]> git.proxmox.com Git - mirror_ubuntu-eoan-kernel.git/blame - drivers/gpu/drm/i915/i915_gem_context.c
Revert "drm/i915: Parsing LFP brightness control from VBT"
[mirror_ubuntu-eoan-kernel.git] / drivers / gpu / drm / i915 / i915_gem_context.c
CommitLineData
254f965c
BW
1/*
2 * Copyright © 2011-2012 Intel Corporation
3 *
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:
10 *
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
13 * Software.
14 *
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
21 * IN THE SOFTWARE.
22 *
23 * Authors:
24 * Ben Widawsky <ben@bwidawsk.net>
25 *
26 */
27
28/*
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.
35 *
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.
46 *
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.
51 *
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
63 *
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
73 *
74 * There are two confusing terms used above:
75 * The "current context" means the context which is currently running on the
508842a0 76 * GPU. The GPU has loaded its state already and has stored away the gtt
254f965c
BW
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.
80 *
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.
85 *
86 */
87
760285e7
DH
88#include <drm/drmP.h>
89#include <drm/i915_drm.h>
254f965c 90#include "i915_drv.h"
198c974d 91#include "i915_trace.h"
254f965c 92
40521054
BW
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.
96 */
b731d33d
BW
97#define GEN6_CONTEXT_ALIGN (64<<10)
98#define GEN7_CONTEXT_ALIGN 4096
40521054 99
b731d33d
BW
100static size_t get_context_alignment(struct drm_device *dev)
101{
102 if (IS_GEN6(dev))
103 return GEN6_CONTEXT_ALIGN;
104
105 return GEN7_CONTEXT_ALIGN;
106}
107
254f965c
BW
108static int get_context_size(struct drm_device *dev)
109{
110 struct drm_i915_private *dev_priv = dev->dev_private;
111 int ret;
112 u32 reg;
113
114 switch (INTEL_INFO(dev)->gen) {
115 case 6:
116 reg = I915_READ(CXT_SIZE);
117 ret = GEN6_CXT_TOTAL_SIZE(reg) * 64;
118 break;
119 case 7:
4f91dd6f 120 reg = I915_READ(GEN7_CXT_SIZE);
2e4291e0 121 if (IS_HASWELL(dev))
a0de80a0 122 ret = HSW_CXT_TOTAL_SIZE;
2e4291e0
BW
123 else
124 ret = GEN7_CXT_TOTAL_SIZE(reg) * 64;
254f965c 125 break;
8897644a
BW
126 case 8:
127 ret = GEN8_CXT_TOTAL_SIZE;
128 break;
254f965c
BW
129 default:
130 BUG();
131 }
132
133 return ret;
134}
135
dce3271b 136void i915_gem_context_free(struct kref *ctx_ref)
40521054 137{
273497e5 138 struct intel_context *ctx = container_of(ctx_ref,
ae6c4806 139 typeof(*ctx), ref);
40521054 140
198c974d
DCS
141 trace_i915_context_free(ctx);
142
ae6c4806 143 if (i915.enable_execlists)
ede7d42b 144 intel_lr_context_free(ctx);
c7c48dfd 145
ae6c4806
DV
146 i915_ppgtt_put(ctx->ppgtt);
147
2f295791
BW
148 if (ctx->legacy_hw_ctx.rcs_state)
149 drm_gem_object_unreference(&ctx->legacy_hw_ctx.rcs_state->base);
c7c48dfd 150 list_del(&ctx->link);
40521054
BW
151 kfree(ctx);
152}
153
8c857917 154struct drm_i915_gem_object *
aa0c13da
OM
155i915_gem_alloc_context_obj(struct drm_device *dev, size_t size)
156{
157 struct drm_i915_gem_object *obj;
158 int ret;
159
160 obj = i915_gem_alloc_object(dev, size);
161 if (obj == NULL)
162 return ERR_PTR(-ENOMEM);
163
164 /*
165 * Try to make the context utilize L3 as well as LLC.
166 *
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.
171 */
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 */
175 if (WARN_ON(ret)) {
176 drm_gem_object_unreference(&obj->base);
177 return ERR_PTR(ret);
178 }
179 }
180
181 return obj;
182}
183
273497e5 184static struct intel_context *
0eea67eb 185__create_hw_context(struct drm_device *dev,
ee960be7 186 struct drm_i915_file_private *file_priv)
40521054
BW
187{
188 struct drm_i915_private *dev_priv = dev->dev_private;
273497e5 189 struct intel_context *ctx;
c8c470af 190 int ret;
40521054 191
f94982b0 192 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
146937e5
BW
193 if (ctx == NULL)
194 return ERR_PTR(-ENOMEM);
40521054 195
dce3271b 196 kref_init(&ctx->ref);
691e6415 197 list_add_tail(&ctx->link, &dev_priv->context_list);
40521054 198
691e6415 199 if (dev_priv->hw_context_size) {
aa0c13da
OM
200 struct drm_i915_gem_object *obj =
201 i915_gem_alloc_context_obj(dev, dev_priv->hw_context_size);
202 if (IS_ERR(obj)) {
203 ret = PTR_ERR(obj);
4615d4c9 204 goto err_out;
691e6415 205 }
ea0c76f8 206 ctx->legacy_hw_ctx.rcs_state = obj;
691e6415 207 }
40521054
BW
208
209 /* Default context will never have a file_priv */
691e6415
CW
210 if (file_priv != NULL) {
211 ret = idr_alloc(&file_priv->context_idr, ctx,
821d66dd 212 DEFAULT_CONTEXT_HANDLE, 0, GFP_KERNEL);
691e6415
CW
213 if (ret < 0)
214 goto err_out;
215 } else
821d66dd 216 ret = DEFAULT_CONTEXT_HANDLE;
dce3271b
MK
217
218 ctx->file_priv = file_priv;
821d66dd 219 ctx->user_handle = ret;
3ccfd19d
BW
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;
40521054 224
146937e5 225 return ctx;
40521054
BW
226
227err_out:
dce3271b 228 i915_gem_context_unreference(ctx);
146937e5 229 return ERR_PTR(ret);
40521054
BW
230}
231
254f965c
BW
232/**
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.
236 */
273497e5 237static struct intel_context *
0eea67eb 238i915_gem_create_context(struct drm_device *dev,
d624d86e 239 struct drm_i915_file_private *file_priv)
254f965c 240{
42c3b603 241 const bool is_global_default_ctx = file_priv == NULL;
273497e5 242 struct intel_context *ctx;
bdf4fd7e 243 int ret = 0;
40521054 244
b731d33d 245 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
40521054 246
0eea67eb 247 ctx = __create_hw_context(dev, file_priv);
146937e5 248 if (IS_ERR(ctx))
a45d0f6a 249 return ctx;
40521054 250
ea0c76f8 251 if (is_global_default_ctx && ctx->legacy_hw_ctx.rcs_state) {
42c3b603
CW
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
257 * context.
258 */
ea0c76f8 259 ret = i915_gem_obj_ggtt_pin(ctx->legacy_hw_ctx.rcs_state,
1ec9e26d 260 get_context_alignment(dev), 0);
42c3b603
CW
261 if (ret) {
262 DRM_DEBUG_DRIVER("Couldn't pin %d\n", ret);
263 goto err_destroy;
264 }
265 }
266
d624d86e 267 if (USES_FULL_PPGTT(dev)) {
4d884705 268 struct i915_hw_ppgtt *ppgtt = i915_ppgtt_create(dev, file_priv);
bdf4fd7e
BW
269
270 if (IS_ERR_OR_NULL(ppgtt)) {
0eea67eb
BW
271 DRM_DEBUG_DRIVER("PPGTT setup failed (%ld)\n",
272 PTR_ERR(ppgtt));
bdf4fd7e 273 ret = PTR_ERR(ppgtt);
42c3b603 274 goto err_unpin;
ae6c4806
DV
275 }
276
277 ctx->ppgtt = ppgtt;
278 }
bdf4fd7e 279
198c974d
DCS
280 trace_i915_context_create(ctx);
281
a45d0f6a 282 return ctx;
9a3b5304 283
42c3b603 284err_unpin:
ea0c76f8
OM
285 if (is_global_default_ctx && ctx->legacy_hw_ctx.rcs_state)
286 i915_gem_object_ggtt_unpin(ctx->legacy_hw_ctx.rcs_state);
9a3b5304 287err_destroy:
dce3271b 288 i915_gem_context_unreference(ctx);
a45d0f6a 289 return ERR_PTR(ret);
254f965c
BW
290}
291
acce9ffa
BW
292void i915_gem_context_reset(struct drm_device *dev)
293{
294 struct drm_i915_private *dev_priv = dev->dev_private;
acce9ffa
BW
295 int i;
296
ecdb5fd8
TD
297 /* In execlists mode we will unreference the context when the execlist
298 * queue is cleared and the requests destroyed.
299 */
300 if (i915.enable_execlists)
301 return;
302
acce9ffa 303 for (i = 0; i < I915_NUM_RINGS; i++) {
a4872ba6 304 struct intel_engine_cs *ring = &dev_priv->ring[i];
ea0c76f8 305 struct intel_context *lctx = ring->last_context;
acce9ffa 306
6689c167
MA
307 if (lctx) {
308 if (lctx->legacy_hw_ctx.rcs_state && i == RCS)
309 i915_gem_object_ggtt_unpin(lctx->legacy_hw_ctx.rcs_state);
acce9ffa 310
6689c167
MA
311 i915_gem_context_unreference(lctx);
312 ring->last_context = NULL;
acce9ffa 313 }
acce9ffa
BW
314 }
315}
316
8245be31 317int i915_gem_context_init(struct drm_device *dev)
254f965c
BW
318{
319 struct drm_i915_private *dev_priv = dev->dev_private;
273497e5 320 struct intel_context *ctx;
a45d0f6a 321 int i;
254f965c 322
2fa48d8d
BW
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))
8245be31 326 return 0;
254f965c 327
ede7d42b
OM
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)) {
691e6415
CW
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;
338 }
254f965c
BW
339 }
340
d624d86e 341 ctx = i915_gem_create_context(dev, NULL);
691e6415
CW
342 if (IS_ERR(ctx)) {
343 DRM_ERROR("Failed to create default global context (error %ld)\n",
344 PTR_ERR(ctx));
345 return PTR_ERR(ctx);
254f965c
BW
346 }
347
ede7d42b
OM
348 for (i = 0; i < I915_NUM_RINGS; i++) {
349 struct intel_engine_cs *ring = &dev_priv->ring[i];
350
351 /* NB: RCS will hold a ref for all rings */
352 ring->default_context = ctx;
ede7d42b 353 }
67e3d297 354
ede7d42b
OM
355 DRM_DEBUG_DRIVER("%s context support initialized\n",
356 i915.enable_execlists ? "LR" :
357 dev_priv->hw_context_size ? "HW" : "fake");
8245be31 358 return 0;
254f965c
BW
359}
360
361void i915_gem_context_fini(struct drm_device *dev)
362{
363 struct drm_i915_private *dev_priv = dev->dev_private;
273497e5 364 struct intel_context *dctx = dev_priv->ring[RCS].default_context;
67e3d297 365 int i;
254f965c 366
ea0c76f8 367 if (dctx->legacy_hw_ctx.rcs_state) {
691e6415
CW
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);
372
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 */
ea0c76f8
OM
382 WARN_ON(dctx->legacy_hw_ctx.rcs_state->active);
383 i915_gem_object_ggtt_unpin(dctx->legacy_hw_ctx.rcs_state);
691e6415
CW
384 i915_gem_context_unreference(dctx);
385 dev_priv->ring[RCS].last_context = NULL;
386 }
d3b448d9 387
ea0c76f8 388 i915_gem_object_ggtt_unpin(dctx->legacy_hw_ctx.rcs_state);
67e3d297
BW
389 }
390
391 for (i = 0; i < I915_NUM_RINGS; i++) {
a4872ba6 392 struct intel_engine_cs *ring = &dev_priv->ring[i];
67e3d297
BW
393
394 if (ring->last_context)
395 i915_gem_context_unreference(ring->last_context);
396
397 ring->default_context = NULL;
0009e46c 398 ring->last_context = NULL;
71b76d00
BW
399 }
400
dce3271b 401 i915_gem_context_unreference(dctx);
254f965c
BW
402}
403
2fa48d8d
BW
404int i915_gem_context_enable(struct drm_i915_private *dev_priv)
405{
a4872ba6 406 struct intel_engine_cs *ring;
2fa48d8d
BW
407 int ret, i;
408
2fa48d8d 409 BUG_ON(!dev_priv->ring[RCS].default_context);
bdf4fd7e 410
e7778be1
TD
411 if (i915.enable_execlists) {
412 for_each_ring(ring, dev_priv, i) {
413 if (ring->init_context) {
414 ret = ring->init_context(ring,
415 ring->default_context);
416 if (ret) {
417 DRM_ERROR("ring init context: %d\n",
418 ret);
419 return ret;
420 }
421 }
422 }
ecdb5fd8 423
e7778be1
TD
424 } else
425 for_each_ring(ring, dev_priv, i) {
426 ret = i915_switch_context(ring, ring->default_context);
427 if (ret)
428 return ret;
429 }
2fa48d8d
BW
430
431 return 0;
432}
433
40521054
BW
434static int context_idr_cleanup(int id, void *p, void *data)
435{
273497e5 436 struct intel_context *ctx = p;
40521054 437
dce3271b 438 i915_gem_context_unreference(ctx);
40521054 439 return 0;
254f965c
BW
440}
441
e422b888
BW
442int i915_gem_context_open(struct drm_device *dev, struct drm_file *file)
443{
444 struct drm_i915_file_private *file_priv = file->driver_priv;
f83d6518 445 struct intel_context *ctx;
e422b888
BW
446
447 idr_init(&file_priv->context_idr);
448
0eea67eb 449 mutex_lock(&dev->struct_mutex);
d624d86e 450 ctx = i915_gem_create_context(dev, file_priv);
0eea67eb
BW
451 mutex_unlock(&dev->struct_mutex);
452
f83d6518 453 if (IS_ERR(ctx)) {
0eea67eb 454 idr_destroy(&file_priv->context_idr);
f83d6518 455 return PTR_ERR(ctx);
0eea67eb
BW
456 }
457
e422b888
BW
458 return 0;
459}
460
254f965c
BW
461void i915_gem_context_close(struct drm_device *dev, struct drm_file *file)
462{
40521054 463 struct drm_i915_file_private *file_priv = file->driver_priv;
254f965c 464
73c273eb 465 idr_for_each(&file_priv->context_idr, context_idr_cleanup, NULL);
40521054 466 idr_destroy(&file_priv->context_idr);
40521054
BW
467}
468
273497e5 469struct intel_context *
40521054
BW
470i915_gem_context_get(struct drm_i915_file_private *file_priv, u32 id)
471{
273497e5 472 struct intel_context *ctx;
72ad5c45 473
273497e5 474 ctx = (struct intel_context *)idr_find(&file_priv->context_idr, id);
72ad5c45
BW
475 if (!ctx)
476 return ERR_PTR(-ENOENT);
477
478 return ctx;
254f965c 479}
e0556841
BW
480
481static inline int
a4872ba6 482mi_set_context(struct intel_engine_cs *ring,
273497e5 483 struct intel_context *new_context,
e0556841
BW
484 u32 hw_flags)
485{
e80f14b6 486 u32 flags = hw_flags | MI_MM_SPACE_GTT;
e0556841
BW
487 int ret;
488
12b0286f
BW
489 /* w/a: If Flush TLB Invalidation Mode is enabled, driver must do a TLB
490 * invalidation prior to MI_SET_CONTEXT. On GEN6 we don't set the value
491 * explicitly, so we rely on the value at ring init, stored in
492 * itlb_before_ctx_switch.
493 */
057f6a8a 494 if (IS_GEN6(ring->dev)) {
ac82ea2e 495 ret = ring->flush(ring, I915_GEM_GPU_DOMAINS, 0);
12b0286f
BW
496 if (ret)
497 return ret;
498 }
499
e80f14b6
BW
500 /* These flags are for resource streamer on HSW+ */
501 if (!IS_HASWELL(ring->dev) && INTEL_INFO(ring->dev)->gen < 8)
502 flags |= (MI_SAVE_EXT_STATE_EN | MI_RESTORE_EXT_STATE_EN);
503
e37ec39b 504 ret = intel_ring_begin(ring, 6);
e0556841
BW
505 if (ret)
506 return ret;
507
b3f797ac 508 /* WaProgramMiArbOnOffAroundMiSetContext:ivb,vlv,hsw,bdw,chv */
64bed788 509 if (INTEL_INFO(ring->dev)->gen >= 7)
e37ec39b
BW
510 intel_ring_emit(ring, MI_ARB_ON_OFF | MI_ARB_DISABLE);
511 else
512 intel_ring_emit(ring, MI_NOOP);
513
e0556841
BW
514 intel_ring_emit(ring, MI_NOOP);
515 intel_ring_emit(ring, MI_SET_CONTEXT);
ea0c76f8 516 intel_ring_emit(ring, i915_gem_obj_ggtt_offset(new_context->legacy_hw_ctx.rcs_state) |
e80f14b6 517 flags);
2b7e8082
VS
518 /*
519 * w/a: MI_SET_CONTEXT must always be followed by MI_NOOP
520 * WaMiSetContext_Hang:snb,ivb,vlv
521 */
e0556841
BW
522 intel_ring_emit(ring, MI_NOOP);
523
64bed788 524 if (INTEL_INFO(ring->dev)->gen >= 7)
e37ec39b
BW
525 intel_ring_emit(ring, MI_ARB_ON_OFF | MI_ARB_ENABLE);
526 else
527 intel_ring_emit(ring, MI_NOOP);
528
e0556841
BW
529 intel_ring_advance(ring);
530
531 return ret;
532}
533
a4872ba6 534static int do_switch(struct intel_engine_cs *ring,
273497e5 535 struct intel_context *to)
e0556841 536{
6f65e29a 537 struct drm_i915_private *dev_priv = ring->dev->dev_private;
273497e5 538 struct intel_context *from = ring->last_context;
e0556841 539 u32 hw_flags = 0;
967ab6b1 540 bool uninitialized = false;
aff43766 541 struct i915_vma *vma;
3ccfd19d 542 int ret, i;
e0556841 543
67e3d297 544 if (from != NULL && ring == &dev_priv->ring[RCS]) {
ea0c76f8
OM
545 BUG_ON(from->legacy_hw_ctx.rcs_state == NULL);
546 BUG_ON(!i915_gem_obj_is_pinned(from->legacy_hw_ctx.rcs_state));
67e3d297 547 }
e0556841 548
14d8ec54 549 if (from == to && !to->remap_slice)
9a3b5304
CW
550 return 0;
551
7e0d96bc
BW
552 /* Trying to pin first makes error handling easier. */
553 if (ring == &dev_priv->ring[RCS]) {
ea0c76f8 554 ret = i915_gem_obj_ggtt_pin(to->legacy_hw_ctx.rcs_state,
1ec9e26d 555 get_context_alignment(ring->dev), 0);
7e0d96bc
BW
556 if (ret)
557 return ret;
67e3d297
BW
558 }
559
acc240d4
DV
560 /*
561 * Pin can switch back to the default context if we end up calling into
562 * evict_everything - as a last ditch gtt defrag effort that also
563 * switches to the default context. Hence we need to reload from here.
564 */
565 from = ring->last_context;
566
ae6c4806 567 if (to->ppgtt) {
198c974d 568 trace_switch_mm(ring, to);
6689c167 569 ret = to->ppgtt->switch_mm(to->ppgtt, ring);
7e0d96bc
BW
570 if (ret)
571 goto unpin_out;
572 }
573
574 if (ring != &dev_priv->ring[RCS]) {
575 if (from)
576 i915_gem_context_unreference(from);
577 goto done;
578 }
579
acc240d4
DV
580 /*
581 * Clear this page out of any CPU caches for coherent swap-in/out. Note
d3373a24
CW
582 * that thanks to write = false in this call and us not setting any gpu
583 * write domains when putting a context object onto the active list
584 * (when switching away from it), this won't block.
acc240d4
DV
585 *
586 * XXX: We need a real interface to do this instead of trickery.
587 */
ea0c76f8 588 ret = i915_gem_object_set_to_gtt_domain(to->legacy_hw_ctx.rcs_state, false);
7e0d96bc
BW
589 if (ret)
590 goto unpin_out;
d3373a24 591
aff43766 592 vma = i915_gem_obj_to_ggtt(to->legacy_hw_ctx.rcs_state);
fe14d5f4
TU
593 if (!(vma->bound & GLOBAL_BIND)) {
594 ret = i915_vma_bind(vma,
595 to->legacy_hw_ctx.rcs_state->cache_level,
596 GLOBAL_BIND);
597 /* This shouldn't ever fail. */
598 if (WARN_ONCE(ret, "GGTT context bind failed!"))
599 goto unpin_out;
600 }
3af7b857 601
ea0c76f8 602 if (!to->legacy_hw_ctx.initialized || i915_gem_context_is_default(to))
e0556841 603 hw_flags |= MI_RESTORE_INHIBIT;
e0556841 604
e0556841 605 ret = mi_set_context(ring, to, hw_flags);
7e0d96bc
BW
606 if (ret)
607 goto unpin_out;
e0556841 608
3ccfd19d
BW
609 for (i = 0; i < MAX_L3_SLICES; i++) {
610 if (!(to->remap_slice & (1<<i)))
611 continue;
612
613 ret = i915_gem_l3_remap(ring, i);
614 /* If it failed, try again next round */
615 if (ret)
616 DRM_DEBUG_DRIVER("L3 remapping failed\n");
617 else
618 to->remap_slice &= ~(1<<i);
619 }
620
e0556841
BW
621 /* The backing object for the context is done after switching to the
622 * *next* context. Therefore we cannot retire the previous context until
623 * the next context has already started running. In fact, the below code
624 * is a bit suboptimal because the retiring can occur simply after the
625 * MI_SET_CONTEXT instead of when the next seqno has completed.
626 */
112522f6 627 if (from != NULL) {
ea0c76f8
OM
628 from->legacy_hw_ctx.rcs_state->base.read_domains = I915_GEM_DOMAIN_INSTRUCTION;
629 i915_vma_move_to_active(i915_gem_obj_to_ggtt(from->legacy_hw_ctx.rcs_state), ring);
e0556841
BW
630 /* As long as MI_SET_CONTEXT is serializing, ie. it flushes the
631 * whole damn pipeline, we don't need to explicitly mark the
632 * object dirty. The only exception is that the context must be
633 * correct in case the object gets swapped out. Ideally we'd be
634 * able to defer doing this until we know the object would be
635 * swapped, but there is no way to do that yet.
636 */
ea0c76f8 637 from->legacy_hw_ctx.rcs_state->dirty = 1;
41c52415
JH
638 BUG_ON(i915_gem_request_get_ring(
639 from->legacy_hw_ctx.rcs_state->last_read_req) != ring);
112522f6 640
c0321e2c 641 /* obj is kept alive until the next request by its active ref */
ea0c76f8 642 i915_gem_object_ggtt_unpin(from->legacy_hw_ctx.rcs_state);
112522f6 643 i915_gem_context_unreference(from);
e0556841
BW
644 }
645
ea0c76f8
OM
646 uninitialized = !to->legacy_hw_ctx.initialized && from == NULL;
647 to->legacy_hw_ctx.initialized = true;
967ab6b1 648
67e3d297 649done:
112522f6
CW
650 i915_gem_context_reference(to);
651 ring->last_context = to;
e0556841 652
967ab6b1 653 if (uninitialized) {
86d7f238 654 if (ring->init_context) {
771b9a53 655 ret = ring->init_context(ring, to);
86d7f238
AS
656 if (ret)
657 DRM_ERROR("ring init context: %d\n", ret);
658 }
46470fc9
MK
659 }
660
e0556841 661 return 0;
7e0d96bc
BW
662
663unpin_out:
664 if (ring->id == RCS)
ea0c76f8 665 i915_gem_object_ggtt_unpin(to->legacy_hw_ctx.rcs_state);
7e0d96bc 666 return ret;
e0556841
BW
667}
668
669/**
670 * i915_switch_context() - perform a GPU context switch.
671 * @ring: ring for which we'll execute the context switch
96a6f0f1 672 * @to: the context to switch to
e0556841
BW
673 *
674 * The context life cycle is simple. The context refcount is incremented and
675 * decremented by 1 and create and destroy. If the context is in use by the GPU,
ecdb5fd8 676 * it will have a refcount > 1. This allows us to destroy the context abstract
e0556841 677 * object while letting the normal object tracking destroy the backing BO.
ecdb5fd8
TD
678 *
679 * This function should not be used in execlists mode. Instead the context is
680 * switched by writing to the ELSP and requests keep a reference to their
681 * context.
e0556841 682 */
a4872ba6 683int i915_switch_context(struct intel_engine_cs *ring,
273497e5 684 struct intel_context *to)
e0556841
BW
685{
686 struct drm_i915_private *dev_priv = ring->dev->dev_private;
e0556841 687
ecdb5fd8 688 WARN_ON(i915.enable_execlists);
0eea67eb
BW
689 WARN_ON(!mutex_is_locked(&dev_priv->dev->struct_mutex));
690
ea0c76f8 691 if (to->legacy_hw_ctx.rcs_state == NULL) { /* We have the fake context */
691e6415
CW
692 if (to != ring->last_context) {
693 i915_gem_context_reference(to);
694 if (ring->last_context)
695 i915_gem_context_unreference(ring->last_context);
696 ring->last_context = to;
697 }
c482972a 698 return 0;
a95f6a00 699 }
c482972a 700
67e3d297 701 return do_switch(ring, to);
e0556841 702}
84624813 703
ec3e9963 704static bool contexts_enabled(struct drm_device *dev)
691e6415 705{
ec3e9963 706 return i915.enable_execlists || to_i915(dev)->hw_context_size;
691e6415
CW
707}
708
84624813
BW
709int i915_gem_context_create_ioctl(struct drm_device *dev, void *data,
710 struct drm_file *file)
711{
84624813
BW
712 struct drm_i915_gem_context_create *args = data;
713 struct drm_i915_file_private *file_priv = file->driver_priv;
273497e5 714 struct intel_context *ctx;
84624813
BW
715 int ret;
716
ec3e9963 717 if (!contexts_enabled(dev))
5fa8be65
DV
718 return -ENODEV;
719
84624813
BW
720 ret = i915_mutex_lock_interruptible(dev);
721 if (ret)
722 return ret;
723
d624d86e 724 ctx = i915_gem_create_context(dev, file_priv);
84624813 725 mutex_unlock(&dev->struct_mutex);
be636387
DC
726 if (IS_ERR(ctx))
727 return PTR_ERR(ctx);
84624813 728
821d66dd 729 args->ctx_id = ctx->user_handle;
84624813
BW
730 DRM_DEBUG_DRIVER("HW context %d created\n", args->ctx_id);
731
be636387 732 return 0;
84624813
BW
733}
734
735int i915_gem_context_destroy_ioctl(struct drm_device *dev, void *data,
736 struct drm_file *file)
737{
738 struct drm_i915_gem_context_destroy *args = data;
739 struct drm_i915_file_private *file_priv = file->driver_priv;
273497e5 740 struct intel_context *ctx;
84624813
BW
741 int ret;
742
821d66dd 743 if (args->ctx_id == DEFAULT_CONTEXT_HANDLE)
c2cf2416 744 return -ENOENT;
0eea67eb 745
84624813
BW
746 ret = i915_mutex_lock_interruptible(dev);
747 if (ret)
748 return ret;
749
750 ctx = i915_gem_context_get(file_priv, args->ctx_id);
72ad5c45 751 if (IS_ERR(ctx)) {
84624813 752 mutex_unlock(&dev->struct_mutex);
72ad5c45 753 return PTR_ERR(ctx);
84624813
BW
754 }
755
821d66dd 756 idr_remove(&ctx->file_priv->context_idr, ctx->user_handle);
dce3271b 757 i915_gem_context_unreference(ctx);
84624813
BW
758 mutex_unlock(&dev->struct_mutex);
759
760 DRM_DEBUG_DRIVER("HW context %d destroyed\n", args->ctx_id);
761 return 0;
762}