2 * SPDX-License-Identifier: MIT
4 * Copyright © 2011-2012 Intel Corporation
8 * This file implements HW context support. On gen5+ a HW context consists of an
9 * opaque GPU object which is referenced at times of context saves and restores.
10 * With RC6 enabled, the context is also referenced as the GPU enters and exists
11 * from RC6 (GPU has it's own internal power context, except on gen5). Though
12 * something like a context does exist for the media ring, the code only
13 * supports contexts for the render ring.
15 * In software, there is a distinction between contexts created by the user,
16 * and the default HW context. The default HW context is used by GPU clients
17 * that do not request setup of their own hardware context. The default
18 * context's state is never restored to help prevent programming errors. This
19 * would happen if a client ran and piggy-backed off another clients GPU state.
20 * The default context only exists to give the GPU some offset to load as the
21 * current to invoke a save of the context we actually care about. In fact, the
22 * code could likely be constructed, albeit in a more complicated fashion, to
23 * never use the default context, though that limits the driver's ability to
24 * swap out, and/or destroy other contexts.
26 * All other contexts are created as a request by the GPU client. These contexts
27 * store GPU state, and thus allow GPU clients to not re-emit state (and
28 * potentially query certain state) at any time. The kernel driver makes
29 * certain that the appropriate commands are inserted.
31 * The context life cycle is semi-complicated in that context BOs may live
32 * longer than the context itself because of the way the hardware, and object
33 * tracking works. Below is a very crude representation of the state machine
34 * describing the context life.
35 * refcount pincount active
36 * S0: initial state 0 0 0
37 * S1: context created 1 0 0
38 * S2: context is currently running 2 1 X
39 * S3: GPU referenced, but not current 2 0 1
40 * S4: context is current, but destroyed 1 1 0
41 * S5: like S3, but destroyed 1 0 1
43 * The most common (but not all) transitions:
44 * S0->S1: client creates a context
45 * S1->S2: client submits execbuf with context
46 * S2->S3: other clients submits execbuf with context
47 * S3->S1: context object was retired
48 * S3->S2: clients submits another execbuf
49 * S2->S4: context destroy called with current context
50 * S3->S5->S0: destroy path
51 * S4->S5->S0: destroy path on current context
53 * There are two confusing terms used above:
54 * The "current context" means the context which is currently running on the
55 * GPU. The GPU has loaded its state already and has stored away the gtt
56 * offset of the BO. The GPU is not actively referencing the data at this
57 * offset, but it will on the next context switch. The only way to avoid this
58 * is to do a GPU reset.
60 * An "active context' is one which was previously the "current context" and is
61 * on the active list waiting for the next context switch to occur. Until this
62 * happens, the object must remain at the same gtt offset. It is therefore
63 * possible to destroy a context, but it is still active.
67 #include <linux/log2.h>
68 #include <linux/nospec.h>
70 #include <drm/i915_drm.h>
72 #include "gt/intel_lrc_reg.h"
74 #include "i915_gem_context.h"
75 #include "i915_globals.h"
76 #include "i915_trace.h"
77 #include "i915_user_extensions.h"
79 #define ALL_L3_SLICES(dev) (1 << NUM_L3_SLICES(dev)) - 1
81 static struct i915_global_gem_context {
82 struct i915_global base;
83 struct kmem_cache *slab_luts;
86 struct i915_lut_handle *i915_lut_handle_alloc(void)
88 return kmem_cache_alloc(global.slab_luts, GFP_KERNEL);
91 void i915_lut_handle_free(struct i915_lut_handle *lut)
93 return kmem_cache_free(global.slab_luts, lut);
96 static void lut_close(struct i915_gem_context *ctx)
98 struct radix_tree_iter iter;
101 lockdep_assert_held(&ctx->mutex);
104 radix_tree_for_each_slot(slot, &ctx->handles_vma, &iter, 0) {
105 struct i915_vma *vma = rcu_dereference_raw(*slot);
106 struct drm_i915_gem_object *obj = vma->obj;
107 struct i915_lut_handle *lut;
109 if (!kref_get_unless_zero(&obj->base.refcount))
113 i915_gem_object_lock(obj);
114 list_for_each_entry(lut, &obj->lut_list, obj_link) {
118 if (lut->handle != iter.index)
121 list_del(&lut->obj_link);
124 i915_gem_object_unlock(obj);
127 if (&lut->obj_link != &obj->lut_list) {
128 i915_lut_handle_free(lut);
129 radix_tree_iter_delete(&ctx->handles_vma, &iter, slot);
130 if (atomic_dec_and_test(&vma->open_count) &&
131 !i915_vma_is_ggtt(vma))
133 i915_gem_object_put(obj);
136 i915_gem_object_put(obj);
141 static struct intel_context *
142 lookup_user_engine(struct i915_gem_context *ctx,
144 const struct i915_engine_class_instance *ci)
145 #define LOOKUP_USER_INDEX BIT(0)
149 if (!!(flags & LOOKUP_USER_INDEX) != i915_gem_context_user_engines(ctx))
150 return ERR_PTR(-EINVAL);
152 if (!i915_gem_context_user_engines(ctx)) {
153 struct intel_engine_cs *engine;
155 engine = intel_engine_lookup_user(ctx->i915,
157 ci->engine_instance);
159 return ERR_PTR(-EINVAL);
163 idx = ci->engine_instance;
166 return i915_gem_context_get_engine(ctx, idx);
169 static inline int new_hw_id(struct drm_i915_private *i915, gfp_t gfp)
173 lockdep_assert_held(&i915->contexts.mutex);
175 if (INTEL_GEN(i915) >= 11)
176 max = GEN11_MAX_CONTEXT_HW_ID;
177 else if (USES_GUC_SUBMISSION(i915))
179 * When using GuC in proxy submission, GuC consumes the
180 * highest bit in the context id to indicate proxy submission.
182 max = MAX_GUC_CONTEXT_HW_ID;
184 max = MAX_CONTEXT_HW_ID;
186 return ida_simple_get(&i915->contexts.hw_ida, 0, max, gfp);
189 static int steal_hw_id(struct drm_i915_private *i915)
191 struct i915_gem_context *ctx, *cn;
195 lockdep_assert_held(&i915->contexts.mutex);
197 list_for_each_entry_safe(ctx, cn,
198 &i915->contexts.hw_id_list, hw_id_link) {
199 if (atomic_read(&ctx->hw_id_pin_count)) {
200 list_move_tail(&ctx->hw_id_link, &pinned);
204 GEM_BUG_ON(!ctx->hw_id); /* perma-pinned kernel context */
205 list_del_init(&ctx->hw_id_link);
211 * Remember how far we got up on the last repossesion scan, so the
212 * list is kept in a "least recently scanned" order.
214 list_splice_tail(&pinned, &i915->contexts.hw_id_list);
218 static int assign_hw_id(struct drm_i915_private *i915, unsigned int *out)
222 lockdep_assert_held(&i915->contexts.mutex);
225 * We prefer to steal/stall ourselves and our users over that of the
226 * entire system. That may be a little unfair to our users, and
227 * even hurt high priority clients. The choice is whether to oomkill
228 * something else, or steal a context id.
230 ret = new_hw_id(i915, GFP_KERNEL | __GFP_RETRY_MAYFAIL | __GFP_NOWARN);
231 if (unlikely(ret < 0)) {
232 ret = steal_hw_id(i915);
233 if (ret < 0) /* once again for the correct errno code */
234 ret = new_hw_id(i915, GFP_KERNEL);
243 static void release_hw_id(struct i915_gem_context *ctx)
245 struct drm_i915_private *i915 = ctx->i915;
247 if (list_empty(&ctx->hw_id_link))
250 mutex_lock(&i915->contexts.mutex);
251 if (!list_empty(&ctx->hw_id_link)) {
252 ida_simple_remove(&i915->contexts.hw_ida, ctx->hw_id);
253 list_del_init(&ctx->hw_id_link);
255 mutex_unlock(&i915->contexts.mutex);
258 static void __free_engines(struct i915_gem_engines *e, unsigned int count)
261 if (!e->engines[count])
264 intel_context_put(e->engines[count]);
269 static void free_engines(struct i915_gem_engines *e)
271 __free_engines(e, e->num_engines);
274 static void free_engines_rcu(struct rcu_head *rcu)
276 free_engines(container_of(rcu, struct i915_gem_engines, rcu));
279 static struct i915_gem_engines *default_engines(struct i915_gem_context *ctx)
281 struct intel_engine_cs *engine;
282 struct i915_gem_engines *e;
283 enum intel_engine_id id;
285 e = kzalloc(struct_size(e, engines, I915_NUM_ENGINES), GFP_KERNEL);
287 return ERR_PTR(-ENOMEM);
289 init_rcu_head(&e->rcu);
290 for_each_engine(engine, ctx->i915, id) {
291 struct intel_context *ce;
293 ce = intel_context_create(ctx, engine);
295 __free_engines(e, id);
306 static void i915_gem_context_free(struct i915_gem_context *ctx)
308 lockdep_assert_held(&ctx->i915->drm.struct_mutex);
309 GEM_BUG_ON(!i915_gem_context_is_closed(ctx));
313 i915_vm_put(ctx->vm);
315 free_engines(rcu_access_pointer(ctx->engines));
316 mutex_destroy(&ctx->engines_mutex);
318 kfree(ctx->jump_whitelist);
321 i915_timeline_put(ctx->timeline);
326 list_del(&ctx->link);
327 mutex_destroy(&ctx->mutex);
332 static void contexts_free(struct drm_i915_private *i915)
334 struct llist_node *freed = llist_del_all(&i915->contexts.free_list);
335 struct i915_gem_context *ctx, *cn;
337 lockdep_assert_held(&i915->drm.struct_mutex);
339 llist_for_each_entry_safe(ctx, cn, freed, free_link)
340 i915_gem_context_free(ctx);
343 static void contexts_free_first(struct drm_i915_private *i915)
345 struct i915_gem_context *ctx;
346 struct llist_node *freed;
348 lockdep_assert_held(&i915->drm.struct_mutex);
350 freed = llist_del_first(&i915->contexts.free_list);
354 ctx = container_of(freed, typeof(*ctx), free_link);
355 i915_gem_context_free(ctx);
358 static void contexts_free_worker(struct work_struct *work)
360 struct drm_i915_private *i915 =
361 container_of(work, typeof(*i915), contexts.free_work);
363 mutex_lock(&i915->drm.struct_mutex);
365 mutex_unlock(&i915->drm.struct_mutex);
368 void i915_gem_context_release(struct kref *ref)
370 struct i915_gem_context *ctx = container_of(ref, typeof(*ctx), ref);
371 struct drm_i915_private *i915 = ctx->i915;
373 trace_i915_context_free(ctx);
374 if (llist_add(&ctx->free_link, &i915->contexts.free_list))
375 queue_work(i915->wq, &i915->contexts.free_work);
378 static void context_close(struct i915_gem_context *ctx)
380 mutex_lock(&ctx->mutex);
382 i915_gem_context_set_closed(ctx);
383 ctx->file_priv = ERR_PTR(-EBADF);
386 * This context will never again be assinged to HW, so we can
387 * reuse its ID for the next context.
392 * The LUT uses the VMA as a backpointer to unref the object,
393 * so we need to clear the LUT before we close all the VMA (inside
398 mutex_unlock(&ctx->mutex);
399 i915_gem_context_put(ctx);
402 static u32 default_desc_template(const struct drm_i915_private *i915,
403 const struct i915_address_space *vm)
408 desc = GEN8_CTX_VALID | GEN8_CTX_PRIVILEGE;
410 address_mode = INTEL_LEGACY_32B_CONTEXT;
411 if (vm && i915_vm_is_4lvl(vm))
412 address_mode = INTEL_LEGACY_64B_CONTEXT;
413 desc |= address_mode << GEN8_CTX_ADDRESSING_MODE_SHIFT;
416 desc |= GEN8_CTX_L3LLC_COHERENT;
418 /* TODO: WaDisableLiteRestore when we start using semaphore
419 * signalling between Command Streamers
420 * ring->ctx_desc_template |= GEN8_CTX_FORCE_RESTORE;
426 static struct i915_gem_context *
427 __create_context(struct drm_i915_private *i915)
429 struct i915_gem_context *ctx;
430 struct i915_gem_engines *e;
434 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
436 return ERR_PTR(-ENOMEM);
438 kref_init(&ctx->ref);
439 list_add_tail(&ctx->link, &i915->contexts.list);
441 ctx->sched.priority = I915_USER_PRIORITY(I915_PRIORITY_NORMAL);
442 mutex_init(&ctx->mutex);
444 mutex_init(&ctx->engines_mutex);
445 e = default_engines(ctx);
450 RCU_INIT_POINTER(ctx->engines, e);
452 INIT_RADIX_TREE(&ctx->handles_vma, GFP_KERNEL);
453 INIT_LIST_HEAD(&ctx->hw_id_link);
455 /* NB: Mark all slices as needing a remap so that when the context first
456 * loads it will restore whatever remap state already exists. If there
457 * is no remap info, it will be a NOP. */
458 ctx->remap_slice = ALL_L3_SLICES(i915);
460 i915_gem_context_set_bannable(ctx);
461 i915_gem_context_set_recoverable(ctx);
463 ctx->ring_size = 4 * PAGE_SIZE;
465 default_desc_template(i915, &i915->mm.aliasing_ppgtt->vm);
467 for (i = 0; i < ARRAY_SIZE(ctx->hang_timestamp); i++)
468 ctx->hang_timestamp[i] = jiffies - CONTEXT_FAST_HANG_JIFFIES;
470 ctx->jump_whitelist = NULL;
471 ctx->jump_whitelist_cmds = 0;
480 static struct i915_address_space *
481 __set_ppgtt(struct i915_gem_context *ctx, struct i915_address_space *vm)
483 struct i915_address_space *old = ctx->vm;
485 ctx->vm = i915_vm_get(vm);
486 ctx->desc_template = default_desc_template(ctx->i915, vm);
491 static void __assign_ppgtt(struct i915_gem_context *ctx,
492 struct i915_address_space *vm)
497 vm = __set_ppgtt(ctx, vm);
502 static struct i915_gem_context *
503 i915_gem_create_context(struct drm_i915_private *dev_priv, unsigned int flags)
505 struct i915_gem_context *ctx;
507 lockdep_assert_held(&dev_priv->drm.struct_mutex);
509 if (flags & I915_CONTEXT_CREATE_FLAGS_SINGLE_TIMELINE &&
510 !HAS_EXECLISTS(dev_priv))
511 return ERR_PTR(-EINVAL);
513 /* Reap the most stale context */
514 contexts_free_first(dev_priv);
516 ctx = __create_context(dev_priv);
520 if (HAS_FULL_PPGTT(dev_priv)) {
521 struct i915_ppgtt *ppgtt;
523 ppgtt = i915_ppgtt_create(dev_priv);
525 DRM_DEBUG_DRIVER("PPGTT setup failed (%ld)\n",
528 return ERR_CAST(ppgtt);
531 __assign_ppgtt(ctx, &ppgtt->vm);
532 i915_vm_put(&ppgtt->vm);
535 if (flags & I915_CONTEXT_CREATE_FLAGS_SINGLE_TIMELINE) {
536 struct i915_timeline *timeline;
538 timeline = i915_timeline_create(dev_priv, NULL);
539 if (IS_ERR(timeline)) {
541 return ERR_CAST(timeline);
544 ctx->timeline = timeline;
547 trace_i915_context_create(ctx);
553 * i915_gem_context_create_gvt - create a GVT GEM context
556 * This function is used to create a GVT specific GEM context.
559 * pointer to i915_gem_context on success, error pointer if failed
562 struct i915_gem_context *
563 i915_gem_context_create_gvt(struct drm_device *dev)
565 struct i915_gem_context *ctx;
568 if (!IS_ENABLED(CONFIG_DRM_I915_GVT))
569 return ERR_PTR(-ENODEV);
571 ret = i915_mutex_lock_interruptible(dev);
575 ctx = i915_gem_create_context(to_i915(dev), 0);
579 ret = i915_gem_context_pin_hw_id(ctx);
586 ctx->file_priv = ERR_PTR(-EBADF);
587 i915_gem_context_set_closed(ctx); /* not user accessible */
588 i915_gem_context_clear_bannable(ctx);
589 i915_gem_context_set_force_single_submission(ctx);
590 if (!USES_GUC_SUBMISSION(to_i915(dev)))
591 ctx->ring_size = 512 * PAGE_SIZE; /* Max ring buffer size */
593 GEM_BUG_ON(i915_gem_context_is_kernel(ctx));
595 mutex_unlock(&dev->struct_mutex);
600 destroy_kernel_context(struct i915_gem_context **ctxp)
602 struct i915_gem_context *ctx;
604 /* Keep the context ref so that we can free it immediately ourselves */
605 ctx = i915_gem_context_get(fetch_and_zero(ctxp));
606 GEM_BUG_ON(!i915_gem_context_is_kernel(ctx));
609 i915_gem_context_free(ctx);
612 struct i915_gem_context *
613 i915_gem_context_create_kernel(struct drm_i915_private *i915, int prio)
615 struct i915_gem_context *ctx;
618 ctx = i915_gem_create_context(i915, 0);
622 err = i915_gem_context_pin_hw_id(ctx);
624 destroy_kernel_context(&ctx);
628 i915_gem_context_clear_bannable(ctx);
629 ctx->sched.priority = I915_USER_PRIORITY(prio);
630 ctx->ring_size = PAGE_SIZE;
632 GEM_BUG_ON(!i915_gem_context_is_kernel(ctx));
637 static void init_contexts(struct drm_i915_private *i915)
639 mutex_init(&i915->contexts.mutex);
640 INIT_LIST_HEAD(&i915->contexts.list);
642 /* Using the simple ida interface, the max is limited by sizeof(int) */
643 BUILD_BUG_ON(MAX_CONTEXT_HW_ID > INT_MAX);
644 BUILD_BUG_ON(GEN11_MAX_CONTEXT_HW_ID > INT_MAX);
645 ida_init(&i915->contexts.hw_ida);
646 INIT_LIST_HEAD(&i915->contexts.hw_id_list);
648 INIT_WORK(&i915->contexts.free_work, contexts_free_worker);
649 init_llist_head(&i915->contexts.free_list);
652 static bool needs_preempt_context(struct drm_i915_private *i915)
654 return HAS_EXECLISTS(i915);
657 int i915_gem_contexts_init(struct drm_i915_private *dev_priv)
659 struct i915_gem_context *ctx;
661 /* Reassure ourselves we are only called once */
662 GEM_BUG_ON(dev_priv->kernel_context);
663 GEM_BUG_ON(dev_priv->preempt_context);
665 intel_engine_init_ctx_wa(dev_priv->engine[RCS0]);
666 init_contexts(dev_priv);
668 /* lowest priority; idle task */
669 ctx = i915_gem_context_create_kernel(dev_priv, I915_PRIORITY_MIN);
671 DRM_ERROR("Failed to create default global context\n");
675 * For easy recognisablity, we want the kernel context to be 0 and then
676 * all user contexts will have non-zero hw_id. Kernel contexts are
677 * permanently pinned, so that we never suffer a stall and can
678 * use them from any allocation context (e.g. for evicting other
679 * contexts and from inside the shrinker).
681 GEM_BUG_ON(ctx->hw_id);
682 GEM_BUG_ON(!atomic_read(&ctx->hw_id_pin_count));
683 dev_priv->kernel_context = ctx;
685 /* highest priority; preempting task */
686 if (needs_preempt_context(dev_priv)) {
687 ctx = i915_gem_context_create_kernel(dev_priv, INT_MAX);
689 dev_priv->preempt_context = ctx;
691 DRM_ERROR("Failed to create preempt context; disabling preemption\n");
694 DRM_DEBUG_DRIVER("%s context support initialized\n",
695 DRIVER_CAPS(dev_priv)->has_logical_contexts ?
700 void i915_gem_contexts_fini(struct drm_i915_private *i915)
702 lockdep_assert_held(&i915->drm.struct_mutex);
704 if (i915->preempt_context)
705 destroy_kernel_context(&i915->preempt_context);
706 destroy_kernel_context(&i915->kernel_context);
708 /* Must free all deferred contexts (via flush_workqueue) first */
709 GEM_BUG_ON(!list_empty(&i915->contexts.hw_id_list));
710 ida_destroy(&i915->contexts.hw_ida);
713 static int context_idr_cleanup(int id, void *p, void *data)
719 static int vm_idr_cleanup(int id, void *p, void *data)
725 static int gem_context_register(struct i915_gem_context *ctx,
726 struct drm_i915_file_private *fpriv)
730 ctx->file_priv = fpriv;
732 ctx->vm->file = fpriv;
734 ctx->pid = get_task_pid(current, PIDTYPE_PID);
735 ctx->name = kasprintf(GFP_KERNEL, "%s[%d]",
736 current->comm, pid_nr(ctx->pid));
742 /* And finally expose ourselves to userspace via the idr */
743 mutex_lock(&fpriv->context_idr_lock);
744 ret = idr_alloc(&fpriv->context_idr, ctx, 0, 0, GFP_KERNEL);
745 mutex_unlock(&fpriv->context_idr_lock);
749 kfree(fetch_and_zero(&ctx->name));
751 put_pid(fetch_and_zero(&ctx->pid));
756 int i915_gem_context_open(struct drm_i915_private *i915,
757 struct drm_file *file)
759 struct drm_i915_file_private *file_priv = file->driver_priv;
760 struct i915_gem_context *ctx;
763 mutex_init(&file_priv->context_idr_lock);
764 mutex_init(&file_priv->vm_idr_lock);
766 idr_init(&file_priv->context_idr);
767 idr_init_base(&file_priv->vm_idr, 1);
769 mutex_lock(&i915->drm.struct_mutex);
770 ctx = i915_gem_create_context(i915, 0);
771 mutex_unlock(&i915->drm.struct_mutex);
777 err = gem_context_register(ctx, file_priv);
781 GEM_BUG_ON(i915_gem_context_is_kernel(ctx));
789 idr_destroy(&file_priv->vm_idr);
790 idr_destroy(&file_priv->context_idr);
791 mutex_destroy(&file_priv->vm_idr_lock);
792 mutex_destroy(&file_priv->context_idr_lock);
796 void i915_gem_context_close(struct drm_file *file)
798 struct drm_i915_file_private *file_priv = file->driver_priv;
800 idr_for_each(&file_priv->context_idr, context_idr_cleanup, NULL);
801 idr_destroy(&file_priv->context_idr);
802 mutex_destroy(&file_priv->context_idr_lock);
804 idr_for_each(&file_priv->vm_idr, vm_idr_cleanup, NULL);
805 idr_destroy(&file_priv->vm_idr);
806 mutex_destroy(&file_priv->vm_idr_lock);
809 int i915_gem_vm_create_ioctl(struct drm_device *dev, void *data,
810 struct drm_file *file)
812 struct drm_i915_private *i915 = to_i915(dev);
813 struct drm_i915_gem_vm_control *args = data;
814 struct drm_i915_file_private *file_priv = file->driver_priv;
815 struct i915_ppgtt *ppgtt;
818 if (!HAS_FULL_PPGTT(i915))
824 ppgtt = i915_ppgtt_create(i915);
826 return PTR_ERR(ppgtt);
828 ppgtt->vm.file = file_priv;
830 if (args->extensions) {
831 err = i915_user_extensions(u64_to_user_ptr(args->extensions),
838 err = mutex_lock_interruptible(&file_priv->vm_idr_lock);
842 err = idr_alloc(&file_priv->vm_idr, &ppgtt->vm, 0, 0, GFP_KERNEL);
846 GEM_BUG_ON(err == 0); /* reserved for invalid/unassigned ppgtt */
848 mutex_unlock(&file_priv->vm_idr_lock);
854 mutex_unlock(&file_priv->vm_idr_lock);
856 i915_vm_put(&ppgtt->vm);
860 int i915_gem_vm_destroy_ioctl(struct drm_device *dev, void *data,
861 struct drm_file *file)
863 struct drm_i915_file_private *file_priv = file->driver_priv;
864 struct drm_i915_gem_vm_control *args = data;
865 struct i915_address_space *vm;
872 if (args->extensions)
879 err = mutex_lock_interruptible(&file_priv->vm_idr_lock);
883 vm = idr_remove(&file_priv->vm_idr, id);
885 mutex_unlock(&file_priv->vm_idr_lock);
893 struct context_barrier_task {
894 struct i915_active base;
895 void (*task)(void *data);
899 static void cb_retire(struct i915_active *base)
901 struct context_barrier_task *cb = container_of(base, typeof(*cb), base);
906 i915_active_fini(&cb->base);
910 I915_SELFTEST_DECLARE(static intel_engine_mask_t context_barrier_inject_fault);
911 static int context_barrier_task(struct i915_gem_context *ctx,
912 intel_engine_mask_t engines,
913 bool (*skip)(struct intel_context *ce, void *data),
914 int (*emit)(struct i915_request *rq, void *data),
915 void (*task)(void *data),
918 struct drm_i915_private *i915 = ctx->i915;
919 struct context_barrier_task *cb;
920 struct i915_gem_engines_iter it;
921 struct intel_context *ce;
924 lockdep_assert_held(&i915->drm.struct_mutex);
927 cb = kmalloc(sizeof(*cb), GFP_KERNEL);
931 i915_active_init(i915, &cb->base, cb_retire);
932 i915_active_acquire(&cb->base);
934 for_each_gem_engine(ce, i915_gem_context_lock_engines(ctx), it) {
935 struct i915_request *rq;
937 if (I915_SELFTEST_ONLY(context_barrier_inject_fault &
943 if (!(ce->engine->mask & engines))
946 if (skip && skip(ce, data))
949 rq = intel_context_create_request(ce);
957 err = emit(rq, data);
959 err = i915_active_ref(&cb->base, rq->fence.context, rq);
961 i915_request_add(rq);
965 i915_gem_context_unlock_engines(ctx);
967 cb->task = err ? NULL : task; /* caller needs to unwind instead */
970 i915_active_release(&cb->base);
975 static int get_ppgtt(struct drm_i915_file_private *file_priv,
976 struct i915_gem_context *ctx,
977 struct drm_i915_gem_context_param *args)
979 struct i915_address_space *vm;
985 /* XXX rcu acquire? */
986 ret = mutex_lock_interruptible(&ctx->i915->drm.struct_mutex);
990 vm = i915_vm_get(ctx->vm);
991 mutex_unlock(&ctx->i915->drm.struct_mutex);
993 ret = mutex_lock_interruptible(&file_priv->vm_idr_lock);
997 ret = idr_alloc(&file_priv->vm_idr, vm, 0, 0, GFP_KERNEL);
1009 mutex_unlock(&file_priv->vm_idr_lock);
1015 static void set_ppgtt_barrier(void *data)
1017 struct i915_address_space *old = data;
1019 if (INTEL_GEN(old->i915) < 8)
1020 gen6_ppgtt_unpin_all(i915_vm_to_ppgtt(old));
1025 static int emit_ppgtt_update(struct i915_request *rq, void *data)
1027 struct i915_address_space *vm = rq->gem_context->vm;
1028 struct intel_engine_cs *engine = rq->engine;
1029 u32 base = engine->mmio_base;
1033 if (i915_vm_is_4lvl(vm)) {
1034 struct i915_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1035 const dma_addr_t pd_daddr = px_dma(ppgtt->pd);
1037 cs = intel_ring_begin(rq, 6);
1041 *cs++ = MI_LOAD_REGISTER_IMM(2);
1043 *cs++ = i915_mmio_reg_offset(GEN8_RING_PDP_UDW(base, 0));
1044 *cs++ = upper_32_bits(pd_daddr);
1045 *cs++ = i915_mmio_reg_offset(GEN8_RING_PDP_LDW(base, 0));
1046 *cs++ = lower_32_bits(pd_daddr);
1049 intel_ring_advance(rq, cs);
1050 } else if (HAS_LOGICAL_RING_CONTEXTS(engine->i915)) {
1051 struct i915_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1053 cs = intel_ring_begin(rq, 4 * GEN8_3LVL_PDPES + 2);
1057 *cs++ = MI_LOAD_REGISTER_IMM(2 * GEN8_3LVL_PDPES);
1058 for (i = GEN8_3LVL_PDPES; i--; ) {
1059 const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
1061 *cs++ = i915_mmio_reg_offset(GEN8_RING_PDP_UDW(base, i));
1062 *cs++ = upper_32_bits(pd_daddr);
1063 *cs++ = i915_mmio_reg_offset(GEN8_RING_PDP_LDW(base, i));
1064 *cs++ = lower_32_bits(pd_daddr);
1067 intel_ring_advance(rq, cs);
1069 /* ppGTT is not part of the legacy context image */
1070 gen6_ppgtt_pin(i915_vm_to_ppgtt(vm));
1076 static bool skip_ppgtt_update(struct intel_context *ce, void *data)
1078 if (HAS_LOGICAL_RING_CONTEXTS(ce->engine->i915))
1081 return !atomic_read(&ce->pin_count);
1084 static int set_ppgtt(struct drm_i915_file_private *file_priv,
1085 struct i915_gem_context *ctx,
1086 struct drm_i915_gem_context_param *args)
1088 struct i915_address_space *vm, *old;
1097 if (upper_32_bits(args->value))
1100 err = mutex_lock_interruptible(&file_priv->vm_idr_lock);
1104 vm = idr_find(&file_priv->vm_idr, args->value);
1107 mutex_unlock(&file_priv->vm_idr_lock);
1111 err = mutex_lock_interruptible(&ctx->i915->drm.struct_mutex);
1118 /* Teardown the existing obj:vma cache, it will have to be rebuilt. */
1119 mutex_lock(&ctx->mutex);
1121 mutex_unlock(&ctx->mutex);
1123 old = __set_ppgtt(ctx, vm);
1126 * We need to flush any requests using the current ppgtt before
1127 * we release it as the requests do not hold a reference themselves,
1128 * only indirectly through the context.
1130 err = context_barrier_task(ctx, ALL_ENGINES,
1137 ctx->desc_template = default_desc_template(ctx->i915, old);
1142 mutex_unlock(&ctx->i915->drm.struct_mutex);
1149 static int gen8_emit_rpcs_config(struct i915_request *rq,
1150 struct intel_context *ce,
1151 struct intel_sseu sseu)
1156 cs = intel_ring_begin(rq, 4);
1160 offset = i915_ggtt_offset(ce->state) +
1161 LRC_STATE_PN * PAGE_SIZE +
1162 (CTX_R_PWR_CLK_STATE + 1) * 4;
1164 *cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
1165 *cs++ = lower_32_bits(offset);
1166 *cs++ = upper_32_bits(offset);
1167 *cs++ = intel_sseu_make_rpcs(rq->i915, &sseu);
1169 intel_ring_advance(rq, cs);
1175 gen8_modify_rpcs(struct intel_context *ce, struct intel_sseu sseu)
1177 struct i915_request *rq;
1180 lockdep_assert_held(&ce->pin_mutex);
1183 * If the context is not idle, we have to submit an ordered request to
1184 * modify its context image via the kernel context (writing to our own
1185 * image, or into the registers directory, does not stick). Pristine
1186 * and idle contexts will be configured on pinning.
1188 if (!intel_context_is_pinned(ce))
1191 rq = i915_request_create(ce->engine->kernel_context);
1195 /* Queue this switch after all other activity by this context. */
1196 ret = i915_active_request_set(&ce->ring->timeline->last_request, rq);
1201 * Guarantee context image and the timeline remains pinned until the
1202 * modifying request is retired by setting the ce activity tracker.
1204 * But we only need to take one pin on the account of it. Or in other
1205 * words transfer the pinned ce object to tracked active request.
1207 GEM_BUG_ON(i915_active_is_idle(&ce->active));
1208 ret = i915_active_ref(&ce->active, rq->fence.context, rq);
1212 ret = gen8_emit_rpcs_config(rq, ce, sseu);
1215 i915_request_add(rq);
1220 __intel_context_reconfigure_sseu(struct intel_context *ce,
1221 struct intel_sseu sseu)
1225 GEM_BUG_ON(INTEL_GEN(ce->gem_context->i915) < 8);
1227 ret = intel_context_lock_pinned(ce);
1231 /* Nothing to do if unmodified. */
1232 if (!memcmp(&ce->sseu, &sseu, sizeof(sseu)))
1235 ret = gen8_modify_rpcs(ce, sseu);
1240 intel_context_unlock_pinned(ce);
1245 intel_context_reconfigure_sseu(struct intel_context *ce, struct intel_sseu sseu)
1247 struct drm_i915_private *i915 = ce->gem_context->i915;
1250 ret = mutex_lock_interruptible(&i915->drm.struct_mutex);
1254 ret = __intel_context_reconfigure_sseu(ce, sseu);
1256 mutex_unlock(&i915->drm.struct_mutex);
1262 user_to_context_sseu(struct drm_i915_private *i915,
1263 const struct drm_i915_gem_context_param_sseu *user,
1264 struct intel_sseu *context)
1266 const struct sseu_dev_info *device = &RUNTIME_INFO(i915)->sseu;
1268 /* No zeros in any field. */
1269 if (!user->slice_mask || !user->subslice_mask ||
1270 !user->min_eus_per_subslice || !user->max_eus_per_subslice)
1274 if (user->max_eus_per_subslice < user->min_eus_per_subslice)
1278 * Some future proofing on the types since the uAPI is wider than the
1279 * current internal implementation.
1281 if (overflows_type(user->slice_mask, context->slice_mask) ||
1282 overflows_type(user->subslice_mask, context->subslice_mask) ||
1283 overflows_type(user->min_eus_per_subslice,
1284 context->min_eus_per_subslice) ||
1285 overflows_type(user->max_eus_per_subslice,
1286 context->max_eus_per_subslice))
1289 /* Check validity against hardware. */
1290 if (user->slice_mask & ~device->slice_mask)
1293 if (user->subslice_mask & ~device->subslice_mask[0])
1296 if (user->max_eus_per_subslice > device->max_eus_per_subslice)
1299 context->slice_mask = user->slice_mask;
1300 context->subslice_mask = user->subslice_mask;
1301 context->min_eus_per_subslice = user->min_eus_per_subslice;
1302 context->max_eus_per_subslice = user->max_eus_per_subslice;
1304 /* Part specific restrictions. */
1305 if (IS_GEN(i915, 11)) {
1306 unsigned int hw_s = hweight8(device->slice_mask);
1307 unsigned int hw_ss_per_s = hweight8(device->subslice_mask[0]);
1308 unsigned int req_s = hweight8(context->slice_mask);
1309 unsigned int req_ss = hweight8(context->subslice_mask);
1312 * Only full subslice enablement is possible if more than one
1313 * slice is turned on.
1315 if (req_s > 1 && req_ss != hw_ss_per_s)
1319 * If more than four (SScount bitfield limit) subslices are
1320 * requested then the number has to be even.
1322 if (req_ss > 4 && (req_ss & 1))
1326 * If only one slice is enabled and subslice count is below the
1327 * device full enablement, it must be at most half of the all
1328 * available subslices.
1330 if (req_s == 1 && req_ss < hw_ss_per_s &&
1331 req_ss > (hw_ss_per_s / 2))
1334 /* ABI restriction - VME use case only. */
1336 /* All slices or one slice only. */
1337 if (req_s != 1 && req_s != hw_s)
1341 * Half subslices or full enablement only when one slice is
1345 (req_ss != hw_ss_per_s && req_ss != (hw_ss_per_s / 2)))
1348 /* No EU configuration changes. */
1349 if ((user->min_eus_per_subslice !=
1350 device->max_eus_per_subslice) ||
1351 (user->max_eus_per_subslice !=
1352 device->max_eus_per_subslice))
1359 static int set_sseu(struct i915_gem_context *ctx,
1360 struct drm_i915_gem_context_param *args)
1362 struct drm_i915_private *i915 = ctx->i915;
1363 struct drm_i915_gem_context_param_sseu user_sseu;
1364 struct intel_context *ce;
1365 struct intel_sseu sseu;
1366 unsigned long lookup;
1369 if (args->size < sizeof(user_sseu))
1372 if (!IS_GEN(i915, 11))
1375 if (copy_from_user(&user_sseu, u64_to_user_ptr(args->value),
1382 if (user_sseu.flags & ~(I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX))
1386 if (user_sseu.flags & I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX)
1387 lookup |= LOOKUP_USER_INDEX;
1389 ce = lookup_user_engine(ctx, lookup, &user_sseu.engine);
1393 /* Only render engine supports RPCS configuration. */
1394 if (ce->engine->class != RENDER_CLASS) {
1399 ret = user_to_context_sseu(i915, &user_sseu, &sseu);
1403 ret = intel_context_reconfigure_sseu(ce, sseu);
1407 args->size = sizeof(user_sseu);
1410 intel_context_put(ce);
1414 struct set_engines {
1415 struct i915_gem_context *ctx;
1416 struct i915_gem_engines *engines;
1420 set_engines__load_balance(struct i915_user_extension __user *base, void *data)
1422 struct i915_context_engines_load_balance __user *ext =
1423 container_of_user(base, typeof(*ext), base);
1424 const struct set_engines *set = data;
1425 struct intel_engine_cs *stack[16];
1426 struct intel_engine_cs **siblings;
1427 struct intel_context *ce;
1428 u16 num_siblings, idx;
1432 if (!HAS_EXECLISTS(set->ctx->i915))
1435 if (USES_GUC_SUBMISSION(set->ctx->i915))
1436 return -ENODEV; /* not implement yet */
1438 if (get_user(idx, &ext->engine_index))
1441 if (idx >= set->engines->num_engines) {
1442 DRM_DEBUG("Invalid placement value, %d >= %d\n",
1443 idx, set->engines->num_engines);
1447 idx = array_index_nospec(idx, set->engines->num_engines);
1448 if (set->engines->engines[idx]) {
1449 DRM_DEBUG("Invalid placement[%d], already occupied\n", idx);
1453 if (get_user(num_siblings, &ext->num_siblings))
1456 err = check_user_mbz(&ext->flags);
1460 err = check_user_mbz(&ext->mbz64);
1465 if (num_siblings > ARRAY_SIZE(stack)) {
1466 siblings = kmalloc_array(num_siblings,
1473 for (n = 0; n < num_siblings; n++) {
1474 struct i915_engine_class_instance ci;
1476 if (copy_from_user(&ci, &ext->engines[n], sizeof(ci))) {
1481 siblings[n] = intel_engine_lookup_user(set->ctx->i915,
1483 ci.engine_instance);
1485 DRM_DEBUG("Invalid sibling[%d]: { class:%d, inst:%d }\n",
1486 n, ci.engine_class, ci.engine_instance);
1492 ce = intel_execlists_create_virtual(set->ctx, siblings, n);
1498 if (cmpxchg(&set->engines->engines[idx], NULL, ce)) {
1499 intel_context_put(ce);
1505 if (siblings != stack)
1512 set_engines__bond(struct i915_user_extension __user *base, void *data)
1514 struct i915_context_engines_bond __user *ext =
1515 container_of_user(base, typeof(*ext), base);
1516 const struct set_engines *set = data;
1517 struct i915_engine_class_instance ci;
1518 struct intel_engine_cs *virtual;
1519 struct intel_engine_cs *master;
1523 if (get_user(idx, &ext->virtual_index))
1526 if (idx >= set->engines->num_engines) {
1527 DRM_DEBUG("Invalid index for virtual engine: %d >= %d\n",
1528 idx, set->engines->num_engines);
1532 idx = array_index_nospec(idx, set->engines->num_engines);
1533 if (!set->engines->engines[idx]) {
1534 DRM_DEBUG("Invalid engine at %d\n", idx);
1537 virtual = set->engines->engines[idx]->engine;
1539 err = check_user_mbz(&ext->flags);
1543 for (n = 0; n < ARRAY_SIZE(ext->mbz64); n++) {
1544 err = check_user_mbz(&ext->mbz64[n]);
1549 if (copy_from_user(&ci, &ext->master, sizeof(ci)))
1552 master = intel_engine_lookup_user(set->ctx->i915,
1553 ci.engine_class, ci.engine_instance);
1555 DRM_DEBUG("Unrecognised master engine: { class:%u, instance:%u }\n",
1556 ci.engine_class, ci.engine_instance);
1560 if (get_user(num_bonds, &ext->num_bonds))
1563 for (n = 0; n < num_bonds; n++) {
1564 struct intel_engine_cs *bond;
1566 if (copy_from_user(&ci, &ext->engines[n], sizeof(ci)))
1569 bond = intel_engine_lookup_user(set->ctx->i915,
1571 ci.engine_instance);
1573 DRM_DEBUG("Unrecognised engine[%d] for bonding: { class:%d, instance: %d }\n",
1574 n, ci.engine_class, ci.engine_instance);
1579 * A non-virtual engine has no siblings to choose between; and
1580 * a submit fence will always be directed to the one engine.
1582 if (intel_engine_is_virtual(virtual)) {
1583 err = intel_virtual_engine_attach_bond(virtual,
1594 static const i915_user_extension_fn set_engines__extensions[] = {
1595 [I915_CONTEXT_ENGINES_EXT_LOAD_BALANCE] = set_engines__load_balance,
1596 [I915_CONTEXT_ENGINES_EXT_BOND] = set_engines__bond,
1600 set_engines(struct i915_gem_context *ctx,
1601 const struct drm_i915_gem_context_param *args)
1603 struct i915_context_param_engines __user *user =
1604 u64_to_user_ptr(args->value);
1605 struct set_engines set = { .ctx = ctx };
1606 unsigned int num_engines, n;
1610 if (!args->size) { /* switch back to legacy user_ring_map */
1611 if (!i915_gem_context_user_engines(ctx))
1614 set.engines = default_engines(ctx);
1615 if (IS_ERR(set.engines))
1616 return PTR_ERR(set.engines);
1621 BUILD_BUG_ON(!IS_ALIGNED(sizeof(*user), sizeof(*user->engines)));
1622 if (args->size < sizeof(*user) ||
1623 !IS_ALIGNED(args->size, sizeof(*user->engines))) {
1624 DRM_DEBUG("Invalid size for engine array: %d\n",
1630 * Note that I915_EXEC_RING_MASK limits execbuf to only using the
1631 * first 64 engines defined here.
1633 num_engines = (args->size - sizeof(*user)) / sizeof(*user->engines);
1635 set.engines = kmalloc(struct_size(set.engines, engines, num_engines),
1640 init_rcu_head(&set.engines->rcu);
1641 for (n = 0; n < num_engines; n++) {
1642 struct i915_engine_class_instance ci;
1643 struct intel_engine_cs *engine;
1645 if (copy_from_user(&ci, &user->engines[n], sizeof(ci))) {
1646 __free_engines(set.engines, n);
1650 if (ci.engine_class == (u16)I915_ENGINE_CLASS_INVALID &&
1651 ci.engine_instance == (u16)I915_ENGINE_CLASS_INVALID_NONE) {
1652 set.engines->engines[n] = NULL;
1656 engine = intel_engine_lookup_user(ctx->i915,
1658 ci.engine_instance);
1660 DRM_DEBUG("Invalid engine[%d]: { class:%d, instance:%d }\n",
1661 n, ci.engine_class, ci.engine_instance);
1662 __free_engines(set.engines, n);
1666 set.engines->engines[n] = intel_context_create(ctx, engine);
1667 if (!set.engines->engines[n]) {
1668 __free_engines(set.engines, n);
1672 set.engines->num_engines = num_engines;
1675 if (!get_user(extensions, &user->extensions))
1676 err = i915_user_extensions(u64_to_user_ptr(extensions),
1677 set_engines__extensions,
1678 ARRAY_SIZE(set_engines__extensions),
1681 free_engines(set.engines);
1686 mutex_lock(&ctx->engines_mutex);
1688 i915_gem_context_set_user_engines(ctx);
1690 i915_gem_context_clear_user_engines(ctx);
1691 rcu_swap_protected(ctx->engines, set.engines, 1);
1692 mutex_unlock(&ctx->engines_mutex);
1694 call_rcu(&set.engines->rcu, free_engines_rcu);
1699 static struct i915_gem_engines *
1700 __copy_engines(struct i915_gem_engines *e)
1702 struct i915_gem_engines *copy;
1705 copy = kmalloc(struct_size(e, engines, e->num_engines), GFP_KERNEL);
1707 return ERR_PTR(-ENOMEM);
1709 init_rcu_head(©->rcu);
1710 for (n = 0; n < e->num_engines; n++) {
1712 copy->engines[n] = intel_context_get(e->engines[n]);
1714 copy->engines[n] = NULL;
1716 copy->num_engines = n;
1722 get_engines(struct i915_gem_context *ctx,
1723 struct drm_i915_gem_context_param *args)
1725 struct i915_context_param_engines __user *user;
1726 struct i915_gem_engines *e;
1727 size_t n, count, size;
1730 err = mutex_lock_interruptible(&ctx->engines_mutex);
1735 if (i915_gem_context_user_engines(ctx))
1736 e = __copy_engines(i915_gem_context_engines(ctx));
1737 mutex_unlock(&ctx->engines_mutex);
1738 if (IS_ERR_OR_NULL(e)) {
1740 return PTR_ERR_OR_ZERO(e);
1743 count = e->num_engines;
1745 /* Be paranoid in case we have an impedance mismatch */
1746 if (!check_struct_size(user, engines, count, &size)) {
1750 if (overflows_type(size, args->size)) {
1760 if (args->size < size) {
1765 user = u64_to_user_ptr(args->value);
1766 if (!access_ok(user, size)) {
1771 if (put_user(0, &user->extensions)) {
1776 for (n = 0; n < count; n++) {
1777 struct i915_engine_class_instance ci = {
1778 .engine_class = I915_ENGINE_CLASS_INVALID,
1779 .engine_instance = I915_ENGINE_CLASS_INVALID_NONE,
1782 if (e->engines[n]) {
1783 ci.engine_class = e->engines[n]->engine->uabi_class;
1784 ci.engine_instance = e->engines[n]->engine->instance;
1787 if (copy_to_user(&user->engines[n], &ci, sizeof(ci))) {
1800 static int ctx_setparam(struct drm_i915_file_private *fpriv,
1801 struct i915_gem_context *ctx,
1802 struct drm_i915_gem_context_param *args)
1806 switch (args->param) {
1807 case I915_CONTEXT_PARAM_NO_ZEROMAP:
1810 else if (args->value)
1811 set_bit(UCONTEXT_NO_ZEROMAP, &ctx->user_flags);
1813 clear_bit(UCONTEXT_NO_ZEROMAP, &ctx->user_flags);
1816 case I915_CONTEXT_PARAM_NO_ERROR_CAPTURE:
1819 else if (args->value)
1820 i915_gem_context_set_no_error_capture(ctx);
1822 i915_gem_context_clear_no_error_capture(ctx);
1825 case I915_CONTEXT_PARAM_BANNABLE:
1828 else if (!capable(CAP_SYS_ADMIN) && !args->value)
1830 else if (args->value)
1831 i915_gem_context_set_bannable(ctx);
1833 i915_gem_context_clear_bannable(ctx);
1836 case I915_CONTEXT_PARAM_RECOVERABLE:
1839 else if (args->value)
1840 i915_gem_context_set_recoverable(ctx);
1842 i915_gem_context_clear_recoverable(ctx);
1845 case I915_CONTEXT_PARAM_PRIORITY:
1847 s64 priority = args->value;
1851 else if (!(ctx->i915->caps.scheduler & I915_SCHEDULER_CAP_PRIORITY))
1853 else if (priority > I915_CONTEXT_MAX_USER_PRIORITY ||
1854 priority < I915_CONTEXT_MIN_USER_PRIORITY)
1856 else if (priority > I915_CONTEXT_DEFAULT_PRIORITY &&
1857 !capable(CAP_SYS_NICE))
1860 ctx->sched.priority =
1861 I915_USER_PRIORITY(priority);
1865 case I915_CONTEXT_PARAM_SSEU:
1866 ret = set_sseu(ctx, args);
1869 case I915_CONTEXT_PARAM_VM:
1870 ret = set_ppgtt(fpriv, ctx, args);
1873 case I915_CONTEXT_PARAM_ENGINES:
1874 ret = set_engines(ctx, args);
1877 case I915_CONTEXT_PARAM_BAN_PERIOD:
1887 struct i915_gem_context *ctx;
1888 struct drm_i915_file_private *fpriv;
1891 static int create_setparam(struct i915_user_extension __user *ext, void *data)
1893 struct drm_i915_gem_context_create_ext_setparam local;
1894 const struct create_ext *arg = data;
1896 if (copy_from_user(&local, ext, sizeof(local)))
1899 if (local.param.ctx_id)
1902 return ctx_setparam(arg->fpriv, arg->ctx, &local.param);
1905 static int clone_engines(struct i915_gem_context *dst,
1906 struct i915_gem_context *src)
1908 struct i915_gem_engines *e = i915_gem_context_lock_engines(src);
1909 struct i915_gem_engines *clone;
1913 clone = kmalloc(struct_size(e, engines, e->num_engines), GFP_KERNEL);
1917 init_rcu_head(&clone->rcu);
1918 for (n = 0; n < e->num_engines; n++) {
1919 struct intel_engine_cs *engine;
1921 if (!e->engines[n]) {
1922 clone->engines[n] = NULL;
1925 engine = e->engines[n]->engine;
1928 * Virtual engines are singletons; they can only exist
1929 * inside a single context, because they embed their
1930 * HW context... As each virtual context implies a single
1931 * timeline (each engine can only dequeue a single request
1932 * at any time), it would be surprising for two contexts
1933 * to use the same engine. So let's create a copy of
1934 * the virtual engine instead.
1936 if (intel_engine_is_virtual(engine))
1938 intel_execlists_clone_virtual(dst, engine);
1940 clone->engines[n] = intel_context_create(dst, engine);
1941 if (IS_ERR_OR_NULL(clone->engines[n])) {
1942 __free_engines(clone, n);
1946 clone->num_engines = n;
1948 user_engines = i915_gem_context_user_engines(src);
1949 i915_gem_context_unlock_engines(src);
1951 free_engines(dst->engines);
1952 RCU_INIT_POINTER(dst->engines, clone);
1954 i915_gem_context_set_user_engines(dst);
1956 i915_gem_context_clear_user_engines(dst);
1960 i915_gem_context_unlock_engines(src);
1964 static int clone_flags(struct i915_gem_context *dst,
1965 struct i915_gem_context *src)
1967 dst->user_flags = src->user_flags;
1971 static int clone_schedattr(struct i915_gem_context *dst,
1972 struct i915_gem_context *src)
1974 dst->sched = src->sched;
1978 static int clone_sseu(struct i915_gem_context *dst,
1979 struct i915_gem_context *src)
1981 struct i915_gem_engines *e = i915_gem_context_lock_engines(src);
1982 struct i915_gem_engines *clone;
1986 clone = dst->engines; /* no locking required; sole access */
1987 if (e->num_engines != clone->num_engines) {
1992 for (n = 0; n < e->num_engines; n++) {
1993 struct intel_context *ce = e->engines[n];
1995 if (clone->engines[n]->engine->class != ce->engine->class) {
1996 /* Must have compatible engine maps! */
2001 /* serialises with set_sseu */
2002 err = intel_context_lock_pinned(ce);
2006 clone->engines[n]->sseu = ce->sseu;
2007 intel_context_unlock_pinned(ce);
2012 i915_gem_context_unlock_engines(src);
2016 static int clone_timeline(struct i915_gem_context *dst,
2017 struct i915_gem_context *src)
2019 if (src->timeline) {
2020 GEM_BUG_ON(src->timeline == dst->timeline);
2023 i915_timeline_put(dst->timeline);
2024 dst->timeline = i915_timeline_get(src->timeline);
2030 static int clone_vm(struct i915_gem_context *dst,
2031 struct i915_gem_context *src)
2033 struct i915_address_space *vm;
2037 vm = READ_ONCE(src->vm);
2041 if (!kref_get_unless_zero(&vm->ref))
2045 * This ppgtt may have be reallocated between
2046 * the read and the kref, and reassigned to a third
2047 * context. In order to avoid inadvertent sharing
2048 * of this ppgtt with that third context (and not
2049 * src), we have to confirm that we have the same
2050 * ppgtt after passing through the strong memory
2051 * barrier implied by a successful
2052 * kref_get_unless_zero().
2054 * Once we have acquired the current ppgtt of src,
2055 * we no longer care if it is released from src, as
2056 * it cannot be reallocated elsewhere.
2059 if (vm == READ_ONCE(src->vm))
2067 __assign_ppgtt(dst, vm);
2074 static int create_clone(struct i915_user_extension __user *ext, void *data)
2076 static int (* const fn[])(struct i915_gem_context *dst,
2077 struct i915_gem_context *src) = {
2078 #define MAP(x, y) [ilog2(I915_CONTEXT_CLONE_##x)] = y
2079 MAP(ENGINES, clone_engines),
2080 MAP(FLAGS, clone_flags),
2081 MAP(SCHEDATTR, clone_schedattr),
2082 MAP(SSEU, clone_sseu),
2083 MAP(TIMELINE, clone_timeline),
2087 struct drm_i915_gem_context_create_ext_clone local;
2088 const struct create_ext *arg = data;
2089 struct i915_gem_context *dst = arg->ctx;
2090 struct i915_gem_context *src;
2093 if (copy_from_user(&local, ext, sizeof(local)))
2096 BUILD_BUG_ON(GENMASK(BITS_PER_TYPE(local.flags) - 1, ARRAY_SIZE(fn)) !=
2097 I915_CONTEXT_CLONE_UNKNOWN);
2099 if (local.flags & I915_CONTEXT_CLONE_UNKNOWN)
2106 src = __i915_gem_context_lookup_rcu(arg->fpriv, local.clone_id);
2111 GEM_BUG_ON(src == dst);
2113 for (bit = 0; bit < ARRAY_SIZE(fn); bit++) {
2114 if (!(local.flags & BIT(bit)))
2117 err = fn[bit](dst, src);
2125 static const i915_user_extension_fn create_extensions[] = {
2126 [I915_CONTEXT_CREATE_EXT_SETPARAM] = create_setparam,
2127 [I915_CONTEXT_CREATE_EXT_CLONE] = create_clone,
2130 static bool client_is_banned(struct drm_i915_file_private *file_priv)
2132 return atomic_read(&file_priv->ban_score) >= I915_CLIENT_SCORE_BANNED;
2135 int i915_gem_context_create_ioctl(struct drm_device *dev, void *data,
2136 struct drm_file *file)
2138 struct drm_i915_private *i915 = to_i915(dev);
2139 struct drm_i915_gem_context_create_ext *args = data;
2140 struct create_ext ext_data;
2143 if (!DRIVER_CAPS(i915)->has_logical_contexts)
2146 if (args->flags & I915_CONTEXT_CREATE_FLAGS_UNKNOWN)
2149 ret = i915_terminally_wedged(i915);
2153 ext_data.fpriv = file->driver_priv;
2154 if (client_is_banned(ext_data.fpriv)) {
2155 DRM_DEBUG("client %s[%d] banned from creating ctx\n",
2157 pid_nr(get_task_pid(current, PIDTYPE_PID)));
2161 ret = i915_mutex_lock_interruptible(dev);
2165 ext_data.ctx = i915_gem_create_context(i915, args->flags);
2166 mutex_unlock(&dev->struct_mutex);
2167 if (IS_ERR(ext_data.ctx))
2168 return PTR_ERR(ext_data.ctx);
2170 if (args->flags & I915_CONTEXT_CREATE_FLAGS_USE_EXTENSIONS) {
2171 ret = i915_user_extensions(u64_to_user_ptr(args->extensions),
2173 ARRAY_SIZE(create_extensions),
2179 ret = gem_context_register(ext_data.ctx, ext_data.fpriv);
2184 DRM_DEBUG("HW context %d created\n", args->ctx_id);
2189 context_close(ext_data.ctx);
2193 int i915_gem_context_destroy_ioctl(struct drm_device *dev, void *data,
2194 struct drm_file *file)
2196 struct drm_i915_gem_context_destroy *args = data;
2197 struct drm_i915_file_private *file_priv = file->driver_priv;
2198 struct i915_gem_context *ctx;
2206 if (mutex_lock_interruptible(&file_priv->context_idr_lock))
2209 ctx = idr_remove(&file_priv->context_idr, args->ctx_id);
2210 mutex_unlock(&file_priv->context_idr_lock);
2218 static int get_sseu(struct i915_gem_context *ctx,
2219 struct drm_i915_gem_context_param *args)
2221 struct drm_i915_gem_context_param_sseu user_sseu;
2222 struct intel_context *ce;
2223 unsigned long lookup;
2226 if (args->size == 0)
2228 else if (args->size < sizeof(user_sseu))
2231 if (copy_from_user(&user_sseu, u64_to_user_ptr(args->value),
2238 if (user_sseu.flags & ~(I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX))
2242 if (user_sseu.flags & I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX)
2243 lookup |= LOOKUP_USER_INDEX;
2245 ce = lookup_user_engine(ctx, lookup, &user_sseu.engine);
2249 err = intel_context_lock_pinned(ce); /* serialises with set_sseu */
2251 intel_context_put(ce);
2255 user_sseu.slice_mask = ce->sseu.slice_mask;
2256 user_sseu.subslice_mask = ce->sseu.subslice_mask;
2257 user_sseu.min_eus_per_subslice = ce->sseu.min_eus_per_subslice;
2258 user_sseu.max_eus_per_subslice = ce->sseu.max_eus_per_subslice;
2260 intel_context_unlock_pinned(ce);
2261 intel_context_put(ce);
2263 if (copy_to_user(u64_to_user_ptr(args->value), &user_sseu,
2268 args->size = sizeof(user_sseu);
2273 int i915_gem_context_getparam_ioctl(struct drm_device *dev, void *data,
2274 struct drm_file *file)
2276 struct drm_i915_file_private *file_priv = file->driver_priv;
2277 struct drm_i915_gem_context_param *args = data;
2278 struct i915_gem_context *ctx;
2281 ctx = i915_gem_context_lookup(file_priv, args->ctx_id);
2285 switch (args->param) {
2286 case I915_CONTEXT_PARAM_NO_ZEROMAP:
2288 args->value = test_bit(UCONTEXT_NO_ZEROMAP, &ctx->user_flags);
2291 case I915_CONTEXT_PARAM_GTT_SIZE:
2294 args->value = ctx->vm->total;
2295 else if (to_i915(dev)->mm.aliasing_ppgtt)
2296 args->value = to_i915(dev)->mm.aliasing_ppgtt->vm.total;
2298 args->value = to_i915(dev)->ggtt.vm.total;
2301 case I915_CONTEXT_PARAM_NO_ERROR_CAPTURE:
2303 args->value = i915_gem_context_no_error_capture(ctx);
2306 case I915_CONTEXT_PARAM_BANNABLE:
2308 args->value = i915_gem_context_is_bannable(ctx);
2311 case I915_CONTEXT_PARAM_RECOVERABLE:
2313 args->value = i915_gem_context_is_recoverable(ctx);
2316 case I915_CONTEXT_PARAM_PRIORITY:
2318 args->value = ctx->sched.priority >> I915_USER_PRIORITY_SHIFT;
2321 case I915_CONTEXT_PARAM_SSEU:
2322 ret = get_sseu(ctx, args);
2325 case I915_CONTEXT_PARAM_VM:
2326 ret = get_ppgtt(file_priv, ctx, args);
2329 case I915_CONTEXT_PARAM_ENGINES:
2330 ret = get_engines(ctx, args);
2333 case I915_CONTEXT_PARAM_BAN_PERIOD:
2339 i915_gem_context_put(ctx);
2343 int i915_gem_context_setparam_ioctl(struct drm_device *dev, void *data,
2344 struct drm_file *file)
2346 struct drm_i915_file_private *file_priv = file->driver_priv;
2347 struct drm_i915_gem_context_param *args = data;
2348 struct i915_gem_context *ctx;
2351 ctx = i915_gem_context_lookup(file_priv, args->ctx_id);
2355 ret = ctx_setparam(file_priv, ctx, args);
2357 i915_gem_context_put(ctx);
2361 int i915_gem_context_reset_stats_ioctl(struct drm_device *dev,
2362 void *data, struct drm_file *file)
2364 struct drm_i915_private *dev_priv = to_i915(dev);
2365 struct drm_i915_reset_stats *args = data;
2366 struct i915_gem_context *ctx;
2369 if (args->flags || args->pad)
2374 ctx = __i915_gem_context_lookup_rcu(file->driver_priv, args->ctx_id);
2379 * We opt for unserialised reads here. This may result in tearing
2380 * in the extremely unlikely event of a GPU hang on this context
2381 * as we are querying them. If we need that extra layer of protection,
2382 * we should wrap the hangstats with a seqlock.
2385 if (capable(CAP_SYS_ADMIN))
2386 args->reset_count = i915_reset_count(&dev_priv->gpu_error);
2388 args->reset_count = 0;
2390 args->batch_active = atomic_read(&ctx->guilty_count);
2391 args->batch_pending = atomic_read(&ctx->active_count);
2399 int __i915_gem_context_pin_hw_id(struct i915_gem_context *ctx)
2401 struct drm_i915_private *i915 = ctx->i915;
2404 mutex_lock(&i915->contexts.mutex);
2406 GEM_BUG_ON(i915_gem_context_is_closed(ctx));
2408 if (list_empty(&ctx->hw_id_link)) {
2409 GEM_BUG_ON(atomic_read(&ctx->hw_id_pin_count));
2411 err = assign_hw_id(i915, &ctx->hw_id);
2415 list_add_tail(&ctx->hw_id_link, &i915->contexts.hw_id_list);
2418 GEM_BUG_ON(atomic_read(&ctx->hw_id_pin_count) == ~0u);
2419 atomic_inc(&ctx->hw_id_pin_count);
2422 mutex_unlock(&i915->contexts.mutex);
2426 /* GEM context-engines iterator: for_each_gem_engine() */
2427 struct intel_context *
2428 i915_gem_engines_iter_next(struct i915_gem_engines_iter *it)
2430 const struct i915_gem_engines *e = it->engines;
2431 struct intel_context *ctx;
2434 if (it->idx >= e->num_engines)
2437 ctx = e->engines[it->idx++];
2443 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
2444 #include "selftests/mock_context.c"
2445 #include "selftests/i915_gem_context.c"
2448 static void i915_global_gem_context_shrink(void)
2450 kmem_cache_shrink(global.slab_luts);
2453 static void i915_global_gem_context_exit(void)
2455 kmem_cache_destroy(global.slab_luts);
2458 static struct i915_global_gem_context global = { {
2459 .shrink = i915_global_gem_context_shrink,
2460 .exit = i915_global_gem_context_exit,
2463 int __init i915_global_gem_context_init(void)
2465 global.slab_luts = KMEM_CACHE(i915_lut_handle, 0);
2466 if (!global.slab_luts)
2469 i915_global_register(&global.base);