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 it's 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 CONTEXT_ALIGN (64<<10)
98 static struct i915_hw_context
*
99 i915_gem_context_get(struct drm_i915_file_private
*file_priv
, u32 id
);
100 static int do_switch(struct i915_hw_context
*to
);
102 static int get_context_size(struct drm_device
*dev
)
104 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
108 switch (INTEL_INFO(dev
)->gen
) {
110 reg
= I915_READ(CXT_SIZE
);
111 ret
= GEN6_CXT_TOTAL_SIZE(reg
) * 64;
114 reg
= I915_READ(GEN7_CXT_SIZE
);
116 ret
= HSW_CXT_TOTAL_SIZE(reg
) * 64;
118 ret
= GEN7_CXT_TOTAL_SIZE(reg
) * 64;
127 static void do_destroy(struct i915_hw_context
*ctx
)
130 idr_remove(&ctx
->file_priv
->context_idr
, ctx
->id
);
132 drm_gem_object_unreference(&ctx
->obj
->base
);
136 static struct i915_hw_context
*
137 create_hw_context(struct drm_device
*dev
,
138 struct drm_i915_file_private
*file_priv
)
140 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
141 struct i915_hw_context
*ctx
;
144 ctx
= kzalloc(sizeof(*ctx
), GFP_KERNEL
);
146 return ERR_PTR(-ENOMEM
);
148 ctx
->obj
= i915_gem_alloc_object(dev
, dev_priv
->hw_context_size
);
149 if (ctx
->obj
== NULL
) {
151 DRM_DEBUG_DRIVER("Context object allocated failed\n");
152 return ERR_PTR(-ENOMEM
);
155 /* The ring associated with the context object is handled by the normal
156 * object tracking code. We give an initial ring value simple to pass an
157 * assertion in the context switch code.
159 ctx
->ring
= &dev_priv
->ring
[RCS
];
161 /* Default context will never have a file_priv */
162 if (file_priv
== NULL
)
165 ctx
->file_priv
= file_priv
;
167 ret
= idr_alloc(&file_priv
->context_idr
, ctx
, DEFAULT_CONTEXT_ID
+ 1, 0,
180 static inline bool is_default_context(struct i915_hw_context
*ctx
)
182 return (ctx
== ctx
->ring
->default_context
);
186 * The default context needs to exist per ring that uses contexts. It stores the
187 * context state of the GPU for applications that don't utilize HW contexts, as
188 * well as an idle case.
190 static int create_default_context(struct drm_i915_private
*dev_priv
)
192 struct i915_hw_context
*ctx
;
195 BUG_ON(!mutex_is_locked(&dev_priv
->dev
->struct_mutex
));
197 ctx
= create_hw_context(dev_priv
->dev
, NULL
);
201 /* We may need to do things with the shrinker which require us to
202 * immediately switch back to the default context. This can cause a
203 * problem as pinning the default context also requires GTT space which
204 * may not be available. To avoid this we always pin the
207 dev_priv
->ring
[RCS
].default_context
= ctx
;
208 ret
= i915_gem_object_pin(ctx
->obj
, CONTEXT_ALIGN
, false, false);
212 ret
= do_switch(ctx
);
216 DRM_DEBUG_DRIVER("Default HW context loaded\n");
220 i915_gem_object_unpin(ctx
->obj
);
226 void i915_gem_context_init(struct drm_device
*dev
)
228 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
230 if (!HAS_HW_CONTEXTS(dev
)) {
231 dev_priv
->hw_contexts_disabled
= true;
235 /* If called from reset, or thaw... we've been here already */
236 if (dev_priv
->hw_contexts_disabled
||
237 dev_priv
->ring
[RCS
].default_context
)
240 dev_priv
->hw_context_size
= round_up(get_context_size(dev
), 4096);
242 if (dev_priv
->hw_context_size
> (1<<20)) {
243 dev_priv
->hw_contexts_disabled
= true;
247 if (create_default_context(dev_priv
)) {
248 dev_priv
->hw_contexts_disabled
= true;
252 DRM_DEBUG_DRIVER("HW context support initialized\n");
255 void i915_gem_context_fini(struct drm_device
*dev
)
257 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
259 if (dev_priv
->hw_contexts_disabled
)
262 /* The only known way to stop the gpu from accessing the hw context is
263 * to reset it. Do this as the very last operation to avoid confusing
264 * other code, leading to spurious errors. */
265 intel_gpu_reset(dev
);
267 i915_gem_object_unpin(dev_priv
->ring
[RCS
].default_context
->obj
);
269 do_destroy(dev_priv
->ring
[RCS
].default_context
);
272 static int context_idr_cleanup(int id
, void *p
, void *data
)
274 struct i915_hw_context
*ctx
= p
;
276 BUG_ON(id
== DEFAULT_CONTEXT_ID
);
283 void i915_gem_context_close(struct drm_device
*dev
, struct drm_file
*file
)
285 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
287 mutex_lock(&dev
->struct_mutex
);
288 idr_for_each(&file_priv
->context_idr
, context_idr_cleanup
, NULL
);
289 idr_destroy(&file_priv
->context_idr
);
290 mutex_unlock(&dev
->struct_mutex
);
293 static struct i915_hw_context
*
294 i915_gem_context_get(struct drm_i915_file_private
*file_priv
, u32 id
)
296 return (struct i915_hw_context
*)idr_find(&file_priv
->context_idr
, id
);
300 mi_set_context(struct intel_ring_buffer
*ring
,
301 struct i915_hw_context
*new_context
,
306 /* w/a: If Flush TLB Invalidation Mode is enabled, driver must do a TLB
307 * invalidation prior to MI_SET_CONTEXT. On GEN6 we don't set the value
308 * explicitly, so we rely on the value at ring init, stored in
309 * itlb_before_ctx_switch.
311 if (IS_GEN6(ring
->dev
) && ring
->itlb_before_ctx_switch
) {
312 ret
= ring
->flush(ring
, I915_GEM_GPU_DOMAINS
, 0);
317 ret
= intel_ring_begin(ring
, 6);
321 if (IS_GEN7(ring
->dev
))
322 intel_ring_emit(ring
, MI_ARB_ON_OFF
| MI_ARB_DISABLE
);
324 intel_ring_emit(ring
, MI_NOOP
);
326 intel_ring_emit(ring
, MI_NOOP
);
327 intel_ring_emit(ring
, MI_SET_CONTEXT
);
328 intel_ring_emit(ring
, new_context
->obj
->gtt_offset
|
330 MI_SAVE_EXT_STATE_EN
|
331 MI_RESTORE_EXT_STATE_EN
|
333 /* w/a: MI_SET_CONTEXT must always be followed by MI_NOOP */
334 intel_ring_emit(ring
, MI_NOOP
);
336 if (IS_GEN7(ring
->dev
))
337 intel_ring_emit(ring
, MI_ARB_ON_OFF
| MI_ARB_ENABLE
);
339 intel_ring_emit(ring
, MI_NOOP
);
341 intel_ring_advance(ring
);
346 static int do_switch(struct i915_hw_context
*to
)
348 struct intel_ring_buffer
*ring
= to
->ring
;
349 struct drm_i915_gem_object
*from_obj
= ring
->last_context_obj
;
353 BUG_ON(from_obj
!= NULL
&& from_obj
->pin_count
== 0);
355 if (from_obj
== to
->obj
)
358 ret
= i915_gem_object_pin(to
->obj
, CONTEXT_ALIGN
, false, false);
362 /* Clear this page out of any CPU caches for coherent swap-in/out. Note
363 * that thanks to write = false in this call and us not setting any gpu
364 * write domains when putting a context object onto the active list
365 * (when switching away from it), this won't block.
366 * XXX: We need a real interface to do this instead of trickery. */
367 ret
= i915_gem_object_set_to_gtt_domain(to
->obj
, false);
369 i915_gem_object_unpin(to
->obj
);
373 if (!to
->obj
->has_global_gtt_mapping
)
374 i915_gem_gtt_bind_object(to
->obj
, to
->obj
->cache_level
);
376 if (!to
->is_initialized
|| is_default_context(to
))
377 hw_flags
|= MI_RESTORE_INHIBIT
;
378 else if (WARN_ON_ONCE(from_obj
== to
->obj
)) /* not yet expected */
379 hw_flags
|= MI_FORCE_RESTORE
;
381 ret
= mi_set_context(ring
, to
, hw_flags
);
383 i915_gem_object_unpin(to
->obj
);
387 /* The backing object for the context is done after switching to the
388 * *next* context. Therefore we cannot retire the previous context until
389 * the next context has already started running. In fact, the below code
390 * is a bit suboptimal because the retiring can occur simply after the
391 * MI_SET_CONTEXT instead of when the next seqno has completed.
393 if (from_obj
!= NULL
) {
394 from_obj
->base
.read_domains
= I915_GEM_DOMAIN_INSTRUCTION
;
395 i915_gem_object_move_to_active(from_obj
, ring
);
396 /* As long as MI_SET_CONTEXT is serializing, ie. it flushes the
397 * whole damn pipeline, we don't need to explicitly mark the
398 * object dirty. The only exception is that the context must be
399 * correct in case the object gets swapped out. Ideally we'd be
400 * able to defer doing this until we know the object would be
401 * swapped, but there is no way to do that yet.
404 BUG_ON(from_obj
->ring
!= ring
);
405 i915_gem_object_unpin(from_obj
);
407 drm_gem_object_unreference(&from_obj
->base
);
410 drm_gem_object_reference(&to
->obj
->base
);
411 ring
->last_context_obj
= to
->obj
;
412 to
->is_initialized
= true;
418 * i915_switch_context() - perform a GPU context switch.
419 * @ring: ring for which we'll execute the context switch
420 * @file_priv: file_priv associated with the context, may be NULL
421 * @id: context id number
422 * @seqno: sequence number by which the new context will be switched to
425 * The context life cycle is simple. The context refcount is incremented and
426 * decremented by 1 and create and destroy. If the context is in use by the GPU,
427 * it will have a refoucnt > 1. This allows us to destroy the context abstract
428 * object while letting the normal object tracking destroy the backing BO.
430 int i915_switch_context(struct intel_ring_buffer
*ring
,
431 struct drm_file
*file
,
434 struct drm_i915_private
*dev_priv
= ring
->dev
->dev_private
;
435 struct i915_hw_context
*to
;
437 if (dev_priv
->hw_contexts_disabled
)
440 if (ring
!= &dev_priv
->ring
[RCS
])
443 if (to_id
== DEFAULT_CONTEXT_ID
) {
444 to
= ring
->default_context
;
449 to
= i915_gem_context_get(file
->driver_priv
, to_id
);
454 return do_switch(to
);
457 int i915_gem_context_create_ioctl(struct drm_device
*dev
, void *data
,
458 struct drm_file
*file
)
460 struct drm_i915_private
*dev_priv
= dev
->dev_private
;
461 struct drm_i915_gem_context_create
*args
= data
;
462 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
463 struct i915_hw_context
*ctx
;
466 if (!(dev
->driver
->driver_features
& DRIVER_GEM
))
469 if (dev_priv
->hw_contexts_disabled
)
472 ret
= i915_mutex_lock_interruptible(dev
);
476 ctx
= create_hw_context(dev
, file_priv
);
477 mutex_unlock(&dev
->struct_mutex
);
481 args
->ctx_id
= ctx
->id
;
482 DRM_DEBUG_DRIVER("HW context %d created\n", args
->ctx_id
);
487 int i915_gem_context_destroy_ioctl(struct drm_device
*dev
, void *data
,
488 struct drm_file
*file
)
490 struct drm_i915_gem_context_destroy
*args
= data
;
491 struct drm_i915_file_private
*file_priv
= file
->driver_priv
;
492 struct i915_hw_context
*ctx
;
495 if (!(dev
->driver
->driver_features
& DRIVER_GEM
))
498 ret
= i915_mutex_lock_interruptible(dev
);
502 ctx
= i915_gem_context_get(file_priv
, args
->ctx_id
);
504 mutex_unlock(&dev
->struct_mutex
);
510 mutex_unlock(&dev
->struct_mutex
);
512 DRM_DEBUG_DRIVER("HW context %d destroyed\n", args
->ctx_id
);