]>
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 | { |
9ea4feec | 138 | struct intel_context *ctx = container_of(ctx_ref, typeof(*ctx), ref); |
40521054 | 139 | |
198c974d DCS |
140 | trace_i915_context_free(ctx); |
141 | ||
ae6c4806 | 142 | if (i915.enable_execlists) |
ede7d42b | 143 | intel_lr_context_free(ctx); |
c7c48dfd | 144 | |
ae6c4806 DV |
145 | i915_ppgtt_put(ctx->ppgtt); |
146 | ||
2f295791 BW |
147 | if (ctx->legacy_hw_ctx.rcs_state) |
148 | drm_gem_object_unreference(&ctx->legacy_hw_ctx.rcs_state->base); | |
c7c48dfd | 149 | list_del(&ctx->link); |
40521054 BW |
150 | kfree(ctx); |
151 | } | |
152 | ||
8c857917 | 153 | struct drm_i915_gem_object * |
aa0c13da OM |
154 | i915_gem_alloc_context_obj(struct drm_device *dev, size_t size) |
155 | { | |
156 | struct drm_i915_gem_object *obj; | |
157 | int ret; | |
158 | ||
52613921 | 159 | obj = i915_gem_alloc_object(dev, size); |
aa0c13da OM |
160 | if (obj == NULL) |
161 | return ERR_PTR(-ENOMEM); | |
162 | ||
163 | /* | |
164 | * Try to make the context utilize L3 as well as LLC. | |
165 | * | |
166 | * On VLV we don't have L3 controls in the PTEs so we | |
167 | * shouldn't touch the cache level, especially as that | |
168 | * would make the object snooped which might have a | |
169 | * negative performance impact. | |
170 | */ | |
171 | if (INTEL_INFO(dev)->gen >= 7 && !IS_VALLEYVIEW(dev)) { | |
172 | ret = i915_gem_object_set_cache_level(obj, I915_CACHE_L3_LLC); | |
173 | /* Failure shouldn't ever happen this early */ | |
174 | if (WARN_ON(ret)) { | |
175 | drm_gem_object_unreference(&obj->base); | |
176 | return ERR_PTR(ret); | |
177 | } | |
178 | } | |
179 | ||
180 | return obj; | |
181 | } | |
182 | ||
273497e5 | 183 | static struct intel_context * |
0eea67eb | 184 | __create_hw_context(struct drm_device *dev, |
ee960be7 | 185 | struct drm_i915_file_private *file_priv) |
40521054 BW |
186 | { |
187 | struct drm_i915_private *dev_priv = dev->dev_private; | |
273497e5 | 188 | struct intel_context *ctx; |
c8c470af | 189 | int ret; |
40521054 | 190 | |
f94982b0 | 191 | ctx = kzalloc(sizeof(*ctx), GFP_KERNEL); |
146937e5 BW |
192 | if (ctx == NULL) |
193 | return ERR_PTR(-ENOMEM); | |
40521054 | 194 | |
dce3271b | 195 | kref_init(&ctx->ref); |
691e6415 | 196 | list_add_tail(&ctx->link, &dev_priv->context_list); |
9ea4feec | 197 | ctx->i915 = dev_priv; |
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 | ||
3e5b6f05 TD |
299 | if (i915.enable_execlists) { |
300 | struct intel_context *ctx; | |
301 | ||
302 | list_for_each_entry(ctx, &dev_priv->context_list, link) { | |
303 | intel_lr_context_reset(dev, ctx); | |
304 | } | |
305 | ||
ecdb5fd8 | 306 | return; |
3e5b6f05 | 307 | } |
ecdb5fd8 | 308 | |
acce9ffa | 309 | for (i = 0; i < I915_NUM_RINGS; i++) { |
a4872ba6 | 310 | struct intel_engine_cs *ring = &dev_priv->ring[i]; |
ea0c76f8 | 311 | struct intel_context *lctx = ring->last_context; |
acce9ffa | 312 | |
6689c167 MA |
313 | if (lctx) { |
314 | if (lctx->legacy_hw_ctx.rcs_state && i == RCS) | |
315 | i915_gem_object_ggtt_unpin(lctx->legacy_hw_ctx.rcs_state); | |
acce9ffa | 316 | |
6689c167 MA |
317 | i915_gem_context_unreference(lctx); |
318 | ring->last_context = NULL; | |
acce9ffa | 319 | } |
acce9ffa BW |
320 | } |
321 | } | |
322 | ||
8245be31 | 323 | int i915_gem_context_init(struct drm_device *dev) |
254f965c BW |
324 | { |
325 | struct drm_i915_private *dev_priv = dev->dev_private; | |
273497e5 | 326 | struct intel_context *ctx; |
a45d0f6a | 327 | int i; |
254f965c | 328 | |
2fa48d8d BW |
329 | /* Init should only be called once per module load. Eventually the |
330 | * restriction on the context_disabled check can be loosened. */ | |
331 | if (WARN_ON(dev_priv->ring[RCS].default_context)) | |
8245be31 | 332 | return 0; |
254f965c | 333 | |
ede7d42b OM |
334 | if (i915.enable_execlists) { |
335 | /* NB: intentionally left blank. We will allocate our own | |
336 | * backing objects as we need them, thank you very much */ | |
337 | dev_priv->hw_context_size = 0; | |
338 | } else if (HAS_HW_CONTEXTS(dev)) { | |
691e6415 CW |
339 | dev_priv->hw_context_size = round_up(get_context_size(dev), 4096); |
340 | if (dev_priv->hw_context_size > (1<<20)) { | |
341 | DRM_DEBUG_DRIVER("Disabling HW Contexts; invalid size %d\n", | |
342 | dev_priv->hw_context_size); | |
343 | dev_priv->hw_context_size = 0; | |
344 | } | |
254f965c BW |
345 | } |
346 | ||
d624d86e | 347 | ctx = i915_gem_create_context(dev, NULL); |
691e6415 CW |
348 | if (IS_ERR(ctx)) { |
349 | DRM_ERROR("Failed to create default global context (error %ld)\n", | |
350 | PTR_ERR(ctx)); | |
351 | return PTR_ERR(ctx); | |
254f965c BW |
352 | } |
353 | ||
ede7d42b OM |
354 | for (i = 0; i < I915_NUM_RINGS; i++) { |
355 | struct intel_engine_cs *ring = &dev_priv->ring[i]; | |
356 | ||
357 | /* NB: RCS will hold a ref for all rings */ | |
358 | ring->default_context = ctx; | |
ede7d42b | 359 | } |
67e3d297 | 360 | |
ede7d42b OM |
361 | DRM_DEBUG_DRIVER("%s context support initialized\n", |
362 | i915.enable_execlists ? "LR" : | |
363 | dev_priv->hw_context_size ? "HW" : "fake"); | |
8245be31 | 364 | return 0; |
254f965c BW |
365 | } |
366 | ||
367 | void i915_gem_context_fini(struct drm_device *dev) | |
368 | { | |
369 | struct drm_i915_private *dev_priv = dev->dev_private; | |
273497e5 | 370 | struct intel_context *dctx = dev_priv->ring[RCS].default_context; |
67e3d297 | 371 | int i; |
254f965c | 372 | |
ea0c76f8 | 373 | if (dctx->legacy_hw_ctx.rcs_state) { |
691e6415 CW |
374 | /* The only known way to stop the gpu from accessing the hw context is |
375 | * to reset it. Do this as the very last operation to avoid confusing | |
376 | * other code, leading to spurious errors. */ | |
377 | intel_gpu_reset(dev); | |
378 | ||
379 | /* When default context is created and switched to, base object refcount | |
380 | * will be 2 (+1 from object creation and +1 from do_switch()). | |
381 | * i915_gem_context_fini() will be called after gpu_idle() has switched | |
382 | * to default context. So we need to unreference the base object once | |
383 | * to offset the do_switch part, so that i915_gem_context_unreference() | |
384 | * can then free the base object correctly. */ | |
385 | WARN_ON(!dev_priv->ring[RCS].last_context); | |
386 | if (dev_priv->ring[RCS].last_context == dctx) { | |
387 | /* Fake switch to NULL context */ | |
ea0c76f8 OM |
388 | WARN_ON(dctx->legacy_hw_ctx.rcs_state->active); |
389 | i915_gem_object_ggtt_unpin(dctx->legacy_hw_ctx.rcs_state); | |
691e6415 CW |
390 | i915_gem_context_unreference(dctx); |
391 | dev_priv->ring[RCS].last_context = NULL; | |
392 | } | |
d3b448d9 | 393 | |
ea0c76f8 | 394 | i915_gem_object_ggtt_unpin(dctx->legacy_hw_ctx.rcs_state); |
67e3d297 BW |
395 | } |
396 | ||
397 | for (i = 0; i < I915_NUM_RINGS; i++) { | |
a4872ba6 | 398 | struct intel_engine_cs *ring = &dev_priv->ring[i]; |
67e3d297 BW |
399 | |
400 | if (ring->last_context) | |
401 | i915_gem_context_unreference(ring->last_context); | |
402 | ||
403 | ring->default_context = NULL; | |
0009e46c | 404 | ring->last_context = NULL; |
71b76d00 BW |
405 | } |
406 | ||
dce3271b | 407 | i915_gem_context_unreference(dctx); |
254f965c BW |
408 | } |
409 | ||
b3dd6b96 | 410 | int i915_gem_context_enable(struct drm_i915_gem_request *req) |
2fa48d8d | 411 | { |
b3dd6b96 | 412 | struct intel_engine_cs *ring = req->ring; |
90638cc1 | 413 | int ret; |
bdf4fd7e | 414 | |
e7778be1 | 415 | if (i915.enable_execlists) { |
90638cc1 JH |
416 | if (ring->init_context == NULL) |
417 | return 0; | |
ecdb5fd8 | 418 | |
8753181e | 419 | ret = ring->init_context(req); |
e7778be1 | 420 | } else |
ba01cc93 | 421 | ret = i915_switch_context(req); |
90638cc1 JH |
422 | |
423 | if (ret) { | |
424 | DRM_ERROR("ring init context: %d\n", ret); | |
425 | return ret; | |
426 | } | |
2fa48d8d BW |
427 | |
428 | return 0; | |
429 | } | |
430 | ||
40521054 BW |
431 | static int context_idr_cleanup(int id, void *p, void *data) |
432 | { | |
273497e5 | 433 | struct intel_context *ctx = p; |
40521054 | 434 | |
dce3271b | 435 | i915_gem_context_unreference(ctx); |
40521054 | 436 | return 0; |
254f965c BW |
437 | } |
438 | ||
e422b888 BW |
439 | int i915_gem_context_open(struct drm_device *dev, struct drm_file *file) |
440 | { | |
441 | struct drm_i915_file_private *file_priv = file->driver_priv; | |
f83d6518 | 442 | struct intel_context *ctx; |
e422b888 BW |
443 | |
444 | idr_init(&file_priv->context_idr); | |
445 | ||
0eea67eb | 446 | mutex_lock(&dev->struct_mutex); |
d624d86e | 447 | ctx = i915_gem_create_context(dev, file_priv); |
0eea67eb BW |
448 | mutex_unlock(&dev->struct_mutex); |
449 | ||
f83d6518 | 450 | if (IS_ERR(ctx)) { |
0eea67eb | 451 | idr_destroy(&file_priv->context_idr); |
f83d6518 | 452 | return PTR_ERR(ctx); |
0eea67eb BW |
453 | } |
454 | ||
e422b888 BW |
455 | return 0; |
456 | } | |
457 | ||
254f965c BW |
458 | void i915_gem_context_close(struct drm_device *dev, struct drm_file *file) |
459 | { | |
40521054 | 460 | struct drm_i915_file_private *file_priv = file->driver_priv; |
254f965c | 461 | |
73c273eb | 462 | idr_for_each(&file_priv->context_idr, context_idr_cleanup, NULL); |
40521054 | 463 | idr_destroy(&file_priv->context_idr); |
40521054 BW |
464 | } |
465 | ||
273497e5 | 466 | struct intel_context * |
40521054 BW |
467 | i915_gem_context_get(struct drm_i915_file_private *file_priv, u32 id) |
468 | { | |
273497e5 | 469 | struct intel_context *ctx; |
72ad5c45 | 470 | |
273497e5 | 471 | ctx = (struct intel_context *)idr_find(&file_priv->context_idr, id); |
72ad5c45 BW |
472 | if (!ctx) |
473 | return ERR_PTR(-ENOENT); | |
474 | ||
475 | return ctx; | |
254f965c | 476 | } |
e0556841 BW |
477 | |
478 | static inline int | |
1d719cda | 479 | mi_set_context(struct drm_i915_gem_request *req, u32 hw_flags) |
e0556841 | 480 | { |
1d719cda | 481 | struct intel_engine_cs *ring = req->ring; |
e80f14b6 | 482 | u32 flags = hw_flags | MI_MM_SPACE_GTT; |
2c550183 CW |
483 | const int num_rings = |
484 | /* Use an extended w/a on ivb+ if signalling from other rings */ | |
485 | i915_semaphore_is_enabled(ring->dev) ? | |
486 | hweight32(INTEL_INFO(ring->dev)->ring_mask) - 1 : | |
487 | 0; | |
488 | int len, i, ret; | |
e0556841 | 489 | |
12b0286f BW |
490 | /* w/a: If Flush TLB Invalidation Mode is enabled, driver must do a TLB |
491 | * invalidation prior to MI_SET_CONTEXT. On GEN6 we don't set the value | |
492 | * explicitly, so we rely on the value at ring init, stored in | |
493 | * itlb_before_ctx_switch. | |
494 | */ | |
057f6a8a | 495 | if (IS_GEN6(ring->dev)) { |
a84c3ae1 | 496 | ret = ring->flush(req, I915_GEM_GPU_DOMAINS, 0); |
12b0286f BW |
497 | if (ret) |
498 | return ret; | |
499 | } | |
500 | ||
e80f14b6 | 501 | /* These flags are for resource streamer on HSW+ */ |
4c436d55 AJ |
502 | if (IS_HASWELL(ring->dev) || INTEL_INFO(ring->dev)->gen >= 8) |
503 | flags |= (HSW_MI_RS_SAVE_STATE_EN | HSW_MI_RS_RESTORE_STATE_EN); | |
504 | else if (INTEL_INFO(ring->dev)->gen < 8) | |
e80f14b6 BW |
505 | flags |= (MI_SAVE_EXT_STATE_EN | MI_RESTORE_EXT_STATE_EN); |
506 | ||
2c550183 CW |
507 | |
508 | len = 4; | |
509 | if (INTEL_INFO(ring->dev)->gen >= 7) | |
510 | len += 2 + (num_rings ? 4*num_rings + 2 : 0); | |
511 | ||
5fb9de1a | 512 | ret = intel_ring_begin(req, len); |
e0556841 BW |
513 | if (ret) |
514 | return ret; | |
515 | ||
b3f797ac | 516 | /* WaProgramMiArbOnOffAroundMiSetContext:ivb,vlv,hsw,bdw,chv */ |
2c550183 | 517 | if (INTEL_INFO(ring->dev)->gen >= 7) { |
e37ec39b | 518 | intel_ring_emit(ring, MI_ARB_ON_OFF | MI_ARB_DISABLE); |
2c550183 CW |
519 | if (num_rings) { |
520 | struct intel_engine_cs *signaller; | |
521 | ||
522 | intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(num_rings)); | |
523 | for_each_ring(signaller, to_i915(ring->dev), i) { | |
524 | if (signaller == ring) | |
525 | continue; | |
526 | ||
527 | intel_ring_emit(ring, RING_PSMI_CTL(signaller->mmio_base)); | |
528 | intel_ring_emit(ring, _MASKED_BIT_ENABLE(GEN6_PSMI_SLEEP_MSG_DISABLE)); | |
529 | } | |
530 | } | |
531 | } | |
e37ec39b | 532 | |
e0556841 BW |
533 | intel_ring_emit(ring, MI_NOOP); |
534 | intel_ring_emit(ring, MI_SET_CONTEXT); | |
1d719cda | 535 | intel_ring_emit(ring, i915_gem_obj_ggtt_offset(req->ctx->legacy_hw_ctx.rcs_state) | |
e80f14b6 | 536 | flags); |
2b7e8082 VS |
537 | /* |
538 | * w/a: MI_SET_CONTEXT must always be followed by MI_NOOP | |
539 | * WaMiSetContext_Hang:snb,ivb,vlv | |
540 | */ | |
e0556841 BW |
541 | intel_ring_emit(ring, MI_NOOP); |
542 | ||
2c550183 CW |
543 | if (INTEL_INFO(ring->dev)->gen >= 7) { |
544 | if (num_rings) { | |
545 | struct intel_engine_cs *signaller; | |
546 | ||
547 | intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(num_rings)); | |
548 | for_each_ring(signaller, to_i915(ring->dev), i) { | |
549 | if (signaller == ring) | |
550 | continue; | |
551 | ||
552 | intel_ring_emit(ring, RING_PSMI_CTL(signaller->mmio_base)); | |
553 | intel_ring_emit(ring, _MASKED_BIT_DISABLE(GEN6_PSMI_SLEEP_MSG_DISABLE)); | |
554 | } | |
555 | } | |
e37ec39b | 556 | intel_ring_emit(ring, MI_ARB_ON_OFF | MI_ARB_ENABLE); |
2c550183 | 557 | } |
e37ec39b | 558 | |
e0556841 BW |
559 | intel_ring_advance(ring); |
560 | ||
561 | return ret; | |
562 | } | |
563 | ||
317b4e90 BW |
564 | static inline bool should_skip_switch(struct intel_engine_cs *ring, |
565 | struct intel_context *from, | |
566 | struct intel_context *to) | |
567 | { | |
563222a7 BW |
568 | if (to->remap_slice) |
569 | return false; | |
570 | ||
9258811c DV |
571 | if (to->ppgtt && from == to && |
572 | !(intel_ring_flag(ring) & to->ppgtt->pd_dirty_rings)) | |
573 | return true; | |
317b4e90 BW |
574 | |
575 | return false; | |
576 | } | |
577 | ||
578 | static bool | |
579 | needs_pd_load_pre(struct intel_engine_cs *ring, struct intel_context *to) | |
580 | { | |
581 | struct drm_i915_private *dev_priv = ring->dev->dev_private; | |
582 | ||
583 | if (!to->ppgtt) | |
584 | return false; | |
585 | ||
586 | if (INTEL_INFO(ring->dev)->gen < 8) | |
587 | return true; | |
588 | ||
589 | if (ring != &dev_priv->ring[RCS]) | |
590 | return true; | |
591 | ||
592 | return false; | |
593 | } | |
594 | ||
595 | static bool | |
6702cf16 BW |
596 | needs_pd_load_post(struct intel_engine_cs *ring, struct intel_context *to, |
597 | u32 hw_flags) | |
317b4e90 BW |
598 | { |
599 | struct drm_i915_private *dev_priv = ring->dev->dev_private; | |
600 | ||
601 | if (!to->ppgtt) | |
602 | return false; | |
603 | ||
604 | if (!IS_GEN8(ring->dev)) | |
605 | return false; | |
606 | ||
607 | if (ring != &dev_priv->ring[RCS]) | |
608 | return false; | |
609 | ||
6702cf16 | 610 | if (hw_flags & MI_RESTORE_INHIBIT) |
317b4e90 BW |
611 | return true; |
612 | ||
613 | return false; | |
614 | } | |
615 | ||
abd68d9e | 616 | static int do_switch(struct drm_i915_gem_request *req) |
e0556841 | 617 | { |
abd68d9e JH |
618 | struct intel_context *to = req->ctx; |
619 | struct intel_engine_cs *ring = req->ring; | |
6f65e29a | 620 | struct drm_i915_private *dev_priv = ring->dev->dev_private; |
273497e5 | 621 | struct intel_context *from = ring->last_context; |
e0556841 | 622 | u32 hw_flags = 0; |
967ab6b1 | 623 | bool uninitialized = false; |
3ccfd19d | 624 | int ret, i; |
e0556841 | 625 | |
67e3d297 | 626 | if (from != NULL && ring == &dev_priv->ring[RCS]) { |
ea0c76f8 OM |
627 | BUG_ON(from->legacy_hw_ctx.rcs_state == NULL); |
628 | BUG_ON(!i915_gem_obj_is_pinned(from->legacy_hw_ctx.rcs_state)); | |
67e3d297 | 629 | } |
e0556841 | 630 | |
317b4e90 | 631 | if (should_skip_switch(ring, from, to)) |
9a3b5304 CW |
632 | return 0; |
633 | ||
7e0d96bc BW |
634 | /* Trying to pin first makes error handling easier. */ |
635 | if (ring == &dev_priv->ring[RCS]) { | |
ea0c76f8 | 636 | ret = i915_gem_obj_ggtt_pin(to->legacy_hw_ctx.rcs_state, |
1ec9e26d | 637 | get_context_alignment(ring->dev), 0); |
7e0d96bc BW |
638 | if (ret) |
639 | return ret; | |
67e3d297 BW |
640 | } |
641 | ||
acc240d4 DV |
642 | /* |
643 | * Pin can switch back to the default context if we end up calling into | |
644 | * evict_everything - as a last ditch gtt defrag effort that also | |
645 | * switches to the default context. Hence we need to reload from here. | |
646 | */ | |
647 | from = ring->last_context; | |
648 | ||
317b4e90 BW |
649 | if (needs_pd_load_pre(ring, to)) { |
650 | /* Older GENs and non render rings still want the load first, | |
651 | * "PP_DCLV followed by PP_DIR_BASE register through Load | |
652 | * Register Immediate commands in Ring Buffer before submitting | |
653 | * a context."*/ | |
198c974d | 654 | trace_switch_mm(ring, to); |
e85b26dc | 655 | ret = to->ppgtt->switch_mm(to->ppgtt, req); |
7e0d96bc BW |
656 | if (ret) |
657 | goto unpin_out; | |
563222a7 BW |
658 | |
659 | /* Doing a PD load always reloads the page dirs */ | |
9258811c | 660 | to->ppgtt->pd_dirty_rings &= ~intel_ring_flag(ring); |
7e0d96bc BW |
661 | } |
662 | ||
663 | if (ring != &dev_priv->ring[RCS]) { | |
664 | if (from) | |
665 | i915_gem_context_unreference(from); | |
666 | goto done; | |
667 | } | |
668 | ||
acc240d4 DV |
669 | /* |
670 | * Clear this page out of any CPU caches for coherent swap-in/out. Note | |
d3373a24 CW |
671 | * that thanks to write = false in this call and us not setting any gpu |
672 | * write domains when putting a context object onto the active list | |
673 | * (when switching away from it), this won't block. | |
acc240d4 DV |
674 | * |
675 | * XXX: We need a real interface to do this instead of trickery. | |
676 | */ | |
ea0c76f8 | 677 | ret = i915_gem_object_set_to_gtt_domain(to->legacy_hw_ctx.rcs_state, false); |
7e0d96bc BW |
678 | if (ret) |
679 | goto unpin_out; | |
d3373a24 | 680 | |
6702cf16 | 681 | if (!to->legacy_hw_ctx.initialized) { |
e0556841 | 682 | hw_flags |= MI_RESTORE_INHIBIT; |
6702cf16 BW |
683 | /* NB: If we inhibit the restore, the context is not allowed to |
684 | * die because future work may end up depending on valid address | |
685 | * space. This means we must enforce that a page table load | |
686 | * occur when this occurs. */ | |
687 | } else if (to->ppgtt && | |
9258811c | 688 | (intel_ring_flag(ring) & to->ppgtt->pd_dirty_rings)) { |
563222a7 | 689 | hw_flags |= MI_FORCE_RESTORE; |
9258811c DV |
690 | to->ppgtt->pd_dirty_rings &= ~intel_ring_flag(ring); |
691 | } | |
e0556841 | 692 | |
6702cf16 BW |
693 | /* We should never emit switch_mm more than once */ |
694 | WARN_ON(needs_pd_load_pre(ring, to) && | |
9258811c | 695 | needs_pd_load_post(ring, to, hw_flags)); |
6702cf16 | 696 | |
1d719cda | 697 | ret = mi_set_context(req, hw_flags); |
7e0d96bc BW |
698 | if (ret) |
699 | goto unpin_out; | |
e0556841 | 700 | |
6702cf16 BW |
701 | /* GEN8 does *not* require an explicit reload if the PDPs have been |
702 | * setup, and we do not wish to move them. | |
703 | */ | |
704 | if (needs_pd_load_post(ring, to, hw_flags)) { | |
317b4e90 | 705 | trace_switch_mm(ring, to); |
e85b26dc | 706 | ret = to->ppgtt->switch_mm(to->ppgtt, req); |
317b4e90 BW |
707 | /* The hardware context switch is emitted, but we haven't |
708 | * actually changed the state - so it's probably safe to bail | |
709 | * here. Still, let the user know something dangerous has | |
710 | * happened. | |
711 | */ | |
712 | if (ret) { | |
713 | DRM_ERROR("Failed to change address space on context switch\n"); | |
714 | goto unpin_out; | |
715 | } | |
716 | } | |
717 | ||
3ccfd19d BW |
718 | for (i = 0; i < MAX_L3_SLICES; i++) { |
719 | if (!(to->remap_slice & (1<<i))) | |
720 | continue; | |
721 | ||
6909a666 | 722 | ret = i915_gem_l3_remap(req, i); |
3ccfd19d BW |
723 | /* If it failed, try again next round */ |
724 | if (ret) | |
725 | DRM_DEBUG_DRIVER("L3 remapping failed\n"); | |
726 | else | |
727 | to->remap_slice &= ~(1<<i); | |
728 | } | |
729 | ||
e0556841 BW |
730 | /* The backing object for the context is done after switching to the |
731 | * *next* context. Therefore we cannot retire the previous context until | |
732 | * the next context has already started running. In fact, the below code | |
733 | * is a bit suboptimal because the retiring can occur simply after the | |
734 | * MI_SET_CONTEXT instead of when the next seqno has completed. | |
735 | */ | |
112522f6 | 736 | if (from != NULL) { |
ea0c76f8 | 737 | from->legacy_hw_ctx.rcs_state->base.read_domains = I915_GEM_DOMAIN_INSTRUCTION; |
b2af0376 | 738 | i915_vma_move_to_active(i915_gem_obj_to_ggtt(from->legacy_hw_ctx.rcs_state), req); |
e0556841 BW |
739 | /* As long as MI_SET_CONTEXT is serializing, ie. it flushes the |
740 | * whole damn pipeline, we don't need to explicitly mark the | |
741 | * object dirty. The only exception is that the context must be | |
742 | * correct in case the object gets swapped out. Ideally we'd be | |
743 | * able to defer doing this until we know the object would be | |
744 | * swapped, but there is no way to do that yet. | |
745 | */ | |
ea0c76f8 | 746 | from->legacy_hw_ctx.rcs_state->dirty = 1; |
112522f6 | 747 | |
c0321e2c | 748 | /* obj is kept alive until the next request by its active ref */ |
ea0c76f8 | 749 | i915_gem_object_ggtt_unpin(from->legacy_hw_ctx.rcs_state); |
112522f6 | 750 | i915_gem_context_unreference(from); |
e0556841 BW |
751 | } |
752 | ||
6702cf16 | 753 | uninitialized = !to->legacy_hw_ctx.initialized; |
ea0c76f8 | 754 | to->legacy_hw_ctx.initialized = true; |
967ab6b1 | 755 | |
67e3d297 | 756 | done: |
112522f6 CW |
757 | i915_gem_context_reference(to); |
758 | ring->last_context = to; | |
e0556841 | 759 | |
967ab6b1 | 760 | if (uninitialized) { |
86d7f238 | 761 | if (ring->init_context) { |
8753181e | 762 | ret = ring->init_context(req); |
86d7f238 AS |
763 | if (ret) |
764 | DRM_ERROR("ring init context: %d\n", ret); | |
765 | } | |
46470fc9 MK |
766 | } |
767 | ||
e0556841 | 768 | return 0; |
7e0d96bc BW |
769 | |
770 | unpin_out: | |
771 | if (ring->id == RCS) | |
ea0c76f8 | 772 | i915_gem_object_ggtt_unpin(to->legacy_hw_ctx.rcs_state); |
7e0d96bc | 773 | return ret; |
e0556841 BW |
774 | } |
775 | ||
776 | /** | |
777 | * i915_switch_context() - perform a GPU context switch. | |
ba01cc93 | 778 | * @req: request for which we'll execute the context switch |
e0556841 BW |
779 | * |
780 | * The context life cycle is simple. The context refcount is incremented and | |
781 | * decremented by 1 and create and destroy. If the context is in use by the GPU, | |
ecdb5fd8 | 782 | * it will have a refcount > 1. This allows us to destroy the context abstract |
e0556841 | 783 | * object while letting the normal object tracking destroy the backing BO. |
ecdb5fd8 TD |
784 | * |
785 | * This function should not be used in execlists mode. Instead the context is | |
786 | * switched by writing to the ELSP and requests keep a reference to their | |
787 | * context. | |
e0556841 | 788 | */ |
ba01cc93 | 789 | int i915_switch_context(struct drm_i915_gem_request *req) |
e0556841 | 790 | { |
ba01cc93 | 791 | struct intel_engine_cs *ring = req->ring; |
e0556841 | 792 | struct drm_i915_private *dev_priv = ring->dev->dev_private; |
e0556841 | 793 | |
ecdb5fd8 | 794 | WARN_ON(i915.enable_execlists); |
0eea67eb BW |
795 | WARN_ON(!mutex_is_locked(&dev_priv->dev->struct_mutex)); |
796 | ||
ba01cc93 JH |
797 | if (req->ctx->legacy_hw_ctx.rcs_state == NULL) { /* We have the fake context */ |
798 | if (req->ctx != ring->last_context) { | |
799 | i915_gem_context_reference(req->ctx); | |
691e6415 CW |
800 | if (ring->last_context) |
801 | i915_gem_context_unreference(ring->last_context); | |
ba01cc93 | 802 | ring->last_context = req->ctx; |
691e6415 | 803 | } |
c482972a | 804 | return 0; |
a95f6a00 | 805 | } |
c482972a | 806 | |
abd68d9e | 807 | return do_switch(req); |
e0556841 | 808 | } |
84624813 | 809 | |
ec3e9963 | 810 | static bool contexts_enabled(struct drm_device *dev) |
691e6415 | 811 | { |
ec3e9963 | 812 | return i915.enable_execlists || to_i915(dev)->hw_context_size; |
691e6415 CW |
813 | } |
814 | ||
84624813 BW |
815 | int i915_gem_context_create_ioctl(struct drm_device *dev, void *data, |
816 | struct drm_file *file) | |
817 | { | |
84624813 BW |
818 | struct drm_i915_gem_context_create *args = data; |
819 | struct drm_i915_file_private *file_priv = file->driver_priv; | |
273497e5 | 820 | struct intel_context *ctx; |
84624813 BW |
821 | int ret; |
822 | ||
ec3e9963 | 823 | if (!contexts_enabled(dev)) |
5fa8be65 DV |
824 | return -ENODEV; |
825 | ||
84624813 BW |
826 | ret = i915_mutex_lock_interruptible(dev); |
827 | if (ret) | |
828 | return ret; | |
829 | ||
d624d86e | 830 | ctx = i915_gem_create_context(dev, file_priv); |
84624813 | 831 | mutex_unlock(&dev->struct_mutex); |
be636387 DC |
832 | if (IS_ERR(ctx)) |
833 | return PTR_ERR(ctx); | |
84624813 | 834 | |
821d66dd | 835 | args->ctx_id = ctx->user_handle; |
84624813 BW |
836 | DRM_DEBUG_DRIVER("HW context %d created\n", args->ctx_id); |
837 | ||
be636387 | 838 | return 0; |
84624813 BW |
839 | } |
840 | ||
841 | int i915_gem_context_destroy_ioctl(struct drm_device *dev, void *data, | |
842 | struct drm_file *file) | |
843 | { | |
844 | struct drm_i915_gem_context_destroy *args = data; | |
845 | struct drm_i915_file_private *file_priv = file->driver_priv; | |
273497e5 | 846 | struct intel_context *ctx; |
84624813 BW |
847 | int ret; |
848 | ||
821d66dd | 849 | if (args->ctx_id == DEFAULT_CONTEXT_HANDLE) |
c2cf2416 | 850 | return -ENOENT; |
0eea67eb | 851 | |
84624813 BW |
852 | ret = i915_mutex_lock_interruptible(dev); |
853 | if (ret) | |
854 | return ret; | |
855 | ||
856 | ctx = i915_gem_context_get(file_priv, args->ctx_id); | |
72ad5c45 | 857 | if (IS_ERR(ctx)) { |
84624813 | 858 | mutex_unlock(&dev->struct_mutex); |
72ad5c45 | 859 | return PTR_ERR(ctx); |
84624813 BW |
860 | } |
861 | ||
821d66dd | 862 | idr_remove(&ctx->file_priv->context_idr, ctx->user_handle); |
dce3271b | 863 | i915_gem_context_unreference(ctx); |
84624813 BW |
864 | mutex_unlock(&dev->struct_mutex); |
865 | ||
866 | DRM_DEBUG_DRIVER("HW context %d destroyed\n", args->ctx_id); | |
867 | return 0; | |
868 | } | |
c9dc0f35 CW |
869 | |
870 | int i915_gem_context_getparam_ioctl(struct drm_device *dev, void *data, | |
871 | struct drm_file *file) | |
872 | { | |
873 | struct drm_i915_file_private *file_priv = file->driver_priv; | |
874 | struct drm_i915_gem_context_param *args = data; | |
875 | struct intel_context *ctx; | |
876 | int ret; | |
877 | ||
878 | ret = i915_mutex_lock_interruptible(dev); | |
879 | if (ret) | |
880 | return ret; | |
881 | ||
882 | ctx = i915_gem_context_get(file_priv, args->ctx_id); | |
883 | if (IS_ERR(ctx)) { | |
884 | mutex_unlock(&dev->struct_mutex); | |
885 | return PTR_ERR(ctx); | |
886 | } | |
887 | ||
888 | args->size = 0; | |
889 | switch (args->param) { | |
890 | case I915_CONTEXT_PARAM_BAN_PERIOD: | |
891 | args->value = ctx->hang_stats.ban_period_seconds; | |
892 | break; | |
b1b38278 DW |
893 | case I915_CONTEXT_PARAM_NO_ZEROMAP: |
894 | args->value = ctx->flags & CONTEXT_NO_ZEROMAP; | |
895 | break; | |
c9dc0f35 CW |
896 | default: |
897 | ret = -EINVAL; | |
898 | break; | |
899 | } | |
900 | mutex_unlock(&dev->struct_mutex); | |
901 | ||
902 | return ret; | |
903 | } | |
904 | ||
905 | int i915_gem_context_setparam_ioctl(struct drm_device *dev, void *data, | |
906 | struct drm_file *file) | |
907 | { | |
908 | struct drm_i915_file_private *file_priv = file->driver_priv; | |
909 | struct drm_i915_gem_context_param *args = data; | |
910 | struct intel_context *ctx; | |
911 | int ret; | |
912 | ||
913 | ret = i915_mutex_lock_interruptible(dev); | |
914 | if (ret) | |
915 | return ret; | |
916 | ||
917 | ctx = i915_gem_context_get(file_priv, args->ctx_id); | |
918 | if (IS_ERR(ctx)) { | |
919 | mutex_unlock(&dev->struct_mutex); | |
920 | return PTR_ERR(ctx); | |
921 | } | |
922 | ||
923 | switch (args->param) { | |
924 | case I915_CONTEXT_PARAM_BAN_PERIOD: | |
925 | if (args->size) | |
926 | ret = -EINVAL; | |
927 | else if (args->value < ctx->hang_stats.ban_period_seconds && | |
928 | !capable(CAP_SYS_ADMIN)) | |
929 | ret = -EPERM; | |
930 | else | |
931 | ctx->hang_stats.ban_period_seconds = args->value; | |
932 | break; | |
b1b38278 DW |
933 | case I915_CONTEXT_PARAM_NO_ZEROMAP: |
934 | if (args->size) { | |
935 | ret = -EINVAL; | |
936 | } else { | |
937 | ctx->flags &= ~CONTEXT_NO_ZEROMAP; | |
938 | ctx->flags |= args->value ? CONTEXT_NO_ZEROMAP : 0; | |
939 | } | |
940 | break; | |
c9dc0f35 CW |
941 | default: |
942 | ret = -EINVAL; | |
943 | break; | |
944 | } | |
945 | mutex_unlock(&dev->struct_mutex); | |
946 | ||
947 | return ret; | |
948 | } |