Loading...
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
76 * GPU. The GPU has loaded its state already and has stored away the gtt
77 * offset of the BO. The GPU is not actively referencing the data at this
78 * offset, but it will on the next context switch. The only way to avoid this
79 * is to do a GPU reset.
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 <drm/drmP.h>
89#include <drm/i915_drm.h>
90#include "i915_drv.h"
91#include "i915_trace.h"
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 */
97#define GEN6_CONTEXT_ALIGN (64<<10)
98#define GEN7_CONTEXT_ALIGN 4096
99
100static size_t get_context_alignment(struct drm_device *dev)
101{
102 if (IS_GEN6(dev))
103 return GEN6_CONTEXT_ALIGN;
104
105 return GEN7_CONTEXT_ALIGN;
106}
107
108static int get_context_size(struct drm_device *dev)
109{
110 struct drm_i915_private *dev_priv = dev->dev_private;
111 int ret;
112 u32 reg;
113
114 switch (INTEL_INFO(dev)->gen) {
115 case 6:
116 reg = I915_READ(CXT_SIZE);
117 ret = GEN6_CXT_TOTAL_SIZE(reg) * 64;
118 break;
119 case 7:
120 reg = I915_READ(GEN7_CXT_SIZE);
121 if (IS_HASWELL(dev))
122 ret = HSW_CXT_TOTAL_SIZE;
123 else
124 ret = GEN7_CXT_TOTAL_SIZE(reg) * 64;
125 break;
126 case 8:
127 ret = GEN8_CXT_TOTAL_SIZE;
128 break;
129 default:
130 BUG();
131 }
132
133 return ret;
134}
135
136static void i915_gem_context_clean(struct intel_context *ctx)
137{
138 struct i915_hw_ppgtt *ppgtt = ctx->ppgtt;
139 struct i915_vma *vma, *next;
140
141 if (!ppgtt)
142 return;
143
144 list_for_each_entry_safe(vma, next, &ppgtt->base.inactive_list,
145 vm_link) {
146 if (WARN_ON(__i915_vma_unbind_no_wait(vma)))
147 break;
148 }
149}
150
151void i915_gem_context_free(struct kref *ctx_ref)
152{
153 struct intel_context *ctx = container_of(ctx_ref, typeof(*ctx), ref);
154
155 trace_i915_context_free(ctx);
156
157 if (i915.enable_execlists)
158 intel_lr_context_free(ctx);
159
160 /*
161 * This context is going away and we need to remove all VMAs still
162 * around. This is to handle imported shared objects for which
163 * destructor did not run when their handles were closed.
164 */
165 i915_gem_context_clean(ctx);
166
167 i915_ppgtt_put(ctx->ppgtt);
168
169 if (ctx->legacy_hw_ctx.rcs_state)
170 drm_gem_object_unreference(&ctx->legacy_hw_ctx.rcs_state->base);
171 list_del(&ctx->link);
172 kfree(ctx);
173}
174
175struct drm_i915_gem_object *
176i915_gem_alloc_context_obj(struct drm_device *dev, size_t size)
177{
178 struct drm_i915_gem_object *obj;
179 int ret;
180
181 obj = i915_gem_alloc_object(dev, size);
182 if (obj == NULL)
183 return ERR_PTR(-ENOMEM);
184
185 /*
186 * Try to make the context utilize L3 as well as LLC.
187 *
188 * On VLV we don't have L3 controls in the PTEs so we
189 * shouldn't touch the cache level, especially as that
190 * would make the object snooped which might have a
191 * negative performance impact.
192 *
193 * Snooping is required on non-llc platforms in execlist
194 * mode, but since all GGTT accesses use PAT entry 0 we
195 * get snooping anyway regardless of cache_level.
196 *
197 * This is only applicable for Ivy Bridge devices since
198 * later platforms don't have L3 control bits in the PTE.
199 */
200 if (IS_IVYBRIDGE(dev)) {
201 ret = i915_gem_object_set_cache_level(obj, I915_CACHE_L3_LLC);
202 /* Failure shouldn't ever happen this early */
203 if (WARN_ON(ret)) {
204 drm_gem_object_unreference(&obj->base);
205 return ERR_PTR(ret);
206 }
207 }
208
209 return obj;
210}
211
212static struct intel_context *
213__create_hw_context(struct drm_device *dev,
214 struct drm_i915_file_private *file_priv)
215{
216 struct drm_i915_private *dev_priv = dev->dev_private;
217 struct intel_context *ctx;
218 int ret;
219
220 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
221 if (ctx == NULL)
222 return ERR_PTR(-ENOMEM);
223
224 kref_init(&ctx->ref);
225 list_add_tail(&ctx->link, &dev_priv->context_list);
226 ctx->i915 = dev_priv;
227
228 if (dev_priv->hw_context_size) {
229 struct drm_i915_gem_object *obj =
230 i915_gem_alloc_context_obj(dev, dev_priv->hw_context_size);
231 if (IS_ERR(obj)) {
232 ret = PTR_ERR(obj);
233 goto err_out;
234 }
235 ctx->legacy_hw_ctx.rcs_state = obj;
236 }
237
238 /* Default context will never have a file_priv */
239 if (file_priv != NULL) {
240 ret = idr_alloc(&file_priv->context_idr, ctx,
241 DEFAULT_CONTEXT_HANDLE, 0, GFP_KERNEL);
242 if (ret < 0)
243 goto err_out;
244 } else
245 ret = DEFAULT_CONTEXT_HANDLE;
246
247 ctx->file_priv = file_priv;
248 ctx->user_handle = ret;
249 /* NB: Mark all slices as needing a remap so that when the context first
250 * loads it will restore whatever remap state already exists. If there
251 * is no remap info, it will be a NOP. */
252 ctx->remap_slice = (1 << NUM_L3_SLICES(dev)) - 1;
253
254 ctx->hang_stats.ban_period_seconds = DRM_I915_CTX_BAN_PERIOD;
255
256 return ctx;
257
258err_out:
259 i915_gem_context_unreference(ctx);
260 return ERR_PTR(ret);
261}
262
263/**
264 * The default context needs to exist per ring that uses contexts. It stores the
265 * context state of the GPU for applications that don't utilize HW contexts, as
266 * well as an idle case.
267 */
268static struct intel_context *
269i915_gem_create_context(struct drm_device *dev,
270 struct drm_i915_file_private *file_priv)
271{
272 const bool is_global_default_ctx = file_priv == NULL;
273 struct intel_context *ctx;
274 int ret = 0;
275
276 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
277
278 ctx = __create_hw_context(dev, file_priv);
279 if (IS_ERR(ctx))
280 return ctx;
281
282 if (is_global_default_ctx && ctx->legacy_hw_ctx.rcs_state) {
283 /* We may need to do things with the shrinker which
284 * require us to immediately switch back to the default
285 * context. This can cause a problem as pinning the
286 * default context also requires GTT space which may not
287 * be available. To avoid this we always pin the default
288 * context.
289 */
290 ret = i915_gem_obj_ggtt_pin(ctx->legacy_hw_ctx.rcs_state,
291 get_context_alignment(dev), 0);
292 if (ret) {
293 DRM_DEBUG_DRIVER("Couldn't pin %d\n", ret);
294 goto err_destroy;
295 }
296 }
297
298 if (USES_FULL_PPGTT(dev)) {
299 struct i915_hw_ppgtt *ppgtt = i915_ppgtt_create(dev, file_priv);
300
301 if (IS_ERR_OR_NULL(ppgtt)) {
302 DRM_DEBUG_DRIVER("PPGTT setup failed (%ld)\n",
303 PTR_ERR(ppgtt));
304 ret = PTR_ERR(ppgtt);
305 goto err_unpin;
306 }
307
308 ctx->ppgtt = ppgtt;
309 }
310
311 trace_i915_context_create(ctx);
312
313 return ctx;
314
315err_unpin:
316 if (is_global_default_ctx && ctx->legacy_hw_ctx.rcs_state)
317 i915_gem_object_ggtt_unpin(ctx->legacy_hw_ctx.rcs_state);
318err_destroy:
319 idr_remove(&file_priv->context_idr, ctx->user_handle);
320 i915_gem_context_unreference(ctx);
321 return ERR_PTR(ret);
322}
323
324static void i915_gem_context_unpin(struct intel_context *ctx,
325 struct intel_engine_cs *engine)
326{
327 if (i915.enable_execlists) {
328 intel_lr_context_unpin(ctx, engine);
329 } else {
330 if (engine->id == RCS && ctx->legacy_hw_ctx.rcs_state)
331 i915_gem_object_ggtt_unpin(ctx->legacy_hw_ctx.rcs_state);
332 i915_gem_context_unreference(ctx);
333 }
334}
335
336void i915_gem_context_reset(struct drm_device *dev)
337{
338 struct drm_i915_private *dev_priv = dev->dev_private;
339 int i;
340
341 if (i915.enable_execlists) {
342 struct intel_context *ctx;
343
344 list_for_each_entry(ctx, &dev_priv->context_list, link)
345 intel_lr_context_reset(dev, ctx);
346 }
347
348 for (i = 0; i < I915_NUM_RINGS; i++) {
349 struct intel_engine_cs *ring = &dev_priv->ring[i];
350
351 if (ring->last_context) {
352 i915_gem_context_unpin(ring->last_context, ring);
353 ring->last_context = NULL;
354 }
355 }
356
357 /* Force the GPU state to be reinitialised on enabling */
358 dev_priv->kernel_context->legacy_hw_ctx.initialized = false;
359}
360
361int i915_gem_context_init(struct drm_device *dev)
362{
363 struct drm_i915_private *dev_priv = dev->dev_private;
364 struct intel_context *ctx;
365
366 /* Init should only be called once per module load. Eventually the
367 * restriction on the context_disabled check can be loosened. */
368 if (WARN_ON(dev_priv->kernel_context))
369 return 0;
370
371 if (intel_vgpu_active(dev) && HAS_LOGICAL_RING_CONTEXTS(dev)) {
372 if (!i915.enable_execlists) {
373 DRM_INFO("Only EXECLIST mode is supported in vgpu.\n");
374 return -EINVAL;
375 }
376 }
377
378 if (i915.enable_execlists) {
379 /* NB: intentionally left blank. We will allocate our own
380 * backing objects as we need them, thank you very much */
381 dev_priv->hw_context_size = 0;
382 } else if (HAS_HW_CONTEXTS(dev)) {
383 dev_priv->hw_context_size = round_up(get_context_size(dev), 4096);
384 if (dev_priv->hw_context_size > (1<<20)) {
385 DRM_DEBUG_DRIVER("Disabling HW Contexts; invalid size %d\n",
386 dev_priv->hw_context_size);
387 dev_priv->hw_context_size = 0;
388 }
389 }
390
391 ctx = i915_gem_create_context(dev, NULL);
392 if (IS_ERR(ctx)) {
393 DRM_ERROR("Failed to create default global context (error %ld)\n",
394 PTR_ERR(ctx));
395 return PTR_ERR(ctx);
396 }
397
398 dev_priv->kernel_context = ctx;
399
400 DRM_DEBUG_DRIVER("%s context support initialized\n",
401 i915.enable_execlists ? "LR" :
402 dev_priv->hw_context_size ? "HW" : "fake");
403 return 0;
404}
405
406void i915_gem_context_fini(struct drm_device *dev)
407{
408 struct drm_i915_private *dev_priv = dev->dev_private;
409 struct intel_context *dctx = dev_priv->kernel_context;
410 int i;
411
412 if (dctx->legacy_hw_ctx.rcs_state) {
413 /* The only known way to stop the gpu from accessing the hw context is
414 * to reset it. Do this as the very last operation to avoid confusing
415 * other code, leading to spurious errors. */
416 intel_gpu_reset(dev);
417
418 /* When default context is created and switched to, base object refcount
419 * will be 2 (+1 from object creation and +1 from do_switch()).
420 * i915_gem_context_fini() will be called after gpu_idle() has switched
421 * to default context. So we need to unreference the base object once
422 * to offset the do_switch part, so that i915_gem_context_unreference()
423 * can then free the base object correctly. */
424 WARN_ON(!dev_priv->ring[RCS].last_context);
425
426 i915_gem_object_ggtt_unpin(dctx->legacy_hw_ctx.rcs_state);
427 }
428
429 for (i = I915_NUM_RINGS; --i >= 0;) {
430 struct intel_engine_cs *ring = &dev_priv->ring[i];
431
432 if (ring->last_context) {
433 i915_gem_context_unpin(ring->last_context, ring);
434 ring->last_context = NULL;
435 }
436 }
437
438 i915_gem_context_unreference(dctx);
439 dev_priv->kernel_context = NULL;
440}
441
442int i915_gem_context_enable(struct drm_i915_gem_request *req)
443{
444 struct intel_engine_cs *ring = req->ring;
445 int ret;
446
447 if (i915.enable_execlists) {
448 if (ring->init_context == NULL)
449 return 0;
450
451 ret = ring->init_context(req);
452 } else
453 ret = i915_switch_context(req);
454
455 if (ret) {
456 DRM_ERROR("ring init context: %d\n", ret);
457 return ret;
458 }
459
460 return 0;
461}
462
463static int context_idr_cleanup(int id, void *p, void *data)
464{
465 struct intel_context *ctx = p;
466
467 i915_gem_context_unreference(ctx);
468 return 0;
469}
470
471int i915_gem_context_open(struct drm_device *dev, struct drm_file *file)
472{
473 struct drm_i915_file_private *file_priv = file->driver_priv;
474 struct intel_context *ctx;
475
476 idr_init(&file_priv->context_idr);
477
478 mutex_lock(&dev->struct_mutex);
479 ctx = i915_gem_create_context(dev, file_priv);
480 mutex_unlock(&dev->struct_mutex);
481
482 if (IS_ERR(ctx)) {
483 idr_destroy(&file_priv->context_idr);
484 return PTR_ERR(ctx);
485 }
486
487 return 0;
488}
489
490void i915_gem_context_close(struct drm_device *dev, struct drm_file *file)
491{
492 struct drm_i915_file_private *file_priv = file->driver_priv;
493
494 idr_for_each(&file_priv->context_idr, context_idr_cleanup, NULL);
495 idr_destroy(&file_priv->context_idr);
496}
497
498struct intel_context *
499i915_gem_context_get(struct drm_i915_file_private *file_priv, u32 id)
500{
501 struct intel_context *ctx;
502
503 ctx = (struct intel_context *)idr_find(&file_priv->context_idr, id);
504 if (!ctx)
505 return ERR_PTR(-ENOENT);
506
507 return ctx;
508}
509
510static inline int
511mi_set_context(struct drm_i915_gem_request *req, u32 hw_flags)
512{
513 struct intel_engine_cs *ring = req->ring;
514 u32 flags = hw_flags | MI_MM_SPACE_GTT;
515 const int num_rings =
516 /* Use an extended w/a on ivb+ if signalling from other rings */
517 i915_semaphore_is_enabled(ring->dev) ?
518 hweight32(INTEL_INFO(ring->dev)->ring_mask) - 1 :
519 0;
520 int len, i, ret;
521
522 /* w/a: If Flush TLB Invalidation Mode is enabled, driver must do a TLB
523 * invalidation prior to MI_SET_CONTEXT. On GEN6 we don't set the value
524 * explicitly, so we rely on the value at ring init, stored in
525 * itlb_before_ctx_switch.
526 */
527 if (IS_GEN6(ring->dev)) {
528 ret = ring->flush(req, I915_GEM_GPU_DOMAINS, 0);
529 if (ret)
530 return ret;
531 }
532
533 /* These flags are for resource streamer on HSW+ */
534 if (IS_HASWELL(ring->dev) || INTEL_INFO(ring->dev)->gen >= 8)
535 flags |= (HSW_MI_RS_SAVE_STATE_EN | HSW_MI_RS_RESTORE_STATE_EN);
536 else if (INTEL_INFO(ring->dev)->gen < 8)
537 flags |= (MI_SAVE_EXT_STATE_EN | MI_RESTORE_EXT_STATE_EN);
538
539
540 len = 4;
541 if (INTEL_INFO(ring->dev)->gen >= 7)
542 len += 2 + (num_rings ? 4*num_rings + 2 : 0);
543
544 ret = intel_ring_begin(req, len);
545 if (ret)
546 return ret;
547
548 /* WaProgramMiArbOnOffAroundMiSetContext:ivb,vlv,hsw,bdw,chv */
549 if (INTEL_INFO(ring->dev)->gen >= 7) {
550 intel_ring_emit(ring, MI_ARB_ON_OFF | MI_ARB_DISABLE);
551 if (num_rings) {
552 struct intel_engine_cs *signaller;
553
554 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(num_rings));
555 for_each_ring(signaller, to_i915(ring->dev), i) {
556 if (signaller == ring)
557 continue;
558
559 intel_ring_emit_reg(ring, RING_PSMI_CTL(signaller->mmio_base));
560 intel_ring_emit(ring, _MASKED_BIT_ENABLE(GEN6_PSMI_SLEEP_MSG_DISABLE));
561 }
562 }
563 }
564
565 intel_ring_emit(ring, MI_NOOP);
566 intel_ring_emit(ring, MI_SET_CONTEXT);
567 intel_ring_emit(ring, i915_gem_obj_ggtt_offset(req->ctx->legacy_hw_ctx.rcs_state) |
568 flags);
569 /*
570 * w/a: MI_SET_CONTEXT must always be followed by MI_NOOP
571 * WaMiSetContext_Hang:snb,ivb,vlv
572 */
573 intel_ring_emit(ring, MI_NOOP);
574
575 if (INTEL_INFO(ring->dev)->gen >= 7) {
576 if (num_rings) {
577 struct intel_engine_cs *signaller;
578
579 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(num_rings));
580 for_each_ring(signaller, to_i915(ring->dev), i) {
581 if (signaller == ring)
582 continue;
583
584 intel_ring_emit_reg(ring, RING_PSMI_CTL(signaller->mmio_base));
585 intel_ring_emit(ring, _MASKED_BIT_DISABLE(GEN6_PSMI_SLEEP_MSG_DISABLE));
586 }
587 }
588 intel_ring_emit(ring, MI_ARB_ON_OFF | MI_ARB_ENABLE);
589 }
590
591 intel_ring_advance(ring);
592
593 return ret;
594}
595
596static inline bool should_skip_switch(struct intel_engine_cs *ring,
597 struct intel_context *from,
598 struct intel_context *to)
599{
600 if (to->remap_slice)
601 return false;
602
603 if (to->ppgtt && from == to &&
604 !(intel_ring_flag(ring) & to->ppgtt->pd_dirty_rings))
605 return true;
606
607 return false;
608}
609
610static bool
611needs_pd_load_pre(struct intel_engine_cs *ring, struct intel_context *to)
612{
613 struct drm_i915_private *dev_priv = ring->dev->dev_private;
614
615 if (!to->ppgtt)
616 return false;
617
618 if (INTEL_INFO(ring->dev)->gen < 8)
619 return true;
620
621 if (ring != &dev_priv->ring[RCS])
622 return true;
623
624 return false;
625}
626
627static bool
628needs_pd_load_post(struct intel_engine_cs *ring, struct intel_context *to,
629 u32 hw_flags)
630{
631 struct drm_i915_private *dev_priv = ring->dev->dev_private;
632
633 if (!to->ppgtt)
634 return false;
635
636 if (!IS_GEN8(ring->dev))
637 return false;
638
639 if (ring != &dev_priv->ring[RCS])
640 return false;
641
642 if (hw_flags & MI_RESTORE_INHIBIT)
643 return true;
644
645 return false;
646}
647
648static int do_switch(struct drm_i915_gem_request *req)
649{
650 struct intel_context *to = req->ctx;
651 struct intel_engine_cs *ring = req->ring;
652 struct drm_i915_private *dev_priv = ring->dev->dev_private;
653 struct intel_context *from = ring->last_context;
654 u32 hw_flags = 0;
655 bool uninitialized = false;
656 int ret, i;
657
658 if (from != NULL && ring == &dev_priv->ring[RCS]) {
659 BUG_ON(from->legacy_hw_ctx.rcs_state == NULL);
660 BUG_ON(!i915_gem_obj_is_pinned(from->legacy_hw_ctx.rcs_state));
661 }
662
663 if (should_skip_switch(ring, from, to))
664 return 0;
665
666 /* Trying to pin first makes error handling easier. */
667 if (ring == &dev_priv->ring[RCS]) {
668 ret = i915_gem_obj_ggtt_pin(to->legacy_hw_ctx.rcs_state,
669 get_context_alignment(ring->dev), 0);
670 if (ret)
671 return ret;
672 }
673
674 /*
675 * Pin can switch back to the default context if we end up calling into
676 * evict_everything - as a last ditch gtt defrag effort that also
677 * switches to the default context. Hence we need to reload from here.
678 */
679 from = ring->last_context;
680
681 if (needs_pd_load_pre(ring, to)) {
682 /* Older GENs and non render rings still want the load first,
683 * "PP_DCLV followed by PP_DIR_BASE register through Load
684 * Register Immediate commands in Ring Buffer before submitting
685 * a context."*/
686 trace_switch_mm(ring, to);
687 ret = to->ppgtt->switch_mm(to->ppgtt, req);
688 if (ret)
689 goto unpin_out;
690
691 /* Doing a PD load always reloads the page dirs */
692 to->ppgtt->pd_dirty_rings &= ~intel_ring_flag(ring);
693 }
694
695 if (ring != &dev_priv->ring[RCS]) {
696 if (from)
697 i915_gem_context_unreference(from);
698 goto done;
699 }
700
701 /*
702 * Clear this page out of any CPU caches for coherent swap-in/out. Note
703 * that thanks to write = false in this call and us not setting any gpu
704 * write domains when putting a context object onto the active list
705 * (when switching away from it), this won't block.
706 *
707 * XXX: We need a real interface to do this instead of trickery.
708 */
709 ret = i915_gem_object_set_to_gtt_domain(to->legacy_hw_ctx.rcs_state, false);
710 if (ret)
711 goto unpin_out;
712
713 if (!to->legacy_hw_ctx.initialized || i915_gem_context_is_default(to)) {
714 hw_flags |= MI_RESTORE_INHIBIT;
715 /* NB: If we inhibit the restore, the context is not allowed to
716 * die because future work may end up depending on valid address
717 * space. This means we must enforce that a page table load
718 * occur when this occurs. */
719 } else if (to->ppgtt &&
720 (intel_ring_flag(ring) & to->ppgtt->pd_dirty_rings)) {
721 hw_flags |= MI_FORCE_RESTORE;
722 to->ppgtt->pd_dirty_rings &= ~intel_ring_flag(ring);
723 }
724
725 /* We should never emit switch_mm more than once */
726 WARN_ON(needs_pd_load_pre(ring, to) &&
727 needs_pd_load_post(ring, to, hw_flags));
728
729 ret = mi_set_context(req, hw_flags);
730 if (ret)
731 goto unpin_out;
732
733 /* GEN8 does *not* require an explicit reload if the PDPs have been
734 * setup, and we do not wish to move them.
735 */
736 if (needs_pd_load_post(ring, to, hw_flags)) {
737 trace_switch_mm(ring, to);
738 ret = to->ppgtt->switch_mm(to->ppgtt, req);
739 /* The hardware context switch is emitted, but we haven't
740 * actually changed the state - so it's probably safe to bail
741 * here. Still, let the user know something dangerous has
742 * happened.
743 */
744 if (ret) {
745 DRM_ERROR("Failed to change address space on context switch\n");
746 goto unpin_out;
747 }
748 }
749
750 for (i = 0; i < MAX_L3_SLICES; i++) {
751 if (!(to->remap_slice & (1<<i)))
752 continue;
753
754 ret = i915_gem_l3_remap(req, i);
755 /* If it failed, try again next round */
756 if (ret)
757 DRM_DEBUG_DRIVER("L3 remapping failed\n");
758 else
759 to->remap_slice &= ~(1<<i);
760 }
761
762 /* The backing object for the context is done after switching to the
763 * *next* context. Therefore we cannot retire the previous context until
764 * the next context has already started running. In fact, the below code
765 * is a bit suboptimal because the retiring can occur simply after the
766 * MI_SET_CONTEXT instead of when the next seqno has completed.
767 */
768 if (from != NULL) {
769 from->legacy_hw_ctx.rcs_state->base.read_domains = I915_GEM_DOMAIN_INSTRUCTION;
770 i915_vma_move_to_active(i915_gem_obj_to_ggtt(from->legacy_hw_ctx.rcs_state), req);
771 /* As long as MI_SET_CONTEXT is serializing, ie. it flushes the
772 * whole damn pipeline, we don't need to explicitly mark the
773 * object dirty. The only exception is that the context must be
774 * correct in case the object gets swapped out. Ideally we'd be
775 * able to defer doing this until we know the object would be
776 * swapped, but there is no way to do that yet.
777 */
778 from->legacy_hw_ctx.rcs_state->dirty = 1;
779
780 /* obj is kept alive until the next request by its active ref */
781 i915_gem_object_ggtt_unpin(from->legacy_hw_ctx.rcs_state);
782 i915_gem_context_unreference(from);
783 }
784
785 uninitialized = !to->legacy_hw_ctx.initialized;
786 to->legacy_hw_ctx.initialized = true;
787
788done:
789 i915_gem_context_reference(to);
790 ring->last_context = to;
791
792 if (uninitialized) {
793 if (ring->init_context) {
794 ret = ring->init_context(req);
795 if (ret)
796 DRM_ERROR("ring init context: %d\n", ret);
797 }
798 }
799
800 return 0;
801
802unpin_out:
803 if (ring->id == RCS)
804 i915_gem_object_ggtt_unpin(to->legacy_hw_ctx.rcs_state);
805 return ret;
806}
807
808/**
809 * i915_switch_context() - perform a GPU context switch.
810 * @req: request for which we'll execute the context switch
811 *
812 * The context life cycle is simple. The context refcount is incremented and
813 * decremented by 1 and create and destroy. If the context is in use by the GPU,
814 * it will have a refcount > 1. This allows us to destroy the context abstract
815 * object while letting the normal object tracking destroy the backing BO.
816 *
817 * This function should not be used in execlists mode. Instead the context is
818 * switched by writing to the ELSP and requests keep a reference to their
819 * context.
820 */
821int i915_switch_context(struct drm_i915_gem_request *req)
822{
823 struct intel_engine_cs *ring = req->ring;
824 struct drm_i915_private *dev_priv = ring->dev->dev_private;
825
826 WARN_ON(i915.enable_execlists);
827 WARN_ON(!mutex_is_locked(&dev_priv->dev->struct_mutex));
828
829 if (req->ctx->legacy_hw_ctx.rcs_state == NULL) { /* We have the fake context */
830 if (req->ctx != ring->last_context) {
831 i915_gem_context_reference(req->ctx);
832 if (ring->last_context)
833 i915_gem_context_unreference(ring->last_context);
834 ring->last_context = req->ctx;
835 }
836 return 0;
837 }
838
839 return do_switch(req);
840}
841
842static bool contexts_enabled(struct drm_device *dev)
843{
844 return i915.enable_execlists || to_i915(dev)->hw_context_size;
845}
846
847int i915_gem_context_create_ioctl(struct drm_device *dev, void *data,
848 struct drm_file *file)
849{
850 struct drm_i915_gem_context_create *args = data;
851 struct drm_i915_file_private *file_priv = file->driver_priv;
852 struct intel_context *ctx;
853 int ret;
854
855 if (!contexts_enabled(dev))
856 return -ENODEV;
857
858 if (args->pad != 0)
859 return -EINVAL;
860
861 ret = i915_mutex_lock_interruptible(dev);
862 if (ret)
863 return ret;
864
865 ctx = i915_gem_create_context(dev, file_priv);
866 mutex_unlock(&dev->struct_mutex);
867 if (IS_ERR(ctx))
868 return PTR_ERR(ctx);
869
870 args->ctx_id = ctx->user_handle;
871 DRM_DEBUG_DRIVER("HW context %d created\n", args->ctx_id);
872
873 return 0;
874}
875
876int i915_gem_context_destroy_ioctl(struct drm_device *dev, void *data,
877 struct drm_file *file)
878{
879 struct drm_i915_gem_context_destroy *args = data;
880 struct drm_i915_file_private *file_priv = file->driver_priv;
881 struct intel_context *ctx;
882 int ret;
883
884 if (args->pad != 0)
885 return -EINVAL;
886
887 if (args->ctx_id == DEFAULT_CONTEXT_HANDLE)
888 return -ENOENT;
889
890 ret = i915_mutex_lock_interruptible(dev);
891 if (ret)
892 return ret;
893
894 ctx = i915_gem_context_get(file_priv, args->ctx_id);
895 if (IS_ERR(ctx)) {
896 mutex_unlock(&dev->struct_mutex);
897 return PTR_ERR(ctx);
898 }
899
900 idr_remove(&ctx->file_priv->context_idr, ctx->user_handle);
901 i915_gem_context_unreference(ctx);
902 mutex_unlock(&dev->struct_mutex);
903
904 DRM_DEBUG_DRIVER("HW context %d destroyed\n", args->ctx_id);
905 return 0;
906}
907
908int i915_gem_context_getparam_ioctl(struct drm_device *dev, void *data,
909 struct drm_file *file)
910{
911 struct drm_i915_file_private *file_priv = file->driver_priv;
912 struct drm_i915_gem_context_param *args = data;
913 struct intel_context *ctx;
914 int ret;
915
916 ret = i915_mutex_lock_interruptible(dev);
917 if (ret)
918 return ret;
919
920 ctx = i915_gem_context_get(file_priv, args->ctx_id);
921 if (IS_ERR(ctx)) {
922 mutex_unlock(&dev->struct_mutex);
923 return PTR_ERR(ctx);
924 }
925
926 args->size = 0;
927 switch (args->param) {
928 case I915_CONTEXT_PARAM_BAN_PERIOD:
929 args->value = ctx->hang_stats.ban_period_seconds;
930 break;
931 case I915_CONTEXT_PARAM_NO_ZEROMAP:
932 args->value = ctx->flags & CONTEXT_NO_ZEROMAP;
933 break;
934 case I915_CONTEXT_PARAM_GTT_SIZE:
935 if (ctx->ppgtt)
936 args->value = ctx->ppgtt->base.total;
937 else if (to_i915(dev)->mm.aliasing_ppgtt)
938 args->value = to_i915(dev)->mm.aliasing_ppgtt->base.total;
939 else
940 args->value = to_i915(dev)->gtt.base.total;
941 break;
942 default:
943 ret = -EINVAL;
944 break;
945 }
946 mutex_unlock(&dev->struct_mutex);
947
948 return ret;
949}
950
951int i915_gem_context_setparam_ioctl(struct drm_device *dev, void *data,
952 struct drm_file *file)
953{
954 struct drm_i915_file_private *file_priv = file->driver_priv;
955 struct drm_i915_gem_context_param *args = data;
956 struct intel_context *ctx;
957 int ret;
958
959 ret = i915_mutex_lock_interruptible(dev);
960 if (ret)
961 return ret;
962
963 ctx = i915_gem_context_get(file_priv, args->ctx_id);
964 if (IS_ERR(ctx)) {
965 mutex_unlock(&dev->struct_mutex);
966 return PTR_ERR(ctx);
967 }
968
969 switch (args->param) {
970 case I915_CONTEXT_PARAM_BAN_PERIOD:
971 if (args->size)
972 ret = -EINVAL;
973 else if (args->value < ctx->hang_stats.ban_period_seconds &&
974 !capable(CAP_SYS_ADMIN))
975 ret = -EPERM;
976 else
977 ctx->hang_stats.ban_period_seconds = args->value;
978 break;
979 case I915_CONTEXT_PARAM_NO_ZEROMAP:
980 if (args->size) {
981 ret = -EINVAL;
982 } else {
983 ctx->flags &= ~CONTEXT_NO_ZEROMAP;
984 ctx->flags |= args->value ? CONTEXT_NO_ZEROMAP : 0;
985 }
986 break;
987 default:
988 ret = -EINVAL;
989 break;
990 }
991 mutex_unlock(&dev->struct_mutex);
992
993 return ret;
994}
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
76 * GPU. The GPU has loaded its state already and has stored away the gtt
77 * offset of the BO. The GPU is not actively referencing the data at this
78 * offset, but it will on the next context switch. The only way to avoid this
79 * is to do a GPU reset.
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 <drm/drmP.h>
89#include <drm/i915_drm.h>
90#include "i915_drv.h"
91
92/* This is a HW constraint. The value below is the largest known requirement
93 * I've seen in a spec to date, and that was a workaround for a non-shipping
94 * part. It should be safe to decrease this, but it's more future proof as is.
95 */
96#define GEN6_CONTEXT_ALIGN (64<<10)
97#define GEN7_CONTEXT_ALIGN 4096
98
99static void do_ppgtt_cleanup(struct i915_hw_ppgtt *ppgtt)
100{
101 struct drm_device *dev = ppgtt->base.dev;
102 struct drm_i915_private *dev_priv = dev->dev_private;
103 struct i915_address_space *vm = &ppgtt->base;
104
105 if (ppgtt == dev_priv->mm.aliasing_ppgtt ||
106 (list_empty(&vm->active_list) && list_empty(&vm->inactive_list))) {
107 ppgtt->base.cleanup(&ppgtt->base);
108 return;
109 }
110
111 /*
112 * Make sure vmas are unbound before we take down the drm_mm
113 *
114 * FIXME: Proper refcounting should take care of this, this shouldn't be
115 * needed at all.
116 */
117 if (!list_empty(&vm->active_list)) {
118 struct i915_vma *vma;
119
120 list_for_each_entry(vma, &vm->active_list, mm_list)
121 if (WARN_ON(list_empty(&vma->vma_link) ||
122 list_is_singular(&vma->vma_link)))
123 break;
124
125 i915_gem_evict_vm(&ppgtt->base, true);
126 } else {
127 i915_gem_retire_requests(dev);
128 i915_gem_evict_vm(&ppgtt->base, false);
129 }
130
131 ppgtt->base.cleanup(&ppgtt->base);
132}
133
134static void ppgtt_release(struct kref *kref)
135{
136 struct i915_hw_ppgtt *ppgtt =
137 container_of(kref, struct i915_hw_ppgtt, ref);
138
139 do_ppgtt_cleanup(ppgtt);
140 kfree(ppgtt);
141}
142
143static size_t get_context_alignment(struct drm_device *dev)
144{
145 if (IS_GEN6(dev))
146 return GEN6_CONTEXT_ALIGN;
147
148 return GEN7_CONTEXT_ALIGN;
149}
150
151static int get_context_size(struct drm_device *dev)
152{
153 struct drm_i915_private *dev_priv = dev->dev_private;
154 int ret;
155 u32 reg;
156
157 switch (INTEL_INFO(dev)->gen) {
158 case 6:
159 reg = I915_READ(CXT_SIZE);
160 ret = GEN6_CXT_TOTAL_SIZE(reg) * 64;
161 break;
162 case 7:
163 reg = I915_READ(GEN7_CXT_SIZE);
164 if (IS_HASWELL(dev))
165 ret = HSW_CXT_TOTAL_SIZE;
166 else
167 ret = GEN7_CXT_TOTAL_SIZE(reg) * 64;
168 break;
169 case 8:
170 ret = GEN8_CXT_TOTAL_SIZE;
171 break;
172 default:
173 BUG();
174 }
175
176 return ret;
177}
178
179void i915_gem_context_free(struct kref *ctx_ref)
180{
181 struct i915_hw_context *ctx = container_of(ctx_ref,
182 typeof(*ctx), ref);
183 struct i915_hw_ppgtt *ppgtt = NULL;
184
185 if (ctx->obj) {
186 /* We refcount even the aliasing PPGTT to keep the code symmetric */
187 if (USES_PPGTT(ctx->obj->base.dev))
188 ppgtt = ctx_to_ppgtt(ctx);
189
190 /* XXX: Free up the object before tearing down the address space, in
191 * case we're bound in the PPGTT */
192 drm_gem_object_unreference(&ctx->obj->base);
193 }
194
195 if (ppgtt)
196 kref_put(&ppgtt->ref, ppgtt_release);
197 list_del(&ctx->link);
198 kfree(ctx);
199}
200
201static struct i915_hw_ppgtt *
202create_vm_for_ctx(struct drm_device *dev, struct i915_hw_context *ctx)
203{
204 struct i915_hw_ppgtt *ppgtt;
205 int ret;
206
207 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
208 if (!ppgtt)
209 return ERR_PTR(-ENOMEM);
210
211 ret = i915_gem_init_ppgtt(dev, ppgtt);
212 if (ret) {
213 kfree(ppgtt);
214 return ERR_PTR(ret);
215 }
216
217 ppgtt->ctx = ctx;
218 return ppgtt;
219}
220
221static struct i915_hw_context *
222__create_hw_context(struct drm_device *dev,
223 struct drm_i915_file_private *file_priv)
224{
225 struct drm_i915_private *dev_priv = dev->dev_private;
226 struct i915_hw_context *ctx;
227 int ret;
228
229 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
230 if (ctx == NULL)
231 return ERR_PTR(-ENOMEM);
232
233 kref_init(&ctx->ref);
234 list_add_tail(&ctx->link, &dev_priv->context_list);
235
236 if (dev_priv->hw_context_size) {
237 ctx->obj = i915_gem_alloc_object(dev, dev_priv->hw_context_size);
238 if (ctx->obj == NULL) {
239 ret = -ENOMEM;
240 goto err_out;
241 }
242
243 if (INTEL_INFO(dev)->gen >= 7) {
244 ret = i915_gem_object_set_cache_level(ctx->obj,
245 I915_CACHE_L3_LLC);
246 /* Failure shouldn't ever happen this early */
247 if (WARN_ON(ret))
248 goto err_out;
249 }
250 }
251
252 /* Default context will never have a file_priv */
253 if (file_priv != NULL) {
254 ret = idr_alloc(&file_priv->context_idr, ctx,
255 DEFAULT_CONTEXT_ID, 0, GFP_KERNEL);
256 if (ret < 0)
257 goto err_out;
258 } else
259 ret = DEFAULT_CONTEXT_ID;
260
261 ctx->file_priv = file_priv;
262 ctx->id = ret;
263 /* NB: Mark all slices as needing a remap so that when the context first
264 * loads it will restore whatever remap state already exists. If there
265 * is no remap info, it will be a NOP. */
266 ctx->remap_slice = (1 << NUM_L3_SLICES(dev)) - 1;
267
268 return ctx;
269
270err_out:
271 i915_gem_context_unreference(ctx);
272 return ERR_PTR(ret);
273}
274
275/**
276 * The default context needs to exist per ring that uses contexts. It stores the
277 * context state of the GPU for applications that don't utilize HW contexts, as
278 * well as an idle case.
279 */
280static struct i915_hw_context *
281i915_gem_create_context(struct drm_device *dev,
282 struct drm_i915_file_private *file_priv,
283 bool create_vm)
284{
285 const bool is_global_default_ctx = file_priv == NULL;
286 struct drm_i915_private *dev_priv = dev->dev_private;
287 struct i915_hw_context *ctx;
288 int ret = 0;
289
290 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
291
292 ctx = __create_hw_context(dev, file_priv);
293 if (IS_ERR(ctx))
294 return ctx;
295
296 if (is_global_default_ctx && ctx->obj) {
297 /* We may need to do things with the shrinker which
298 * require us to immediately switch back to the default
299 * context. This can cause a problem as pinning the
300 * default context also requires GTT space which may not
301 * be available. To avoid this we always pin the default
302 * context.
303 */
304 ret = i915_gem_obj_ggtt_pin(ctx->obj,
305 get_context_alignment(dev), 0);
306 if (ret) {
307 DRM_DEBUG_DRIVER("Couldn't pin %d\n", ret);
308 goto err_destroy;
309 }
310 }
311
312 if (create_vm) {
313 struct i915_hw_ppgtt *ppgtt = create_vm_for_ctx(dev, ctx);
314
315 if (IS_ERR_OR_NULL(ppgtt)) {
316 DRM_DEBUG_DRIVER("PPGTT setup failed (%ld)\n",
317 PTR_ERR(ppgtt));
318 ret = PTR_ERR(ppgtt);
319 goto err_unpin;
320 } else
321 ctx->vm = &ppgtt->base;
322
323 /* This case is reserved for the global default context and
324 * should only happen once. */
325 if (is_global_default_ctx) {
326 if (WARN_ON(dev_priv->mm.aliasing_ppgtt)) {
327 ret = -EEXIST;
328 goto err_unpin;
329 }
330
331 dev_priv->mm.aliasing_ppgtt = ppgtt;
332 }
333 } else if (USES_PPGTT(dev)) {
334 /* For platforms which only have aliasing PPGTT, we fake the
335 * address space and refcounting. */
336 ctx->vm = &dev_priv->mm.aliasing_ppgtt->base;
337 kref_get(&dev_priv->mm.aliasing_ppgtt->ref);
338 } else
339 ctx->vm = &dev_priv->gtt.base;
340
341 return ctx;
342
343err_unpin:
344 if (is_global_default_ctx && ctx->obj)
345 i915_gem_object_ggtt_unpin(ctx->obj);
346err_destroy:
347 i915_gem_context_unreference(ctx);
348 return ERR_PTR(ret);
349}
350
351void i915_gem_context_reset(struct drm_device *dev)
352{
353 struct drm_i915_private *dev_priv = dev->dev_private;
354 int i;
355
356 /* Prevent the hardware from restoring the last context (which hung) on
357 * the next switch */
358 for (i = 0; i < I915_NUM_RINGS; i++) {
359 struct intel_ring_buffer *ring = &dev_priv->ring[i];
360 struct i915_hw_context *dctx = ring->default_context;
361
362 /* Do a fake switch to the default context */
363 if (ring->last_context == dctx)
364 continue;
365
366 if (!ring->last_context)
367 continue;
368
369 if (dctx->obj && i == RCS) {
370 WARN_ON(i915_gem_obj_ggtt_pin(dctx->obj,
371 get_context_alignment(dev), 0));
372 /* Fake a finish/inactive */
373 dctx->obj->base.write_domain = 0;
374 dctx->obj->active = 0;
375 }
376
377 i915_gem_context_unreference(ring->last_context);
378 i915_gem_context_reference(dctx);
379 ring->last_context = dctx;
380 }
381}
382
383int i915_gem_context_init(struct drm_device *dev)
384{
385 struct drm_i915_private *dev_priv = dev->dev_private;
386 struct i915_hw_context *ctx;
387 int i;
388
389 /* Init should only be called once per module load. Eventually the
390 * restriction on the context_disabled check can be loosened. */
391 if (WARN_ON(dev_priv->ring[RCS].default_context))
392 return 0;
393
394 if (HAS_HW_CONTEXTS(dev)) {
395 dev_priv->hw_context_size = round_up(get_context_size(dev), 4096);
396 if (dev_priv->hw_context_size > (1<<20)) {
397 DRM_DEBUG_DRIVER("Disabling HW Contexts; invalid size %d\n",
398 dev_priv->hw_context_size);
399 dev_priv->hw_context_size = 0;
400 }
401 }
402
403 ctx = i915_gem_create_context(dev, NULL, USES_PPGTT(dev));
404 if (IS_ERR(ctx)) {
405 DRM_ERROR("Failed to create default global context (error %ld)\n",
406 PTR_ERR(ctx));
407 return PTR_ERR(ctx);
408 }
409
410 /* NB: RCS will hold a ref for all rings */
411 for (i = 0; i < I915_NUM_RINGS; i++)
412 dev_priv->ring[i].default_context = ctx;
413
414 DRM_DEBUG_DRIVER("%s context support initialized\n", dev_priv->hw_context_size ? "HW" : "fake");
415 return 0;
416}
417
418void i915_gem_context_fini(struct drm_device *dev)
419{
420 struct drm_i915_private *dev_priv = dev->dev_private;
421 struct i915_hw_context *dctx = dev_priv->ring[RCS].default_context;
422 int i;
423
424 if (dctx->obj) {
425 /* The only known way to stop the gpu from accessing the hw context is
426 * to reset it. Do this as the very last operation to avoid confusing
427 * other code, leading to spurious errors. */
428 intel_gpu_reset(dev);
429
430 /* When default context is created and switched to, base object refcount
431 * will be 2 (+1 from object creation and +1 from do_switch()).
432 * i915_gem_context_fini() will be called after gpu_idle() has switched
433 * to default context. So we need to unreference the base object once
434 * to offset the do_switch part, so that i915_gem_context_unreference()
435 * can then free the base object correctly. */
436 WARN_ON(!dev_priv->ring[RCS].last_context);
437 if (dev_priv->ring[RCS].last_context == dctx) {
438 /* Fake switch to NULL context */
439 WARN_ON(dctx->obj->active);
440 i915_gem_object_ggtt_unpin(dctx->obj);
441 i915_gem_context_unreference(dctx);
442 dev_priv->ring[RCS].last_context = NULL;
443 }
444 }
445
446 for (i = 0; i < I915_NUM_RINGS; i++) {
447 struct intel_ring_buffer *ring = &dev_priv->ring[i];
448
449 if (ring->last_context)
450 i915_gem_context_unreference(ring->last_context);
451
452 ring->default_context = NULL;
453 ring->last_context = NULL;
454 }
455
456 i915_gem_object_ggtt_unpin(dctx->obj);
457 i915_gem_context_unreference(dctx);
458}
459
460int i915_gem_context_enable(struct drm_i915_private *dev_priv)
461{
462 struct intel_ring_buffer *ring;
463 int ret, i;
464
465 /* This is the only place the aliasing PPGTT gets enabled, which means
466 * it has to happen before we bail on reset */
467 if (dev_priv->mm.aliasing_ppgtt) {
468 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
469 ppgtt->enable(ppgtt);
470 }
471
472 /* FIXME: We should make this work, even in reset */
473 if (i915_reset_in_progress(&dev_priv->gpu_error))
474 return 0;
475
476 BUG_ON(!dev_priv->ring[RCS].default_context);
477
478 for_each_ring(ring, dev_priv, i) {
479 ret = i915_switch_context(ring, ring->default_context);
480 if (ret)
481 return ret;
482 }
483
484 return 0;
485}
486
487static int context_idr_cleanup(int id, void *p, void *data)
488{
489 struct i915_hw_context *ctx = p;
490
491 /* Ignore the default context because close will handle it */
492 if (i915_gem_context_is_default(ctx))
493 return 0;
494
495 i915_gem_context_unreference(ctx);
496 return 0;
497}
498
499int i915_gem_context_open(struct drm_device *dev, struct drm_file *file)
500{
501 struct drm_i915_file_private *file_priv = file->driver_priv;
502
503 idr_init(&file_priv->context_idr);
504
505 mutex_lock(&dev->struct_mutex);
506 file_priv->private_default_ctx =
507 i915_gem_create_context(dev, file_priv, USES_FULL_PPGTT(dev));
508 mutex_unlock(&dev->struct_mutex);
509
510 if (IS_ERR(file_priv->private_default_ctx)) {
511 idr_destroy(&file_priv->context_idr);
512 return PTR_ERR(file_priv->private_default_ctx);
513 }
514
515 return 0;
516}
517
518void i915_gem_context_close(struct drm_device *dev, struct drm_file *file)
519{
520 struct drm_i915_file_private *file_priv = file->driver_priv;
521
522 idr_for_each(&file_priv->context_idr, context_idr_cleanup, NULL);
523 idr_destroy(&file_priv->context_idr);
524
525 i915_gem_context_unreference(file_priv->private_default_ctx);
526}
527
528struct i915_hw_context *
529i915_gem_context_get(struct drm_i915_file_private *file_priv, u32 id)
530{
531 struct i915_hw_context *ctx;
532
533 ctx = (struct i915_hw_context *)idr_find(&file_priv->context_idr, id);
534 if (!ctx)
535 return ERR_PTR(-ENOENT);
536
537 return ctx;
538}
539
540static inline int
541mi_set_context(struct intel_ring_buffer *ring,
542 struct i915_hw_context *new_context,
543 u32 hw_flags)
544{
545 int ret;
546
547 /* w/a: If Flush TLB Invalidation Mode is enabled, driver must do a TLB
548 * invalidation prior to MI_SET_CONTEXT. On GEN6 we don't set the value
549 * explicitly, so we rely on the value at ring init, stored in
550 * itlb_before_ctx_switch.
551 */
552 if (IS_GEN6(ring->dev) && ring->itlb_before_ctx_switch) {
553 ret = ring->flush(ring, I915_GEM_GPU_DOMAINS, 0);
554 if (ret)
555 return ret;
556 }
557
558 ret = intel_ring_begin(ring, 6);
559 if (ret)
560 return ret;
561
562 /* WaProgramMiArbOnOffAroundMiSetContext:ivb,vlv,hsw */
563 if (IS_GEN7(ring->dev))
564 intel_ring_emit(ring, MI_ARB_ON_OFF | MI_ARB_DISABLE);
565 else
566 intel_ring_emit(ring, MI_NOOP);
567
568 intel_ring_emit(ring, MI_NOOP);
569 intel_ring_emit(ring, MI_SET_CONTEXT);
570 intel_ring_emit(ring, i915_gem_obj_ggtt_offset(new_context->obj) |
571 MI_MM_SPACE_GTT |
572 MI_SAVE_EXT_STATE_EN |
573 MI_RESTORE_EXT_STATE_EN |
574 hw_flags);
575 /*
576 * w/a: MI_SET_CONTEXT must always be followed by MI_NOOP
577 * WaMiSetContext_Hang:snb,ivb,vlv
578 */
579 intel_ring_emit(ring, MI_NOOP);
580
581 if (IS_GEN7(ring->dev))
582 intel_ring_emit(ring, MI_ARB_ON_OFF | MI_ARB_ENABLE);
583 else
584 intel_ring_emit(ring, MI_NOOP);
585
586 intel_ring_advance(ring);
587
588 return ret;
589}
590
591static int do_switch(struct intel_ring_buffer *ring,
592 struct i915_hw_context *to)
593{
594 struct drm_i915_private *dev_priv = ring->dev->dev_private;
595 struct i915_hw_context *from = ring->last_context;
596 struct i915_hw_ppgtt *ppgtt = ctx_to_ppgtt(to);
597 u32 hw_flags = 0;
598 int ret, i;
599
600 if (from != NULL && ring == &dev_priv->ring[RCS]) {
601 BUG_ON(from->obj == NULL);
602 BUG_ON(!i915_gem_obj_is_pinned(from->obj));
603 }
604
605 if (from == to && from->last_ring == ring && !to->remap_slice)
606 return 0;
607
608 /* Trying to pin first makes error handling easier. */
609 if (ring == &dev_priv->ring[RCS]) {
610 ret = i915_gem_obj_ggtt_pin(to->obj,
611 get_context_alignment(ring->dev), 0);
612 if (ret)
613 return ret;
614 }
615
616 /*
617 * Pin can switch back to the default context if we end up calling into
618 * evict_everything - as a last ditch gtt defrag effort that also
619 * switches to the default context. Hence we need to reload from here.
620 */
621 from = ring->last_context;
622
623 if (USES_FULL_PPGTT(ring->dev)) {
624 ret = ppgtt->switch_mm(ppgtt, ring, false);
625 if (ret)
626 goto unpin_out;
627 }
628
629 if (ring != &dev_priv->ring[RCS]) {
630 if (from)
631 i915_gem_context_unreference(from);
632 goto done;
633 }
634
635 /*
636 * Clear this page out of any CPU caches for coherent swap-in/out. Note
637 * that thanks to write = false in this call and us not setting any gpu
638 * write domains when putting a context object onto the active list
639 * (when switching away from it), this won't block.
640 *
641 * XXX: We need a real interface to do this instead of trickery.
642 */
643 ret = i915_gem_object_set_to_gtt_domain(to->obj, false);
644 if (ret)
645 goto unpin_out;
646
647 if (!to->obj->has_global_gtt_mapping) {
648 struct i915_vma *vma = i915_gem_obj_to_vma(to->obj,
649 &dev_priv->gtt.base);
650 vma->bind_vma(vma, to->obj->cache_level, GLOBAL_BIND);
651 }
652
653 if (!to->is_initialized || i915_gem_context_is_default(to))
654 hw_flags |= MI_RESTORE_INHIBIT;
655
656 ret = mi_set_context(ring, to, hw_flags);
657 if (ret)
658 goto unpin_out;
659
660 for (i = 0; i < MAX_L3_SLICES; i++) {
661 if (!(to->remap_slice & (1<<i)))
662 continue;
663
664 ret = i915_gem_l3_remap(ring, i);
665 /* If it failed, try again next round */
666 if (ret)
667 DRM_DEBUG_DRIVER("L3 remapping failed\n");
668 else
669 to->remap_slice &= ~(1<<i);
670 }
671
672 /* The backing object for the context is done after switching to the
673 * *next* context. Therefore we cannot retire the previous context until
674 * the next context has already started running. In fact, the below code
675 * is a bit suboptimal because the retiring can occur simply after the
676 * MI_SET_CONTEXT instead of when the next seqno has completed.
677 */
678 if (from != NULL) {
679 from->obj->base.read_domains = I915_GEM_DOMAIN_INSTRUCTION;
680 i915_vma_move_to_active(i915_gem_obj_to_ggtt(from->obj), ring);
681 /* As long as MI_SET_CONTEXT is serializing, ie. it flushes the
682 * whole damn pipeline, we don't need to explicitly mark the
683 * object dirty. The only exception is that the context must be
684 * correct in case the object gets swapped out. Ideally we'd be
685 * able to defer doing this until we know the object would be
686 * swapped, but there is no way to do that yet.
687 */
688 from->obj->dirty = 1;
689 BUG_ON(from->obj->ring != ring);
690
691 /* obj is kept alive until the next request by its active ref */
692 i915_gem_object_ggtt_unpin(from->obj);
693 i915_gem_context_unreference(from);
694 }
695
696 to->is_initialized = true;
697
698done:
699 i915_gem_context_reference(to);
700 ring->last_context = to;
701 to->last_ring = ring;
702
703 return 0;
704
705unpin_out:
706 if (ring->id == RCS)
707 i915_gem_object_ggtt_unpin(to->obj);
708 return ret;
709}
710
711/**
712 * i915_switch_context() - perform a GPU context switch.
713 * @ring: ring for which we'll execute the context switch
714 * @to: the context to switch to
715 *
716 * The context life cycle is simple. The context refcount is incremented and
717 * decremented by 1 and create and destroy. If the context is in use by the GPU,
718 * it will have a refoucnt > 1. This allows us to destroy the context abstract
719 * object while letting the normal object tracking destroy the backing BO.
720 */
721int i915_switch_context(struct intel_ring_buffer *ring,
722 struct i915_hw_context *to)
723{
724 struct drm_i915_private *dev_priv = ring->dev->dev_private;
725
726 WARN_ON(!mutex_is_locked(&dev_priv->dev->struct_mutex));
727
728 if (to->obj == NULL) { /* We have the fake context */
729 if (to != ring->last_context) {
730 i915_gem_context_reference(to);
731 if (ring->last_context)
732 i915_gem_context_unreference(ring->last_context);
733 ring->last_context = to;
734 }
735 return 0;
736 }
737
738 return do_switch(ring, to);
739}
740
741static bool hw_context_enabled(struct drm_device *dev)
742{
743 return to_i915(dev)->hw_context_size;
744}
745
746int i915_gem_context_create_ioctl(struct drm_device *dev, void *data,
747 struct drm_file *file)
748{
749 struct drm_i915_gem_context_create *args = data;
750 struct drm_i915_file_private *file_priv = file->driver_priv;
751 struct i915_hw_context *ctx;
752 int ret;
753
754 if (!hw_context_enabled(dev))
755 return -ENODEV;
756
757 ret = i915_mutex_lock_interruptible(dev);
758 if (ret)
759 return ret;
760
761 ctx = i915_gem_create_context(dev, file_priv, USES_FULL_PPGTT(dev));
762 mutex_unlock(&dev->struct_mutex);
763 if (IS_ERR(ctx))
764 return PTR_ERR(ctx);
765
766 args->ctx_id = ctx->id;
767 DRM_DEBUG_DRIVER("HW context %d created\n", args->ctx_id);
768
769 return 0;
770}
771
772int i915_gem_context_destroy_ioctl(struct drm_device *dev, void *data,
773 struct drm_file *file)
774{
775 struct drm_i915_gem_context_destroy *args = data;
776 struct drm_i915_file_private *file_priv = file->driver_priv;
777 struct i915_hw_context *ctx;
778 int ret;
779
780 if (args->ctx_id == DEFAULT_CONTEXT_ID)
781 return -ENOENT;
782
783 ret = i915_mutex_lock_interruptible(dev);
784 if (ret)
785 return ret;
786
787 ctx = i915_gem_context_get(file_priv, args->ctx_id);
788 if (IS_ERR(ctx)) {
789 mutex_unlock(&dev->struct_mutex);
790 return PTR_ERR(ctx);
791 }
792
793 idr_remove(&ctx->file_priv->context_idr, ctx->id);
794 i915_gem_context_unreference(ctx);
795 mutex_unlock(&dev->struct_mutex);
796
797 DRM_DEBUG_DRIVER("HW context %d destroyed\n", args->ctx_id);
798 return 0;
799}