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 #define ALL_L3_SLICES(dev) (1 << NUM_L3_SLICES(dev)) - 1
95 /* This is a HW constraint. The value below is the largest known requirement
96 * I've seen in a spec to date, and that was a workaround for a non-shipping
97 * part. It should be safe to decrease this, but it's more future proof as is.
99 #define GEN6_CONTEXT_ALIGN (64<<10)
100 #define GEN7_CONTEXT_ALIGN 4096
102 static size_t get_context_alignment(struct drm_device
*dev
)
105 return GEN6_CONTEXT_ALIGN
;
107 return GEN7_CONTEXT_ALIGN
;
110 static int get_context_size(struct drm_device
*dev
)
112 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
116 switch (INTEL_INFO(dev
)->gen
) {
118 reg
= I915_READ(CXT_SIZE
);
119 ret
= GEN6_CXT_TOTAL_SIZE(reg
) * 64;
122 reg
= I915_READ(GEN7_CXT_SIZE
);
124 ret
= HSW_CXT_TOTAL_SIZE
;
126 ret
= GEN7_CXT_TOTAL_SIZE(reg
) * 64;
129 ret
= GEN8_CXT_TOTAL_SIZE
;
138 static void i915_gem_context_clean(struct intel_context
*ctx
)
140 struct i915_hw_ppgtt
*ppgtt
= ctx
->ppgtt
;
141 struct i915_vma
*vma
, *next
;
146 list_for_each_entry_safe(vma
, next
, &ppgtt
->base
.inactive_list
,
148 if (WARN_ON(__i915_vma_unbind_no_wait(vma
)))
153 void i915_gem_context_free(struct kref
*ctx_ref
)
155 struct intel_context
*ctx
= container_of(ctx_ref
, typeof(*ctx
), ref
);
157 trace_i915_context_free(ctx
);
159 if (i915
.enable_execlists
)
160 intel_lr_context_free(ctx
);
163 * This context is going away and we need to remove all VMAs still
164 * around. This is to handle imported shared objects for which
165 * destructor did not run when their handles were closed.
167 i915_gem_context_clean(ctx
);
169 i915_ppgtt_put(ctx
->ppgtt
);
171 if (ctx
->legacy_hw_ctx
.rcs_state
)
172 drm_gem_object_unreference(&ctx
->legacy_hw_ctx
.rcs_state
->base
);
173 list_del(&ctx
->link
);
177 struct drm_i915_gem_object
*
178 i915_gem_alloc_context_obj(struct drm_device
*dev
, size_t size
)
180 struct drm_i915_gem_object
*obj
;
183 obj
= i915_gem_object_create(dev
, size
);
188 * Try to make the context utilize L3 as well as LLC.
190 * On VLV we don't have L3 controls in the PTEs so we
191 * shouldn't touch the cache level, especially as that
192 * would make the object snooped which might have a
193 * negative performance impact.
195 * Snooping is required on non-llc platforms in execlist
196 * mode, but since all GGTT accesses use PAT entry 0 we
197 * get snooping anyway regardless of cache_level.
199 * This is only applicable for Ivy Bridge devices since
200 * later platforms don't have L3 control bits in the PTE.
202 if (IS_IVYBRIDGE(dev
)) {
203 ret
= i915_gem_object_set_cache_level(obj
, I915_CACHE_L3_LLC
);
204 /* Failure shouldn't ever happen this early */
206 drm_gem_object_unreference(&obj
->base
);
214 static struct intel_context
*
215 __create_hw_context(struct drm_device
*dev
,
216 struct drm_i915_file_private
*file_priv
)
218 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
219 struct intel_context
*ctx
;
222 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
224 return ERR_PTR(-ENOMEM
);
226 kref_init(&ctx
->ref
);
227 list_add_tail(&ctx
->link
, &dev_priv
->context_list
);
228 ctx
->i915
= dev_priv
;
230 if (dev_priv
->hw_context_size
) {
231 struct drm_i915_gem_object
*obj
=
232 i915_gem_alloc_context_obj(dev
, dev_priv
->hw_context_size
);
237 ctx
->legacy_hw_ctx
.rcs_state
= obj
;
240 /* Default context will never have a file_priv */
241 if (file_priv
!= NULL
) {
242 ret
= idr_alloc(&file_priv
->context_idr
, ctx
,
243 DEFAULT_CONTEXT_HANDLE
, 0, GFP_KERNEL
);
247 ret
= DEFAULT_CONTEXT_HANDLE
;
249 ctx
->file_priv
= file_priv
;
250 ctx
->user_handle
= ret
;
251 /* NB: Mark all slices as needing a remap so that when the context first
252 * loads it will restore whatever remap state already exists. If there
253 * is no remap info, it will be a NOP. */
254 ctx
->remap_slice
= ALL_L3_SLICES(dev_priv
);
256 ctx
->hang_stats
.ban_period_seconds
= DRM_I915_CTX_BAN_PERIOD
;
261 i915_gem_context_unreference(ctx
);
266 * The default context needs to exist per ring that uses contexts. It stores the
267 * context state of the GPU for applications that don't utilize HW contexts, as
268 * well as an idle case.
270 static struct intel_context
*
271 i915_gem_create_context(struct drm_device
*dev
,
272 struct drm_i915_file_private
*file_priv
)
274 const bool is_global_default_ctx
= file_priv
== NULL
;
275 struct intel_context
*ctx
;
278 BUG_ON(!mutex_is_locked(&dev
->struct_mutex
));
280 ctx
= __create_hw_context(dev
, file_priv
);
284 if (is_global_default_ctx
&& ctx
->legacy_hw_ctx
.rcs_state
) {
285 /* We may need to do things with the shrinker which
286 * require us to immediately switch back to the default
287 * context. This can cause a problem as pinning the
288 * default context also requires GTT space which may not
289 * be available. To avoid this we always pin the default
292 ret
= i915_gem_obj_ggtt_pin(ctx
->legacy_hw_ctx
.rcs_state
,
293 get_context_alignment(dev
), 0);
295 DRM_DEBUG_DRIVER("Couldn't pin %d\n", ret
);
300 if (USES_FULL_PPGTT(dev
)) {
301 struct i915_hw_ppgtt
*ppgtt
= i915_ppgtt_create(dev
, file_priv
);
303 if (IS_ERR_OR_NULL(ppgtt
)) {
304 DRM_DEBUG_DRIVER("PPGTT setup failed (%ld)\n",
306 ret
= PTR_ERR(ppgtt
);
313 trace_i915_context_create(ctx
);
318 if (is_global_default_ctx
&& ctx
->legacy_hw_ctx
.rcs_state
)
319 i915_gem_object_ggtt_unpin(ctx
->legacy_hw_ctx
.rcs_state
);
321 idr_remove(&file_priv
->context_idr
, ctx
->user_handle
);
322 i915_gem_context_unreference(ctx
);
326 static void i915_gem_context_unpin(struct intel_context
*ctx
,
327 struct intel_engine_cs
*engine
)
329 if (i915
.enable_execlists
) {
330 intel_lr_context_unpin(ctx
, engine
);
332 if (engine
->id
== RCS
&& ctx
->legacy_hw_ctx
.rcs_state
)
333 i915_gem_object_ggtt_unpin(ctx
->legacy_hw_ctx
.rcs_state
);
334 i915_gem_context_unreference(ctx
);
338 void i915_gem_context_reset(struct drm_device
*dev
)
340 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
342 if (i915
.enable_execlists
) {
343 struct intel_context
*ctx
;
345 list_for_each_entry(ctx
, &dev_priv
->context_list
, link
)
346 intel_lr_context_reset(dev_priv
, ctx
);
349 i915_gem_context_lost(dev_priv
);
352 int i915_gem_context_init(struct drm_device
*dev
)
354 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
355 struct intel_context
*ctx
;
357 /* Init should only be called once per module load. Eventually the
358 * restriction on the context_disabled check can be loosened. */
359 if (WARN_ON(dev_priv
->kernel_context
))
362 if (intel_vgpu_active(dev
) && HAS_LOGICAL_RING_CONTEXTS(dev
)) {
363 if (!i915
.enable_execlists
) {
364 DRM_INFO("Only EXECLIST mode is supported in vgpu.\n");
369 if (i915
.enable_execlists
) {
370 /* NB: intentionally left blank. We will allocate our own
371 * backing objects as we need them, thank you very much */
372 dev_priv
->hw_context_size
= 0;
373 } else if (HAS_HW_CONTEXTS(dev
)) {
374 dev_priv
->hw_context_size
= round_up(get_context_size(dev
), 4096);
375 if (dev_priv
->hw_context_size
> (1<<20)) {
376 DRM_DEBUG_DRIVER("Disabling HW Contexts; invalid size %d\n",
377 dev_priv
->hw_context_size
);
378 dev_priv
->hw_context_size
= 0;
382 ctx
= i915_gem_create_context(dev
, NULL
);
384 DRM_ERROR("Failed to create default global context (error %ld)\n",
389 dev_priv
->kernel_context
= ctx
;
391 DRM_DEBUG_DRIVER("%s context support initialized\n",
392 i915
.enable_execlists
? "LR" :
393 dev_priv
->hw_context_size
? "HW" : "fake");
397 void i915_gem_context_lost(struct drm_i915_private
*dev_priv
)
399 struct intel_engine_cs
*engine
;
401 for_each_engine(engine
, dev_priv
) {
402 if (engine
->last_context
== NULL
)
405 i915_gem_context_unpin(engine
->last_context
, engine
);
406 engine
->last_context
= NULL
;
409 /* Force the GPU state to be reinitialised on enabling */
410 dev_priv
->kernel_context
->legacy_hw_ctx
.initialized
= false;
411 dev_priv
->kernel_context
->remap_slice
= ALL_L3_SLICES(dev_priv
);
414 void i915_gem_context_fini(struct drm_device
*dev
)
416 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
417 struct intel_context
*dctx
= dev_priv
->kernel_context
;
419 i915_gem_context_lost(dev_priv
);
421 if (dctx
->legacy_hw_ctx
.rcs_state
) {
422 /* The only known way to stop the gpu from accessing the hw context is
423 * to reset it. Do this as the very last operation to avoid confusing
424 * other code, leading to spurious errors. */
425 intel_gpu_reset(dev
, ALL_ENGINES
);
427 i915_gem_object_ggtt_unpin(dctx
->legacy_hw_ctx
.rcs_state
);
430 i915_gem_context_unreference(dctx
);
431 dev_priv
->kernel_context
= NULL
;
434 static int context_idr_cleanup(int id
, void *p
, void *data
)
436 struct intel_context
*ctx
= p
;
438 i915_gem_context_unreference(ctx
);
442 int i915_gem_context_open(struct drm_device
*dev
, struct drm_file
*file
)
444 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
445 struct intel_context
*ctx
;
447 idr_init(&file_priv
->context_idr
);
449 mutex_lock(&dev
->struct_mutex
);
450 ctx
= i915_gem_create_context(dev
, file_priv
);
451 mutex_unlock(&dev
->struct_mutex
);
454 idr_destroy(&file_priv
->context_idr
);
461 void i915_gem_context_close(struct drm_device
*dev
, struct drm_file
*file
)
463 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
465 idr_for_each(&file_priv
->context_idr
, context_idr_cleanup
, NULL
);
466 idr_destroy(&file_priv
->context_idr
);
469 struct intel_context
*
470 i915_gem_context_get(struct drm_i915_file_private
*file_priv
, u32 id
)
472 struct intel_context
*ctx
;
474 ctx
= (struct intel_context
*)idr_find(&file_priv
->context_idr
, id
);
476 return ERR_PTR(-ENOENT
);
482 mi_set_context(struct drm_i915_gem_request
*req
, u32 hw_flags
)
484 struct intel_engine_cs
*engine
= req
->engine
;
485 u32 flags
= hw_flags
| MI_MM_SPACE_GTT
;
486 const int num_rings
=
487 /* Use an extended w/a on ivb+ if signalling from other rings */
488 i915_semaphore_is_enabled(engine
->dev
) ?
489 hweight32(INTEL_INFO(engine
->dev
)->ring_mask
) - 1 :
493 /* w/a: If Flush TLB Invalidation Mode is enabled, driver must do a TLB
494 * invalidation prior to MI_SET_CONTEXT. On GEN6 we don't set the value
495 * explicitly, so we rely on the value at ring init, stored in
496 * itlb_before_ctx_switch.
498 if (IS_GEN6(engine
->dev
)) {
499 ret
= engine
->flush(req
, I915_GEM_GPU_DOMAINS
, 0);
504 /* These flags are for resource streamer on HSW+ */
505 if (IS_HASWELL(engine
->dev
) || INTEL_INFO(engine
->dev
)->gen
>= 8)
506 flags
|= (HSW_MI_RS_SAVE_STATE_EN
| HSW_MI_RS_RESTORE_STATE_EN
);
507 else if (INTEL_INFO(engine
->dev
)->gen
< 8)
508 flags
|= (MI_SAVE_EXT_STATE_EN
| MI_RESTORE_EXT_STATE_EN
);
512 if (INTEL_INFO(engine
->dev
)->gen
>= 7)
513 len
+= 2 + (num_rings
? 4*num_rings
+ 6 : 0);
515 ret
= intel_ring_begin(req
, len
);
519 /* WaProgramMiArbOnOffAroundMiSetContext:ivb,vlv,hsw,bdw,chv */
520 if (INTEL_INFO(engine
->dev
)->gen
>= 7) {
521 intel_ring_emit(engine
, MI_ARB_ON_OFF
| MI_ARB_DISABLE
);
523 struct intel_engine_cs
*signaller
;
525 intel_ring_emit(engine
,
526 MI_LOAD_REGISTER_IMM(num_rings
));
527 for_each_engine(signaller
, to_i915(engine
->dev
)) {
528 if (signaller
== engine
)
531 intel_ring_emit_reg(engine
,
532 RING_PSMI_CTL(signaller
->mmio_base
));
533 intel_ring_emit(engine
,
534 _MASKED_BIT_ENABLE(GEN6_PSMI_SLEEP_MSG_DISABLE
));
539 intel_ring_emit(engine
, MI_NOOP
);
540 intel_ring_emit(engine
, MI_SET_CONTEXT
);
541 intel_ring_emit(engine
,
542 i915_gem_obj_ggtt_offset(req
->ctx
->legacy_hw_ctx
.rcs_state
) |
545 * w/a: MI_SET_CONTEXT must always be followed by MI_NOOP
546 * WaMiSetContext_Hang:snb,ivb,vlv
548 intel_ring_emit(engine
, MI_NOOP
);
550 if (INTEL_INFO(engine
->dev
)->gen
>= 7) {
552 struct intel_engine_cs
*signaller
;
553 i915_reg_t last_reg
= {}; /* keep gcc quiet */
555 intel_ring_emit(engine
,
556 MI_LOAD_REGISTER_IMM(num_rings
));
557 for_each_engine(signaller
, to_i915(engine
->dev
)) {
558 if (signaller
== engine
)
561 last_reg
= RING_PSMI_CTL(signaller
->mmio_base
);
562 intel_ring_emit_reg(engine
, last_reg
);
563 intel_ring_emit(engine
,
564 _MASKED_BIT_DISABLE(GEN6_PSMI_SLEEP_MSG_DISABLE
));
567 /* Insert a delay before the next switch! */
568 intel_ring_emit(engine
,
569 MI_STORE_REGISTER_MEM
|
570 MI_SRM_LRM_GLOBAL_GTT
);
571 intel_ring_emit_reg(engine
, last_reg
);
572 intel_ring_emit(engine
, engine
->scratch
.gtt_offset
);
573 intel_ring_emit(engine
, MI_NOOP
);
575 intel_ring_emit(engine
, MI_ARB_ON_OFF
| MI_ARB_ENABLE
);
578 intel_ring_advance(engine
);
583 static int remap_l3(struct drm_i915_gem_request
*req
, int slice
)
585 u32
*remap_info
= req
->i915
->l3_parity
.remap_info
[slice
];
586 struct intel_engine_cs
*engine
= req
->engine
;
592 ret
= intel_ring_begin(req
, GEN7_L3LOG_SIZE
/4 * 2 + 2);
597 * Note: We do not worry about the concurrent register cacheline hang
598 * here because no other code should access these registers other than
599 * at initialization time.
601 intel_ring_emit(engine
, MI_LOAD_REGISTER_IMM(GEN7_L3LOG_SIZE
/4));
602 for (i
= 0; i
< GEN7_L3LOG_SIZE
/4; i
++) {
603 intel_ring_emit_reg(engine
, GEN7_L3LOG(slice
, i
));
604 intel_ring_emit(engine
, remap_info
[i
]);
606 intel_ring_emit(engine
, MI_NOOP
);
607 intel_ring_advance(engine
);
612 static inline bool skip_rcs_switch(struct i915_hw_ppgtt
*ppgtt
,
613 struct intel_engine_cs
*engine
,
614 struct intel_context
*to
)
619 if (!to
->legacy_hw_ctx
.initialized
)
622 if (ppgtt
&& (intel_engine_flag(engine
) & ppgtt
->pd_dirty_rings
))
625 return to
== engine
->last_context
;
629 needs_pd_load_pre(struct i915_hw_ppgtt
*ppgtt
,
630 struct intel_engine_cs
*engine
,
631 struct intel_context
*to
)
636 /* Always load the ppgtt on first use */
637 if (!engine
->last_context
)
640 /* Same context without new entries, skip */
641 if (engine
->last_context
== to
&&
642 !(intel_engine_flag(engine
) & ppgtt
->pd_dirty_rings
))
645 if (engine
->id
!= RCS
)
648 if (INTEL_INFO(engine
->dev
)->gen
< 8)
655 needs_pd_load_post(struct i915_hw_ppgtt
*ppgtt
,
656 struct intel_context
*to
,
662 if (!IS_GEN8(to
->i915
))
665 if (hw_flags
& MI_RESTORE_INHIBIT
)
671 static int do_rcs_switch(struct drm_i915_gem_request
*req
)
673 struct intel_context
*to
= req
->ctx
;
674 struct intel_engine_cs
*engine
= req
->engine
;
675 struct i915_hw_ppgtt
*ppgtt
= to
->ppgtt
?: req
->i915
->mm
.aliasing_ppgtt
;
676 struct intel_context
*from
;
680 if (skip_rcs_switch(ppgtt
, engine
, to
))
683 /* Trying to pin first makes error handling easier. */
684 ret
= i915_gem_obj_ggtt_pin(to
->legacy_hw_ctx
.rcs_state
,
685 get_context_alignment(engine
->dev
),
691 * Pin can switch back to the default context if we end up calling into
692 * evict_everything - as a last ditch gtt defrag effort that also
693 * switches to the default context. Hence we need to reload from here.
695 * XXX: Doing so is painfully broken!
697 from
= engine
->last_context
;
700 * Clear this page out of any CPU caches for coherent swap-in/out. Note
701 * that thanks to write = false in this call and us not setting any gpu
702 * write domains when putting a context object onto the active list
703 * (when switching away from it), this won't block.
705 * XXX: We need a real interface to do this instead of trickery.
707 ret
= i915_gem_object_set_to_gtt_domain(to
->legacy_hw_ctx
.rcs_state
, false);
711 if (needs_pd_load_pre(ppgtt
, engine
, to
)) {
712 /* Older GENs and non render rings still want the load first,
713 * "PP_DCLV followed by PP_DIR_BASE register through Load
714 * Register Immediate commands in Ring Buffer before submitting
716 trace_switch_mm(engine
, to
);
717 ret
= ppgtt
->switch_mm(ppgtt
, req
);
722 if (!to
->legacy_hw_ctx
.initialized
|| i915_gem_context_is_default(to
))
723 /* NB: If we inhibit the restore, the context is not allowed to
724 * die because future work may end up depending on valid address
725 * space. This means we must enforce that a page table load
726 * occur when this occurs. */
727 hw_flags
= MI_RESTORE_INHIBIT
;
728 else if (ppgtt
&& intel_engine_flag(engine
) & ppgtt
->pd_dirty_rings
)
729 hw_flags
= MI_FORCE_RESTORE
;
733 if (to
!= from
|| (hw_flags
& MI_FORCE_RESTORE
)) {
734 ret
= mi_set_context(req
, hw_flags
);
739 /* The backing object for the context is done after switching to the
740 * *next* context. Therefore we cannot retire the previous context until
741 * the next context has already started running. In fact, the below code
742 * is a bit suboptimal because the retiring can occur simply after the
743 * MI_SET_CONTEXT instead of when the next seqno has completed.
746 from
->legacy_hw_ctx
.rcs_state
->base
.read_domains
= I915_GEM_DOMAIN_INSTRUCTION
;
747 i915_vma_move_to_active(i915_gem_obj_to_ggtt(from
->legacy_hw_ctx
.rcs_state
), req
);
748 /* As long as MI_SET_CONTEXT is serializing, ie. it flushes the
749 * whole damn pipeline, we don't need to explicitly mark the
750 * object dirty. The only exception is that the context must be
751 * correct in case the object gets swapped out. Ideally we'd be
752 * able to defer doing this until we know the object would be
753 * swapped, but there is no way to do that yet.
755 from
->legacy_hw_ctx
.rcs_state
->dirty
= 1;
757 /* obj is kept alive until the next request by its active ref */
758 i915_gem_object_ggtt_unpin(from
->legacy_hw_ctx
.rcs_state
);
759 i915_gem_context_unreference(from
);
761 i915_gem_context_reference(to
);
762 engine
->last_context
= to
;
764 /* GEN8 does *not* require an explicit reload if the PDPs have been
765 * setup, and we do not wish to move them.
767 if (needs_pd_load_post(ppgtt
, to
, hw_flags
)) {
768 trace_switch_mm(engine
, to
);
769 ret
= ppgtt
->switch_mm(ppgtt
, req
);
770 /* The hardware context switch is emitted, but we haven't
771 * actually changed the state - so it's probably safe to bail
772 * here. Still, let the user know something dangerous has
780 ppgtt
->pd_dirty_rings
&= ~intel_engine_flag(engine
);
782 for (i
= 0; i
< MAX_L3_SLICES
; i
++) {
783 if (!(to
->remap_slice
& (1<<i
)))
786 ret
= remap_l3(req
, i
);
790 to
->remap_slice
&= ~(1<<i
);
793 if (!to
->legacy_hw_ctx
.initialized
) {
794 if (engine
->init_context
) {
795 ret
= engine
->init_context(req
);
799 to
->legacy_hw_ctx
.initialized
= true;
805 i915_gem_object_ggtt_unpin(to
->legacy_hw_ctx
.rcs_state
);
810 * i915_switch_context() - perform a GPU context switch.
811 * @req: request for which we'll execute the context switch
813 * The context life cycle is simple. The context refcount is incremented and
814 * decremented by 1 and create and destroy. If the context is in use by the GPU,
815 * it will have a refcount > 1. This allows us to destroy the context abstract
816 * object while letting the normal object tracking destroy the backing BO.
818 * This function should not be used in execlists mode. Instead the context is
819 * switched by writing to the ELSP and requests keep a reference to their
822 int i915_switch_context(struct drm_i915_gem_request
*req
)
824 struct intel_engine_cs
*engine
= req
->engine
;
825 struct drm_i915_private
*dev_priv
= req
->i915
;
827 WARN_ON(i915
.enable_execlists
);
828 WARN_ON(!mutex_is_locked(&dev_priv
->dev
->struct_mutex
));
830 if (engine
->id
!= RCS
||
831 req
->ctx
->legacy_hw_ctx
.rcs_state
== NULL
) {
832 struct intel_context
*to
= req
->ctx
;
833 struct i915_hw_ppgtt
*ppgtt
=
834 to
->ppgtt
?: req
->i915
->mm
.aliasing_ppgtt
;
836 if (needs_pd_load_pre(ppgtt
, engine
, to
)) {
839 trace_switch_mm(engine
, to
);
840 ret
= ppgtt
->switch_mm(ppgtt
, req
);
844 ppgtt
->pd_dirty_rings
&= ~intel_engine_flag(engine
);
847 if (to
!= engine
->last_context
) {
848 i915_gem_context_reference(to
);
849 if (engine
->last_context
)
850 i915_gem_context_unreference(engine
->last_context
);
851 engine
->last_context
= to
;
857 return do_rcs_switch(req
);
860 static bool contexts_enabled(struct drm_device
*dev
)
862 return i915
.enable_execlists
|| to_i915(dev
)->hw_context_size
;
865 int i915_gem_context_create_ioctl(struct drm_device
*dev
, void *data
,
866 struct drm_file
*file
)
868 struct drm_i915_gem_context_create
*args
= data
;
869 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
870 struct intel_context
*ctx
;
873 if (!contexts_enabled(dev
))
879 ret
= i915_mutex_lock_interruptible(dev
);
883 ctx
= i915_gem_create_context(dev
, file_priv
);
884 mutex_unlock(&dev
->struct_mutex
);
888 args
->ctx_id
= ctx
->user_handle
;
889 DRM_DEBUG_DRIVER("HW context %d created\n", args
->ctx_id
);
894 int i915_gem_context_destroy_ioctl(struct drm_device
*dev
, void *data
,
895 struct drm_file
*file
)
897 struct drm_i915_gem_context_destroy
*args
= data
;
898 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
899 struct intel_context
*ctx
;
905 if (args
->ctx_id
== DEFAULT_CONTEXT_HANDLE
)
908 ret
= i915_mutex_lock_interruptible(dev
);
912 ctx
= i915_gem_context_get(file_priv
, args
->ctx_id
);
914 mutex_unlock(&dev
->struct_mutex
);
918 idr_remove(&ctx
->file_priv
->context_idr
, ctx
->user_handle
);
919 i915_gem_context_unreference(ctx
);
920 mutex_unlock(&dev
->struct_mutex
);
922 DRM_DEBUG_DRIVER("HW context %d destroyed\n", args
->ctx_id
);
926 int i915_gem_context_getparam_ioctl(struct drm_device
*dev
, void *data
,
927 struct drm_file
*file
)
929 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
930 struct drm_i915_gem_context_param
*args
= data
;
931 struct intel_context
*ctx
;
934 ret
= i915_mutex_lock_interruptible(dev
);
938 ctx
= i915_gem_context_get(file_priv
, args
->ctx_id
);
940 mutex_unlock(&dev
->struct_mutex
);
945 switch (args
->param
) {
946 case I915_CONTEXT_PARAM_BAN_PERIOD
:
947 args
->value
= ctx
->hang_stats
.ban_period_seconds
;
949 case I915_CONTEXT_PARAM_NO_ZEROMAP
:
950 args
->value
= ctx
->flags
& CONTEXT_NO_ZEROMAP
;
952 case I915_CONTEXT_PARAM_GTT_SIZE
:
954 args
->value
= ctx
->ppgtt
->base
.total
;
955 else if (to_i915(dev
)->mm
.aliasing_ppgtt
)
956 args
->value
= to_i915(dev
)->mm
.aliasing_ppgtt
->base
.total
;
958 args
->value
= to_i915(dev
)->ggtt
.base
.total
;
964 mutex_unlock(&dev
->struct_mutex
);
969 int i915_gem_context_setparam_ioctl(struct drm_device
*dev
, void *data
,
970 struct drm_file
*file
)
972 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
973 struct drm_i915_gem_context_param
*args
= data
;
974 struct intel_context
*ctx
;
977 ret
= i915_mutex_lock_interruptible(dev
);
981 ctx
= i915_gem_context_get(file_priv
, args
->ctx_id
);
983 mutex_unlock(&dev
->struct_mutex
);
987 switch (args
->param
) {
988 case I915_CONTEXT_PARAM_BAN_PERIOD
:
991 else if (args
->value
< ctx
->hang_stats
.ban_period_seconds
&&
992 !capable(CAP_SYS_ADMIN
))
995 ctx
->hang_stats
.ban_period_seconds
= args
->value
;
997 case I915_CONTEXT_PARAM_NO_ZEROMAP
:
1001 ctx
->flags
&= ~CONTEXT_NO_ZEROMAP
;
1002 ctx
->flags
|= args
->value
? CONTEXT_NO_ZEROMAP
: 0;
1009 mutex_unlock(&dev
->struct_mutex
);