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>
91 #include "i915_trace.h"
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.
97 #define GEN6_CONTEXT_ALIGN (64<<10)
98 #define GEN7_CONTEXT_ALIGN 4096
100 static size_t get_context_alignment(struct drm_device *dev)
103 return GEN6_CONTEXT_ALIGN;
105 return GEN7_CONTEXT_ALIGN;
108 static int get_context_size(struct drm_device *dev)
110 struct drm_i915_private *dev_priv = dev->dev_private;
114 switch (INTEL_INFO(dev)->gen) {
116 reg = I915_READ(CXT_SIZE);
117 ret = GEN6_CXT_TOTAL_SIZE(reg) * 64;
120 reg = I915_READ(GEN7_CXT_SIZE);
122 ret = HSW_CXT_TOTAL_SIZE;
124 ret = GEN7_CXT_TOTAL_SIZE(reg) * 64;
127 ret = GEN8_CXT_TOTAL_SIZE;
136 void i915_gem_context_free(struct kref *ctx_ref)
138 struct intel_context *ctx = container_of(ctx_ref,
141 trace_i915_context_free(ctx);
143 if (i915.enable_execlists)
144 intel_lr_context_free(ctx);
146 i915_ppgtt_put(ctx->ppgtt);
148 if (ctx->legacy_hw_ctx.rcs_state)
149 drm_gem_object_unreference(&ctx->legacy_hw_ctx.rcs_state->base);
150 list_del(&ctx->link);
154 struct drm_i915_gem_object *
155 i915_gem_alloc_context_obj(struct drm_device *dev, size_t size)
157 struct drm_i915_gem_object *obj;
160 obj = i915_gem_alloc_object(dev, size);
162 return ERR_PTR(-ENOMEM);
165 * Try to make the context utilize L3 as well as LLC.
167 * On VLV we don't have L3 controls in the PTEs so we
168 * shouldn't touch the cache level, especially as that
169 * would make the object snooped which might have a
170 * negative performance impact.
172 if (INTEL_INFO(dev)->gen >= 7 && !IS_VALLEYVIEW(dev)) {
173 ret = i915_gem_object_set_cache_level(obj, I915_CACHE_L3_LLC);
174 /* Failure shouldn't ever happen this early */
176 drm_gem_object_unreference(&obj->base);
184 static struct intel_context *
185 __create_hw_context(struct drm_device *dev,
186 struct drm_i915_file_private *file_priv)
188 struct drm_i915_private *dev_priv = dev->dev_private;
189 struct intel_context *ctx;
192 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
194 return ERR_PTR(-ENOMEM);
196 kref_init(&ctx->ref);
197 list_add_tail(&ctx->link, &dev_priv->context_list);
199 if (dev_priv->hw_context_size) {
200 struct drm_i915_gem_object *obj =
201 i915_gem_alloc_context_obj(dev, dev_priv->hw_context_size);
206 ctx->legacy_hw_ctx.rcs_state = obj;
209 /* Default context will never have a file_priv */
210 if (file_priv != NULL) {
211 ret = idr_alloc(&file_priv->context_idr, ctx,
212 DEFAULT_CONTEXT_HANDLE, 0, GFP_KERNEL);
216 ret = DEFAULT_CONTEXT_HANDLE;
218 ctx->file_priv = file_priv;
219 ctx->user_handle = ret;
220 /* NB: Mark all slices as needing a remap so that when the context first
221 * loads it will restore whatever remap state already exists. If there
222 * is no remap info, it will be a NOP. */
223 ctx->remap_slice = (1 << NUM_L3_SLICES(dev)) - 1;
228 i915_gem_context_unreference(ctx);
233 * The default context needs to exist per ring that uses contexts. It stores the
234 * context state of the GPU for applications that don't utilize HW contexts, as
235 * well as an idle case.
237 static struct intel_context *
238 i915_gem_create_context(struct drm_device *dev,
239 struct drm_i915_file_private *file_priv)
241 const bool is_global_default_ctx = file_priv == NULL;
242 struct intel_context *ctx;
245 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
247 ctx = __create_hw_context(dev, file_priv);
251 if (is_global_default_ctx && ctx->legacy_hw_ctx.rcs_state) {
252 /* We may need to do things with the shrinker which
253 * require us to immediately switch back to the default
254 * context. This can cause a problem as pinning the
255 * default context also requires GTT space which may not
256 * be available. To avoid this we always pin the default
259 ret = i915_gem_obj_ggtt_pin(ctx->legacy_hw_ctx.rcs_state,
260 get_context_alignment(dev), 0);
262 DRM_DEBUG_DRIVER("Couldn't pin %d\n", ret);
267 if (USES_FULL_PPGTT(dev)) {
268 struct i915_hw_ppgtt *ppgtt = i915_ppgtt_create(dev, file_priv);
270 if (IS_ERR_OR_NULL(ppgtt)) {
271 DRM_DEBUG_DRIVER("PPGTT setup failed (%ld)\n",
273 ret = PTR_ERR(ppgtt);
280 trace_i915_context_create(ctx);
285 if (is_global_default_ctx && ctx->legacy_hw_ctx.rcs_state)
286 i915_gem_object_ggtt_unpin(ctx->legacy_hw_ctx.rcs_state);
288 i915_gem_context_unreference(ctx);
292 void i915_gem_context_reset(struct drm_device *dev)
294 struct drm_i915_private *dev_priv = dev->dev_private;
297 /* In execlists mode we will unreference the context when the execlist
298 * queue is cleared and the requests destroyed.
300 if (i915.enable_execlists)
303 for (i = 0; i < I915_NUM_RINGS; i++) {
304 struct intel_engine_cs *ring = &dev_priv->ring[i];
305 struct intel_context *lctx = ring->last_context;
308 if (lctx->legacy_hw_ctx.rcs_state && i == RCS)
309 i915_gem_object_ggtt_unpin(lctx->legacy_hw_ctx.rcs_state);
311 i915_gem_context_unreference(lctx);
312 ring->last_context = NULL;
317 int i915_gem_context_init(struct drm_device *dev)
319 struct drm_i915_private *dev_priv = dev->dev_private;
320 struct intel_context *ctx;
323 /* Init should only be called once per module load. Eventually the
324 * restriction on the context_disabled check can be loosened. */
325 if (WARN_ON(dev_priv->ring[RCS].default_context))
328 if (i915.enable_execlists) {
329 /* NB: intentionally left blank. We will allocate our own
330 * backing objects as we need them, thank you very much */
331 dev_priv->hw_context_size = 0;
332 } else if (HAS_HW_CONTEXTS(dev)) {
333 dev_priv->hw_context_size = round_up(get_context_size(dev), 4096);
334 if (dev_priv->hw_context_size > (1<<20)) {
335 DRM_DEBUG_DRIVER("Disabling HW Contexts; invalid size %d\n",
336 dev_priv->hw_context_size);
337 dev_priv->hw_context_size = 0;
341 ctx = i915_gem_create_context(dev, NULL);
343 DRM_ERROR("Failed to create default global context (error %ld)\n",
348 for (i = 0; i < I915_NUM_RINGS; i++) {
349 struct intel_engine_cs *ring = &dev_priv->ring[i];
351 /* NB: RCS will hold a ref for all rings */
352 ring->default_context = ctx;
355 DRM_DEBUG_DRIVER("%s context support initialized\n",
356 i915.enable_execlists ? "LR" :
357 dev_priv->hw_context_size ? "HW" : "fake");
361 void i915_gem_context_fini(struct drm_device *dev)
363 struct drm_i915_private *dev_priv = dev->dev_private;
364 struct intel_context *dctx = dev_priv->ring[RCS].default_context;
367 if (dctx->legacy_hw_ctx.rcs_state) {
368 /* The only known way to stop the gpu from accessing the hw context is
369 * to reset it. Do this as the very last operation to avoid confusing
370 * other code, leading to spurious errors. */
371 intel_gpu_reset(dev);
373 /* When default context is created and switched to, base object refcount
374 * will be 2 (+1 from object creation and +1 from do_switch()).
375 * i915_gem_context_fini() will be called after gpu_idle() has switched
376 * to default context. So we need to unreference the base object once
377 * to offset the do_switch part, so that i915_gem_context_unreference()
378 * can then free the base object correctly. */
379 WARN_ON(!dev_priv->ring[RCS].last_context);
380 if (dev_priv->ring[RCS].last_context == dctx) {
381 /* Fake switch to NULL context */
382 WARN_ON(dctx->legacy_hw_ctx.rcs_state->active);
383 i915_gem_object_ggtt_unpin(dctx->legacy_hw_ctx.rcs_state);
384 i915_gem_context_unreference(dctx);
385 dev_priv->ring[RCS].last_context = NULL;
388 i915_gem_object_ggtt_unpin(dctx->legacy_hw_ctx.rcs_state);
391 for (i = 0; i < I915_NUM_RINGS; i++) {
392 struct intel_engine_cs *ring = &dev_priv->ring[i];
394 if (ring->last_context)
395 i915_gem_context_unreference(ring->last_context);
397 ring->default_context = NULL;
398 ring->last_context = NULL;
401 i915_gem_context_unreference(dctx);
404 int i915_gem_context_enable(struct drm_i915_private *dev_priv)
406 struct intel_engine_cs *ring;
409 BUG_ON(!dev_priv->ring[RCS].default_context);
411 if (i915.enable_execlists)
414 for_each_ring(ring, dev_priv, i) {
415 ret = i915_switch_context(ring, ring->default_context);
423 static int context_idr_cleanup(int id, void *p, void *data)
425 struct intel_context *ctx = p;
427 i915_gem_context_unreference(ctx);
431 int i915_gem_context_open(struct drm_device *dev, struct drm_file *file)
433 struct drm_i915_file_private *file_priv = file->driver_priv;
434 struct intel_context *ctx;
436 idr_init(&file_priv->context_idr);
438 mutex_lock(&dev->struct_mutex);
439 ctx = i915_gem_create_context(dev, file_priv);
440 mutex_unlock(&dev->struct_mutex);
443 idr_destroy(&file_priv->context_idr);
450 void i915_gem_context_close(struct drm_device *dev, struct drm_file *file)
452 struct drm_i915_file_private *file_priv = file->driver_priv;
454 idr_for_each(&file_priv->context_idr, context_idr_cleanup, NULL);
455 idr_destroy(&file_priv->context_idr);
458 struct intel_context *
459 i915_gem_context_get(struct drm_i915_file_private *file_priv, u32 id)
461 struct intel_context *ctx;
463 ctx = (struct intel_context *)idr_find(&file_priv->context_idr, id);
465 return ERR_PTR(-ENOENT);
471 mi_set_context(struct intel_engine_cs *ring,
472 struct intel_context *new_context,
475 u32 flags = hw_flags | MI_MM_SPACE_GTT;
476 const int num_rings =
477 /* Use an extended w/a on ivb+ if signalling from other rings */
478 i915_semaphore_is_enabled(ring->dev) ?
479 hweight32(INTEL_INFO(ring->dev)->ring_mask) - 1 :
483 /* w/a: If Flush TLB Invalidation Mode is enabled, driver must do a TLB
484 * invalidation prior to MI_SET_CONTEXT. On GEN6 we don't set the value
485 * explicitly, so we rely on the value at ring init, stored in
486 * itlb_before_ctx_switch.
488 if (IS_GEN6(ring->dev)) {
489 ret = ring->flush(ring, I915_GEM_GPU_DOMAINS, 0);
494 /* These flags are for resource streamer on HSW+ */
495 if (!IS_HASWELL(ring->dev) && INTEL_INFO(ring->dev)->gen < 8)
496 flags |= (MI_SAVE_EXT_STATE_EN | MI_RESTORE_EXT_STATE_EN);
500 if (INTEL_INFO(ring->dev)->gen >= 7)
501 len += 2 + (num_rings ? 4*num_rings + 2 : 0);
503 ret = intel_ring_begin(ring, len);
507 /* WaProgramMiArbOnOffAroundMiSetContext:ivb,vlv,hsw,bdw,chv */
508 if (INTEL_INFO(ring->dev)->gen >= 7) {
509 intel_ring_emit(ring, MI_ARB_ON_OFF | MI_ARB_DISABLE);
511 struct intel_engine_cs *signaller;
513 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(num_rings));
514 for_each_ring(signaller, to_i915(ring->dev), i) {
515 if (signaller == ring)
518 intel_ring_emit(ring, RING_PSMI_CTL(signaller->mmio_base));
519 intel_ring_emit(ring, _MASKED_BIT_ENABLE(GEN6_PSMI_SLEEP_MSG_DISABLE));
524 intel_ring_emit(ring, MI_NOOP);
525 intel_ring_emit(ring, MI_SET_CONTEXT);
526 intel_ring_emit(ring, i915_gem_obj_ggtt_offset(new_context->legacy_hw_ctx.rcs_state) |
529 * w/a: MI_SET_CONTEXT must always be followed by MI_NOOP
530 * WaMiSetContext_Hang:snb,ivb,vlv
532 intel_ring_emit(ring, MI_NOOP);
534 if (INTEL_INFO(ring->dev)->gen >= 7) {
536 struct intel_engine_cs *signaller;
538 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(num_rings));
539 for_each_ring(signaller, to_i915(ring->dev), i) {
540 if (signaller == ring)
543 intel_ring_emit(ring, RING_PSMI_CTL(signaller->mmio_base));
544 intel_ring_emit(ring, _MASKED_BIT_DISABLE(GEN6_PSMI_SLEEP_MSG_DISABLE));
547 intel_ring_emit(ring, MI_ARB_ON_OFF | MI_ARB_ENABLE);
550 intel_ring_advance(ring);
555 static int do_switch(struct intel_engine_cs *ring,
556 struct intel_context *to)
558 struct drm_i915_private *dev_priv = ring->dev->dev_private;
559 struct intel_context *from = ring->last_context;
561 bool uninitialized = false;
562 struct i915_vma *vma;
565 if (from != NULL && ring == &dev_priv->ring[RCS]) {
566 BUG_ON(from->legacy_hw_ctx.rcs_state == NULL);
567 BUG_ON(!i915_gem_obj_is_pinned(from->legacy_hw_ctx.rcs_state));
570 if (from == to && !to->remap_slice)
573 /* Trying to pin first makes error handling easier. */
574 if (ring == &dev_priv->ring[RCS]) {
575 ret = i915_gem_obj_ggtt_pin(to->legacy_hw_ctx.rcs_state,
576 get_context_alignment(ring->dev), 0);
582 * Pin can switch back to the default context if we end up calling into
583 * evict_everything - as a last ditch gtt defrag effort that also
584 * switches to the default context. Hence we need to reload from here.
586 from = ring->last_context;
589 trace_switch_mm(ring, to);
590 ret = to->ppgtt->switch_mm(to->ppgtt, ring);
595 if (ring != &dev_priv->ring[RCS]) {
597 i915_gem_context_unreference(from);
602 * Clear this page out of any CPU caches for coherent swap-in/out. Note
603 * that thanks to write = false in this call and us not setting any gpu
604 * write domains when putting a context object onto the active list
605 * (when switching away from it), this won't block.
607 * XXX: We need a real interface to do this instead of trickery.
609 ret = i915_gem_object_set_to_gtt_domain(to->legacy_hw_ctx.rcs_state, false);
613 vma = i915_gem_obj_to_ggtt(to->legacy_hw_ctx.rcs_state);
614 if (!(vma->bound & GLOBAL_BIND))
615 vma->bind_vma(vma, to->legacy_hw_ctx.rcs_state->cache_level,
618 if (!to->legacy_hw_ctx.initialized || i915_gem_context_is_default(to))
619 hw_flags |= MI_RESTORE_INHIBIT;
621 ret = mi_set_context(ring, to, hw_flags);
625 for (i = 0; i < MAX_L3_SLICES; i++) {
626 if (!(to->remap_slice & (1<<i)))
629 ret = i915_gem_l3_remap(ring, i);
630 /* If it failed, try again next round */
632 DRM_DEBUG_DRIVER("L3 remapping failed\n");
634 to->remap_slice &= ~(1<<i);
637 /* The backing object for the context is done after switching to the
638 * *next* context. Therefore we cannot retire the previous context until
639 * the next context has already started running. In fact, the below code
640 * is a bit suboptimal because the retiring can occur simply after the
641 * MI_SET_CONTEXT instead of when the next seqno has completed.
644 from->legacy_hw_ctx.rcs_state->base.read_domains = I915_GEM_DOMAIN_INSTRUCTION;
645 i915_vma_move_to_active(i915_gem_obj_to_ggtt(from->legacy_hw_ctx.rcs_state), ring);
646 /* As long as MI_SET_CONTEXT is serializing, ie. it flushes the
647 * whole damn pipeline, we don't need to explicitly mark the
648 * object dirty. The only exception is that the context must be
649 * correct in case the object gets swapped out. Ideally we'd be
650 * able to defer doing this until we know the object would be
651 * swapped, but there is no way to do that yet.
653 from->legacy_hw_ctx.rcs_state->dirty = 1;
654 BUG_ON(from->legacy_hw_ctx.rcs_state->ring != ring);
656 /* obj is kept alive until the next request by its active ref */
657 i915_gem_object_ggtt_unpin(from->legacy_hw_ctx.rcs_state);
658 i915_gem_context_unreference(from);
661 uninitialized = !to->legacy_hw_ctx.initialized && from == NULL;
662 to->legacy_hw_ctx.initialized = true;
665 i915_gem_context_reference(to);
666 ring->last_context = to;
669 if (ring->init_context) {
670 ret = ring->init_context(ring, to);
672 DRM_ERROR("ring init context: %d\n", ret);
675 ret = i915_gem_render_state_init(ring);
677 DRM_ERROR("init render state: %d\n", ret);
684 i915_gem_object_ggtt_unpin(to->legacy_hw_ctx.rcs_state);
689 * i915_switch_context() - perform a GPU context switch.
690 * @ring: ring for which we'll execute the context switch
691 * @to: the context to switch to
693 * The context life cycle is simple. The context refcount is incremented and
694 * decremented by 1 and create and destroy. If the context is in use by the GPU,
695 * it will have a refcount > 1. This allows us to destroy the context abstract
696 * object while letting the normal object tracking destroy the backing BO.
698 * This function should not be used in execlists mode. Instead the context is
699 * switched by writing to the ELSP and requests keep a reference to their
702 int i915_switch_context(struct intel_engine_cs *ring,
703 struct intel_context *to)
705 struct drm_i915_private *dev_priv = ring->dev->dev_private;
707 WARN_ON(i915.enable_execlists);
708 WARN_ON(!mutex_is_locked(&dev_priv->dev->struct_mutex));
710 if (to->legacy_hw_ctx.rcs_state == NULL) { /* We have the fake context */
711 if (to != ring->last_context) {
712 i915_gem_context_reference(to);
713 if (ring->last_context)
714 i915_gem_context_unreference(ring->last_context);
715 ring->last_context = to;
720 return do_switch(ring, to);
723 static bool contexts_enabled(struct drm_device *dev)
725 return i915.enable_execlists || to_i915(dev)->hw_context_size;
728 int i915_gem_context_create_ioctl(struct drm_device *dev, void *data,
729 struct drm_file *file)
731 struct drm_i915_gem_context_create *args = data;
732 struct drm_i915_file_private *file_priv = file->driver_priv;
733 struct intel_context *ctx;
736 if (!contexts_enabled(dev))
739 ret = i915_mutex_lock_interruptible(dev);
743 ctx = i915_gem_create_context(dev, file_priv);
744 mutex_unlock(&dev->struct_mutex);
748 args->ctx_id = ctx->user_handle;
749 DRM_DEBUG_DRIVER("HW context %d created\n", args->ctx_id);
754 int i915_gem_context_destroy_ioctl(struct drm_device *dev, void *data,
755 struct drm_file *file)
757 struct drm_i915_gem_context_destroy *args = data;
758 struct drm_i915_file_private *file_priv = file->driver_priv;
759 struct intel_context *ctx;
762 if (args->ctx_id == DEFAULT_CONTEXT_HANDLE)
765 ret = i915_mutex_lock_interruptible(dev);
769 ctx = i915_gem_context_get(file_priv, args->ctx_id);
771 mutex_unlock(&dev->struct_mutex);
775 idr_remove(&ctx->file_priv->context_idr, ctx->user_handle);
776 i915_gem_context_unreference(ctx);
777 mutex_unlock(&dev->struct_mutex);
779 DRM_DEBUG_DRIVER("HW context %d destroyed\n", args->ctx_id);