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"
73 #include "gt/intel_engine_user.h"
75 #include "i915_gem_context.h"
76 #include "i915_globals.h"
77 #include "i915_trace.h"
78 #include "i915_user_extensions.h"
80 #define ALL_L3_SLICES(dev) (1 << NUM_L3_SLICES(dev)) - 1
82 static struct i915_global_gem_context {
83 struct i915_global base;
84 struct kmem_cache *slab_luts;
87 struct i915_lut_handle *i915_lut_handle_alloc(void)
89 return kmem_cache_alloc(global.slab_luts, GFP_KERNEL);
92 void i915_lut_handle_free(struct i915_lut_handle *lut)
94 return kmem_cache_free(global.slab_luts, lut);
97 static void lut_close(struct i915_gem_context *ctx)
99 struct radix_tree_iter iter;
102 lockdep_assert_held(&ctx->mutex);
105 radix_tree_for_each_slot(slot, &ctx->handles_vma, &iter, 0) {
106 struct i915_vma *vma = rcu_dereference_raw(*slot);
107 struct drm_i915_gem_object *obj = vma->obj;
108 struct i915_lut_handle *lut;
110 if (!kref_get_unless_zero(&obj->base.refcount))
114 i915_gem_object_lock(obj);
115 list_for_each_entry(lut, &obj->lut_list, obj_link) {
119 if (lut->handle != iter.index)
122 list_del(&lut->obj_link);
125 i915_gem_object_unlock(obj);
128 if (&lut->obj_link != &obj->lut_list) {
129 i915_lut_handle_free(lut);
130 radix_tree_iter_delete(&ctx->handles_vma, &iter, slot);
131 if (atomic_dec_and_test(&vma->open_count) &&
132 !i915_vma_is_ggtt(vma))
134 i915_gem_object_put(obj);
137 i915_gem_object_put(obj);
142 static struct intel_context *
143 lookup_user_engine(struct i915_gem_context *ctx,
145 const struct i915_engine_class_instance *ci)
146 #define LOOKUP_USER_INDEX BIT(0)
150 if (!!(flags & LOOKUP_USER_INDEX) != i915_gem_context_user_engines(ctx))
151 return ERR_PTR(-EINVAL);
153 if (!i915_gem_context_user_engines(ctx)) {
154 struct intel_engine_cs *engine;
156 engine = intel_engine_lookup_user(ctx->i915,
158 ci->engine_instance);
160 return ERR_PTR(-EINVAL);
162 idx = engine->legacy_idx;
164 idx = ci->engine_instance;
167 return i915_gem_context_get_engine(ctx, idx);
170 static inline int new_hw_id(struct drm_i915_private *i915, gfp_t gfp)
174 lockdep_assert_held(&i915->contexts.mutex);
176 if (INTEL_GEN(i915) >= 12)
177 max = GEN12_MAX_CONTEXT_HW_ID;
178 else if (INTEL_GEN(i915) >= 11)
179 max = GEN11_MAX_CONTEXT_HW_ID;
180 else if (USES_GUC_SUBMISSION(i915))
182 * When using GuC in proxy submission, GuC consumes the
183 * highest bit in the context id to indicate proxy submission.
185 max = MAX_GUC_CONTEXT_HW_ID;
187 max = MAX_CONTEXT_HW_ID;
189 return ida_simple_get(&i915->contexts.hw_ida, 0, max, gfp);
192 static int steal_hw_id(struct drm_i915_private *i915)
194 struct i915_gem_context *ctx, *cn;
198 lockdep_assert_held(&i915->contexts.mutex);
200 list_for_each_entry_safe(ctx, cn,
201 &i915->contexts.hw_id_list, hw_id_link) {
202 if (atomic_read(&ctx->hw_id_pin_count)) {
203 list_move_tail(&ctx->hw_id_link, &pinned);
207 GEM_BUG_ON(!ctx->hw_id); /* perma-pinned kernel context */
208 list_del_init(&ctx->hw_id_link);
214 * Remember how far we got up on the last repossesion scan, so the
215 * list is kept in a "least recently scanned" order.
217 list_splice_tail(&pinned, &i915->contexts.hw_id_list);
221 static int assign_hw_id(struct drm_i915_private *i915, unsigned int *out)
225 lockdep_assert_held(&i915->contexts.mutex);
228 * We prefer to steal/stall ourselves and our users over that of the
229 * entire system. That may be a little unfair to our users, and
230 * even hurt high priority clients. The choice is whether to oomkill
231 * something else, or steal a context id.
233 ret = new_hw_id(i915, GFP_KERNEL | __GFP_RETRY_MAYFAIL | __GFP_NOWARN);
234 if (unlikely(ret < 0)) {
235 ret = steal_hw_id(i915);
236 if (ret < 0) /* once again for the correct errno code */
237 ret = new_hw_id(i915, GFP_KERNEL);
246 static void release_hw_id(struct i915_gem_context *ctx)
248 struct drm_i915_private *i915 = ctx->i915;
250 if (list_empty(&ctx->hw_id_link))
253 mutex_lock(&i915->contexts.mutex);
254 if (!list_empty(&ctx->hw_id_link)) {
255 ida_simple_remove(&i915->contexts.hw_ida, ctx->hw_id);
256 list_del_init(&ctx->hw_id_link);
258 mutex_unlock(&i915->contexts.mutex);
261 static void __free_engines(struct i915_gem_engines *e, unsigned int count)
264 if (!e->engines[count])
267 intel_context_put(e->engines[count]);
272 static void free_engines(struct i915_gem_engines *e)
274 __free_engines(e, e->num_engines);
277 static void free_engines_rcu(struct rcu_head *rcu)
279 free_engines(container_of(rcu, struct i915_gem_engines, rcu));
282 static struct i915_gem_engines *default_engines(struct i915_gem_context *ctx)
284 const struct intel_gt *gt = &ctx->i915->gt;
285 struct intel_engine_cs *engine;
286 struct i915_gem_engines *e;
287 enum intel_engine_id id;
289 e = kzalloc(struct_size(e, engines, I915_NUM_ENGINES), GFP_KERNEL);
291 return ERR_PTR(-ENOMEM);
293 init_rcu_head(&e->rcu);
294 for_each_engine(engine, gt, id) {
295 struct intel_context *ce;
297 ce = intel_context_create(ctx, engine);
299 __free_engines(e, id);
304 e->num_engines = id + 1;
310 static void i915_gem_context_free(struct i915_gem_context *ctx)
312 lockdep_assert_held(&ctx->i915->drm.struct_mutex);
313 GEM_BUG_ON(!i915_gem_context_is_closed(ctx));
317 i915_vm_put(ctx->vm);
319 free_engines(rcu_access_pointer(ctx->engines));
320 mutex_destroy(&ctx->engines_mutex);
323 intel_timeline_put(ctx->timeline);
328 list_del(&ctx->link);
329 mutex_destroy(&ctx->mutex);
334 static void contexts_free(struct drm_i915_private *i915)
336 struct llist_node *freed = llist_del_all(&i915->contexts.free_list);
337 struct i915_gem_context *ctx, *cn;
339 lockdep_assert_held(&i915->drm.struct_mutex);
341 llist_for_each_entry_safe(ctx, cn, freed, free_link)
342 i915_gem_context_free(ctx);
345 static void contexts_free_first(struct drm_i915_private *i915)
347 struct i915_gem_context *ctx;
348 struct llist_node *freed;
350 lockdep_assert_held(&i915->drm.struct_mutex);
352 freed = llist_del_first(&i915->contexts.free_list);
356 ctx = container_of(freed, typeof(*ctx), free_link);
357 i915_gem_context_free(ctx);
360 static void contexts_free_worker(struct work_struct *work)
362 struct drm_i915_private *i915 =
363 container_of(work, typeof(*i915), contexts.free_work);
365 mutex_lock(&i915->drm.struct_mutex);
367 mutex_unlock(&i915->drm.struct_mutex);
370 void i915_gem_context_release(struct kref *ref)
372 struct i915_gem_context *ctx = container_of(ref, typeof(*ctx), ref);
373 struct drm_i915_private *i915 = ctx->i915;
375 trace_i915_context_free(ctx);
376 if (llist_add(&ctx->free_link, &i915->contexts.free_list))
377 queue_work(i915->wq, &i915->contexts.free_work);
380 static void context_close(struct i915_gem_context *ctx)
382 mutex_lock(&ctx->mutex);
384 i915_gem_context_set_closed(ctx);
385 ctx->file_priv = ERR_PTR(-EBADF);
388 * This context will never again be assinged to HW, so we can
389 * reuse its ID for the next context.
394 * The LUT uses the VMA as a backpointer to unref the object,
395 * so we need to clear the LUT before we close all the VMA (inside
400 mutex_unlock(&ctx->mutex);
401 i915_gem_context_put(ctx);
404 static struct i915_gem_context *
405 __create_context(struct drm_i915_private *i915)
407 struct i915_gem_context *ctx;
408 struct i915_gem_engines *e;
412 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
414 return ERR_PTR(-ENOMEM);
416 kref_init(&ctx->ref);
417 list_add_tail(&ctx->link, &i915->contexts.list);
419 ctx->sched.priority = I915_USER_PRIORITY(I915_PRIORITY_NORMAL);
420 mutex_init(&ctx->mutex);
422 mutex_init(&ctx->engines_mutex);
423 e = default_engines(ctx);
428 RCU_INIT_POINTER(ctx->engines, e);
430 INIT_RADIX_TREE(&ctx->handles_vma, GFP_KERNEL);
431 INIT_LIST_HEAD(&ctx->hw_id_link);
433 /* NB: Mark all slices as needing a remap so that when the context first
434 * loads it will restore whatever remap state already exists. If there
435 * is no remap info, it will be a NOP. */
436 ctx->remap_slice = ALL_L3_SLICES(i915);
438 i915_gem_context_set_bannable(ctx);
439 i915_gem_context_set_recoverable(ctx);
441 for (i = 0; i < ARRAY_SIZE(ctx->hang_timestamp); i++)
442 ctx->hang_timestamp[i] = jiffies - CONTEXT_FAST_HANG_JIFFIES;
452 context_apply_all(struct i915_gem_context *ctx,
453 void (*fn)(struct intel_context *ce, void *data),
456 struct i915_gem_engines_iter it;
457 struct intel_context *ce;
459 for_each_gem_engine(ce, i915_gem_context_lock_engines(ctx), it)
461 i915_gem_context_unlock_engines(ctx);
464 static void __apply_ppgtt(struct intel_context *ce, void *vm)
467 ce->vm = i915_vm_get(vm);
470 static struct i915_address_space *
471 __set_ppgtt(struct i915_gem_context *ctx, struct i915_address_space *vm)
473 struct i915_address_space *old = ctx->vm;
475 GEM_BUG_ON(old && i915_vm_is_4lvl(vm) != i915_vm_is_4lvl(old));
477 ctx->vm = i915_vm_get(vm);
478 context_apply_all(ctx, __apply_ppgtt, vm);
483 static void __assign_ppgtt(struct i915_gem_context *ctx,
484 struct i915_address_space *vm)
489 vm = __set_ppgtt(ctx, vm);
494 static void __set_timeline(struct intel_timeline **dst,
495 struct intel_timeline *src)
497 struct intel_timeline *old = *dst;
499 *dst = src ? intel_timeline_get(src) : NULL;
502 intel_timeline_put(old);
505 static void __apply_timeline(struct intel_context *ce, void *timeline)
507 __set_timeline(&ce->timeline, timeline);
510 static void __assign_timeline(struct i915_gem_context *ctx,
511 struct intel_timeline *timeline)
513 __set_timeline(&ctx->timeline, timeline);
514 context_apply_all(ctx, __apply_timeline, timeline);
517 static struct i915_gem_context *
518 i915_gem_create_context(struct drm_i915_private *dev_priv, unsigned int flags)
520 struct i915_gem_context *ctx;
522 lockdep_assert_held(&dev_priv->drm.struct_mutex);
524 if (flags & I915_CONTEXT_CREATE_FLAGS_SINGLE_TIMELINE &&
525 !HAS_EXECLISTS(dev_priv))
526 return ERR_PTR(-EINVAL);
528 /* Reap the most stale context */
529 contexts_free_first(dev_priv);
531 ctx = __create_context(dev_priv);
535 if (HAS_FULL_PPGTT(dev_priv)) {
536 struct i915_ppgtt *ppgtt;
538 ppgtt = i915_ppgtt_create(dev_priv);
540 DRM_DEBUG_DRIVER("PPGTT setup failed (%ld)\n",
543 return ERR_CAST(ppgtt);
546 __assign_ppgtt(ctx, &ppgtt->vm);
547 i915_vm_put(&ppgtt->vm);
550 if (flags & I915_CONTEXT_CREATE_FLAGS_SINGLE_TIMELINE) {
551 struct intel_timeline *timeline;
553 timeline = intel_timeline_create(&dev_priv->gt, NULL);
554 if (IS_ERR(timeline)) {
556 return ERR_CAST(timeline);
559 __assign_timeline(ctx, timeline);
560 intel_timeline_put(timeline);
563 trace_i915_context_create(ctx);
569 destroy_kernel_context(struct i915_gem_context **ctxp)
571 struct i915_gem_context *ctx;
573 /* Keep the context ref so that we can free it immediately ourselves */
574 ctx = i915_gem_context_get(fetch_and_zero(ctxp));
575 GEM_BUG_ON(!i915_gem_context_is_kernel(ctx));
578 i915_gem_context_free(ctx);
581 struct i915_gem_context *
582 i915_gem_context_create_kernel(struct drm_i915_private *i915, int prio)
584 struct i915_gem_context *ctx;
587 ctx = i915_gem_create_context(i915, 0);
591 err = i915_gem_context_pin_hw_id(ctx);
593 destroy_kernel_context(&ctx);
597 i915_gem_context_clear_bannable(ctx);
598 ctx->sched.priority = I915_USER_PRIORITY(prio);
600 GEM_BUG_ON(!i915_gem_context_is_kernel(ctx));
605 static void init_contexts(struct drm_i915_private *i915)
607 mutex_init(&i915->contexts.mutex);
608 INIT_LIST_HEAD(&i915->contexts.list);
610 /* Using the simple ida interface, the max is limited by sizeof(int) */
611 BUILD_BUG_ON(MAX_CONTEXT_HW_ID > INT_MAX);
612 BUILD_BUG_ON(GEN11_MAX_CONTEXT_HW_ID > INT_MAX);
613 ida_init(&i915->contexts.hw_ida);
614 INIT_LIST_HEAD(&i915->contexts.hw_id_list);
616 INIT_WORK(&i915->contexts.free_work, contexts_free_worker);
617 init_llist_head(&i915->contexts.free_list);
620 int i915_gem_contexts_init(struct drm_i915_private *dev_priv)
622 struct i915_gem_context *ctx;
624 /* Reassure ourselves we are only called once */
625 GEM_BUG_ON(dev_priv->kernel_context);
627 init_contexts(dev_priv);
629 /* lowest priority; idle task */
630 ctx = i915_gem_context_create_kernel(dev_priv, I915_PRIORITY_MIN);
632 DRM_ERROR("Failed to create default global context\n");
636 * For easy recognisablity, we want the kernel context to be 0 and then
637 * all user contexts will have non-zero hw_id. Kernel contexts are
638 * permanently pinned, so that we never suffer a stall and can
639 * use them from any allocation context (e.g. for evicting other
640 * contexts and from inside the shrinker).
642 GEM_BUG_ON(ctx->hw_id);
643 GEM_BUG_ON(!atomic_read(&ctx->hw_id_pin_count));
644 dev_priv->kernel_context = ctx;
646 DRM_DEBUG_DRIVER("%s context support initialized\n",
647 DRIVER_CAPS(dev_priv)->has_logical_contexts ?
652 void i915_gem_contexts_fini(struct drm_i915_private *i915)
654 lockdep_assert_held(&i915->drm.struct_mutex);
656 destroy_kernel_context(&i915->kernel_context);
658 /* Must free all deferred contexts (via flush_workqueue) first */
659 GEM_BUG_ON(!list_empty(&i915->contexts.hw_id_list));
660 ida_destroy(&i915->contexts.hw_ida);
663 static int context_idr_cleanup(int id, void *p, void *data)
669 static int vm_idr_cleanup(int id, void *p, void *data)
675 static int gem_context_register(struct i915_gem_context *ctx,
676 struct drm_i915_file_private *fpriv)
680 ctx->file_priv = fpriv;
682 ctx->vm->file = fpriv;
684 ctx->pid = get_task_pid(current, PIDTYPE_PID);
685 ctx->name = kasprintf(GFP_KERNEL, "%s[%d]",
686 current->comm, pid_nr(ctx->pid));
692 /* And finally expose ourselves to userspace via the idr */
693 mutex_lock(&fpriv->context_idr_lock);
694 ret = idr_alloc(&fpriv->context_idr, ctx, 0, 0, GFP_KERNEL);
695 mutex_unlock(&fpriv->context_idr_lock);
699 kfree(fetch_and_zero(&ctx->name));
701 put_pid(fetch_and_zero(&ctx->pid));
706 int i915_gem_context_open(struct drm_i915_private *i915,
707 struct drm_file *file)
709 struct drm_i915_file_private *file_priv = file->driver_priv;
710 struct i915_gem_context *ctx;
713 mutex_init(&file_priv->context_idr_lock);
714 mutex_init(&file_priv->vm_idr_lock);
716 idr_init(&file_priv->context_idr);
717 idr_init_base(&file_priv->vm_idr, 1);
719 mutex_lock(&i915->drm.struct_mutex);
720 ctx = i915_gem_create_context(i915, 0);
721 mutex_unlock(&i915->drm.struct_mutex);
727 err = gem_context_register(ctx, file_priv);
731 GEM_BUG_ON(i915_gem_context_is_kernel(ctx));
739 idr_destroy(&file_priv->vm_idr);
740 idr_destroy(&file_priv->context_idr);
741 mutex_destroy(&file_priv->vm_idr_lock);
742 mutex_destroy(&file_priv->context_idr_lock);
746 void i915_gem_context_close(struct drm_file *file)
748 struct drm_i915_file_private *file_priv = file->driver_priv;
750 idr_for_each(&file_priv->context_idr, context_idr_cleanup, NULL);
751 idr_destroy(&file_priv->context_idr);
752 mutex_destroy(&file_priv->context_idr_lock);
754 idr_for_each(&file_priv->vm_idr, vm_idr_cleanup, NULL);
755 idr_destroy(&file_priv->vm_idr);
756 mutex_destroy(&file_priv->vm_idr_lock);
759 int i915_gem_vm_create_ioctl(struct drm_device *dev, void *data,
760 struct drm_file *file)
762 struct drm_i915_private *i915 = to_i915(dev);
763 struct drm_i915_gem_vm_control *args = data;
764 struct drm_i915_file_private *file_priv = file->driver_priv;
765 struct i915_ppgtt *ppgtt;
768 if (!HAS_FULL_PPGTT(i915))
774 ppgtt = i915_ppgtt_create(i915);
776 return PTR_ERR(ppgtt);
778 ppgtt->vm.file = file_priv;
780 if (args->extensions) {
781 err = i915_user_extensions(u64_to_user_ptr(args->extensions),
788 err = mutex_lock_interruptible(&file_priv->vm_idr_lock);
792 err = idr_alloc(&file_priv->vm_idr, &ppgtt->vm, 0, 0, GFP_KERNEL);
796 GEM_BUG_ON(err == 0); /* reserved for invalid/unassigned ppgtt */
798 mutex_unlock(&file_priv->vm_idr_lock);
804 mutex_unlock(&file_priv->vm_idr_lock);
806 i915_vm_put(&ppgtt->vm);
810 int i915_gem_vm_destroy_ioctl(struct drm_device *dev, void *data,
811 struct drm_file *file)
813 struct drm_i915_file_private *file_priv = file->driver_priv;
814 struct drm_i915_gem_vm_control *args = data;
815 struct i915_address_space *vm;
822 if (args->extensions)
829 err = mutex_lock_interruptible(&file_priv->vm_idr_lock);
833 vm = idr_remove(&file_priv->vm_idr, id);
835 mutex_unlock(&file_priv->vm_idr_lock);
843 struct context_barrier_task {
844 struct i915_active base;
845 void (*task)(void *data);
849 static void cb_retire(struct i915_active *base)
851 struct context_barrier_task *cb = container_of(base, typeof(*cb), base);
856 i915_active_fini(&cb->base);
860 I915_SELFTEST_DECLARE(static intel_engine_mask_t context_barrier_inject_fault);
861 static int context_barrier_task(struct i915_gem_context *ctx,
862 intel_engine_mask_t engines,
863 bool (*skip)(struct intel_context *ce, void *data),
864 int (*emit)(struct i915_request *rq, void *data),
865 void (*task)(void *data),
868 struct drm_i915_private *i915 = ctx->i915;
869 struct context_barrier_task *cb;
870 struct i915_gem_engines_iter it;
871 struct intel_context *ce;
874 lockdep_assert_held(&i915->drm.struct_mutex);
877 cb = kmalloc(sizeof(*cb), GFP_KERNEL);
881 i915_active_init(i915, &cb->base, NULL, cb_retire);
882 err = i915_active_acquire(&cb->base);
888 for_each_gem_engine(ce, i915_gem_context_lock_engines(ctx), it) {
889 struct i915_request *rq;
891 if (I915_SELFTEST_ONLY(context_barrier_inject_fault &
897 if (!(ce->engine->mask & engines))
900 if (skip && skip(ce, data))
903 rq = intel_context_create_request(ce);
911 err = emit(rq, data);
913 err = i915_active_ref(&cb->base, rq->timeline, rq);
915 i915_request_add(rq);
919 i915_gem_context_unlock_engines(ctx);
921 cb->task = err ? NULL : task; /* caller needs to unwind instead */
924 i915_active_release(&cb->base);
929 static int get_ppgtt(struct drm_i915_file_private *file_priv,
930 struct i915_gem_context *ctx,
931 struct drm_i915_gem_context_param *args)
933 struct i915_address_space *vm;
939 /* XXX rcu acquire? */
940 ret = mutex_lock_interruptible(&ctx->i915->drm.struct_mutex);
944 vm = i915_vm_get(ctx->vm);
945 mutex_unlock(&ctx->i915->drm.struct_mutex);
947 ret = mutex_lock_interruptible(&file_priv->vm_idr_lock);
951 ret = idr_alloc(&file_priv->vm_idr, vm, 0, 0, GFP_KERNEL);
963 mutex_unlock(&file_priv->vm_idr_lock);
969 static void set_ppgtt_barrier(void *data)
971 struct i915_address_space *old = data;
973 if (INTEL_GEN(old->i915) < 8)
974 gen6_ppgtt_unpin_all(i915_vm_to_ppgtt(old));
979 static int emit_ppgtt_update(struct i915_request *rq, void *data)
981 struct i915_address_space *vm = rq->hw_context->vm;
982 struct intel_engine_cs *engine = rq->engine;
983 u32 base = engine->mmio_base;
987 if (i915_vm_is_4lvl(vm)) {
988 struct i915_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
989 const dma_addr_t pd_daddr = px_dma(ppgtt->pd);
991 cs = intel_ring_begin(rq, 6);
995 *cs++ = MI_LOAD_REGISTER_IMM(2);
997 *cs++ = i915_mmio_reg_offset(GEN8_RING_PDP_UDW(base, 0));
998 *cs++ = upper_32_bits(pd_daddr);
999 *cs++ = i915_mmio_reg_offset(GEN8_RING_PDP_LDW(base, 0));
1000 *cs++ = lower_32_bits(pd_daddr);
1003 intel_ring_advance(rq, cs);
1004 } else if (HAS_LOGICAL_RING_CONTEXTS(engine->i915)) {
1005 struct i915_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1007 cs = intel_ring_begin(rq, 4 * GEN8_3LVL_PDPES + 2);
1011 *cs++ = MI_LOAD_REGISTER_IMM(2 * GEN8_3LVL_PDPES);
1012 for (i = GEN8_3LVL_PDPES; i--; ) {
1013 const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
1015 *cs++ = i915_mmio_reg_offset(GEN8_RING_PDP_UDW(base, i));
1016 *cs++ = upper_32_bits(pd_daddr);
1017 *cs++ = i915_mmio_reg_offset(GEN8_RING_PDP_LDW(base, i));
1018 *cs++ = lower_32_bits(pd_daddr);
1021 intel_ring_advance(rq, cs);
1023 /* ppGTT is not part of the legacy context image */
1024 gen6_ppgtt_pin(i915_vm_to_ppgtt(vm));
1030 static bool skip_ppgtt_update(struct intel_context *ce, void *data)
1032 if (HAS_LOGICAL_RING_CONTEXTS(ce->engine->i915))
1035 return !atomic_read(&ce->pin_count);
1038 static int set_ppgtt(struct drm_i915_file_private *file_priv,
1039 struct i915_gem_context *ctx,
1040 struct drm_i915_gem_context_param *args)
1042 struct i915_address_space *vm, *old;
1051 if (upper_32_bits(args->value))
1054 err = mutex_lock_interruptible(&file_priv->vm_idr_lock);
1058 vm = idr_find(&file_priv->vm_idr, args->value);
1061 mutex_unlock(&file_priv->vm_idr_lock);
1065 err = mutex_lock_interruptible(&ctx->i915->drm.struct_mutex);
1072 /* Teardown the existing obj:vma cache, it will have to be rebuilt. */
1073 mutex_lock(&ctx->mutex);
1075 mutex_unlock(&ctx->mutex);
1077 old = __set_ppgtt(ctx, vm);
1080 * We need to flush any requests using the current ppgtt before
1081 * we release it as the requests do not hold a reference themselves,
1082 * only indirectly through the context.
1084 err = context_barrier_task(ctx, ALL_ENGINES,
1090 i915_vm_put(__set_ppgtt(ctx, old));
1095 mutex_unlock(&ctx->i915->drm.struct_mutex);
1102 static int gen8_emit_rpcs_config(struct i915_request *rq,
1103 struct intel_context *ce,
1104 struct intel_sseu sseu)
1109 cs = intel_ring_begin(rq, 4);
1113 offset = i915_ggtt_offset(ce->state) +
1114 LRC_STATE_PN * PAGE_SIZE +
1115 (CTX_R_PWR_CLK_STATE + 1) * 4;
1117 *cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
1118 *cs++ = lower_32_bits(offset);
1119 *cs++ = upper_32_bits(offset);
1120 *cs++ = intel_sseu_make_rpcs(rq->i915, &sseu);
1122 intel_ring_advance(rq, cs);
1128 gen8_modify_rpcs(struct intel_context *ce, struct intel_sseu sseu)
1130 struct i915_request *rq;
1133 lockdep_assert_held(&ce->pin_mutex);
1136 * If the context is not idle, we have to submit an ordered request to
1137 * modify its context image via the kernel context (writing to our own
1138 * image, or into the registers directory, does not stick). Pristine
1139 * and idle contexts will be configured on pinning.
1141 if (!intel_context_is_pinned(ce))
1144 rq = i915_request_create(ce->engine->kernel_context);
1148 /* Serialise with the remote context */
1149 ret = intel_context_prepare_remote_request(ce, rq);
1151 ret = gen8_emit_rpcs_config(rq, ce, sseu);
1153 i915_request_add(rq);
1158 __intel_context_reconfigure_sseu(struct intel_context *ce,
1159 struct intel_sseu sseu)
1163 GEM_BUG_ON(INTEL_GEN(ce->engine->i915) < 8);
1165 ret = intel_context_lock_pinned(ce);
1169 /* Nothing to do if unmodified. */
1170 if (!memcmp(&ce->sseu, &sseu, sizeof(sseu)))
1173 ret = gen8_modify_rpcs(ce, sseu);
1178 intel_context_unlock_pinned(ce);
1183 intel_context_reconfigure_sseu(struct intel_context *ce, struct intel_sseu sseu)
1185 struct drm_i915_private *i915 = ce->engine->i915;
1188 ret = mutex_lock_interruptible(&i915->drm.struct_mutex);
1192 ret = __intel_context_reconfigure_sseu(ce, sseu);
1194 mutex_unlock(&i915->drm.struct_mutex);
1200 user_to_context_sseu(struct drm_i915_private *i915,
1201 const struct drm_i915_gem_context_param_sseu *user,
1202 struct intel_sseu *context)
1204 const struct sseu_dev_info *device = &RUNTIME_INFO(i915)->sseu;
1206 /* No zeros in any field. */
1207 if (!user->slice_mask || !user->subslice_mask ||
1208 !user->min_eus_per_subslice || !user->max_eus_per_subslice)
1212 if (user->max_eus_per_subslice < user->min_eus_per_subslice)
1216 * Some future proofing on the types since the uAPI is wider than the
1217 * current internal implementation.
1219 if (overflows_type(user->slice_mask, context->slice_mask) ||
1220 overflows_type(user->subslice_mask, context->subslice_mask) ||
1221 overflows_type(user->min_eus_per_subslice,
1222 context->min_eus_per_subslice) ||
1223 overflows_type(user->max_eus_per_subslice,
1224 context->max_eus_per_subslice))
1227 /* Check validity against hardware. */
1228 if (user->slice_mask & ~device->slice_mask)
1231 if (user->subslice_mask & ~device->subslice_mask[0])
1234 if (user->max_eus_per_subslice > device->max_eus_per_subslice)
1237 context->slice_mask = user->slice_mask;
1238 context->subslice_mask = user->subslice_mask;
1239 context->min_eus_per_subslice = user->min_eus_per_subslice;
1240 context->max_eus_per_subslice = user->max_eus_per_subslice;
1242 /* Part specific restrictions. */
1243 if (IS_GEN(i915, 11)) {
1244 unsigned int hw_s = hweight8(device->slice_mask);
1245 unsigned int hw_ss_per_s = hweight8(device->subslice_mask[0]);
1246 unsigned int req_s = hweight8(context->slice_mask);
1247 unsigned int req_ss = hweight8(context->subslice_mask);
1250 * Only full subslice enablement is possible if more than one
1251 * slice is turned on.
1253 if (req_s > 1 && req_ss != hw_ss_per_s)
1257 * If more than four (SScount bitfield limit) subslices are
1258 * requested then the number has to be even.
1260 if (req_ss > 4 && (req_ss & 1))
1264 * If only one slice is enabled and subslice count is below the
1265 * device full enablement, it must be at most half of the all
1266 * available subslices.
1268 if (req_s == 1 && req_ss < hw_ss_per_s &&
1269 req_ss > (hw_ss_per_s / 2))
1272 /* ABI restriction - VME use case only. */
1274 /* All slices or one slice only. */
1275 if (req_s != 1 && req_s != hw_s)
1279 * Half subslices or full enablement only when one slice is
1283 (req_ss != hw_ss_per_s && req_ss != (hw_ss_per_s / 2)))
1286 /* No EU configuration changes. */
1287 if ((user->min_eus_per_subslice !=
1288 device->max_eus_per_subslice) ||
1289 (user->max_eus_per_subslice !=
1290 device->max_eus_per_subslice))
1297 static int set_sseu(struct i915_gem_context *ctx,
1298 struct drm_i915_gem_context_param *args)
1300 struct drm_i915_private *i915 = ctx->i915;
1301 struct drm_i915_gem_context_param_sseu user_sseu;
1302 struct intel_context *ce;
1303 struct intel_sseu sseu;
1304 unsigned long lookup;
1307 if (args->size < sizeof(user_sseu))
1310 if (!IS_GEN(i915, 11))
1313 if (copy_from_user(&user_sseu, u64_to_user_ptr(args->value),
1320 if (user_sseu.flags & ~(I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX))
1324 if (user_sseu.flags & I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX)
1325 lookup |= LOOKUP_USER_INDEX;
1327 ce = lookup_user_engine(ctx, lookup, &user_sseu.engine);
1331 /* Only render engine supports RPCS configuration. */
1332 if (ce->engine->class != RENDER_CLASS) {
1337 ret = user_to_context_sseu(i915, &user_sseu, &sseu);
1341 ret = intel_context_reconfigure_sseu(ce, sseu);
1345 args->size = sizeof(user_sseu);
1348 intel_context_put(ce);
1352 struct set_engines {
1353 struct i915_gem_context *ctx;
1354 struct i915_gem_engines *engines;
1358 set_engines__load_balance(struct i915_user_extension __user *base, void *data)
1360 struct i915_context_engines_load_balance __user *ext =
1361 container_of_user(base, typeof(*ext), base);
1362 const struct set_engines *set = data;
1363 struct intel_engine_cs *stack[16];
1364 struct intel_engine_cs **siblings;
1365 struct intel_context *ce;
1366 u16 num_siblings, idx;
1370 if (!HAS_EXECLISTS(set->ctx->i915))
1373 if (USES_GUC_SUBMISSION(set->ctx->i915))
1374 return -ENODEV; /* not implement yet */
1376 if (get_user(idx, &ext->engine_index))
1379 if (idx >= set->engines->num_engines) {
1380 DRM_DEBUG("Invalid placement value, %d >= %d\n",
1381 idx, set->engines->num_engines);
1385 idx = array_index_nospec(idx, set->engines->num_engines);
1386 if (set->engines->engines[idx]) {
1387 DRM_DEBUG("Invalid placement[%d], already occupied\n", idx);
1391 if (get_user(num_siblings, &ext->num_siblings))
1394 err = check_user_mbz(&ext->flags);
1398 err = check_user_mbz(&ext->mbz64);
1403 if (num_siblings > ARRAY_SIZE(stack)) {
1404 siblings = kmalloc_array(num_siblings,
1411 for (n = 0; n < num_siblings; n++) {
1412 struct i915_engine_class_instance ci;
1414 if (copy_from_user(&ci, &ext->engines[n], sizeof(ci))) {
1419 siblings[n] = intel_engine_lookup_user(set->ctx->i915,
1421 ci.engine_instance);
1423 DRM_DEBUG("Invalid sibling[%d]: { class:%d, inst:%d }\n",
1424 n, ci.engine_class, ci.engine_instance);
1430 ce = intel_execlists_create_virtual(set->ctx, siblings, n);
1436 if (cmpxchg(&set->engines->engines[idx], NULL, ce)) {
1437 intel_context_put(ce);
1443 if (siblings != stack)
1450 set_engines__bond(struct i915_user_extension __user *base, void *data)
1452 struct i915_context_engines_bond __user *ext =
1453 container_of_user(base, typeof(*ext), base);
1454 const struct set_engines *set = data;
1455 struct i915_engine_class_instance ci;
1456 struct intel_engine_cs *virtual;
1457 struct intel_engine_cs *master;
1461 if (get_user(idx, &ext->virtual_index))
1464 if (idx >= set->engines->num_engines) {
1465 DRM_DEBUG("Invalid index for virtual engine: %d >= %d\n",
1466 idx, set->engines->num_engines);
1470 idx = array_index_nospec(idx, set->engines->num_engines);
1471 if (!set->engines->engines[idx]) {
1472 DRM_DEBUG("Invalid engine at %d\n", idx);
1475 virtual = set->engines->engines[idx]->engine;
1477 err = check_user_mbz(&ext->flags);
1481 for (n = 0; n < ARRAY_SIZE(ext->mbz64); n++) {
1482 err = check_user_mbz(&ext->mbz64[n]);
1487 if (copy_from_user(&ci, &ext->master, sizeof(ci)))
1490 master = intel_engine_lookup_user(set->ctx->i915,
1491 ci.engine_class, ci.engine_instance);
1493 DRM_DEBUG("Unrecognised master engine: { class:%u, instance:%u }\n",
1494 ci.engine_class, ci.engine_instance);
1498 if (get_user(num_bonds, &ext->num_bonds))
1501 for (n = 0; n < num_bonds; n++) {
1502 struct intel_engine_cs *bond;
1504 if (copy_from_user(&ci, &ext->engines[n], sizeof(ci)))
1507 bond = intel_engine_lookup_user(set->ctx->i915,
1509 ci.engine_instance);
1511 DRM_DEBUG("Unrecognised engine[%d] for bonding: { class:%d, instance: %d }\n",
1512 n, ci.engine_class, ci.engine_instance);
1517 * A non-virtual engine has no siblings to choose between; and
1518 * a submit fence will always be directed to the one engine.
1520 if (intel_engine_is_virtual(virtual)) {
1521 err = intel_virtual_engine_attach_bond(virtual,
1532 static const i915_user_extension_fn set_engines__extensions[] = {
1533 [I915_CONTEXT_ENGINES_EXT_LOAD_BALANCE] = set_engines__load_balance,
1534 [I915_CONTEXT_ENGINES_EXT_BOND] = set_engines__bond,
1538 set_engines(struct i915_gem_context *ctx,
1539 const struct drm_i915_gem_context_param *args)
1541 struct i915_context_param_engines __user *user =
1542 u64_to_user_ptr(args->value);
1543 struct set_engines set = { .ctx = ctx };
1544 unsigned int num_engines, n;
1548 if (!args->size) { /* switch back to legacy user_ring_map */
1549 if (!i915_gem_context_user_engines(ctx))
1552 set.engines = default_engines(ctx);
1553 if (IS_ERR(set.engines))
1554 return PTR_ERR(set.engines);
1559 BUILD_BUG_ON(!IS_ALIGNED(sizeof(*user), sizeof(*user->engines)));
1560 if (args->size < sizeof(*user) ||
1561 !IS_ALIGNED(args->size, sizeof(*user->engines))) {
1562 DRM_DEBUG("Invalid size for engine array: %d\n",
1568 * Note that I915_EXEC_RING_MASK limits execbuf to only using the
1569 * first 64 engines defined here.
1571 num_engines = (args->size - sizeof(*user)) / sizeof(*user->engines);
1573 set.engines = kmalloc(struct_size(set.engines, engines, num_engines),
1578 init_rcu_head(&set.engines->rcu);
1579 for (n = 0; n < num_engines; n++) {
1580 struct i915_engine_class_instance ci;
1581 struct intel_engine_cs *engine;
1582 struct intel_context *ce;
1584 if (copy_from_user(&ci, &user->engines[n], sizeof(ci))) {
1585 __free_engines(set.engines, n);
1589 if (ci.engine_class == (u16)I915_ENGINE_CLASS_INVALID &&
1590 ci.engine_instance == (u16)I915_ENGINE_CLASS_INVALID_NONE) {
1591 set.engines->engines[n] = NULL;
1595 engine = intel_engine_lookup_user(ctx->i915,
1597 ci.engine_instance);
1599 DRM_DEBUG("Invalid engine[%d]: { class:%d, instance:%d }\n",
1600 n, ci.engine_class, ci.engine_instance);
1601 __free_engines(set.engines, n);
1605 ce = intel_context_create(ctx, engine);
1607 __free_engines(set.engines, n);
1611 set.engines->engines[n] = ce;
1613 set.engines->num_engines = num_engines;
1616 if (!get_user(extensions, &user->extensions))
1617 err = i915_user_extensions(u64_to_user_ptr(extensions),
1618 set_engines__extensions,
1619 ARRAY_SIZE(set_engines__extensions),
1622 free_engines(set.engines);
1627 mutex_lock(&ctx->engines_mutex);
1629 i915_gem_context_set_user_engines(ctx);
1631 i915_gem_context_clear_user_engines(ctx);
1632 set.engines = rcu_replace_pointer(ctx->engines, set.engines, 1);
1633 mutex_unlock(&ctx->engines_mutex);
1635 call_rcu(&set.engines->rcu, free_engines_rcu);
1640 static struct i915_gem_engines *
1641 __copy_engines(struct i915_gem_engines *e)
1643 struct i915_gem_engines *copy;
1646 copy = kmalloc(struct_size(e, engines, e->num_engines), GFP_KERNEL);
1648 return ERR_PTR(-ENOMEM);
1650 init_rcu_head(©->rcu);
1651 for (n = 0; n < e->num_engines; n++) {
1653 copy->engines[n] = intel_context_get(e->engines[n]);
1655 copy->engines[n] = NULL;
1657 copy->num_engines = n;
1663 get_engines(struct i915_gem_context *ctx,
1664 struct drm_i915_gem_context_param *args)
1666 struct i915_context_param_engines __user *user;
1667 struct i915_gem_engines *e;
1668 size_t n, count, size;
1671 err = mutex_lock_interruptible(&ctx->engines_mutex);
1676 if (i915_gem_context_user_engines(ctx))
1677 e = __copy_engines(i915_gem_context_engines(ctx));
1678 mutex_unlock(&ctx->engines_mutex);
1679 if (IS_ERR_OR_NULL(e)) {
1681 return PTR_ERR_OR_ZERO(e);
1684 count = e->num_engines;
1686 /* Be paranoid in case we have an impedance mismatch */
1687 if (!check_struct_size(user, engines, count, &size)) {
1691 if (overflows_type(size, args->size)) {
1701 if (args->size < size) {
1706 user = u64_to_user_ptr(args->value);
1707 if (!access_ok(user, size)) {
1712 if (put_user(0, &user->extensions)) {
1717 for (n = 0; n < count; n++) {
1718 struct i915_engine_class_instance ci = {
1719 .engine_class = I915_ENGINE_CLASS_INVALID,
1720 .engine_instance = I915_ENGINE_CLASS_INVALID_NONE,
1723 if (e->engines[n]) {
1724 ci.engine_class = e->engines[n]->engine->uabi_class;
1725 ci.engine_instance = e->engines[n]->engine->uabi_instance;
1728 if (copy_to_user(&user->engines[n], &ci, sizeof(ci))) {
1741 static int ctx_setparam(struct drm_i915_file_private *fpriv,
1742 struct i915_gem_context *ctx,
1743 struct drm_i915_gem_context_param *args)
1747 switch (args->param) {
1748 case I915_CONTEXT_PARAM_NO_ZEROMAP:
1751 else if (args->value)
1752 set_bit(UCONTEXT_NO_ZEROMAP, &ctx->user_flags);
1754 clear_bit(UCONTEXT_NO_ZEROMAP, &ctx->user_flags);
1757 case I915_CONTEXT_PARAM_NO_ERROR_CAPTURE:
1760 else if (args->value)
1761 i915_gem_context_set_no_error_capture(ctx);
1763 i915_gem_context_clear_no_error_capture(ctx);
1766 case I915_CONTEXT_PARAM_BANNABLE:
1769 else if (!capable(CAP_SYS_ADMIN) && !args->value)
1771 else if (args->value)
1772 i915_gem_context_set_bannable(ctx);
1774 i915_gem_context_clear_bannable(ctx);
1777 case I915_CONTEXT_PARAM_RECOVERABLE:
1780 else if (args->value)
1781 i915_gem_context_set_recoverable(ctx);
1783 i915_gem_context_clear_recoverable(ctx);
1786 case I915_CONTEXT_PARAM_PRIORITY:
1788 s64 priority = args->value;
1792 else if (!(ctx->i915->caps.scheduler & I915_SCHEDULER_CAP_PRIORITY))
1794 else if (priority > I915_CONTEXT_MAX_USER_PRIORITY ||
1795 priority < I915_CONTEXT_MIN_USER_PRIORITY)
1797 else if (priority > I915_CONTEXT_DEFAULT_PRIORITY &&
1798 !capable(CAP_SYS_NICE))
1801 ctx->sched.priority =
1802 I915_USER_PRIORITY(priority);
1806 case I915_CONTEXT_PARAM_SSEU:
1807 ret = set_sseu(ctx, args);
1810 case I915_CONTEXT_PARAM_VM:
1811 ret = set_ppgtt(fpriv, ctx, args);
1814 case I915_CONTEXT_PARAM_ENGINES:
1815 ret = set_engines(ctx, args);
1818 case I915_CONTEXT_PARAM_BAN_PERIOD:
1828 struct i915_gem_context *ctx;
1829 struct drm_i915_file_private *fpriv;
1832 static int create_setparam(struct i915_user_extension __user *ext, void *data)
1834 struct drm_i915_gem_context_create_ext_setparam local;
1835 const struct create_ext *arg = data;
1837 if (copy_from_user(&local, ext, sizeof(local)))
1840 if (local.param.ctx_id)
1843 return ctx_setparam(arg->fpriv, arg->ctx, &local.param);
1846 static int clone_engines(struct i915_gem_context *dst,
1847 struct i915_gem_context *src)
1849 struct i915_gem_engines *e = i915_gem_context_lock_engines(src);
1850 struct i915_gem_engines *clone;
1854 clone = kmalloc(struct_size(e, engines, e->num_engines), GFP_KERNEL);
1858 init_rcu_head(&clone->rcu);
1859 for (n = 0; n < e->num_engines; n++) {
1860 struct intel_engine_cs *engine;
1862 if (!e->engines[n]) {
1863 clone->engines[n] = NULL;
1866 engine = e->engines[n]->engine;
1869 * Virtual engines are singletons; they can only exist
1870 * inside a single context, because they embed their
1871 * HW context... As each virtual context implies a single
1872 * timeline (each engine can only dequeue a single request
1873 * at any time), it would be surprising for two contexts
1874 * to use the same engine. So let's create a copy of
1875 * the virtual engine instead.
1877 if (intel_engine_is_virtual(engine))
1879 intel_execlists_clone_virtual(dst, engine);
1881 clone->engines[n] = intel_context_create(dst, engine);
1882 if (IS_ERR_OR_NULL(clone->engines[n])) {
1883 __free_engines(clone, n);
1887 clone->num_engines = n;
1889 user_engines = i915_gem_context_user_engines(src);
1890 i915_gem_context_unlock_engines(src);
1892 free_engines(dst->engines);
1893 RCU_INIT_POINTER(dst->engines, clone);
1895 i915_gem_context_set_user_engines(dst);
1897 i915_gem_context_clear_user_engines(dst);
1901 i915_gem_context_unlock_engines(src);
1905 static int clone_flags(struct i915_gem_context *dst,
1906 struct i915_gem_context *src)
1908 dst->user_flags = src->user_flags;
1912 static int clone_schedattr(struct i915_gem_context *dst,
1913 struct i915_gem_context *src)
1915 dst->sched = src->sched;
1919 static int clone_sseu(struct i915_gem_context *dst,
1920 struct i915_gem_context *src)
1922 struct i915_gem_engines *e = i915_gem_context_lock_engines(src);
1923 struct i915_gem_engines *clone;
1927 clone = dst->engines; /* no locking required; sole access */
1928 if (e->num_engines != clone->num_engines) {
1933 for (n = 0; n < e->num_engines; n++) {
1934 struct intel_context *ce = e->engines[n];
1936 if (clone->engines[n]->engine->class != ce->engine->class) {
1937 /* Must have compatible engine maps! */
1942 /* serialises with set_sseu */
1943 err = intel_context_lock_pinned(ce);
1947 clone->engines[n]->sseu = ce->sseu;
1948 intel_context_unlock_pinned(ce);
1953 i915_gem_context_unlock_engines(src);
1957 static int clone_timeline(struct i915_gem_context *dst,
1958 struct i915_gem_context *src)
1961 __assign_timeline(dst, src->timeline);
1966 static int clone_vm(struct i915_gem_context *dst,
1967 struct i915_gem_context *src)
1969 struct i915_address_space *vm;
1973 vm = READ_ONCE(src->vm);
1977 if (!kref_get_unless_zero(&vm->ref))
1981 * This ppgtt may have be reallocated between
1982 * the read and the kref, and reassigned to a third
1983 * context. In order to avoid inadvertent sharing
1984 * of this ppgtt with that third context (and not
1985 * src), we have to confirm that we have the same
1986 * ppgtt after passing through the strong memory
1987 * barrier implied by a successful
1988 * kref_get_unless_zero().
1990 * Once we have acquired the current ppgtt of src,
1991 * we no longer care if it is released from src, as
1992 * it cannot be reallocated elsewhere.
1995 if (vm == READ_ONCE(src->vm))
2003 __assign_ppgtt(dst, vm);
2010 static int create_clone(struct i915_user_extension __user *ext, void *data)
2012 static int (* const fn[])(struct i915_gem_context *dst,
2013 struct i915_gem_context *src) = {
2014 #define MAP(x, y) [ilog2(I915_CONTEXT_CLONE_##x)] = y
2015 MAP(ENGINES, clone_engines),
2016 MAP(FLAGS, clone_flags),
2017 MAP(SCHEDATTR, clone_schedattr),
2018 MAP(SSEU, clone_sseu),
2019 MAP(TIMELINE, clone_timeline),
2023 struct drm_i915_gem_context_create_ext_clone local;
2024 const struct create_ext *arg = data;
2025 struct i915_gem_context *dst = arg->ctx;
2026 struct i915_gem_context *src;
2029 if (copy_from_user(&local, ext, sizeof(local)))
2032 BUILD_BUG_ON(GENMASK(BITS_PER_TYPE(local.flags) - 1, ARRAY_SIZE(fn)) !=
2033 I915_CONTEXT_CLONE_UNKNOWN);
2035 if (local.flags & I915_CONTEXT_CLONE_UNKNOWN)
2042 src = __i915_gem_context_lookup_rcu(arg->fpriv, local.clone_id);
2047 GEM_BUG_ON(src == dst);
2049 for (bit = 0; bit < ARRAY_SIZE(fn); bit++) {
2050 if (!(local.flags & BIT(bit)))
2053 err = fn[bit](dst, src);
2061 static const i915_user_extension_fn create_extensions[] = {
2062 [I915_CONTEXT_CREATE_EXT_SETPARAM] = create_setparam,
2063 [I915_CONTEXT_CREATE_EXT_CLONE] = create_clone,
2066 static bool client_is_banned(struct drm_i915_file_private *file_priv)
2068 return atomic_read(&file_priv->ban_score) >= I915_CLIENT_SCORE_BANNED;
2071 int i915_gem_context_create_ioctl(struct drm_device *dev, void *data,
2072 struct drm_file *file)
2074 struct drm_i915_private *i915 = to_i915(dev);
2075 struct drm_i915_gem_context_create_ext *args = data;
2076 struct create_ext ext_data;
2079 if (!DRIVER_CAPS(i915)->has_logical_contexts)
2082 if (args->flags & I915_CONTEXT_CREATE_FLAGS_UNKNOWN)
2085 ret = intel_gt_terminally_wedged(&i915->gt);
2089 ext_data.fpriv = file->driver_priv;
2090 if (client_is_banned(ext_data.fpriv)) {
2091 DRM_DEBUG("client %s[%d] banned from creating ctx\n",
2093 pid_nr(get_task_pid(current, PIDTYPE_PID)));
2097 ret = i915_mutex_lock_interruptible(dev);
2101 ext_data.ctx = i915_gem_create_context(i915, args->flags);
2102 mutex_unlock(&dev->struct_mutex);
2103 if (IS_ERR(ext_data.ctx))
2104 return PTR_ERR(ext_data.ctx);
2106 if (args->flags & I915_CONTEXT_CREATE_FLAGS_USE_EXTENSIONS) {
2107 ret = i915_user_extensions(u64_to_user_ptr(args->extensions),
2109 ARRAY_SIZE(create_extensions),
2115 ret = gem_context_register(ext_data.ctx, ext_data.fpriv);
2120 DRM_DEBUG("HW context %d created\n", args->ctx_id);
2125 context_close(ext_data.ctx);
2129 int i915_gem_context_destroy_ioctl(struct drm_device *dev, void *data,
2130 struct drm_file *file)
2132 struct drm_i915_gem_context_destroy *args = data;
2133 struct drm_i915_file_private *file_priv = file->driver_priv;
2134 struct i915_gem_context *ctx;
2142 if (mutex_lock_interruptible(&file_priv->context_idr_lock))
2145 ctx = idr_remove(&file_priv->context_idr, args->ctx_id);
2146 mutex_unlock(&file_priv->context_idr_lock);
2154 static int get_sseu(struct i915_gem_context *ctx,
2155 struct drm_i915_gem_context_param *args)
2157 struct drm_i915_gem_context_param_sseu user_sseu;
2158 struct intel_context *ce;
2159 unsigned long lookup;
2162 if (args->size == 0)
2164 else if (args->size < sizeof(user_sseu))
2167 if (copy_from_user(&user_sseu, u64_to_user_ptr(args->value),
2174 if (user_sseu.flags & ~(I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX))
2178 if (user_sseu.flags & I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX)
2179 lookup |= LOOKUP_USER_INDEX;
2181 ce = lookup_user_engine(ctx, lookup, &user_sseu.engine);
2185 err = intel_context_lock_pinned(ce); /* serialises with set_sseu */
2187 intel_context_put(ce);
2191 user_sseu.slice_mask = ce->sseu.slice_mask;
2192 user_sseu.subslice_mask = ce->sseu.subslice_mask;
2193 user_sseu.min_eus_per_subslice = ce->sseu.min_eus_per_subslice;
2194 user_sseu.max_eus_per_subslice = ce->sseu.max_eus_per_subslice;
2196 intel_context_unlock_pinned(ce);
2197 intel_context_put(ce);
2199 if (copy_to_user(u64_to_user_ptr(args->value), &user_sseu,
2204 args->size = sizeof(user_sseu);
2209 int i915_gem_context_getparam_ioctl(struct drm_device *dev, void *data,
2210 struct drm_file *file)
2212 struct drm_i915_file_private *file_priv = file->driver_priv;
2213 struct drm_i915_gem_context_param *args = data;
2214 struct i915_gem_context *ctx;
2217 ctx = i915_gem_context_lookup(file_priv, args->ctx_id);
2221 switch (args->param) {
2222 case I915_CONTEXT_PARAM_NO_ZEROMAP:
2224 args->value = test_bit(UCONTEXT_NO_ZEROMAP, &ctx->user_flags);
2227 case I915_CONTEXT_PARAM_GTT_SIZE:
2230 args->value = ctx->vm->total;
2231 else if (to_i915(dev)->ggtt.alias)
2232 args->value = to_i915(dev)->ggtt.alias->vm.total;
2234 args->value = to_i915(dev)->ggtt.vm.total;
2237 case I915_CONTEXT_PARAM_NO_ERROR_CAPTURE:
2239 args->value = i915_gem_context_no_error_capture(ctx);
2242 case I915_CONTEXT_PARAM_BANNABLE:
2244 args->value = i915_gem_context_is_bannable(ctx);
2247 case I915_CONTEXT_PARAM_RECOVERABLE:
2249 args->value = i915_gem_context_is_recoverable(ctx);
2252 case I915_CONTEXT_PARAM_PRIORITY:
2254 args->value = ctx->sched.priority >> I915_USER_PRIORITY_SHIFT;
2257 case I915_CONTEXT_PARAM_SSEU:
2258 ret = get_sseu(ctx, args);
2261 case I915_CONTEXT_PARAM_VM:
2262 ret = get_ppgtt(file_priv, ctx, args);
2265 case I915_CONTEXT_PARAM_ENGINES:
2266 ret = get_engines(ctx, args);
2269 case I915_CONTEXT_PARAM_BAN_PERIOD:
2275 i915_gem_context_put(ctx);
2279 int i915_gem_context_setparam_ioctl(struct drm_device *dev, void *data,
2280 struct drm_file *file)
2282 struct drm_i915_file_private *file_priv = file->driver_priv;
2283 struct drm_i915_gem_context_param *args = data;
2284 struct i915_gem_context *ctx;
2287 ctx = i915_gem_context_lookup(file_priv, args->ctx_id);
2291 ret = ctx_setparam(file_priv, ctx, args);
2293 i915_gem_context_put(ctx);
2297 int i915_gem_context_reset_stats_ioctl(struct drm_device *dev,
2298 void *data, struct drm_file *file)
2300 struct drm_i915_private *dev_priv = to_i915(dev);
2301 struct drm_i915_reset_stats *args = data;
2302 struct i915_gem_context *ctx;
2305 if (args->flags || args->pad)
2310 ctx = __i915_gem_context_lookup_rcu(file->driver_priv, args->ctx_id);
2315 * We opt for unserialised reads here. This may result in tearing
2316 * in the extremely unlikely event of a GPU hang on this context
2317 * as we are querying them. If we need that extra layer of protection,
2318 * we should wrap the hangstats with a seqlock.
2321 if (capable(CAP_SYS_ADMIN))
2322 args->reset_count = i915_reset_count(&dev_priv->gpu_error);
2324 args->reset_count = 0;
2326 args->batch_active = atomic_read(&ctx->guilty_count);
2327 args->batch_pending = atomic_read(&ctx->active_count);
2335 int __i915_gem_context_pin_hw_id(struct i915_gem_context *ctx)
2337 struct drm_i915_private *i915 = ctx->i915;
2340 mutex_lock(&i915->contexts.mutex);
2342 GEM_BUG_ON(i915_gem_context_is_closed(ctx));
2344 if (list_empty(&ctx->hw_id_link)) {
2345 GEM_BUG_ON(atomic_read(&ctx->hw_id_pin_count));
2347 err = assign_hw_id(i915, &ctx->hw_id);
2351 list_add_tail(&ctx->hw_id_link, &i915->contexts.hw_id_list);
2354 GEM_BUG_ON(atomic_read(&ctx->hw_id_pin_count) == ~0u);
2355 atomic_inc(&ctx->hw_id_pin_count);
2358 mutex_unlock(&i915->contexts.mutex);
2362 /* GEM context-engines iterator: for_each_gem_engine() */
2363 struct intel_context *
2364 i915_gem_engines_iter_next(struct i915_gem_engines_iter *it)
2366 const struct i915_gem_engines *e = it->engines;
2367 struct intel_context *ctx;
2370 if (it->idx >= e->num_engines)
2373 ctx = e->engines[it->idx++];
2379 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
2380 #include "selftests/mock_context.c"
2381 #include "selftests/i915_gem_context.c"
2384 static void i915_global_gem_context_shrink(void)
2386 kmem_cache_shrink(global.slab_luts);
2389 static void i915_global_gem_context_exit(void)
2391 kmem_cache_destroy(global.slab_luts);
2394 static struct i915_global_gem_context global = { {
2395 .shrink = i915_global_gem_context_shrink,
2396 .exit = i915_global_gem_context_exit,
2399 int __init i915_global_gem_context_init(void)
2401 global.slab_luts = KMEM_CACHE(i915_lut_handle, 0);
2402 if (!global.slab_luts)
2405 i915_global_register(&global.base);