Rev 5060 | Rev 6084 | Go to most recent revision | Details | Compare with Previous | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
3031 | serge | 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 |
||
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 |
||
4560 | Serge | 76 | * GPU. The GPU has loaded its state already and has stored away the gtt |
3031 | serge | 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 | |||
88 | #include |
||
89 | #include |
||
90 | #include "i915_drv.h" |
||
5354 | serge | 91 | #include "i915_trace.h" |
3031 | serge | 92 | |
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 | */ |
||
5060 | serge | 97 | #define GEN6_CONTEXT_ALIGN (64<<10) |
98 | #define GEN7_CONTEXT_ALIGN 4096 |
||
3031 | serge | 99 | |
5060 | serge | 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 | |||
3031 | serge | 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: |
||
120 | reg = I915_READ(GEN7_CXT_SIZE); |
||
121 | if (IS_HASWELL(dev)) |
||
4104 | Serge | 122 | ret = HSW_CXT_TOTAL_SIZE; |
3031 | serge | 123 | else |
124 | ret = GEN7_CXT_TOTAL_SIZE(reg) * 64; |
||
125 | break; |
||
4560 | Serge | 126 | case 8: |
127 | ret = GEN8_CXT_TOTAL_SIZE; |
||
128 | break; |
||
3031 | serge | 129 | default: |
130 | BUG(); |
||
131 | } |
||
132 | |||
133 | return ret; |
||
134 | } |
||
135 | |||
4104 | Serge | 136 | void i915_gem_context_free(struct kref *ctx_ref) |
3031 | serge | 137 | { |
5060 | serge | 138 | struct intel_context *ctx = container_of(ctx_ref, |
4104 | Serge | 139 | typeof(*ctx), ref); |
3031 | serge | 140 | |
5354 | serge | 141 | trace_i915_context_free(ctx); |
5060 | serge | 142 | |
5354 | serge | 143 | if (i915.enable_execlists) |
144 | intel_lr_context_free(ctx); |
||
145 | |||
146 | i915_ppgtt_put(ctx->ppgtt); |
||
147 | |||
5060 | serge | 148 | if (ctx->legacy_hw_ctx.rcs_state) |
149 | drm_gem_object_unreference(&ctx->legacy_hw_ctx.rcs_state->base); |
||
4560 | Serge | 150 | list_del(&ctx->link); |
3031 | serge | 151 | kfree(ctx); |
152 | } |
||
153 | |||
5354 | serge | 154 | struct drm_i915_gem_object * |
5060 | serge | 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 | |||
184 | static struct intel_context * |
||
185 | __create_hw_context(struct drm_device *dev, |
||
3031 | serge | 186 | struct drm_i915_file_private *file_priv) |
187 | { |
||
188 | struct drm_i915_private *dev_priv = dev->dev_private; |
||
5060 | serge | 189 | struct intel_context *ctx; |
3480 | Serge | 190 | int ret; |
3031 | serge | 191 | |
3243 | Serge | 192 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); |
3031 | serge | 193 | if (ctx == NULL) |
194 | return ERR_PTR(-ENOMEM); |
||
195 | |||
4104 | Serge | 196 | kref_init(&ctx->ref); |
5060 | serge | 197 | list_add_tail(&ctx->link, &dev_priv->context_list); |
3031 | serge | 198 | |
5060 | serge | 199 | if (dev_priv->hw_context_size) { |
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); |
||
3746 | Serge | 204 | goto err_out; |
205 | } |
||
5060 | serge | 206 | ctx->legacy_hw_ctx.rcs_state = obj; |
207 | } |
||
3746 | Serge | 208 | |
3031 | serge | 209 | /* Default context will never have a file_priv */ |
5060 | serge | 210 | if (file_priv != NULL) { |
211 | ret = idr_alloc(&file_priv->context_idr, ctx, |
||
212 | DEFAULT_CONTEXT_HANDLE, 0, GFP_KERNEL); |
||
3480 | Serge | 213 | if (ret < 0) |
3031 | serge | 214 | goto err_out; |
5060 | serge | 215 | } else |
216 | ret = DEFAULT_CONTEXT_HANDLE; |
||
4104 | Serge | 217 | |
218 | ctx->file_priv = file_priv; |
||
5060 | serge | 219 | ctx->user_handle = ret; |
4560 | Serge | 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; |
||
3031 | serge | 224 | |
225 | return ctx; |
||
226 | |||
227 | err_out: |
||
4104 | Serge | 228 | i915_gem_context_unreference(ctx); |
3031 | serge | 229 | return ERR_PTR(ret); |
230 | } |
||
231 | |||
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 | */ |
||
5060 | serge | 237 | static struct intel_context * |
238 | i915_gem_create_context(struct drm_device *dev, |
||
5354 | serge | 239 | struct drm_i915_file_private *file_priv) |
3031 | serge | 240 | { |
5060 | serge | 241 | const bool is_global_default_ctx = file_priv == NULL; |
242 | struct intel_context *ctx; |
||
243 | int ret = 0; |
||
3031 | serge | 244 | |
5060 | serge | 245 | BUG_ON(!mutex_is_locked(&dev->struct_mutex)); |
3031 | serge | 246 | |
5060 | serge | 247 | ctx = __create_hw_context(dev, file_priv); |
3031 | serge | 248 | if (IS_ERR(ctx)) |
5060 | serge | 249 | return ctx; |
3031 | serge | 250 | |
5060 | serge | 251 | if (is_global_default_ctx && ctx->legacy_hw_ctx.rcs_state) { |
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. |
||
3031 | serge | 258 | */ |
5060 | serge | 259 | ret = i915_gem_obj_ggtt_pin(ctx->legacy_hw_ctx.rcs_state, |
260 | get_context_alignment(dev), 0); |
||
4104 | Serge | 261 | if (ret) { |
262 | DRM_DEBUG_DRIVER("Couldn't pin %d\n", ret); |
||
3031 | serge | 263 | goto err_destroy; |
4104 | Serge | 264 | } |
5060 | serge | 265 | } |
3031 | serge | 266 | |
5354 | serge | 267 | if (USES_FULL_PPGTT(dev)) { |
268 | struct i915_hw_ppgtt *ppgtt = i915_ppgtt_create(dev, file_priv); |
||
5060 | serge | 269 | |
270 | if (IS_ERR_OR_NULL(ppgtt)) { |
||
271 | DRM_DEBUG_DRIVER("PPGTT setup failed (%ld)\n", |
||
272 | PTR_ERR(ppgtt)); |
||
273 | ret = PTR_ERR(ppgtt); |
||
274 | goto err_unpin; |
||
5354 | serge | 275 | } |
5060 | serge | 276 | |
5354 | serge | 277 | ctx->ppgtt = ppgtt; |
4104 | Serge | 278 | } |
3031 | serge | 279 | |
5354 | serge | 280 | trace_i915_context_create(ctx); |
4560 | Serge | 281 | |
5060 | serge | 282 | return ctx; |
3031 | serge | 283 | |
284 | err_unpin: |
||
5060 | serge | 285 | if (is_global_default_ctx && ctx->legacy_hw_ctx.rcs_state) |
286 | i915_gem_object_ggtt_unpin(ctx->legacy_hw_ctx.rcs_state); |
||
3031 | serge | 287 | err_destroy: |
4104 | Serge | 288 | i915_gem_context_unreference(ctx); |
5060 | serge | 289 | return ERR_PTR(ret); |
3031 | serge | 290 | } |
291 | |||
5060 | serge | 292 | void i915_gem_context_reset(struct drm_device *dev) |
293 | { |
||
294 | struct drm_i915_private *dev_priv = dev->dev_private; |
||
295 | int i; |
||
296 | |||
5354 | serge | 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 | |||
5060 | serge | 303 | for (i = 0; i < I915_NUM_RINGS; i++) { |
304 | struct intel_engine_cs *ring = &dev_priv->ring[i]; |
||
305 | struct intel_context *lctx = ring->last_context; |
||
306 | |||
5354 | serge | 307 | if (lctx) { |
5060 | serge | 308 | if (lctx->legacy_hw_ctx.rcs_state && i == RCS) |
309 | i915_gem_object_ggtt_unpin(lctx->legacy_hw_ctx.rcs_state); |
||
310 | |||
311 | i915_gem_context_unreference(lctx); |
||
5354 | serge | 312 | ring->last_context = NULL; |
313 | } |
||
5060 | serge | 314 | } |
315 | } |
||
316 | |||
4560 | Serge | 317 | int i915_gem_context_init(struct drm_device *dev) |
3031 | serge | 318 | { |
319 | struct drm_i915_private *dev_priv = dev->dev_private; |
||
5060 | serge | 320 | struct intel_context *ctx; |
321 | int i; |
||
3031 | serge | 322 | |
5060 | serge | 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)) |
||
4560 | Serge | 326 | return 0; |
3031 | serge | 327 | |
5354 | serge | 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)) { |
||
3480 | Serge | 333 | dev_priv->hw_context_size = round_up(get_context_size(dev), 4096); |
334 | if (dev_priv->hw_context_size > (1<<20)) { |
||
5060 | serge | 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 | } |
||
3031 | serge | 339 | } |
340 | |||
5354 | serge | 341 | ctx = i915_gem_create_context(dev, NULL); |
5060 | serge | 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); |
||
3031 | serge | 346 | } |
347 | |||
5354 | serge | 348 | for (i = 0; i < I915_NUM_RINGS; i++) { |
349 | struct intel_engine_cs *ring = &dev_priv->ring[i]; |
||
350 | |||
5060 | serge | 351 | /* NB: RCS will hold a ref for all rings */ |
5354 | serge | 352 | ring->default_context = ctx; |
353 | } |
||
5060 | serge | 354 | |
5354 | serge | 355 | DRM_DEBUG_DRIVER("%s context support initialized\n", |
356 | i915.enable_execlists ? "LR" : |
||
357 | dev_priv->hw_context_size ? "HW" : "fake"); |
||
4560 | Serge | 358 | return 0; |
3031 | serge | 359 | } |
360 | |||
361 | void i915_gem_context_fini(struct drm_device *dev) |
||
362 | { |
||
363 | struct drm_i915_private *dev_priv = dev->dev_private; |
||
5060 | serge | 364 | struct intel_context *dctx = dev_priv->ring[RCS].default_context; |
365 | int i; |
||
3031 | serge | 366 | |
5060 | serge | 367 | if (dctx->legacy_hw_ctx.rcs_state) { |
3031 | serge | 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. */ |
||
4104 | Serge | 371 | intel_gpu_reset(dev); |
3031 | serge | 372 | |
4104 | Serge | 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. */ |
||
4560 | Serge | 379 | WARN_ON(!dev_priv->ring[RCS].last_context); |
380 | if (dev_priv->ring[RCS].last_context == dctx) { |
||
381 | /* Fake switch to NULL context */ |
||
5060 | serge | 382 | WARN_ON(dctx->legacy_hw_ctx.rcs_state->active); |
383 | i915_gem_object_ggtt_unpin(dctx->legacy_hw_ctx.rcs_state); |
||
4246 | Serge | 384 | i915_gem_context_unreference(dctx); |
5060 | serge | 385 | dev_priv->ring[RCS].last_context = NULL; |
4560 | Serge | 386 | } |
387 | |||
5060 | serge | 388 | i915_gem_object_ggtt_unpin(dctx->legacy_hw_ctx.rcs_state); |
389 | } |
||
390 | |||
391 | for (i = 0; i < I915_NUM_RINGS; i++) { |
||
392 | struct intel_engine_cs *ring = &dev_priv->ring[i]; |
||
393 | |||
394 | if (ring->last_context) |
||
395 | i915_gem_context_unreference(ring->last_context); |
||
396 | |||
397 | ring->default_context = NULL; |
||
398 | ring->last_context = NULL; |
||
399 | } |
||
400 | |||
4560 | Serge | 401 | i915_gem_context_unreference(dctx); |
3031 | serge | 402 | } |
403 | |||
5060 | serge | 404 | int i915_gem_context_enable(struct drm_i915_private *dev_priv) |
405 | { |
||
406 | struct intel_engine_cs *ring; |
||
407 | int ret, i; |
||
408 | |||
5354 | serge | 409 | BUG_ON(!dev_priv->ring[RCS].default_context); |
5060 | serge | 410 | |
5354 | serge | 411 | if (i915.enable_execlists) |
5060 | serge | 412 | return 0; |
413 | |||
414 | for_each_ring(ring, dev_priv, i) { |
||
415 | ret = i915_switch_context(ring, ring->default_context); |
||
416 | if (ret) |
||
417 | return ret; |
||
418 | } |
||
419 | |||
420 | return 0; |
||
421 | } |
||
422 | |||
3031 | serge | 423 | static int context_idr_cleanup(int id, void *p, void *data) |
424 | { |
||
5060 | serge | 425 | struct intel_context *ctx = p; |
3031 | serge | 426 | |
4280 | Serge | 427 | i915_gem_context_unreference(ctx); |
428 | return 0; |
||
429 | } |
||
3031 | serge | 430 | |
5060 | serge | 431 | int i915_gem_context_open(struct drm_device *dev, struct drm_file *file) |
4280 | Serge | 432 | { |
433 | struct drm_i915_file_private *file_priv = file->driver_priv; |
||
5060 | serge | 434 | struct intel_context *ctx; |
4104 | Serge | 435 | |
5060 | serge | 436 | idr_init(&file_priv->context_idr); |
4280 | Serge | 437 | |
5060 | serge | 438 | mutex_lock(&dev->struct_mutex); |
5354 | serge | 439 | ctx = i915_gem_create_context(dev, file_priv); |
5060 | serge | 440 | mutex_unlock(&dev->struct_mutex); |
4560 | Serge | 441 | |
5060 | serge | 442 | if (IS_ERR(ctx)) { |
443 | idr_destroy(&file_priv->context_idr); |
||
444 | return PTR_ERR(ctx); |
||
445 | } |
||
4280 | Serge | 446 | |
5060 | serge | 447 | return 0; |
3031 | serge | 448 | } |
449 | |||
450 | void i915_gem_context_close(struct drm_device *dev, struct drm_file *file) |
||
451 | { |
||
452 | struct drm_i915_file_private *file_priv = file->driver_priv; |
||
453 | |||
4246 | Serge | 454 | idr_for_each(&file_priv->context_idr, context_idr_cleanup, NULL); |
3031 | serge | 455 | idr_destroy(&file_priv->context_idr); |
456 | } |
||
457 | |||
5060 | serge | 458 | struct intel_context * |
3031 | serge | 459 | i915_gem_context_get(struct drm_i915_file_private *file_priv, u32 id) |
460 | { |
||
5060 | serge | 461 | struct intel_context *ctx; |
462 | |||
463 | ctx = (struct intel_context *)idr_find(&file_priv->context_idr, id); |
||
464 | if (!ctx) |
||
465 | return ERR_PTR(-ENOENT); |
||
466 | |||
467 | return ctx; |
||
3031 | serge | 468 | } |
469 | |||
470 | static inline int |
||
5060 | serge | 471 | mi_set_context(struct intel_engine_cs *ring, |
472 | struct intel_context *new_context, |
||
3031 | serge | 473 | u32 hw_flags) |
474 | { |
||
5354 | serge | 475 | u32 flags = hw_flags | MI_MM_SPACE_GTT; |
476 | const int num_rings = |
||
477 | /* Use an extended w/a on ivb+ if signalling from other rings */ |
||
478 | i915_semaphore_is_enabled(ring->dev) ? |
||
479 | hweight32(INTEL_INFO(ring->dev)->ring_mask) - 1 : |
||
480 | 0; |
||
481 | int len, i, ret; |
||
3031 | serge | 482 | |
483 | /* w/a: If Flush TLB Invalidation Mode is enabled, driver must do a TLB |
||
484 | * invalidation prior to MI_SET_CONTEXT. On GEN6 we don't set the value |
||
485 | * explicitly, so we rely on the value at ring init, stored in |
||
486 | * itlb_before_ctx_switch. |
||
487 | */ |
||
5060 | serge | 488 | if (IS_GEN6(ring->dev)) { |
3031 | serge | 489 | ret = ring->flush(ring, I915_GEM_GPU_DOMAINS, 0); |
490 | if (ret) |
||
491 | return ret; |
||
492 | } |
||
493 | |||
5354 | serge | 494 | /* These flags are for resource streamer on HSW+ */ |
495 | if (!IS_HASWELL(ring->dev) && INTEL_INFO(ring->dev)->gen < 8) |
||
496 | flags |= (MI_SAVE_EXT_STATE_EN | MI_RESTORE_EXT_STATE_EN); |
||
497 | |||
498 | |||
499 | len = 4; |
||
500 | if (INTEL_INFO(ring->dev)->gen >= 7) |
||
501 | len += 2 + (num_rings ? 4*num_rings + 2 : 0); |
||
502 | |||
503 | ret = intel_ring_begin(ring, len); |
||
3031 | serge | 504 | if (ret) |
505 | return ret; |
||
506 | |||
5060 | serge | 507 | /* WaProgramMiArbOnOffAroundMiSetContext:ivb,vlv,hsw,bdw,chv */ |
5354 | serge | 508 | if (INTEL_INFO(ring->dev)->gen >= 7) { |
3031 | serge | 509 | intel_ring_emit(ring, MI_ARB_ON_OFF | MI_ARB_DISABLE); |
5354 | serge | 510 | if (num_rings) { |
511 | struct intel_engine_cs *signaller; |
||
3031 | serge | 512 | |
5354 | serge | 513 | intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(num_rings)); |
514 | for_each_ring(signaller, to_i915(ring->dev), i) { |
||
515 | if (signaller == ring) |
||
516 | continue; |
||
517 | |||
518 | intel_ring_emit(ring, RING_PSMI_CTL(signaller->mmio_base)); |
||
519 | intel_ring_emit(ring, _MASKED_BIT_ENABLE(GEN6_PSMI_SLEEP_MSG_DISABLE)); |
||
520 | } |
||
521 | } |
||
522 | } |
||
523 | |||
3031 | serge | 524 | intel_ring_emit(ring, MI_NOOP); |
525 | intel_ring_emit(ring, MI_SET_CONTEXT); |
||
5060 | serge | 526 | intel_ring_emit(ring, i915_gem_obj_ggtt_offset(new_context->legacy_hw_ctx.rcs_state) | |
5354 | serge | 527 | flags); |
5060 | serge | 528 | /* |
529 | * w/a: MI_SET_CONTEXT must always be followed by MI_NOOP |
||
530 | * WaMiSetContext_Hang:snb,ivb,vlv |
||
531 | */ |
||
3031 | serge | 532 | intel_ring_emit(ring, MI_NOOP); |
533 | |||
5354 | serge | 534 | if (INTEL_INFO(ring->dev)->gen >= 7) { |
535 | if (num_rings) { |
||
536 | struct intel_engine_cs *signaller; |
||
537 | |||
538 | intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(num_rings)); |
||
539 | for_each_ring(signaller, to_i915(ring->dev), i) { |
||
540 | if (signaller == ring) |
||
541 | continue; |
||
542 | |||
543 | intel_ring_emit(ring, RING_PSMI_CTL(signaller->mmio_base)); |
||
544 | intel_ring_emit(ring, _MASKED_BIT_DISABLE(GEN6_PSMI_SLEEP_MSG_DISABLE)); |
||
545 | } |
||
546 | } |
||
3031 | serge | 547 | intel_ring_emit(ring, MI_ARB_ON_OFF | MI_ARB_ENABLE); |
5354 | serge | 548 | } |
3031 | serge | 549 | |
550 | intel_ring_advance(ring); |
||
551 | |||
552 | return ret; |
||
553 | } |
||
554 | |||
5060 | serge | 555 | static int do_switch(struct intel_engine_cs *ring, |
556 | struct intel_context *to) |
||
3031 | serge | 557 | { |
5060 | serge | 558 | struct drm_i915_private *dev_priv = ring->dev->dev_private; |
559 | struct intel_context *from = ring->last_context; |
||
3031 | serge | 560 | u32 hw_flags = 0; |
5060 | serge | 561 | bool uninitialized = false; |
5354 | serge | 562 | struct i915_vma *vma; |
4560 | Serge | 563 | int ret, i; |
3031 | serge | 564 | |
5060 | serge | 565 | if (from != NULL && ring == &dev_priv->ring[RCS]) { |
566 | BUG_ON(from->legacy_hw_ctx.rcs_state == NULL); |
||
567 | BUG_ON(!i915_gem_obj_is_pinned(from->legacy_hw_ctx.rcs_state)); |
||
568 | } |
||
3031 | serge | 569 | |
4560 | Serge | 570 | if (from == to && !to->remap_slice) |
3031 | serge | 571 | return 0; |
572 | |||
5060 | serge | 573 | /* Trying to pin first makes error handling easier. */ |
574 | if (ring == &dev_priv->ring[RCS]) { |
||
575 | ret = i915_gem_obj_ggtt_pin(to->legacy_hw_ctx.rcs_state, |
||
576 | get_context_alignment(ring->dev), 0); |
||
3031 | serge | 577 | if (ret) |
578 | return ret; |
||
5060 | serge | 579 | } |
3031 | serge | 580 | |
4539 | Serge | 581 | /* |
582 | * Pin can switch back to the default context if we end up calling into |
||
583 | * evict_everything - as a last ditch gtt defrag effort that also |
||
584 | * switches to the default context. Hence we need to reload from here. |
||
585 | */ |
||
586 | from = ring->last_context; |
||
587 | |||
5354 | serge | 588 | if (to->ppgtt) { |
589 | trace_switch_mm(ring, to); |
||
590 | ret = to->ppgtt->switch_mm(to->ppgtt, ring); |
||
5060 | serge | 591 | if (ret) |
592 | goto unpin_out; |
||
593 | } |
||
594 | |||
595 | if (ring != &dev_priv->ring[RCS]) { |
||
596 | if (from) |
||
597 | i915_gem_context_unreference(from); |
||
598 | goto done; |
||
599 | } |
||
600 | |||
4539 | Serge | 601 | /* |
602 | * Clear this page out of any CPU caches for coherent swap-in/out. Note |
||
3031 | serge | 603 | * that thanks to write = false in this call and us not setting any gpu |
604 | * write domains when putting a context object onto the active list |
||
605 | * (when switching away from it), this won't block. |
||
4539 | Serge | 606 | * |
607 | * XXX: We need a real interface to do this instead of trickery. |
||
608 | */ |
||
5060 | serge | 609 | ret = i915_gem_object_set_to_gtt_domain(to->legacy_hw_ctx.rcs_state, false); |
610 | if (ret) |
||
611 | goto unpin_out; |
||
612 | |||
5354 | serge | 613 | vma = i915_gem_obj_to_ggtt(to->legacy_hw_ctx.rcs_state); |
614 | if (!(vma->bound & GLOBAL_BIND)) |
||
615 | vma->bind_vma(vma, to->legacy_hw_ctx.rcs_state->cache_level, |
||
616 | GLOBAL_BIND); |
||
3031 | serge | 617 | |
5060 | serge | 618 | if (!to->legacy_hw_ctx.initialized || i915_gem_context_is_default(to)) |
3031 | serge | 619 | hw_flags |= MI_RESTORE_INHIBIT; |
620 | |||
621 | ret = mi_set_context(ring, to, hw_flags); |
||
5060 | serge | 622 | if (ret) |
623 | goto unpin_out; |
||
3031 | serge | 624 | |
4560 | Serge | 625 | for (i = 0; i < MAX_L3_SLICES; i++) { |
626 | if (!(to->remap_slice & (1< |
||
627 | continue; |
||
628 | |||
629 | ret = i915_gem_l3_remap(ring, i); |
||
630 | /* If it failed, try again next round */ |
||
631 | if (ret) |
||
632 | DRM_DEBUG_DRIVER("L3 remapping failed\n"); |
||
633 | else |
||
634 | to->remap_slice &= ~(1< |
||
635 | } |
||
636 | |||
3031 | serge | 637 | /* The backing object for the context is done after switching to the |
638 | * *next* context. Therefore we cannot retire the previous context until |
||
639 | * the next context has already started running. In fact, the below code |
||
640 | * is a bit suboptimal because the retiring can occur simply after the |
||
641 | * MI_SET_CONTEXT instead of when the next seqno has completed. |
||
642 | */ |
||
4104 | Serge | 643 | if (from != NULL) { |
5060 | serge | 644 | from->legacy_hw_ctx.rcs_state->base.read_domains = I915_GEM_DOMAIN_INSTRUCTION; |
645 | i915_vma_move_to_active(i915_gem_obj_to_ggtt(from->legacy_hw_ctx.rcs_state), ring); |
||
3031 | serge | 646 | /* As long as MI_SET_CONTEXT is serializing, ie. it flushes the |
647 | * whole damn pipeline, we don't need to explicitly mark the |
||
648 | * object dirty. The only exception is that the context must be |
||
649 | * correct in case the object gets swapped out. Ideally we'd be |
||
650 | * able to defer doing this until we know the object would be |
||
651 | * swapped, but there is no way to do that yet. |
||
652 | */ |
||
5060 | serge | 653 | from->legacy_hw_ctx.rcs_state->dirty = 1; |
654 | BUG_ON(from->legacy_hw_ctx.rcs_state->ring != ring); |
||
3031 | serge | 655 | |
4560 | Serge | 656 | /* obj is kept alive until the next request by its active ref */ |
5060 | serge | 657 | i915_gem_object_ggtt_unpin(from->legacy_hw_ctx.rcs_state); |
4104 | Serge | 658 | i915_gem_context_unreference(from); |
3031 | serge | 659 | } |
660 | |||
5060 | serge | 661 | uninitialized = !to->legacy_hw_ctx.initialized && from == NULL; |
662 | to->legacy_hw_ctx.initialized = true; |
||
663 | |||
664 | done: |
||
4104 | Serge | 665 | i915_gem_context_reference(to); |
666 | ring->last_context = to; |
||
3031 | serge | 667 | |
5060 | serge | 668 | if (uninitialized) { |
5354 | serge | 669 | if (ring->init_context) { |
670 | ret = ring->init_context(ring, to); |
||
671 | if (ret) |
||
672 | DRM_ERROR("ring init context: %d\n", ret); |
||
673 | } |
||
674 | |||
5060 | serge | 675 | ret = i915_gem_render_state_init(ring); |
676 | if (ret) |
||
677 | DRM_ERROR("init render state: %d\n", ret); |
||
678 | } |
||
679 | |||
3031 | serge | 680 | return 0; |
5060 | serge | 681 | |
682 | unpin_out: |
||
683 | if (ring->id == RCS) |
||
684 | i915_gem_object_ggtt_unpin(to->legacy_hw_ctx.rcs_state); |
||
685 | return ret; |
||
3031 | serge | 686 | } |
687 | |||
688 | /** |
||
689 | * i915_switch_context() - perform a GPU context switch. |
||
690 | * @ring: ring for which we'll execute the context switch |
||
5060 | serge | 691 | * @to: the context to switch to |
3031 | serge | 692 | * |
693 | * The context life cycle is simple. The context refcount is incremented and |
||
694 | * decremented by 1 and create and destroy. If the context is in use by the GPU, |
||
5354 | serge | 695 | * it will have a refcount > 1. This allows us to destroy the context abstract |
3031 | serge | 696 | * object while letting the normal object tracking destroy the backing BO. |
5354 | serge | 697 | * |
698 | * This function should not be used in execlists mode. Instead the context is |
||
699 | * switched by writing to the ELSP and requests keep a reference to their |
||
700 | * context. |
||
3031 | serge | 701 | */ |
5060 | serge | 702 | int i915_switch_context(struct intel_engine_cs *ring, |
703 | struct intel_context *to) |
||
3031 | serge | 704 | { |
705 | struct drm_i915_private *dev_priv = ring->dev->dev_private; |
||
706 | |||
5354 | serge | 707 | WARN_ON(i915.enable_execlists); |
4104 | Serge | 708 | WARN_ON(!mutex_is_locked(&dev_priv->dev->struct_mutex)); |
709 | |||
5060 | serge | 710 | if (to->legacy_hw_ctx.rcs_state == NULL) { /* We have the fake context */ |
711 | if (to != ring->last_context) { |
||
712 | i915_gem_context_reference(to); |
||
713 | if (ring->last_context) |
||
714 | i915_gem_context_unreference(ring->last_context); |
||
715 | ring->last_context = to; |
||
716 | } |
||
3031 | serge | 717 | return 0; |
718 | } |
||
719 | |||
5060 | serge | 720 | return do_switch(ring, to); |
3031 | serge | 721 | } |
722 | |||
5354 | serge | 723 | static bool contexts_enabled(struct drm_device *dev) |
5060 | serge | 724 | { |
5354 | serge | 725 | return i915.enable_execlists || to_i915(dev)->hw_context_size; |
5060 | serge | 726 | } |
727 | |||
3031 | serge | 728 | int i915_gem_context_create_ioctl(struct drm_device *dev, void *data, |
729 | struct drm_file *file) |
||
730 | { |
||
731 | struct drm_i915_gem_context_create *args = data; |
||
732 | struct drm_i915_file_private *file_priv = file->driver_priv; |
||
5060 | serge | 733 | struct intel_context *ctx; |
3031 | serge | 734 | int ret; |
735 | |||
5354 | serge | 736 | if (!contexts_enabled(dev)) |
3031 | serge | 737 | return -ENODEV; |
738 | |||
739 | ret = i915_mutex_lock_interruptible(dev); |
||
740 | if (ret) |
||
741 | return ret; |
||
742 | |||
5354 | serge | 743 | ctx = i915_gem_create_context(dev, file_priv); |
3031 | serge | 744 | mutex_unlock(&dev->struct_mutex); |
745 | if (IS_ERR(ctx)) |
||
746 | return PTR_ERR(ctx); |
||
747 | |||
5060 | serge | 748 | args->ctx_id = ctx->user_handle; |
3031 | serge | 749 | DRM_DEBUG_DRIVER("HW context %d created\n", args->ctx_id); |
750 | |||
751 | return 0; |
||
752 | } |
||
753 | |||
754 | int i915_gem_context_destroy_ioctl(struct drm_device *dev, void *data, |
||
755 | struct drm_file *file) |
||
756 | { |
||
757 | struct drm_i915_gem_context_destroy *args = data; |
||
758 | struct drm_i915_file_private *file_priv = file->driver_priv; |
||
5060 | serge | 759 | struct intel_context *ctx; |
3031 | serge | 760 | int ret; |
761 | |||
5060 | serge | 762 | if (args->ctx_id == DEFAULT_CONTEXT_HANDLE) |
763 | return -ENOENT; |
||
3031 | serge | 764 | |
765 | ret = i915_mutex_lock_interruptible(dev); |
||
766 | if (ret) |
||
767 | return ret; |
||
768 | |||
769 | ctx = i915_gem_context_get(file_priv, args->ctx_id); |
||
5060 | serge | 770 | if (IS_ERR(ctx)) { |
3031 | serge | 771 | mutex_unlock(&dev->struct_mutex); |
5060 | serge | 772 | return PTR_ERR(ctx); |
3031 | serge | 773 | } |
774 | |||
5060 | serge | 775 | idr_remove(&ctx->file_priv->context_idr, ctx->user_handle); |
4246 | Serge | 776 | i915_gem_context_unreference(ctx); |
3031 | serge | 777 | mutex_unlock(&dev->struct_mutex); |
778 | |||
779 | DRM_DEBUG_DRIVER("HW context %d destroyed\n", args->ctx_id); |
||
780 | return 0; |
||
781 | } |