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 struct i915_hw_context
*
100 i915_gem_context_get(struct drm_i915_file_private
*file_priv
, u32 id
);
101 static int do_switch(struct intel_ring_buffer
*ring
,
102 struct i915_hw_context
*to
);
104 static size_t get_context_alignment(struct drm_device
*dev
)
107 return GEN6_CONTEXT_ALIGN
;
109 return GEN7_CONTEXT_ALIGN
;
112 static int get_context_size(struct drm_device
*dev
)
114 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
118 switch (INTEL_INFO(dev
)->gen
) {
120 reg
= I915_READ(CXT_SIZE
);
121 ret
= GEN6_CXT_TOTAL_SIZE(reg
) * 64;
124 reg
= I915_READ(GEN7_CXT_SIZE
);
126 ret
= HSW_CXT_TOTAL_SIZE
;
128 ret
= GEN7_CXT_TOTAL_SIZE(reg
) * 64;
131 ret
= GEN8_CXT_TOTAL_SIZE
;
140 void i915_gem_context_free(struct kref
*ctx_ref
)
142 struct i915_hw_context
*ctx
= container_of(ctx_ref
,
145 list_del(&ctx
->link
);
146 drm_gem_object_unreference(&ctx
->obj
->base
);
150 static struct i915_hw_context
*
151 create_hw_context(struct drm_device
*dev
,
152 struct drm_i915_file_private
*file_priv
)
154 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
155 struct i915_hw_context
*ctx
;
158 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
160 return ERR_PTR(-ENOMEM
);
162 kref_init(&ctx
->ref
);
163 ctx
->obj
= i915_gem_alloc_object(dev
, dev_priv
->hw_context_size
);
164 INIT_LIST_HEAD(&ctx
->link
);
165 if (ctx
->obj
== NULL
) {
167 DRM_DEBUG_DRIVER("Context object allocated failed\n");
168 return ERR_PTR(-ENOMEM
);
171 if (INTEL_INFO(dev
)->gen
>= 7) {
172 ret
= i915_gem_object_set_cache_level(ctx
->obj
,
174 /* Failure shouldn't ever happen this early */
179 list_add_tail(&ctx
->link
, &dev_priv
->context_list
);
181 /* Default context will never have a file_priv */
182 if (file_priv
== NULL
)
185 ret
= idr_alloc(&file_priv
->context_idr
, ctx
, DEFAULT_CONTEXT_ID
+ 1, 0,
190 ctx
->file_priv
= file_priv
;
192 /* NB: Mark all slices as needing a remap so that when the context first
193 * loads it will restore whatever remap state already exists. If there
194 * is no remap info, it will be a NOP. */
195 ctx
->remap_slice
= (1 << NUM_L3_SLICES(dev
)) - 1;
200 i915_gem_context_unreference(ctx
);
204 static inline bool is_default_context(struct i915_hw_context
*ctx
)
206 /* Cheap trick to determine default contexts */
207 return ctx
->file_priv
? false : true;
211 * The default context needs to exist per ring that uses contexts. It stores the
212 * context state of the GPU for applications that don't utilize HW contexts, as
213 * well as an idle case.
215 static int create_default_context(struct drm_device
*dev
)
217 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
218 struct i915_hw_context
*ctx
;
221 BUG_ON(!mutex_is_locked(&dev
->struct_mutex
));
223 ctx
= create_hw_context(dev
, NULL
);
227 /* We may need to do things with the shrinker which require us to
228 * immediately switch back to the default context. This can cause a
229 * problem as pinning the default context also requires GTT space which
230 * may not be available. To avoid this we always pin the
233 ret
= i915_gem_obj_ggtt_pin(ctx
->obj
, get_context_alignment(dev
),
236 DRM_DEBUG_DRIVER("Couldn't pin %d\n", ret
);
240 dev_priv
->ring
[RCS
].default_context
= ctx
;
242 DRM_DEBUG_DRIVER("Default HW context loaded\n");
246 i915_gem_context_unreference(ctx
);
250 void i915_gem_context_reset(struct drm_device
*dev
)
252 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
253 struct intel_ring_buffer
*ring
;
256 if (!HAS_HW_CONTEXTS(dev
))
259 /* Prevent the hardware from restoring the last context (which hung) on
261 for (i
= 0; i
< I915_NUM_RINGS
; i
++) {
262 struct i915_hw_context
*dctx
;
263 if (!(INTEL_INFO(dev
)->ring_mask
& (1<<i
)))
266 /* Do a fake switch to the default context */
267 ring
= &dev_priv
->ring
[i
];
268 dctx
= ring
->default_context
;
272 if (!ring
->last_context
)
275 if (ring
->last_context
== dctx
)
279 WARN_ON(i915_gem_obj_ggtt_pin(dctx
->obj
,
280 get_context_alignment(dev
),
282 /* Fake a finish/inactive */
283 dctx
->obj
->base
.write_domain
= 0;
284 dctx
->obj
->active
= 0;
287 i915_gem_context_unreference(ring
->last_context
);
288 i915_gem_context_reference(dctx
);
289 ring
->last_context
= dctx
;
293 int i915_gem_context_init(struct drm_device
*dev
)
295 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
296 struct intel_ring_buffer
*ring
;
299 if (!HAS_HW_CONTEXTS(dev
))
302 /* Init should only be called once per module load. Eventually the
303 * restriction on the context_disabled check can be loosened. */
304 if (WARN_ON(dev_priv
->ring
[RCS
].default_context
))
307 dev_priv
->hw_context_size
= round_up(get_context_size(dev
), 4096);
309 if (dev_priv
->hw_context_size
> (1<<20)) {
310 DRM_DEBUG_DRIVER("Disabling HW Contexts; invalid size\n");
314 ret
= create_default_context(dev
);
316 DRM_DEBUG_DRIVER("Disabling HW Contexts; create failed %d\n",
321 for (i
= RCS
+ 1; i
< I915_NUM_RINGS
; i
++) {
322 if (!(INTEL_INFO(dev
)->ring_mask
& (1<<i
)))
325 ring
= &dev_priv
->ring
[i
];
327 /* NB: RCS will hold a ref for all rings */
328 ring
->default_context
= dev_priv
->ring
[RCS
].default_context
;
331 DRM_DEBUG_DRIVER("HW context support initialized\n");
335 void i915_gem_context_fini(struct drm_device
*dev
)
337 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
338 struct i915_hw_context
*dctx
= dev_priv
->ring
[RCS
].default_context
;
341 if (!HAS_HW_CONTEXTS(dev
))
344 /* The only known way to stop the gpu from accessing the hw context is
345 * to reset it. Do this as the very last operation to avoid confusing
346 * other code, leading to spurious errors. */
347 intel_gpu_reset(dev
);
349 /* When default context is created and switched to, base object refcount
350 * will be 2 (+1 from object creation and +1 from do_switch()).
351 * i915_gem_context_fini() will be called after gpu_idle() has switched
352 * to default context. So we need to unreference the base object once
353 * to offset the do_switch part, so that i915_gem_context_unreference()
354 * can then free the base object correctly. */
355 WARN_ON(!dev_priv
->ring
[RCS
].last_context
);
356 if (dev_priv
->ring
[RCS
].last_context
== dctx
) {
357 /* Fake switch to NULL context */
358 WARN_ON(dctx
->obj
->active
);
359 i915_gem_object_ggtt_unpin(dctx
->obj
);
360 i915_gem_context_unreference(dctx
);
361 dev_priv
->ring
[RCS
].last_context
= NULL
;
364 for (i
= 0; i
< I915_NUM_RINGS
; i
++) {
365 struct intel_ring_buffer
*ring
= &dev_priv
->ring
[i
];
366 if (!(INTEL_INFO(dev
)->ring_mask
& (1<<i
)))
369 if (ring
->last_context
)
370 i915_gem_context_unreference(ring
->last_context
);
372 ring
->default_context
= NULL
;
373 ring
->last_context
= NULL
;
376 i915_gem_object_ggtt_unpin(dctx
->obj
);
377 i915_gem_context_unreference(dctx
);
380 int i915_gem_context_enable(struct drm_i915_private
*dev_priv
)
382 struct intel_ring_buffer
*ring
;
385 if (!HAS_HW_CONTEXTS(dev_priv
->dev
))
388 /* FIXME: We should make this work, even in reset */
389 if (i915_reset_in_progress(&dev_priv
->gpu_error
))
392 BUG_ON(!dev_priv
->ring
[RCS
].default_context
);
393 for_each_ring(ring
, dev_priv
, i
) {
394 ret
= do_switch(ring
, ring
->default_context
);
402 static int context_idr_cleanup(int id
, void *p
, void *data
)
404 struct i915_hw_context
*ctx
= p
;
406 BUG_ON(id
== DEFAULT_CONTEXT_ID
);
408 i915_gem_context_unreference(ctx
);
412 struct i915_ctx_hang_stats
*
413 i915_gem_context_get_hang_stats(struct drm_device
*dev
,
414 struct drm_file
*file
,
417 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
418 struct i915_hw_context
*ctx
;
420 if (id
== DEFAULT_CONTEXT_ID
)
421 return &file_priv
->hang_stats
;
423 if (!HAS_HW_CONTEXTS(dev
))
424 return ERR_PTR(-ENOENT
);
426 ctx
= i915_gem_context_get(file
->driver_priv
, id
);
428 return ERR_PTR(-ENOENT
);
430 return &ctx
->hang_stats
;
433 int i915_gem_context_open(struct drm_device
*dev
, struct drm_file
*file
)
435 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
437 if (!HAS_HW_CONTEXTS(dev
))
440 idr_init(&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 if (!HAS_HW_CONTEXTS(dev
))
452 mutex_lock(&dev
->struct_mutex
);
453 idr_for_each(&file_priv
->context_idr
, context_idr_cleanup
, NULL
);
454 idr_destroy(&file_priv
->context_idr
);
455 mutex_unlock(&dev
->struct_mutex
);
458 static struct i915_hw_context
*
459 i915_gem_context_get(struct drm_i915_file_private
*file_priv
, u32 id
)
461 return (struct i915_hw_context
*)idr_find(&file_priv
->context_idr
, id
);
465 mi_set_context(struct intel_ring_buffer
*ring
,
466 struct i915_hw_context
*new_context
,
471 /* w/a: If Flush TLB Invalidation Mode is enabled, driver must do a TLB
472 * invalidation prior to MI_SET_CONTEXT. On GEN6 we don't set the value
473 * explicitly, so we rely on the value at ring init, stored in
474 * itlb_before_ctx_switch.
476 if (IS_GEN6(ring
->dev
) && ring
->itlb_before_ctx_switch
) {
477 ret
= ring
->flush(ring
, I915_GEM_GPU_DOMAINS
, 0);
482 ret
= intel_ring_begin(ring
, 6);
486 /* WaProgramMiArbOnOffAroundMiSetContext:ivb,vlv,hsw */
487 if (IS_GEN7(ring
->dev
))
488 intel_ring_emit(ring
, MI_ARB_ON_OFF
| MI_ARB_DISABLE
);
490 intel_ring_emit(ring
, MI_NOOP
);
492 intel_ring_emit(ring
, MI_NOOP
);
493 intel_ring_emit(ring
, MI_SET_CONTEXT
);
494 intel_ring_emit(ring
, i915_gem_obj_ggtt_offset(new_context
->obj
) |
496 MI_SAVE_EXT_STATE_EN
|
497 MI_RESTORE_EXT_STATE_EN
|
499 /* w/a: MI_SET_CONTEXT must always be followed by MI_NOOP */
500 intel_ring_emit(ring
, MI_NOOP
);
502 if (IS_GEN7(ring
->dev
))
503 intel_ring_emit(ring
, MI_ARB_ON_OFF
| MI_ARB_ENABLE
);
505 intel_ring_emit(ring
, MI_NOOP
);
507 intel_ring_advance(ring
);
512 static int do_switch(struct intel_ring_buffer
*ring
,
513 struct i915_hw_context
*to
)
515 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
516 struct i915_hw_context
*from
= ring
->last_context
;
520 if (from
!= NULL
&& ring
== &dev_priv
->ring
[RCS
]) {
521 BUG_ON(from
->obj
== NULL
);
522 BUG_ON(!i915_gem_obj_is_pinned(from
->obj
));
525 if (from
== to
&& from
->last_ring
== ring
&& !to
->remap_slice
)
528 if (ring
!= &dev_priv
->ring
[RCS
]) {
530 i915_gem_context_unreference(from
);
534 ret
= i915_gem_obj_ggtt_pin(to
->obj
, get_context_alignment(ring
->dev
),
539 /* Clear this page out of any CPU caches for coherent swap-in/out. Note
540 * that thanks to write = false in this call and us not setting any gpu
541 * write domains when putting a context object onto the active list
542 * (when switching away from it), this won't block.
543 * XXX: We need a real interface to do this instead of trickery. */
544 ret
= i915_gem_object_set_to_gtt_domain(to
->obj
, false);
546 i915_gem_object_ggtt_unpin(to
->obj
);
550 if (!to
->obj
->has_global_gtt_mapping
) {
551 struct i915_vma
*vma
= i915_gem_obj_to_vma(to
->obj
,
552 &dev_priv
->gtt
.base
);
553 vma
->bind_vma(vma
, to
->obj
->cache_level
, GLOBAL_BIND
);
556 if (!to
->is_initialized
|| is_default_context(to
))
557 hw_flags
|= MI_RESTORE_INHIBIT
;
559 ret
= mi_set_context(ring
, to
, hw_flags
);
561 i915_gem_object_ggtt_unpin(to
->obj
);
565 for (i
= 0; i
< MAX_L3_SLICES
; i
++) {
566 if (!(to
->remap_slice
& (1<<i
)))
569 ret
= i915_gem_l3_remap(ring
, i
);
570 /* If it failed, try again next round */
572 DRM_DEBUG_DRIVER("L3 remapping failed\n");
574 to
->remap_slice
&= ~(1<<i
);
577 /* The backing object for the context is done after switching to the
578 * *next* context. Therefore we cannot retire the previous context until
579 * the next context has already started running. In fact, the below code
580 * is a bit suboptimal because the retiring can occur simply after the
581 * MI_SET_CONTEXT instead of when the next seqno has completed.
584 from
->obj
->base
.read_domains
= I915_GEM_DOMAIN_INSTRUCTION
;
585 i915_vma_move_to_active(i915_gem_obj_to_ggtt(from
->obj
), ring
);
586 /* As long as MI_SET_CONTEXT is serializing, ie. it flushes the
587 * whole damn pipeline, we don't need to explicitly mark the
588 * object dirty. The only exception is that the context must be
589 * correct in case the object gets swapped out. Ideally we'd be
590 * able to defer doing this until we know the object would be
591 * swapped, but there is no way to do that yet.
593 from
->obj
->dirty
= 1;
594 BUG_ON(from
->obj
->ring
!= ring
);
596 /* obj is kept alive until the next request by its active ref */
597 i915_gem_object_ggtt_unpin(from
->obj
);
598 i915_gem_context_unreference(from
);
602 i915_gem_context_reference(to
);
603 ring
->last_context
= to
;
604 to
->is_initialized
= true;
605 to
->last_ring
= ring
;
611 * i915_switch_context() - perform a GPU context switch.
612 * @ring: ring for which we'll execute the context switch
613 * @file_priv: file_priv associated with the context, may be NULL
614 * @id: context id number
616 * The context life cycle is simple. The context refcount is incremented and
617 * decremented by 1 and create and destroy. If the context is in use by the GPU,
618 * it will have a refoucnt > 1. This allows us to destroy the context abstract
619 * object while letting the normal object tracking destroy the backing BO.
621 int i915_switch_context(struct intel_ring_buffer
*ring
,
622 struct drm_file
*file
,
625 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
626 struct i915_hw_context
*to
;
628 if (!HAS_HW_CONTEXTS(ring
->dev
))
631 WARN_ON(!mutex_is_locked(&dev_priv
->dev
->struct_mutex
));
633 if (to_id
== DEFAULT_CONTEXT_ID
) {
634 to
= ring
->default_context
;
639 to
= i915_gem_context_get(file
->driver_priv
, to_id
);
644 return do_switch(ring
, to
);
647 int i915_gem_context_create_ioctl(struct drm_device
*dev
, void *data
,
648 struct drm_file
*file
)
650 struct drm_i915_gem_context_create
*args
= data
;
651 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
652 struct i915_hw_context
*ctx
;
655 if (!(dev
->driver
->driver_features
& DRIVER_GEM
))
658 if (!HAS_HW_CONTEXTS(dev
))
661 ret
= i915_mutex_lock_interruptible(dev
);
665 ctx
= create_hw_context(dev
, file_priv
);
666 mutex_unlock(&dev
->struct_mutex
);
670 args
->ctx_id
= ctx
->id
;
671 DRM_DEBUG_DRIVER("HW context %d created\n", args
->ctx_id
);
676 int i915_gem_context_destroy_ioctl(struct drm_device
*dev
, void *data
,
677 struct drm_file
*file
)
679 struct drm_i915_gem_context_destroy
*args
= data
;
680 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
681 struct i915_hw_context
*ctx
;
684 if (!(dev
->driver
->driver_features
& DRIVER_GEM
))
687 ret
= i915_mutex_lock_interruptible(dev
);
691 ctx
= i915_gem_context_get(file_priv
, args
->ctx_id
);
693 mutex_unlock(&dev
->struct_mutex
);
697 idr_remove(&ctx
->file_priv
->context_idr
, ctx
->id
);
698 i915_gem_context_unreference(ctx
);
699 mutex_unlock(&dev
->struct_mutex
);
701 DRM_DEBUG_DRIVER("HW context %d destroyed\n", args
->ctx_id
);