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 static void i915_gem_context_clean(struct intel_context
*ctx
)
138 struct i915_hw_ppgtt
*ppgtt
= ctx
->ppgtt
;
139 struct i915_vma
*vma
, *next
;
144 list_for_each_entry_safe(vma
, next
, &ppgtt
->base
.inactive_list
,
146 if (WARN_ON(__i915_vma_unbind_no_wait(vma
)))
151 void i915_gem_context_free(struct kref
*ctx_ref
)
153 struct intel_context
*ctx
= container_of(ctx_ref
, typeof(*ctx
), ref
);
155 trace_i915_context_free(ctx
);
157 if (i915
.enable_execlists
)
158 intel_lr_context_free(ctx
);
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.
165 i915_gem_context_clean(ctx
);
167 i915_ppgtt_put(ctx
->ppgtt
);
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
);
175 struct drm_i915_gem_object
*
176 i915_gem_alloc_context_obj(struct drm_device
*dev
, size_t size
)
178 struct drm_i915_gem_object
*obj
;
181 obj
= i915_gem_alloc_object(dev
, size
);
183 return ERR_PTR(-ENOMEM
);
186 * Try to make the context utilize L3 as well as LLC.
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.
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.
197 * This is only applicable for Ivy Bridge devices since
198 * later platforms don't have L3 control bits in the PTE.
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 */
204 drm_gem_object_unreference(&obj
->base
);
212 static struct intel_context
*
213 __create_hw_context(struct drm_device
*dev
,
214 struct drm_i915_file_private
*file_priv
)
216 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
217 struct intel_context
*ctx
;
220 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
222 return ERR_PTR(-ENOMEM
);
224 kref_init(&ctx
->ref
);
225 list_add_tail(&ctx
->link
, &dev_priv
->context_list
);
226 ctx
->i915
= dev_priv
;
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
);
235 ctx
->legacy_hw_ctx
.rcs_state
= obj
;
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
);
245 ret
= DEFAULT_CONTEXT_HANDLE
;
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;
254 ctx
->hang_stats
.ban_period_seconds
= DRM_I915_CTX_BAN_PERIOD
;
259 i915_gem_context_unreference(ctx
);
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.
268 static struct intel_context
*
269 i915_gem_create_context(struct drm_device
*dev
,
270 struct drm_i915_file_private
*file_priv
)
272 const bool is_global_default_ctx
= file_priv
== NULL
;
273 struct intel_context
*ctx
;
276 BUG_ON(!mutex_is_locked(&dev
->struct_mutex
));
278 ctx
= __create_hw_context(dev
, file_priv
);
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
290 ret
= i915_gem_obj_ggtt_pin(ctx
->legacy_hw_ctx
.rcs_state
,
291 get_context_alignment(dev
), 0);
293 DRM_DEBUG_DRIVER("Couldn't pin %d\n", ret
);
298 if (USES_FULL_PPGTT(dev
)) {
299 struct i915_hw_ppgtt
*ppgtt
= i915_ppgtt_create(dev
, file_priv
);
301 if (IS_ERR_OR_NULL(ppgtt
)) {
302 DRM_DEBUG_DRIVER("PPGTT setup failed (%ld)\n",
304 ret
= PTR_ERR(ppgtt
);
311 trace_i915_context_create(ctx
);
316 if (is_global_default_ctx
&& ctx
->legacy_hw_ctx
.rcs_state
)
317 i915_gem_object_ggtt_unpin(ctx
->legacy_hw_ctx
.rcs_state
);
319 idr_remove(&file_priv
->context_idr
, ctx
->user_handle
);
320 i915_gem_context_unreference(ctx
);
324 static void i915_gem_context_unpin(struct intel_context
*ctx
,
325 struct intel_engine_cs
*engine
)
327 if (i915
.enable_execlists
) {
328 intel_lr_context_unpin(ctx
, engine
);
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
);
336 void i915_gem_context_reset(struct drm_device
*dev
)
338 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
341 if (i915
.enable_execlists
) {
342 struct intel_context
*ctx
;
344 list_for_each_entry(ctx
, &dev_priv
->context_list
, link
)
345 intel_lr_context_reset(dev_priv
, ctx
);
348 for (i
= 0; i
< I915_NUM_ENGINES
; i
++) {
349 struct intel_engine_cs
*engine
= &dev_priv
->engine
[i
];
351 if (engine
->last_context
) {
352 i915_gem_context_unpin(engine
->last_context
, engine
);
353 engine
->last_context
= NULL
;
357 /* Force the GPU state to be reinitialised on enabling */
358 dev_priv
->kernel_context
->legacy_hw_ctx
.initialized
= false;
361 int i915_gem_context_init(struct drm_device
*dev
)
363 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
364 struct intel_context
*ctx
;
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
))
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");
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;
391 ctx
= i915_gem_create_context(dev
, NULL
);
393 DRM_ERROR("Failed to create default global context (error %ld)\n",
398 dev_priv
->kernel_context
= ctx
;
400 DRM_DEBUG_DRIVER("%s context support initialized\n",
401 i915
.enable_execlists
? "LR" :
402 dev_priv
->hw_context_size
? "HW" : "fake");
406 void i915_gem_context_fini(struct drm_device
*dev
)
408 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
409 struct intel_context
*dctx
= dev_priv
->kernel_context
;
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
, ALL_ENGINES
);
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
->engine
[RCS
].last_context
);
426 i915_gem_object_ggtt_unpin(dctx
->legacy_hw_ctx
.rcs_state
);
429 for (i
= I915_NUM_ENGINES
; --i
>= 0;) {
430 struct intel_engine_cs
*engine
= &dev_priv
->engine
[i
];
432 if (engine
->last_context
) {
433 i915_gem_context_unpin(engine
->last_context
, engine
);
434 engine
->last_context
= NULL
;
438 i915_gem_context_unreference(dctx
);
439 dev_priv
->kernel_context
= NULL
;
442 int i915_gem_context_enable(struct drm_i915_gem_request
*req
)
444 struct intel_engine_cs
*engine
= req
->engine
;
447 if (i915
.enable_execlists
) {
448 if (engine
->init_context
== NULL
)
451 ret
= engine
->init_context(req
);
453 ret
= i915_switch_context(req
);
456 DRM_ERROR("ring init context: %d\n", ret
);
463 static int context_idr_cleanup(int id
, void *p
, void *data
)
465 struct intel_context
*ctx
= p
;
467 i915_gem_context_unreference(ctx
);
471 int i915_gem_context_open(struct drm_device
*dev
, struct drm_file
*file
)
473 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
474 struct intel_context
*ctx
;
476 idr_init(&file_priv
->context_idr
);
478 mutex_lock(&dev
->struct_mutex
);
479 ctx
= i915_gem_create_context(dev
, file_priv
);
480 mutex_unlock(&dev
->struct_mutex
);
483 idr_destroy(&file_priv
->context_idr
);
490 void i915_gem_context_close(struct drm_device
*dev
, struct drm_file
*file
)
492 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
494 idr_for_each(&file_priv
->context_idr
, context_idr_cleanup
, NULL
);
495 idr_destroy(&file_priv
->context_idr
);
498 struct intel_context
*
499 i915_gem_context_get(struct drm_i915_file_private
*file_priv
, u32 id
)
501 struct intel_context
*ctx
;
503 ctx
= (struct intel_context
*)idr_find(&file_priv
->context_idr
, id
);
505 return ERR_PTR(-ENOENT
);
511 mi_set_context(struct drm_i915_gem_request
*req
, u32 hw_flags
)
513 struct intel_engine_cs
*engine
= req
->engine
;
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(engine
->dev
) ?
518 hweight32(INTEL_INFO(engine
->dev
)->ring_mask
) - 1 :
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.
527 if (IS_GEN6(engine
->dev
)) {
528 ret
= engine
->flush(req
, I915_GEM_GPU_DOMAINS
, 0);
533 /* These flags are for resource streamer on HSW+ */
534 if (IS_HASWELL(engine
->dev
) || INTEL_INFO(engine
->dev
)->gen
>= 8)
535 flags
|= (HSW_MI_RS_SAVE_STATE_EN
| HSW_MI_RS_RESTORE_STATE_EN
);
536 else if (INTEL_INFO(engine
->dev
)->gen
< 8)
537 flags
|= (MI_SAVE_EXT_STATE_EN
| MI_RESTORE_EXT_STATE_EN
);
541 if (INTEL_INFO(engine
->dev
)->gen
>= 7)
542 len
+= 2 + (num_rings
? 4*num_rings
+ 6 : 0);
544 ret
= intel_ring_begin(req
, len
);
548 /* WaProgramMiArbOnOffAroundMiSetContext:ivb,vlv,hsw,bdw,chv */
549 if (INTEL_INFO(engine
->dev
)->gen
>= 7) {
550 intel_ring_emit(engine
, MI_ARB_ON_OFF
| MI_ARB_DISABLE
);
552 struct intel_engine_cs
*signaller
;
554 intel_ring_emit(engine
,
555 MI_LOAD_REGISTER_IMM(num_rings
));
556 for_each_engine(signaller
, to_i915(engine
->dev
)) {
557 if (signaller
== engine
)
560 intel_ring_emit_reg(engine
,
561 RING_PSMI_CTL(signaller
->mmio_base
));
562 intel_ring_emit(engine
,
563 _MASKED_BIT_ENABLE(GEN6_PSMI_SLEEP_MSG_DISABLE
));
568 intel_ring_emit(engine
, MI_NOOP
);
569 intel_ring_emit(engine
, MI_SET_CONTEXT
);
570 intel_ring_emit(engine
,
571 i915_gem_obj_ggtt_offset(req
->ctx
->legacy_hw_ctx
.rcs_state
) |
574 * w/a: MI_SET_CONTEXT must always be followed by MI_NOOP
575 * WaMiSetContext_Hang:snb,ivb,vlv
577 intel_ring_emit(engine
, MI_NOOP
);
579 if (INTEL_INFO(engine
->dev
)->gen
>= 7) {
581 struct intel_engine_cs
*signaller
;
582 i915_reg_t last_reg
= {}; /* keep gcc quiet */
584 intel_ring_emit(engine
,
585 MI_LOAD_REGISTER_IMM(num_rings
));
586 for_each_engine(signaller
, to_i915(engine
->dev
)) {
587 if (signaller
== engine
)
590 last_reg
= RING_PSMI_CTL(signaller
->mmio_base
);
591 intel_ring_emit_reg(engine
, last_reg
);
592 intel_ring_emit(engine
,
593 _MASKED_BIT_DISABLE(GEN6_PSMI_SLEEP_MSG_DISABLE
));
596 /* Insert a delay before the next switch! */
597 intel_ring_emit(engine
,
598 MI_STORE_REGISTER_MEM
|
599 MI_SRM_LRM_GLOBAL_GTT
);
600 intel_ring_emit_reg(engine
, last_reg
);
601 intel_ring_emit(engine
, engine
->scratch
.gtt_offset
);
602 intel_ring_emit(engine
, MI_NOOP
);
604 intel_ring_emit(engine
, MI_ARB_ON_OFF
| MI_ARB_ENABLE
);
607 intel_ring_advance(engine
);
612 static inline bool skip_rcs_switch(struct intel_engine_cs
*engine
,
613 struct intel_context
*to
)
618 if (!to
->legacy_hw_ctx
.initialized
)
622 !(intel_engine_flag(engine
) & to
->ppgtt
->pd_dirty_rings
))
625 return to
== engine
->last_context
;
629 needs_pd_load_pre(struct intel_engine_cs
*engine
, struct intel_context
*to
)
634 if (engine
->last_context
== to
&&
635 !(intel_engine_flag(engine
) & to
->ppgtt
->pd_dirty_rings
))
638 if (engine
->id
!= RCS
)
641 if (INTEL_INFO(engine
->dev
)->gen
< 8)
648 needs_pd_load_post(struct intel_context
*to
, u32 hw_flags
)
653 if (!IS_GEN8(to
->i915
))
656 if (hw_flags
& MI_RESTORE_INHIBIT
)
662 static int do_rcs_switch(struct drm_i915_gem_request
*req
)
664 struct intel_context
*to
= req
->ctx
;
665 struct intel_engine_cs
*engine
= req
->engine
;
666 struct intel_context
*from
;
670 if (skip_rcs_switch(engine
, to
))
673 /* Trying to pin first makes error handling easier. */
674 ret
= i915_gem_obj_ggtt_pin(to
->legacy_hw_ctx
.rcs_state
,
675 get_context_alignment(engine
->dev
),
681 * Pin can switch back to the default context if we end up calling into
682 * evict_everything - as a last ditch gtt defrag effort that also
683 * switches to the default context. Hence we need to reload from here.
685 * XXX: Doing so is painfully broken!
687 from
= engine
->last_context
;
690 * Clear this page out of any CPU caches for coherent swap-in/out. Note
691 * that thanks to write = false in this call and us not setting any gpu
692 * write domains when putting a context object onto the active list
693 * (when switching away from it), this won't block.
695 * XXX: We need a real interface to do this instead of trickery.
697 ret
= i915_gem_object_set_to_gtt_domain(to
->legacy_hw_ctx
.rcs_state
, false);
701 if (needs_pd_load_pre(engine
, to
)) {
702 /* Older GENs and non render rings still want the load first,
703 * "PP_DCLV followed by PP_DIR_BASE register through Load
704 * Register Immediate commands in Ring Buffer before submitting
706 trace_switch_mm(engine
, to
);
707 ret
= to
->ppgtt
->switch_mm(to
->ppgtt
, req
);
712 if (!to
->legacy_hw_ctx
.initialized
|| i915_gem_context_is_default(to
))
713 /* NB: If we inhibit the restore, the context is not allowed to
714 * die because future work may end up depending on valid address
715 * space. This means we must enforce that a page table load
716 * occur when this occurs. */
717 hw_flags
= MI_RESTORE_INHIBIT
;
718 else if (to
->ppgtt
&&
719 intel_engine_flag(engine
) & to
->ppgtt
->pd_dirty_rings
)
720 hw_flags
= MI_FORCE_RESTORE
;
724 /* We should never emit switch_mm more than once */
725 WARN_ON(needs_pd_load_pre(engine
, to
) &&
726 needs_pd_load_post(to
, hw_flags
));
728 if (to
!= from
|| (hw_flags
& MI_FORCE_RESTORE
)) {
729 ret
= mi_set_context(req
, hw_flags
);
734 /* The backing object for the context is done after switching to the
735 * *next* context. Therefore we cannot retire the previous context until
736 * the next context has already started running. In fact, the below code
737 * is a bit suboptimal because the retiring can occur simply after the
738 * MI_SET_CONTEXT instead of when the next seqno has completed.
741 from
->legacy_hw_ctx
.rcs_state
->base
.read_domains
= I915_GEM_DOMAIN_INSTRUCTION
;
742 i915_vma_move_to_active(i915_gem_obj_to_ggtt(from
->legacy_hw_ctx
.rcs_state
), req
);
743 /* As long as MI_SET_CONTEXT is serializing, ie. it flushes the
744 * whole damn pipeline, we don't need to explicitly mark the
745 * object dirty. The only exception is that the context must be
746 * correct in case the object gets swapped out. Ideally we'd be
747 * able to defer doing this until we know the object would be
748 * swapped, but there is no way to do that yet.
750 from
->legacy_hw_ctx
.rcs_state
->dirty
= 1;
752 /* obj is kept alive until the next request by its active ref */
753 i915_gem_object_ggtt_unpin(from
->legacy_hw_ctx
.rcs_state
);
754 i915_gem_context_unreference(from
);
756 i915_gem_context_reference(to
);
757 engine
->last_context
= to
;
759 /* GEN8 does *not* require an explicit reload if the PDPs have been
760 * setup, and we do not wish to move them.
762 if (needs_pd_load_post(to
, hw_flags
)) {
763 trace_switch_mm(engine
, to
);
764 ret
= to
->ppgtt
->switch_mm(to
->ppgtt
, req
);
765 /* The hardware context switch is emitted, but we haven't
766 * actually changed the state - so it's probably safe to bail
767 * here. Still, let the user know something dangerous has
775 to
->ppgtt
->pd_dirty_rings
&= ~intel_engine_flag(engine
);
777 for (i
= 0; i
< MAX_L3_SLICES
; i
++) {
778 if (!(to
->remap_slice
& (1<<i
)))
781 ret
= i915_gem_l3_remap(req
, i
);
785 to
->remap_slice
&= ~(1<<i
);
788 if (!to
->legacy_hw_ctx
.initialized
) {
789 if (engine
->init_context
) {
790 ret
= engine
->init_context(req
);
794 to
->legacy_hw_ctx
.initialized
= true;
800 i915_gem_object_ggtt_unpin(to
->legacy_hw_ctx
.rcs_state
);
805 * i915_switch_context() - perform a GPU context switch.
806 * @req: request for which we'll execute the context switch
808 * The context life cycle is simple. The context refcount is incremented and
809 * decremented by 1 and create and destroy. If the context is in use by the GPU,
810 * it will have a refcount > 1. This allows us to destroy the context abstract
811 * object while letting the normal object tracking destroy the backing BO.
813 * This function should not be used in execlists mode. Instead the context is
814 * switched by writing to the ELSP and requests keep a reference to their
817 int i915_switch_context(struct drm_i915_gem_request
*req
)
819 struct intel_engine_cs
*engine
= req
->engine
;
820 struct drm_i915_private
*dev_priv
= req
->i915
;
822 WARN_ON(i915
.enable_execlists
);
823 WARN_ON(!mutex_is_locked(&dev_priv
->dev
->struct_mutex
));
825 if (engine
->id
!= RCS
||
826 req
->ctx
->legacy_hw_ctx
.rcs_state
== NULL
) {
827 struct intel_context
*to
= req
->ctx
;
829 if (needs_pd_load_pre(engine
, to
)) {
832 trace_switch_mm(engine
, to
);
833 ret
= to
->ppgtt
->switch_mm(to
->ppgtt
, req
);
837 /* Doing a PD load always reloads the page dirs */
838 to
->ppgtt
->pd_dirty_rings
&= ~intel_engine_flag(engine
);
841 if (to
!= engine
->last_context
) {
842 i915_gem_context_reference(to
);
843 if (engine
->last_context
)
844 i915_gem_context_unreference(engine
->last_context
);
845 engine
->last_context
= to
;
851 return do_rcs_switch(req
);
854 static bool contexts_enabled(struct drm_device
*dev
)
856 return i915
.enable_execlists
|| to_i915(dev
)->hw_context_size
;
859 int i915_gem_context_create_ioctl(struct drm_device
*dev
, void *data
,
860 struct drm_file
*file
)
862 struct drm_i915_gem_context_create
*args
= data
;
863 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
864 struct intel_context
*ctx
;
867 if (!contexts_enabled(dev
))
873 ret
= i915_mutex_lock_interruptible(dev
);
877 ctx
= i915_gem_create_context(dev
, file_priv
);
878 mutex_unlock(&dev
->struct_mutex
);
882 args
->ctx_id
= ctx
->user_handle
;
883 DRM_DEBUG_DRIVER("HW context %d created\n", args
->ctx_id
);
888 int i915_gem_context_destroy_ioctl(struct drm_device
*dev
, void *data
,
889 struct drm_file
*file
)
891 struct drm_i915_gem_context_destroy
*args
= data
;
892 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
893 struct intel_context
*ctx
;
899 if (args
->ctx_id
== DEFAULT_CONTEXT_HANDLE
)
902 ret
= i915_mutex_lock_interruptible(dev
);
906 ctx
= i915_gem_context_get(file_priv
, args
->ctx_id
);
908 mutex_unlock(&dev
->struct_mutex
);
912 idr_remove(&ctx
->file_priv
->context_idr
, ctx
->user_handle
);
913 i915_gem_context_unreference(ctx
);
914 mutex_unlock(&dev
->struct_mutex
);
916 DRM_DEBUG_DRIVER("HW context %d destroyed\n", args
->ctx_id
);
920 int i915_gem_context_getparam_ioctl(struct drm_device
*dev
, void *data
,
921 struct drm_file
*file
)
923 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
924 struct drm_i915_gem_context_param
*args
= data
;
925 struct intel_context
*ctx
;
928 ret
= i915_mutex_lock_interruptible(dev
);
932 ctx
= i915_gem_context_get(file_priv
, args
->ctx_id
);
934 mutex_unlock(&dev
->struct_mutex
);
939 switch (args
->param
) {
940 case I915_CONTEXT_PARAM_BAN_PERIOD
:
941 args
->value
= ctx
->hang_stats
.ban_period_seconds
;
943 case I915_CONTEXT_PARAM_NO_ZEROMAP
:
944 args
->value
= ctx
->flags
& CONTEXT_NO_ZEROMAP
;
946 case I915_CONTEXT_PARAM_GTT_SIZE
:
948 args
->value
= ctx
->ppgtt
->base
.total
;
949 else if (to_i915(dev
)->mm
.aliasing_ppgtt
)
950 args
->value
= to_i915(dev
)->mm
.aliasing_ppgtt
->base
.total
;
952 args
->value
= to_i915(dev
)->ggtt
.base
.total
;
958 mutex_unlock(&dev
->struct_mutex
);
963 int i915_gem_context_setparam_ioctl(struct drm_device
*dev
, void *data
,
964 struct drm_file
*file
)
966 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
967 struct drm_i915_gem_context_param
*args
= data
;
968 struct intel_context
*ctx
;
971 ret
= i915_mutex_lock_interruptible(dev
);
975 ctx
= i915_gem_context_get(file_priv
, args
->ctx_id
);
977 mutex_unlock(&dev
->struct_mutex
);
981 switch (args
->param
) {
982 case I915_CONTEXT_PARAM_BAN_PERIOD
:
985 else if (args
->value
< ctx
->hang_stats
.ban_period_seconds
&&
986 !capable(CAP_SYS_ADMIN
))
989 ctx
->hang_stats
.ban_period_seconds
= args
->value
;
991 case I915_CONTEXT_PARAM_NO_ZEROMAP
:
995 ctx
->flags
&= ~CONTEXT_NO_ZEROMAP
;
996 ctx
->flags
|= args
->value
? CONTEXT_NO_ZEROMAP
: 0;
1003 mutex_unlock(&dev
->struct_mutex
);