I915_CACHE_L3_LLC);
171 /* Failure shouldn't ever happen this early */
173 drm_gem_object_unreference(&obj->base);
181 static struct intel_context *
182 __create_hw_context(struct drm_device *dev,
183 struct drm_i915_file_private *file_priv)
185 struct drm_i915_private *dev_priv = dev->dev_private;
186 struct intel_context *ctx;
189 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
191 return ERR_PTR(-ENOMEM);
193 kref_init(&ctx->ref);
194 list_add_tail(&ctx->link, &dev_priv->context_list);
196 if (dev_priv->hw_context_size) {
197 struct drm_i915_gem_object *obj =
198 i915_gem_alloc_context_obj(dev, dev_priv->hw_context_size);
203 ctx->legacy_hw_ctx.rcs_state = obj;
206 /* Default context will never have a file_priv */
207 if (file_priv != NULL) {
208 ret = idr_alloc(&file_priv->context_idr, ctx,
209 DEFAULT_CONTEXT_HANDLE, 0, GFP_KERNEL);
213 ret = DEFAULT_CONTEXT_HANDLE;
215 ctx->file_priv = file_priv;
216 ctx->user_handle = ret;
217 /* NB: Mark all slices as needing a remap so that when the context first
218 * loads it will restore whatever remap state already exists. If there
219 * is no remap info, it will be a NOP. */
220 ctx->remap_slice = (1 << NUM_L3_SLICES(dev)) - 1;
225 i915_gem_context_unreference(ctx);
230 * The default context needs to exist per ring that uses contexts. It stores the
231 * context state of the GPU for applications that don't utilize HW contexts, as
232 * well as an idle case.
234 static struct intel_context *
235 i915_gem_create_context(struct drm_device *dev,
236 struct drm_i915_file_private *file_priv)
238 const bool is_global_default_ctx = file_priv == NULL;
239 struct intel_context *ctx;
242 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
244 ctx = __create_hw_context(dev, file_priv);
248 if (is_global_default_ctx && ctx->legacy_hw_ctx.rcs_state) {
249 /* We may need to do things with the shrinker which
250 * require us to immediately switch back to the default
251 * context. This can cause a problem as pinning the
252 * default context also requires GTT space which may not
253 * be available. To avoid this we always pin the default
256 ret = i915_gem_obj_ggtt_pin(ctx->legacy_hw_ctx.rcs_state,
257 get_context_alignment(dev), 0);
259 DRM_DEBUG_DRIVER("Couldn't pin %d\n", ret);
264 if (USES_FULL_PPGTT(dev)) {
265 struct i915_hw_ppgtt *ppgtt = i915_ppgtt_create(dev, file_priv);
267 if (IS_ERR_OR_NULL(ppgtt)) {
268 DRM_DEBUG_DRIVER("PPGTT setup failed (%ld)\n",
270 ret = PTR_ERR(ppgtt);
280 if (is_global_default_ctx && ctx->legacy_hw_ctx.rcs_state)
281 i915_gem_object_ggtt_unpin(ctx->legacy_hw_ctx.rcs_state);
283 i915_gem_context_unreference(ctx);
287 void i915_gem_context_reset(struct drm_device *dev)
289 struct drm_i915_private *dev_priv = dev->dev_private;
292 /* In execlists mode we will unreference the context when the execlist
293 * queue is cleared and the requests destroyed.
295 if (i915.enable_execlists)
298 for (i = 0; i < I915_NUM_RINGS; i++) {
299 struct intel_engine_cs *ring = &dev_priv->ring[i];
300 struct intel_context *lctx = ring->last_context;
303 if (lctx->legacy_hw_ctx.rcs_state && i == RCS)
304 i915_gem_object_ggtt_unpin(lctx->legacy_hw_ctx.rcs_state);
306 i915_gem_context_unreference(lctx);
307 ring->last_context = NULL;
312 int i915_gem_context_init(struct drm_device *dev)
314 struct drm_i915_private *dev_priv = dev->dev_private;
315 struct intel_context *ctx;
318 /* Init should only be called once per module load. Eventually the
319 * restriction on the context_disabled check can be loosened. */
320 if (WARN_ON(dev_priv->ring[RCS].default_context))
323 if (i915.enable_execlists) {
324 /* NB: intentionally left blank. We will allocate our own
325 * backing objects as we need them, thank you very much */
326 dev_priv->hw_context_size = 0;
327 } else if (HAS_HW_CONTEXTS(dev)) {
328 dev_priv->hw_context_size = round_up(get_context_size(dev), 4096);
329 if (dev_priv->hw_context_size > (1<<20)) {
330 DRM_DEBUG_DRIVER("Disabling HW Contexts; invalid size %d\n",
331 dev_priv->hw_context_size);
332 dev_priv->hw_context_size = 0;
336 ctx = i915_gem_create_context(dev, NULL);
338 DRM_ERROR("Failed to create default global context (error %ld)\n",
343 for (i = 0; i < I915_NUM_RINGS; i++) {
344 struct intel_engine_cs *ring = &dev_priv->ring[i];
346 /* NB: RCS will hold a ref for all rings */
347 ring->default_context = ctx;
350 DRM_DEBUG_DRIVER("%s context support initialized\n",
351 i915.enable_execlists ? "LR" :
352 dev_priv->hw_context_size ? "HW" : "fake");
356 void i915_gem_context_fini(struct drm_device *dev)
358 struct drm_i915_private *dev_priv = dev->dev_private;
359 struct intel_context *dctx = dev_priv->ring[RCS].default_context;
362 if (dctx->legacy_hw_ctx.rcs_state) {
363 /* The only known way to stop the gpu from accessing the hw context is
364 * to reset it. Do this as the very last operation to avoid confusing
365 * other code, leading to spurious errors. */
366 intel_gpu_reset(dev);
368 /* When default context is created and switched to, base object refcount
369 * will be 2 (+1 from object creation and +1 from do_switch()).
370 * i915_gem_context_fini() will be called after gpu_idle() has switched
371 * to default context. So we need to unreference the base object once
372 * to offset the do_switch part, so that i915_gem_context_unreference()
373 * can then free the base object correctly. */
374 WARN_ON(!dev_priv->ring[RCS].last_context);
375 if (dev_priv->ring[RCS].last_context == dctx) {
376 /* Fake switch to NULL context */
377 WARN_ON(dctx->legacy_hw_ctx.rcs_state->active);
378 i915_gem_object_ggtt_unpin(dctx->legacy_hw_ctx.rcs_state);
379 i915_gem_context_unreference(dctx);
380 dev_priv->ring[RCS].last_context = NULL;
383 i915_gem_object_ggtt_unpin(dctx->legacy_hw_ctx.rcs_state);
386 for (i = 0; i < I915_NUM_RINGS; i++) {
387 struct intel_engine_cs *ring = &dev_priv->ring[i];
389 if (ring->last_context)
390 i915_gem_context_unreference(ring->last_context);
392 ring->default_context = NULL;
393 ring->last_context = NULL;
396 i915_gem_context_unreference(dctx);
399 int i915_gem_context_enable(struct drm_i915_private *dev_priv)
401 struct intel_engine_cs *ring;
404 BUG_ON(!dev_priv->ring[RCS].default_context);
406 if (i915.enable_execlists)
409 for_each_ring(ring, dev_priv, i) {
410 ret = i915_switch_context(ring, ring->default_context);
418 static int context_idr_cleanup(int id, void *p, void *data)
420 struct intel_context *ctx = p;
422 i915_gem_context_unreference(ctx);
426 int i915_gem_context_open(struct drm_device *dev, struct drm_file *file)
428 struct drm_i915_file_private *file_priv = file->driver_priv;
429 struct intel_context *ctx;
431 idr_init(&file_priv->context_idr);
433 mutex_lock(&dev->struct_mutex);
434 ctx = i915_gem_create_context(dev, file_priv);
435 mutex_unlock(&dev->struct_mutex);
438 idr_destroy(&file_priv->context_idr);
445 void i915_gem_context_close(struct drm_device *dev, struct drm_file *file)
447 struct drm_i915_file_private *file_priv = file->driver_priv;
449 idr_for_each(&file_priv->context_idr, context_idr_cleanup, NULL);
450 idr_destroy(&file_priv->context_idr);
453 struct intel_context *
454 i915_gem_context_get(struct drm_i915_file_private *file_priv, u32 id)
456 struct intel_context *ctx;
458 ctx = (struct intel_context *)idr_find(&file_priv->context_idr, id);
460 return ERR_PTR(-ENOENT);
466 mi_set_context(struct intel_engine_cs *ring,
467 struct intel_context *new_context,
470 u32 flags = hw_flags | MI_MM_SPACE_GTT;
471 const int num_rings =
472 /* Use an extended w/a on ivb+ if signalling from other rings */
473 i915_semaphore_is_enabled(ring->dev) ?
474 hweight32(INTEL_INFO(ring->dev)->ring_mask) - 1 :
478 /* w/a: If Flush TLB Invalidation Mode is enabled, driver must do a TLB
479 * invalidation prior to MI_SET_CONTEXT. On GEN6 we don't set the value
480 * explicitly, so we rely on the value at ring init, stored in
481 * itlb_before_ctx_switch.
483 if (IS_GEN6(ring->dev)) {
484 ret = ring->flush(ring, I915_GEM_GPU_DOMAINS, 0);
489 /* These flags are for resource streamer on HSW+ */
490 if (!IS_HASWELL(ring->dev) && INTEL_INFO(ring->dev)->gen < 8)
491 flags |= (MI_SAVE_EXT_STATE_EN | MI_RESTORE_EXT_STATE_EN);
494 if (INTEL_INFO(ring->dev)->gen >= 7)
495 len += 2 + (num_rings ? 4*num_rings + 2 : 0);
497 ret = intel_ring_begin(ring, len);
501 /* WaProgramMiArbOnOffAroundMiSetContext:ivb,vlv,hsw,bdw,chv */
502 if (INTEL_INFO(ring->dev)->gen >= 7) {
503 intel_ring_emit(ring, MI_ARB_ON_OFF | MI_ARB_DISABLE);
505 struct intel_engine_cs *signaller;
507 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(num_rings));
508 for_each_ring(signaller, to_i915(ring->dev), i) {
509 if (signaller == ring)
512 intel_ring_emit(ring, RING_PSMI_CTL(signaller->mmio_base));
513 intel_ring_emit(ring, _MASKED_BIT_ENABLE(GEN6_PSMI_SLEEP_MSG_DISABLE));
518 intel_ring_emit(ring, MI_NOOP);
519 intel_ring_emit(ring, MI_SET_CONTEXT);
520 intel_ring_emit(ring, i915_gem_obj_ggtt_offset(new_context->legacy_hw_ctx.rcs_state) |
523 * w/a: MI_SET_CONTEXT must always be followed by MI_NOOP
524 * WaMiSetContext_Hang:snb,ivb,vlv
526 intel_ring_emit(ring, MI_NOOP);
528 if (INTEL_INFO(ring->dev)->gen >= 7) {
530 struct intel_engine_cs *signaller;
532 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(num_rings));
533 for_each_ring(signaller, to_i915(ring->dev), i) {
534 if (signaller == ring)
537 intel_ring_emit(ring, RING_PSMI_CTL(signaller->mmio_base));
538 intel_ring_emit(ring, _MASKED_BIT_DISABLE(GEN6_PSMI_SLEEP_MSG_DISABLE));
541 intel_ring_emit(ring, MI_ARB_ON_OFF | MI_ARB_ENABLE);
544 intel_ring_advance(ring);
549 static int do_switch(struct intel_engine_cs *ring,
550 struct intel_context *to)
552 struct drm_i915_private *dev_priv = ring->dev->dev_private;
553 struct intel_context *from = ring->last_context;
555 bool uninitialized = false;
558 if (from != NULL && ring == &dev_priv->ring[RCS]) {
559 BUG_ON(from->legacy_hw_ctx.rcs_state == NULL);
560 BUG_ON(!i915_gem_obj_is_pinned(from->legacy_hw_ctx.rcs_state));
563 if (from == to && !to->remap_slice)
566 /* Trying to pin first makes error handling easier. */
567 if (ring == &dev_priv->ring[RCS]) {
568 ret = i915_gem_obj_ggtt_pin(to->legacy_hw_ctx.rcs_state,
569 get_context_alignment(ring->dev), 0);
575 * Pin can switch back to the default context if we end up calling into
576 * evict_everything - as a last ditch gtt defrag effort that also
577 * switches to the default context. Hence we need to reload from here.
579 from = ring->last_context;
582 ret = to->ppgtt->switch_mm(to->ppgtt, ring);
587 if (ring != &dev_priv->ring[RCS]) {
589 i915_gem_context_unreference(from);
594 * Clear this page out of any CPU caches for coherent swap-in/out. Note
595 * that thanks to write = false in this call and us not setting any gpu
596 * write domains when putting a context object onto the active list
597 * (when switching away from it), this won't block.
599 * XXX: We need a real interface to do this instead of trickery.
601 ret = i915_gem_object_set_to_gtt_domain(to->legacy_hw_ctx.rcs_state, false);
605 if (!to->legacy_hw_ctx.rcs_state->has_global_gtt_mapping) {
606 struct i915_vma *vma = i915_gem_obj_to_vma(to->legacy_hw_ctx.rcs_state,
607 &dev_priv->gtt.base);
608 vma->bind_vma(vma, to->legacy_hw_ctx.rcs_state->cache_level, GLOBAL_BIND);
611 if (!to->legacy_hw_ctx.initialized || i915_gem_context_is_default(to))
612 hw_flags |= MI_RESTORE_INHIBIT;
614 ret = mi_set_context(ring, to, hw_flags);
618 for (i = 0; i < MAX_L3_SLICES; i++) {
619 if (!(to->remap_slice & (1<<i)))
622 ret = i915_gem_l3_remap(ring, i);
623 /* If it failed, try again next round */
625 DRM_DEBUG_DRIVER("L3 remapping failed\n");
627 to->remap_slice &= ~(1<<i);
630 /* The backing object for the context is done after switching to the
631 * *next* context. Therefore we cannot retire the previous context until
632 * the next context has already started running. In fact, the below code
633 * is a bit suboptimal because the retiring can occur simply after the
634 * MI_SET_CONTEXT instead of when the next seqno has completed.
637 from->legacy_hw_ctx.rcs_state->base.read_domains = I915_GEM_DOMAIN_INSTRUCTION;
638 i915_vma_move_to_active(i915_gem_obj_to_ggtt(from->legacy_hw_ctx.rcs_state), ring);
639 /* As long as MI_SET_CONTEXT is serializing, ie. it flushes the
640 * whole damn pipeline, we don't need to explicitly mark the
641 * object dirty. The only exception is that the context must be
642 * correct in case the object gets swapped out. Ideally we'd be
643 * able to defer doing this until we know the object would be
644 * swapped, but there is no way to do that yet.
646 from->legacy_hw_ctx.rcs_state->dirty = 1;
647 BUG_ON(from->legacy_hw_ctx.rcs_state->ring != ring);
649 /* obj is kept alive until the next request by its active ref */
650 i915_gem_object_ggtt_unpin(from->legacy_hw_ctx.rcs_state);
651 i915_gem_context_unreference(from);
654 uninitialized = !to->legacy_hw_ctx.initialized && from == NULL;
655 to->legacy_hw_ctx.initialized = true;
658 i915_gem_context_reference(to);
659 ring->last_context = to;
662 if (ring->init_context) {
663 ret = ring->init_context(ring);
665 DRM_ERROR("ring init context: %d\n", ret);
668 ret = i915_gem_render_state_init(ring);
670 DRM_ERROR("init render state: %d\n", ret);
677 i915_gem_object_ggtt_unpin(to->legacy_hw_ctx.rcs_state);
682 * i915_switch_context() - perform a GPU context switch.
683 * @ring: ring for which we'll execute the context switch
684 * @to: the context to switch to
686 * The context life cycle is simple. The context refcount is incremented and
687 * decremented by 1 and create and destroy. If the context is in use by the GPU,
688 * it will have a refcount > 1. This allows us to destroy the context abstract
689 * object while letting the normal object tracking destroy the backing BO.
691 * This function should not be used in execlists mode. Instead the context is
692 * switched by writing to the ELSP and requests keep a reference to their
695 int i915_switch_context(struct intel_engine_cs *ring,
696 struct intel_context *to)
698 struct drm_i915_private *dev_priv = ring->dev->dev_private;
700 WARN_ON(i915.enable_execlists);
701 WARN_ON(!mutex_is_locked(&dev_priv->dev->struct_mutex));
703 if (to->legacy_hw_ctx.rcs_state == NULL) { /* We have the fake context */
704 if (to != ring->last_context) {
705 i915_gem_context_reference(to);
706 if (ring->last_context)
707 i915_gem_context_unreference(ring->last_context);
708 ring->last_context = to;
713 return do_switch(ring, to);
716 static bool contexts_enabled(struct drm_device *dev)
718 return i915.enable_execlists || to_i915(dev)->hw_context_size;
721 int i915_gem_context_create_ioctl(struct drm_device *dev, void *data,
722 struct drm_file *file)
724 struct drm_i915_gem_context_create *args = data;
725 struct drm_i915_file_private *file_priv = file->driver_priv;
726 struct intel_context *ctx;
729 if (!contexts_enabled(dev))
732 ret = i915_mutex_lock_interruptible(dev);
736 ctx = i915_gem_create_context(dev, file_priv);
737 mutex_unlock(&dev->struct_mutex);
741 args->ctx_id = ctx->user_handle;
742 DRM_DEBUG_DRIVER("HW context %d created\n", args->ctx_id);
747 int i915_gem_context_destroy_ioctl(struct drm_device *dev, void *data,
748 struct drm_file *file)
750 struct drm_i915_gem_context_destroy *args = data;
751 struct drm_i915_file_private *file_priv = file->driver_priv;
752 struct intel_context *ctx;
755 if (args->ctx_id == DEFAULT_CONTEXT_HANDLE)
758 ret = i915_mutex_lock_interruptible(dev);
762 ctx = i915_gem_context_get(file_priv, args->ctx_id);
764 mutex_unlock(&dev->struct_mutex);
768 idr_remove(&ctx->file_priv->context_idr, ctx->user_handle);
769 i915_gem_context_unreference(ctx);
770 mutex_unlock(&dev->struct_mutex);
772 DRM_DEBUG_DRIVER("HW context %d destroyed\n", args->ctx_id);