]>
Commit | Line | Data |
---|---|---|
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 |
100 | static 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 |
108 | static 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 | 136 | void 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 | 154 | struct drm_i915_gem_object * |
aa0c13da OM |
155 | i915_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 | 184 | static 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 | |
676fa572 CW |
225 | ctx->hang_stats.ban_period_seconds = DRM_I915_CTX_BAN_PERIOD; |
226 | ||
146937e5 | 227 | return ctx; |
40521054 BW |
228 | |
229 | err_out: | |
dce3271b | 230 | i915_gem_context_unreference(ctx); |
146937e5 | 231 | return ERR_PTR(ret); |
40521054 BW |
232 | } |
233 | ||
254f965c BW |
234 | /** |
235 | * The default context needs to exist per ring that uses contexts. It stores the | |
236 | * context state of the GPU for applications that don't utilize HW contexts, as | |
237 | * well as an idle case. | |
238 | */ | |
273497e5 | 239 | static struct intel_context * |
0eea67eb | 240 | i915_gem_create_context(struct drm_device *dev, |
d624d86e | 241 | struct drm_i915_file_private *file_priv) |
254f965c | 242 | { |
42c3b603 | 243 | const bool is_global_default_ctx = file_priv == NULL; |
273497e5 | 244 | struct intel_context *ctx; |
bdf4fd7e | 245 | int ret = 0; |
40521054 | 246 | |
b731d33d | 247 | BUG_ON(!mutex_is_locked(&dev->struct_mutex)); |
40521054 | 248 | |
0eea67eb | 249 | ctx = __create_hw_context(dev, file_priv); |
146937e5 | 250 | if (IS_ERR(ctx)) |
a45d0f6a | 251 | return ctx; |
40521054 | 252 | |
ea0c76f8 | 253 | if (is_global_default_ctx && ctx->legacy_hw_ctx.rcs_state) { |
42c3b603 CW |
254 | /* We may need to do things with the shrinker which |
255 | * require us to immediately switch back to the default | |
256 | * context. This can cause a problem as pinning the | |
257 | * default context also requires GTT space which may not | |
258 | * be available. To avoid this we always pin the default | |
259 | * context. | |
260 | */ | |
ea0c76f8 | 261 | ret = i915_gem_obj_ggtt_pin(ctx->legacy_hw_ctx.rcs_state, |
1ec9e26d | 262 | get_context_alignment(dev), 0); |
42c3b603 CW |
263 | if (ret) { |
264 | DRM_DEBUG_DRIVER("Couldn't pin %d\n", ret); | |
265 | goto err_destroy; | |
266 | } | |
267 | } | |
268 | ||
d624d86e | 269 | if (USES_FULL_PPGTT(dev)) { |
4d884705 | 270 | struct i915_hw_ppgtt *ppgtt = i915_ppgtt_create(dev, file_priv); |
bdf4fd7e BW |
271 | |
272 | if (IS_ERR_OR_NULL(ppgtt)) { | |
0eea67eb BW |
273 | DRM_DEBUG_DRIVER("PPGTT setup failed (%ld)\n", |
274 | PTR_ERR(ppgtt)); | |
bdf4fd7e | 275 | ret = PTR_ERR(ppgtt); |
42c3b603 | 276 | goto err_unpin; |
ae6c4806 DV |
277 | } |
278 | ||
279 | ctx->ppgtt = ppgtt; | |
280 | } | |
bdf4fd7e | 281 | |
198c974d DCS |
282 | trace_i915_context_create(ctx); |
283 | ||
a45d0f6a | 284 | return ctx; |
9a3b5304 | 285 | |
42c3b603 | 286 | err_unpin: |
ea0c76f8 OM |
287 | if (is_global_default_ctx && ctx->legacy_hw_ctx.rcs_state) |
288 | i915_gem_object_ggtt_unpin(ctx->legacy_hw_ctx.rcs_state); | |
9a3b5304 | 289 | err_destroy: |
dce3271b | 290 | i915_gem_context_unreference(ctx); |
a45d0f6a | 291 | return ERR_PTR(ret); |
254f965c BW |
292 | } |
293 | ||
acce9ffa BW |
294 | void i915_gem_context_reset(struct drm_device *dev) |
295 | { | |
296 | struct drm_i915_private *dev_priv = dev->dev_private; | |
acce9ffa BW |
297 | int i; |
298 | ||
ecdb5fd8 TD |
299 | /* In execlists mode we will unreference the context when the execlist |
300 | * queue is cleared and the requests destroyed. | |
301 | */ | |
302 | if (i915.enable_execlists) | |
303 | return; | |
304 | ||
acce9ffa | 305 | for (i = 0; i < I915_NUM_RINGS; i++) { |
a4872ba6 | 306 | struct intel_engine_cs *ring = &dev_priv->ring[i]; |
ea0c76f8 | 307 | struct intel_context *lctx = ring->last_context; |
acce9ffa | 308 | |
6689c167 MA |
309 | if (lctx) { |
310 | if (lctx->legacy_hw_ctx.rcs_state && i == RCS) | |
311 | i915_gem_object_ggtt_unpin(lctx->legacy_hw_ctx.rcs_state); | |
acce9ffa | 312 | |
6689c167 MA |
313 | i915_gem_context_unreference(lctx); |
314 | ring->last_context = NULL; | |
acce9ffa | 315 | } |
acce9ffa BW |
316 | } |
317 | } | |
318 | ||
8245be31 | 319 | int i915_gem_context_init(struct drm_device *dev) |
254f965c BW |
320 | { |
321 | struct drm_i915_private *dev_priv = dev->dev_private; | |
273497e5 | 322 | struct intel_context *ctx; |
a45d0f6a | 323 | int i; |
254f965c | 324 | |
2fa48d8d BW |
325 | /* Init should only be called once per module load. Eventually the |
326 | * restriction on the context_disabled check can be loosened. */ | |
327 | if (WARN_ON(dev_priv->ring[RCS].default_context)) | |
8245be31 | 328 | return 0; |
254f965c | 329 | |
ede7d42b OM |
330 | if (i915.enable_execlists) { |
331 | /* NB: intentionally left blank. We will allocate our own | |
332 | * backing objects as we need them, thank you very much */ | |
333 | dev_priv->hw_context_size = 0; | |
334 | } else if (HAS_HW_CONTEXTS(dev)) { | |
691e6415 CW |
335 | dev_priv->hw_context_size = round_up(get_context_size(dev), 4096); |
336 | if (dev_priv->hw_context_size > (1<<20)) { | |
337 | DRM_DEBUG_DRIVER("Disabling HW Contexts; invalid size %d\n", | |
338 | dev_priv->hw_context_size); | |
339 | dev_priv->hw_context_size = 0; | |
340 | } | |
254f965c BW |
341 | } |
342 | ||
d624d86e | 343 | ctx = i915_gem_create_context(dev, NULL); |
691e6415 CW |
344 | if (IS_ERR(ctx)) { |
345 | DRM_ERROR("Failed to create default global context (error %ld)\n", | |
346 | PTR_ERR(ctx)); | |
347 | return PTR_ERR(ctx); | |
254f965c BW |
348 | } |
349 | ||
ede7d42b OM |
350 | for (i = 0; i < I915_NUM_RINGS; i++) { |
351 | struct intel_engine_cs *ring = &dev_priv->ring[i]; | |
352 | ||
353 | /* NB: RCS will hold a ref for all rings */ | |
354 | ring->default_context = ctx; | |
ede7d42b | 355 | } |
67e3d297 | 356 | |
ede7d42b OM |
357 | DRM_DEBUG_DRIVER("%s context support initialized\n", |
358 | i915.enable_execlists ? "LR" : | |
359 | dev_priv->hw_context_size ? "HW" : "fake"); | |
8245be31 | 360 | return 0; |
254f965c BW |
361 | } |
362 | ||
363 | void i915_gem_context_fini(struct drm_device *dev) | |
364 | { | |
365 | struct drm_i915_private *dev_priv = dev->dev_private; | |
273497e5 | 366 | struct intel_context *dctx = dev_priv->ring[RCS].default_context; |
67e3d297 | 367 | int i; |
254f965c | 368 | |
ea0c76f8 | 369 | if (dctx->legacy_hw_ctx.rcs_state) { |
691e6415 CW |
370 | /* The only known way to stop the gpu from accessing the hw context is |
371 | * to reset it. Do this as the very last operation to avoid confusing | |
372 | * other code, leading to spurious errors. */ | |
373 | intel_gpu_reset(dev); | |
374 | ||
375 | /* When default context is created and switched to, base object refcount | |
376 | * will be 2 (+1 from object creation and +1 from do_switch()). | |
377 | * i915_gem_context_fini() will be called after gpu_idle() has switched | |
378 | * to default context. So we need to unreference the base object once | |
379 | * to offset the do_switch part, so that i915_gem_context_unreference() | |
380 | * can then free the base object correctly. */ | |
381 | WARN_ON(!dev_priv->ring[RCS].last_context); | |
382 | if (dev_priv->ring[RCS].last_context == dctx) { | |
383 | /* Fake switch to NULL context */ | |
ea0c76f8 OM |
384 | WARN_ON(dctx->legacy_hw_ctx.rcs_state->active); |
385 | i915_gem_object_ggtt_unpin(dctx->legacy_hw_ctx.rcs_state); | |
691e6415 CW |
386 | i915_gem_context_unreference(dctx); |
387 | dev_priv->ring[RCS].last_context = NULL; | |
388 | } | |
d3b448d9 | 389 | |
ea0c76f8 | 390 | i915_gem_object_ggtt_unpin(dctx->legacy_hw_ctx.rcs_state); |
67e3d297 BW |
391 | } |
392 | ||
393 | for (i = 0; i < I915_NUM_RINGS; i++) { | |
a4872ba6 | 394 | struct intel_engine_cs *ring = &dev_priv->ring[i]; |
67e3d297 BW |
395 | |
396 | if (ring->last_context) | |
397 | i915_gem_context_unreference(ring->last_context); | |
398 | ||
399 | ring->default_context = NULL; | |
0009e46c | 400 | ring->last_context = NULL; |
71b76d00 BW |
401 | } |
402 | ||
dce3271b | 403 | i915_gem_context_unreference(dctx); |
254f965c BW |
404 | } |
405 | ||
2fa48d8d BW |
406 | int i915_gem_context_enable(struct drm_i915_private *dev_priv) |
407 | { | |
a4872ba6 | 408 | struct intel_engine_cs *ring; |
2fa48d8d BW |
409 | int ret, i; |
410 | ||
2fa48d8d | 411 | BUG_ON(!dev_priv->ring[RCS].default_context); |
bdf4fd7e | 412 | |
e7778be1 TD |
413 | if (i915.enable_execlists) { |
414 | for_each_ring(ring, dev_priv, i) { | |
415 | if (ring->init_context) { | |
416 | ret = ring->init_context(ring, | |
417 | ring->default_context); | |
418 | if (ret) { | |
419 | DRM_ERROR("ring init context: %d\n", | |
420 | ret); | |
421 | return ret; | |
422 | } | |
423 | } | |
424 | } | |
ecdb5fd8 | 425 | |
e7778be1 TD |
426 | } else |
427 | for_each_ring(ring, dev_priv, i) { | |
428 | ret = i915_switch_context(ring, ring->default_context); | |
429 | if (ret) | |
430 | return ret; | |
431 | } | |
2fa48d8d BW |
432 | |
433 | return 0; | |
434 | } | |
435 | ||
40521054 BW |
436 | static int context_idr_cleanup(int id, void *p, void *data) |
437 | { | |
273497e5 | 438 | struct intel_context *ctx = p; |
40521054 | 439 | |
dce3271b | 440 | i915_gem_context_unreference(ctx); |
40521054 | 441 | return 0; |
254f965c BW |
442 | } |
443 | ||
e422b888 BW |
444 | int i915_gem_context_open(struct drm_device *dev, struct drm_file *file) |
445 | { | |
446 | struct drm_i915_file_private *file_priv = file->driver_priv; | |
f83d6518 | 447 | struct intel_context *ctx; |
e422b888 BW |
448 | |
449 | idr_init(&file_priv->context_idr); | |
450 | ||
0eea67eb | 451 | mutex_lock(&dev->struct_mutex); |
d624d86e | 452 | ctx = i915_gem_create_context(dev, file_priv); |
0eea67eb BW |
453 | mutex_unlock(&dev->struct_mutex); |
454 | ||
f83d6518 | 455 | if (IS_ERR(ctx)) { |
0eea67eb | 456 | idr_destroy(&file_priv->context_idr); |
f83d6518 | 457 | return PTR_ERR(ctx); |
0eea67eb BW |
458 | } |
459 | ||
e422b888 BW |
460 | return 0; |
461 | } | |
462 | ||
254f965c BW |
463 | void i915_gem_context_close(struct drm_device *dev, struct drm_file *file) |
464 | { | |
40521054 | 465 | struct drm_i915_file_private *file_priv = file->driver_priv; |
254f965c | 466 | |
73c273eb | 467 | idr_for_each(&file_priv->context_idr, context_idr_cleanup, NULL); |
40521054 | 468 | idr_destroy(&file_priv->context_idr); |
40521054 BW |
469 | } |
470 | ||
273497e5 | 471 | struct intel_context * |
40521054 BW |
472 | i915_gem_context_get(struct drm_i915_file_private *file_priv, u32 id) |
473 | { | |
273497e5 | 474 | struct intel_context *ctx; |
72ad5c45 | 475 | |
273497e5 | 476 | ctx = (struct intel_context *)idr_find(&file_priv->context_idr, id); |
72ad5c45 BW |
477 | if (!ctx) |
478 | return ERR_PTR(-ENOENT); | |
479 | ||
480 | return ctx; | |
254f965c | 481 | } |
e0556841 BW |
482 | |
483 | static inline int | |
a4872ba6 | 484 | mi_set_context(struct intel_engine_cs *ring, |
273497e5 | 485 | struct intel_context *new_context, |
e0556841 BW |
486 | u32 hw_flags) |
487 | { | |
e80f14b6 | 488 | u32 flags = hw_flags | MI_MM_SPACE_GTT; |
e0556841 BW |
489 | int ret; |
490 | ||
12b0286f BW |
491 | /* w/a: If Flush TLB Invalidation Mode is enabled, driver must do a TLB |
492 | * invalidation prior to MI_SET_CONTEXT. On GEN6 we don't set the value | |
493 | * explicitly, so we rely on the value at ring init, stored in | |
494 | * itlb_before_ctx_switch. | |
495 | */ | |
057f6a8a | 496 | if (IS_GEN6(ring->dev)) { |
ac82ea2e | 497 | ret = ring->flush(ring, I915_GEM_GPU_DOMAINS, 0); |
12b0286f BW |
498 | if (ret) |
499 | return ret; | |
500 | } | |
501 | ||
e80f14b6 BW |
502 | /* These flags are for resource streamer on HSW+ */ |
503 | if (!IS_HASWELL(ring->dev) && INTEL_INFO(ring->dev)->gen < 8) | |
504 | flags |= (MI_SAVE_EXT_STATE_EN | MI_RESTORE_EXT_STATE_EN); | |
505 | ||
e37ec39b | 506 | ret = intel_ring_begin(ring, 6); |
e0556841 BW |
507 | if (ret) |
508 | return ret; | |
509 | ||
b3f797ac | 510 | /* WaProgramMiArbOnOffAroundMiSetContext:ivb,vlv,hsw,bdw,chv */ |
64bed788 | 511 | if (INTEL_INFO(ring->dev)->gen >= 7) |
e37ec39b BW |
512 | intel_ring_emit(ring, MI_ARB_ON_OFF | MI_ARB_DISABLE); |
513 | else | |
514 | intel_ring_emit(ring, MI_NOOP); | |
515 | ||
e0556841 BW |
516 | intel_ring_emit(ring, MI_NOOP); |
517 | intel_ring_emit(ring, MI_SET_CONTEXT); | |
ea0c76f8 | 518 | intel_ring_emit(ring, i915_gem_obj_ggtt_offset(new_context->legacy_hw_ctx.rcs_state) | |
e80f14b6 | 519 | flags); |
2b7e8082 VS |
520 | /* |
521 | * w/a: MI_SET_CONTEXT must always be followed by MI_NOOP | |
522 | * WaMiSetContext_Hang:snb,ivb,vlv | |
523 | */ | |
e0556841 BW |
524 | intel_ring_emit(ring, MI_NOOP); |
525 | ||
64bed788 | 526 | if (INTEL_INFO(ring->dev)->gen >= 7) |
e37ec39b BW |
527 | intel_ring_emit(ring, MI_ARB_ON_OFF | MI_ARB_ENABLE); |
528 | else | |
529 | intel_ring_emit(ring, MI_NOOP); | |
530 | ||
e0556841 BW |
531 | intel_ring_advance(ring); |
532 | ||
533 | return ret; | |
534 | } | |
535 | ||
a4872ba6 | 536 | static int do_switch(struct intel_engine_cs *ring, |
273497e5 | 537 | struct intel_context *to) |
e0556841 | 538 | { |
6f65e29a | 539 | struct drm_i915_private *dev_priv = ring->dev->dev_private; |
273497e5 | 540 | struct intel_context *from = ring->last_context; |
e0556841 | 541 | u32 hw_flags = 0; |
967ab6b1 | 542 | bool uninitialized = false; |
aff43766 | 543 | struct i915_vma *vma; |
3ccfd19d | 544 | int ret, i; |
e0556841 | 545 | |
67e3d297 | 546 | if (from != NULL && ring == &dev_priv->ring[RCS]) { |
ea0c76f8 OM |
547 | BUG_ON(from->legacy_hw_ctx.rcs_state == NULL); |
548 | BUG_ON(!i915_gem_obj_is_pinned(from->legacy_hw_ctx.rcs_state)); | |
67e3d297 | 549 | } |
e0556841 | 550 | |
14d8ec54 | 551 | if (from == to && !to->remap_slice) |
9a3b5304 CW |
552 | return 0; |
553 | ||
7e0d96bc BW |
554 | /* Trying to pin first makes error handling easier. */ |
555 | if (ring == &dev_priv->ring[RCS]) { | |
ea0c76f8 | 556 | ret = i915_gem_obj_ggtt_pin(to->legacy_hw_ctx.rcs_state, |
1ec9e26d | 557 | get_context_alignment(ring->dev), 0); |
7e0d96bc BW |
558 | if (ret) |
559 | return ret; | |
67e3d297 BW |
560 | } |
561 | ||
acc240d4 DV |
562 | /* |
563 | * Pin can switch back to the default context if we end up calling into | |
564 | * evict_everything - as a last ditch gtt defrag effort that also | |
565 | * switches to the default context. Hence we need to reload from here. | |
566 | */ | |
567 | from = ring->last_context; | |
568 | ||
ae6c4806 | 569 | if (to->ppgtt) { |
198c974d | 570 | trace_switch_mm(ring, to); |
6689c167 | 571 | ret = to->ppgtt->switch_mm(to->ppgtt, ring); |
7e0d96bc BW |
572 | if (ret) |
573 | goto unpin_out; | |
574 | } | |
575 | ||
576 | if (ring != &dev_priv->ring[RCS]) { | |
577 | if (from) | |
578 | i915_gem_context_unreference(from); | |
579 | goto done; | |
580 | } | |
581 | ||
acc240d4 DV |
582 | /* |
583 | * Clear this page out of any CPU caches for coherent swap-in/out. Note | |
d3373a24 CW |
584 | * that thanks to write = false in this call and us not setting any gpu |
585 | * write domains when putting a context object onto the active list | |
586 | * (when switching away from it), this won't block. | |
acc240d4 DV |
587 | * |
588 | * XXX: We need a real interface to do this instead of trickery. | |
589 | */ | |
ea0c76f8 | 590 | ret = i915_gem_object_set_to_gtt_domain(to->legacy_hw_ctx.rcs_state, false); |
7e0d96bc BW |
591 | if (ret) |
592 | goto unpin_out; | |
d3373a24 | 593 | |
aff43766 | 594 | vma = i915_gem_obj_to_ggtt(to->legacy_hw_ctx.rcs_state); |
fe14d5f4 TU |
595 | if (!(vma->bound & GLOBAL_BIND)) { |
596 | ret = i915_vma_bind(vma, | |
597 | to->legacy_hw_ctx.rcs_state->cache_level, | |
598 | GLOBAL_BIND); | |
599 | /* This shouldn't ever fail. */ | |
600 | if (WARN_ONCE(ret, "GGTT context bind failed!")) | |
601 | goto unpin_out; | |
602 | } | |
3af7b857 | 603 | |
ea0c76f8 | 604 | if (!to->legacy_hw_ctx.initialized || i915_gem_context_is_default(to)) |
e0556841 | 605 | hw_flags |= MI_RESTORE_INHIBIT; |
e0556841 | 606 | |
e0556841 | 607 | ret = mi_set_context(ring, to, hw_flags); |
7e0d96bc BW |
608 | if (ret) |
609 | goto unpin_out; | |
e0556841 | 610 | |
3ccfd19d BW |
611 | for (i = 0; i < MAX_L3_SLICES; i++) { |
612 | if (!(to->remap_slice & (1<<i))) | |
613 | continue; | |
614 | ||
615 | ret = i915_gem_l3_remap(ring, i); | |
616 | /* If it failed, try again next round */ | |
617 | if (ret) | |
618 | DRM_DEBUG_DRIVER("L3 remapping failed\n"); | |
619 | else | |
620 | to->remap_slice &= ~(1<<i); | |
621 | } | |
622 | ||
e0556841 BW |
623 | /* The backing object for the context is done after switching to the |
624 | * *next* context. Therefore we cannot retire the previous context until | |
625 | * the next context has already started running. In fact, the below code | |
626 | * is a bit suboptimal because the retiring can occur simply after the | |
627 | * MI_SET_CONTEXT instead of when the next seqno has completed. | |
628 | */ | |
112522f6 | 629 | if (from != NULL) { |
ea0c76f8 OM |
630 | from->legacy_hw_ctx.rcs_state->base.read_domains = I915_GEM_DOMAIN_INSTRUCTION; |
631 | i915_vma_move_to_active(i915_gem_obj_to_ggtt(from->legacy_hw_ctx.rcs_state), ring); | |
e0556841 BW |
632 | /* As long as MI_SET_CONTEXT is serializing, ie. it flushes the |
633 | * whole damn pipeline, we don't need to explicitly mark the | |
634 | * object dirty. The only exception is that the context must be | |
635 | * correct in case the object gets swapped out. Ideally we'd be | |
636 | * able to defer doing this until we know the object would be | |
637 | * swapped, but there is no way to do that yet. | |
638 | */ | |
ea0c76f8 | 639 | from->legacy_hw_ctx.rcs_state->dirty = 1; |
41c52415 JH |
640 | BUG_ON(i915_gem_request_get_ring( |
641 | from->legacy_hw_ctx.rcs_state->last_read_req) != ring); | |
112522f6 | 642 | |
c0321e2c | 643 | /* obj is kept alive until the next request by its active ref */ |
ea0c76f8 | 644 | i915_gem_object_ggtt_unpin(from->legacy_hw_ctx.rcs_state); |
112522f6 | 645 | i915_gem_context_unreference(from); |
e0556841 BW |
646 | } |
647 | ||
ea0c76f8 OM |
648 | uninitialized = !to->legacy_hw_ctx.initialized && from == NULL; |
649 | to->legacy_hw_ctx.initialized = true; | |
967ab6b1 | 650 | |
67e3d297 | 651 | done: |
112522f6 CW |
652 | i915_gem_context_reference(to); |
653 | ring->last_context = to; | |
e0556841 | 654 | |
967ab6b1 | 655 | if (uninitialized) { |
86d7f238 | 656 | if (ring->init_context) { |
771b9a53 | 657 | ret = ring->init_context(ring, to); |
86d7f238 AS |
658 | if (ret) |
659 | DRM_ERROR("ring init context: %d\n", ret); | |
660 | } | |
46470fc9 MK |
661 | } |
662 | ||
e0556841 | 663 | return 0; |
7e0d96bc BW |
664 | |
665 | unpin_out: | |
666 | if (ring->id == RCS) | |
ea0c76f8 | 667 | i915_gem_object_ggtt_unpin(to->legacy_hw_ctx.rcs_state); |
7e0d96bc | 668 | return ret; |
e0556841 BW |
669 | } |
670 | ||
671 | /** | |
672 | * i915_switch_context() - perform a GPU context switch. | |
673 | * @ring: ring for which we'll execute the context switch | |
96a6f0f1 | 674 | * @to: the context to switch to |
e0556841 BW |
675 | * |
676 | * The context life cycle is simple. The context refcount is incremented and | |
677 | * decremented by 1 and create and destroy. If the context is in use by the GPU, | |
ecdb5fd8 | 678 | * it will have a refcount > 1. This allows us to destroy the context abstract |
e0556841 | 679 | * object while letting the normal object tracking destroy the backing BO. |
ecdb5fd8 TD |
680 | * |
681 | * This function should not be used in execlists mode. Instead the context is | |
682 | * switched by writing to the ELSP and requests keep a reference to their | |
683 | * context. | |
e0556841 | 684 | */ |
a4872ba6 | 685 | int i915_switch_context(struct intel_engine_cs *ring, |
273497e5 | 686 | struct intel_context *to) |
e0556841 BW |
687 | { |
688 | struct drm_i915_private *dev_priv = ring->dev->dev_private; | |
e0556841 | 689 | |
ecdb5fd8 | 690 | WARN_ON(i915.enable_execlists); |
0eea67eb BW |
691 | WARN_ON(!mutex_is_locked(&dev_priv->dev->struct_mutex)); |
692 | ||
ea0c76f8 | 693 | if (to->legacy_hw_ctx.rcs_state == NULL) { /* We have the fake context */ |
691e6415 CW |
694 | if (to != ring->last_context) { |
695 | i915_gem_context_reference(to); | |
696 | if (ring->last_context) | |
697 | i915_gem_context_unreference(ring->last_context); | |
698 | ring->last_context = to; | |
699 | } | |
c482972a | 700 | return 0; |
a95f6a00 | 701 | } |
c482972a | 702 | |
67e3d297 | 703 | return do_switch(ring, to); |
e0556841 | 704 | } |
84624813 | 705 | |
ec3e9963 | 706 | static bool contexts_enabled(struct drm_device *dev) |
691e6415 | 707 | { |
ec3e9963 | 708 | return i915.enable_execlists || to_i915(dev)->hw_context_size; |
691e6415 CW |
709 | } |
710 | ||
84624813 BW |
711 | int i915_gem_context_create_ioctl(struct drm_device *dev, void *data, |
712 | struct drm_file *file) | |
713 | { | |
84624813 BW |
714 | struct drm_i915_gem_context_create *args = data; |
715 | struct drm_i915_file_private *file_priv = file->driver_priv; | |
273497e5 | 716 | struct intel_context *ctx; |
84624813 BW |
717 | int ret; |
718 | ||
ec3e9963 | 719 | if (!contexts_enabled(dev)) |
5fa8be65 DV |
720 | return -ENODEV; |
721 | ||
84624813 BW |
722 | ret = i915_mutex_lock_interruptible(dev); |
723 | if (ret) | |
724 | return ret; | |
725 | ||
d624d86e | 726 | ctx = i915_gem_create_context(dev, file_priv); |
84624813 | 727 | mutex_unlock(&dev->struct_mutex); |
be636387 DC |
728 | if (IS_ERR(ctx)) |
729 | return PTR_ERR(ctx); | |
84624813 | 730 | |
821d66dd | 731 | args->ctx_id = ctx->user_handle; |
84624813 BW |
732 | DRM_DEBUG_DRIVER("HW context %d created\n", args->ctx_id); |
733 | ||
be636387 | 734 | return 0; |
84624813 BW |
735 | } |
736 | ||
737 | int i915_gem_context_destroy_ioctl(struct drm_device *dev, void *data, | |
738 | struct drm_file *file) | |
739 | { | |
740 | struct drm_i915_gem_context_destroy *args = data; | |
741 | struct drm_i915_file_private *file_priv = file->driver_priv; | |
273497e5 | 742 | struct intel_context *ctx; |
84624813 BW |
743 | int ret; |
744 | ||
821d66dd | 745 | if (args->ctx_id == DEFAULT_CONTEXT_HANDLE) |
c2cf2416 | 746 | return -ENOENT; |
0eea67eb | 747 | |
84624813 BW |
748 | ret = i915_mutex_lock_interruptible(dev); |
749 | if (ret) | |
750 | return ret; | |
751 | ||
752 | ctx = i915_gem_context_get(file_priv, args->ctx_id); | |
72ad5c45 | 753 | if (IS_ERR(ctx)) { |
84624813 | 754 | mutex_unlock(&dev->struct_mutex); |
72ad5c45 | 755 | return PTR_ERR(ctx); |
84624813 BW |
756 | } |
757 | ||
821d66dd | 758 | idr_remove(&ctx->file_priv->context_idr, ctx->user_handle); |
dce3271b | 759 | i915_gem_context_unreference(ctx); |
84624813 BW |
760 | mutex_unlock(&dev->struct_mutex); |
761 | ||
762 | DRM_DEBUG_DRIVER("HW context %d destroyed\n", args->ctx_id); | |
763 | return 0; | |
764 | } |