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) >= 11)
177 max = GEN11_MAX_CONTEXT_HW_ID;
178 else if (USES_GUC_SUBMISSION(i915))
180 * When using GuC in proxy submission, GuC consumes the
181 * highest bit in the context id to indicate proxy submission.
183 max = MAX_GUC_CONTEXT_HW_ID;
185 max = MAX_CONTEXT_HW_ID;
187 return ida_simple_get(&i915->contexts.hw_ida, 0, max, gfp);
190 static int steal_hw_id(struct drm_i915_private *i915)
192 struct i915_gem_context *ctx, *cn;
196 lockdep_assert_held(&i915->contexts.mutex);
198 list_for_each_entry_safe(ctx, cn,
199 &i915->contexts.hw_id_list, hw_id_link) {
200 if (atomic_read(&ctx->hw_id_pin_count)) {
201 list_move_tail(&ctx->hw_id_link, &pinned);
205 GEM_BUG_ON(!ctx->hw_id); /* perma-pinned kernel context */
206 list_del_init(&ctx->hw_id_link);
212 * Remember how far we got up on the last repossesion scan, so the
213 * list is kept in a "least recently scanned" order.
215 list_splice_tail(&pinned, &i915->contexts.hw_id_list);
219 static int assign_hw_id(struct drm_i915_private *i915, unsigned int *out)
223 lockdep_assert_held(&i915->contexts.mutex);
226 * We prefer to steal/stall ourselves and our users over that of the
227 * entire system. That may be a little unfair to our users, and
228 * even hurt high priority clients. The choice is whether to oomkill
229 * something else, or steal a context id.
231 ret = new_hw_id(i915, GFP_KERNEL | __GFP_RETRY_MAYFAIL | __GFP_NOWARN);
232 if (unlikely(ret < 0)) {
233 ret = steal_hw_id(i915);
234 if (ret < 0) /* once again for the correct errno code */
235 ret = new_hw_id(i915, GFP_KERNEL);
244 static void release_hw_id(struct i915_gem_context *ctx)
246 struct drm_i915_private *i915 = ctx->i915;
248 if (list_empty(&ctx->hw_id_link))
251 mutex_lock(&i915->contexts.mutex);
252 if (!list_empty(&ctx->hw_id_link)) {
253 ida_simple_remove(&i915->contexts.hw_ida, ctx->hw_id);
254 list_del_init(&ctx->hw_id_link);
256 mutex_unlock(&i915->contexts.mutex);
259 static void __free_engines(struct i915_gem_engines *e, unsigned int count)
262 if (!e->engines[count])
265 intel_context_put(e->engines[count]);
270 static void free_engines(struct i915_gem_engines *e)
272 __free_engines(e, e->num_engines);
275 static void free_engines_rcu(struct rcu_head *rcu)
277 free_engines(container_of(rcu, struct i915_gem_engines, rcu));
280 static struct i915_gem_engines *default_engines(struct i915_gem_context *ctx)
282 const struct intel_gt *gt = &ctx->i915->gt;
283 struct intel_engine_cs *engine;
284 struct i915_gem_engines *e;
285 enum intel_engine_id id;
287 e = kzalloc(struct_size(e, engines, I915_NUM_ENGINES), GFP_KERNEL);
289 return ERR_PTR(-ENOMEM);
291 init_rcu_head(&e->rcu);
292 for_each_engine(engine, gt, id) {
293 struct intel_context *ce;
295 ce = intel_context_create(ctx, engine);
297 __free_engines(e, id);
302 e->num_engines = id + 1;
308 static void i915_gem_context_free(struct i915_gem_context *ctx)
310 lockdep_assert_held(&ctx->i915->drm.struct_mutex);
311 GEM_BUG_ON(!i915_gem_context_is_closed(ctx));
315 i915_vm_put(ctx->vm);
317 free_engines(rcu_access_pointer(ctx->engines));
318 mutex_destroy(&ctx->engines_mutex);
321 intel_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 struct i915_gem_context *
403 __create_context(struct drm_i915_private *i915)
405 struct i915_gem_context *ctx;
406 struct i915_gem_engines *e;
410 ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
412 return ERR_PTR(-ENOMEM);
414 kref_init(&ctx->ref);
415 list_add_tail(&ctx->link, &i915->contexts.list);
417 ctx->sched.priority = I915_USER_PRIORITY(I915_PRIORITY_NORMAL);
418 mutex_init(&ctx->mutex);
420 mutex_init(&ctx->engines_mutex);
421 e = default_engines(ctx);
426 RCU_INIT_POINTER(ctx->engines, e);
428 INIT_RADIX_TREE(&ctx->handles_vma, GFP_KERNEL);
429 INIT_LIST_HEAD(&ctx->hw_id_link);
431 /* NB: Mark all slices as needing a remap so that when the context first
432 * loads it will restore whatever remap state already exists. If there
433 * is no remap info, it will be a NOP. */
434 ctx->remap_slice = ALL_L3_SLICES(i915);
436 i915_gem_context_set_bannable(ctx);
437 i915_gem_context_set_recoverable(ctx);
439 ctx->ring_size = 4 * PAGE_SIZE;
441 for (i = 0; i < ARRAY_SIZE(ctx->hang_timestamp); i++)
442 ctx->hang_timestamp[i] = jiffies - CONTEXT_FAST_HANG_JIFFIES;
451 static struct i915_address_space *
452 __set_ppgtt(struct i915_gem_context *ctx, struct i915_address_space *vm)
454 struct i915_address_space *old = ctx->vm;
455 struct i915_gem_engines_iter it;
456 struct intel_context *ce;
458 GEM_BUG_ON(old && i915_vm_is_4lvl(vm) != i915_vm_is_4lvl(old));
460 ctx->vm = i915_vm_get(vm);
462 for_each_gem_engine(ce, i915_gem_context_lock_engines(ctx), it) {
464 ce->vm = i915_vm_get(vm);
466 i915_gem_context_unlock_engines(ctx);
471 static void __assign_ppgtt(struct i915_gem_context *ctx,
472 struct i915_address_space *vm)
477 vm = __set_ppgtt(ctx, vm);
482 static struct i915_gem_context *
483 i915_gem_create_context(struct drm_i915_private *dev_priv, unsigned int flags)
485 struct i915_gem_context *ctx;
487 lockdep_assert_held(&dev_priv->drm.struct_mutex);
489 if (flags & I915_CONTEXT_CREATE_FLAGS_SINGLE_TIMELINE &&
490 !HAS_EXECLISTS(dev_priv))
491 return ERR_PTR(-EINVAL);
493 /* Reap the most stale context */
494 contexts_free_first(dev_priv);
496 ctx = __create_context(dev_priv);
500 if (HAS_FULL_PPGTT(dev_priv)) {
501 struct i915_ppgtt *ppgtt;
503 ppgtt = i915_ppgtt_create(dev_priv);
505 DRM_DEBUG_DRIVER("PPGTT setup failed (%ld)\n",
508 return ERR_CAST(ppgtt);
511 __assign_ppgtt(ctx, &ppgtt->vm);
512 i915_vm_put(&ppgtt->vm);
515 if (flags & I915_CONTEXT_CREATE_FLAGS_SINGLE_TIMELINE) {
516 struct intel_timeline *timeline;
518 timeline = intel_timeline_create(&dev_priv->gt, NULL);
519 if (IS_ERR(timeline)) {
521 return ERR_CAST(timeline);
524 ctx->timeline = timeline;
527 trace_i915_context_create(ctx);
533 * i915_gem_context_create_gvt - create a GVT GEM context
536 * This function is used to create a GVT specific GEM context.
539 * pointer to i915_gem_context on success, error pointer if failed
542 struct i915_gem_context *
543 i915_gem_context_create_gvt(struct drm_device *dev)
545 struct i915_gem_context *ctx;
548 if (!IS_ENABLED(CONFIG_DRM_I915_GVT))
549 return ERR_PTR(-ENODEV);
551 ret = i915_mutex_lock_interruptible(dev);
555 ctx = i915_gem_create_context(to_i915(dev), 0);
559 ret = i915_gem_context_pin_hw_id(ctx);
566 ctx->file_priv = ERR_PTR(-EBADF);
567 i915_gem_context_set_closed(ctx); /* not user accessible */
568 i915_gem_context_clear_bannable(ctx);
569 i915_gem_context_set_force_single_submission(ctx);
570 if (!USES_GUC_SUBMISSION(to_i915(dev)))
571 ctx->ring_size = 512 * PAGE_SIZE; /* Max ring buffer size */
573 GEM_BUG_ON(i915_gem_context_is_kernel(ctx));
575 mutex_unlock(&dev->struct_mutex);
580 destroy_kernel_context(struct i915_gem_context **ctxp)
582 struct i915_gem_context *ctx;
584 /* Keep the context ref so that we can free it immediately ourselves */
585 ctx = i915_gem_context_get(fetch_and_zero(ctxp));
586 GEM_BUG_ON(!i915_gem_context_is_kernel(ctx));
589 i915_gem_context_free(ctx);
592 struct i915_gem_context *
593 i915_gem_context_create_kernel(struct drm_i915_private *i915, int prio)
595 struct i915_gem_context *ctx;
598 ctx = i915_gem_create_context(i915, 0);
602 err = i915_gem_context_pin_hw_id(ctx);
604 destroy_kernel_context(&ctx);
608 i915_gem_context_clear_bannable(ctx);
609 ctx->sched.priority = I915_USER_PRIORITY(prio);
610 ctx->ring_size = PAGE_SIZE;
612 GEM_BUG_ON(!i915_gem_context_is_kernel(ctx));
617 static void init_contexts(struct drm_i915_private *i915)
619 mutex_init(&i915->contexts.mutex);
620 INIT_LIST_HEAD(&i915->contexts.list);
622 /* Using the simple ida interface, the max is limited by sizeof(int) */
623 BUILD_BUG_ON(MAX_CONTEXT_HW_ID > INT_MAX);
624 BUILD_BUG_ON(GEN11_MAX_CONTEXT_HW_ID > INT_MAX);
625 ida_init(&i915->contexts.hw_ida);
626 INIT_LIST_HEAD(&i915->contexts.hw_id_list);
628 INIT_WORK(&i915->contexts.free_work, contexts_free_worker);
629 init_llist_head(&i915->contexts.free_list);
632 int i915_gem_contexts_init(struct drm_i915_private *dev_priv)
634 struct i915_gem_context *ctx;
636 /* Reassure ourselves we are only called once */
637 GEM_BUG_ON(dev_priv->kernel_context);
639 init_contexts(dev_priv);
641 /* lowest priority; idle task */
642 ctx = i915_gem_context_create_kernel(dev_priv, I915_PRIORITY_MIN);
644 DRM_ERROR("Failed to create default global context\n");
648 * For easy recognisablity, we want the kernel context to be 0 and then
649 * all user contexts will have non-zero hw_id. Kernel contexts are
650 * permanently pinned, so that we never suffer a stall and can
651 * use them from any allocation context (e.g. for evicting other
652 * contexts and from inside the shrinker).
654 GEM_BUG_ON(ctx->hw_id);
655 GEM_BUG_ON(!atomic_read(&ctx->hw_id_pin_count));
656 dev_priv->kernel_context = ctx;
658 DRM_DEBUG_DRIVER("%s context support initialized\n",
659 DRIVER_CAPS(dev_priv)->has_logical_contexts ?
664 void i915_gem_contexts_fini(struct drm_i915_private *i915)
666 lockdep_assert_held(&i915->drm.struct_mutex);
668 destroy_kernel_context(&i915->kernel_context);
670 /* Must free all deferred contexts (via flush_workqueue) first */
671 GEM_BUG_ON(!list_empty(&i915->contexts.hw_id_list));
672 ida_destroy(&i915->contexts.hw_ida);
675 static int context_idr_cleanup(int id, void *p, void *data)
681 static int vm_idr_cleanup(int id, void *p, void *data)
687 static int gem_context_register(struct i915_gem_context *ctx,
688 struct drm_i915_file_private *fpriv)
692 ctx->file_priv = fpriv;
694 ctx->vm->file = fpriv;
696 ctx->pid = get_task_pid(current, PIDTYPE_PID);
697 ctx->name = kasprintf(GFP_KERNEL, "%s[%d]",
698 current->comm, pid_nr(ctx->pid));
704 /* And finally expose ourselves to userspace via the idr */
705 mutex_lock(&fpriv->context_idr_lock);
706 ret = idr_alloc(&fpriv->context_idr, ctx, 0, 0, GFP_KERNEL);
707 mutex_unlock(&fpriv->context_idr_lock);
711 kfree(fetch_and_zero(&ctx->name));
713 put_pid(fetch_and_zero(&ctx->pid));
718 int i915_gem_context_open(struct drm_i915_private *i915,
719 struct drm_file *file)
721 struct drm_i915_file_private *file_priv = file->driver_priv;
722 struct i915_gem_context *ctx;
725 mutex_init(&file_priv->context_idr_lock);
726 mutex_init(&file_priv->vm_idr_lock);
728 idr_init(&file_priv->context_idr);
729 idr_init_base(&file_priv->vm_idr, 1);
731 mutex_lock(&i915->drm.struct_mutex);
732 ctx = i915_gem_create_context(i915, 0);
733 mutex_unlock(&i915->drm.struct_mutex);
739 err = gem_context_register(ctx, file_priv);
743 GEM_BUG_ON(i915_gem_context_is_kernel(ctx));
751 idr_destroy(&file_priv->vm_idr);
752 idr_destroy(&file_priv->context_idr);
753 mutex_destroy(&file_priv->vm_idr_lock);
754 mutex_destroy(&file_priv->context_idr_lock);
758 void i915_gem_context_close(struct drm_file *file)
760 struct drm_i915_file_private *file_priv = file->driver_priv;
762 idr_for_each(&file_priv->context_idr, context_idr_cleanup, NULL);
763 idr_destroy(&file_priv->context_idr);
764 mutex_destroy(&file_priv->context_idr_lock);
766 idr_for_each(&file_priv->vm_idr, vm_idr_cleanup, NULL);
767 idr_destroy(&file_priv->vm_idr);
768 mutex_destroy(&file_priv->vm_idr_lock);
771 int i915_gem_vm_create_ioctl(struct drm_device *dev, void *data,
772 struct drm_file *file)
774 struct drm_i915_private *i915 = to_i915(dev);
775 struct drm_i915_gem_vm_control *args = data;
776 struct drm_i915_file_private *file_priv = file->driver_priv;
777 struct i915_ppgtt *ppgtt;
780 if (!HAS_FULL_PPGTT(i915))
786 ppgtt = i915_ppgtt_create(i915);
788 return PTR_ERR(ppgtt);
790 ppgtt->vm.file = file_priv;
792 if (args->extensions) {
793 err = i915_user_extensions(u64_to_user_ptr(args->extensions),
800 err = mutex_lock_interruptible(&file_priv->vm_idr_lock);
804 err = idr_alloc(&file_priv->vm_idr, &ppgtt->vm, 0, 0, GFP_KERNEL);
808 GEM_BUG_ON(err == 0); /* reserved for invalid/unassigned ppgtt */
810 mutex_unlock(&file_priv->vm_idr_lock);
816 mutex_unlock(&file_priv->vm_idr_lock);
818 i915_vm_put(&ppgtt->vm);
822 int i915_gem_vm_destroy_ioctl(struct drm_device *dev, void *data,
823 struct drm_file *file)
825 struct drm_i915_file_private *file_priv = file->driver_priv;
826 struct drm_i915_gem_vm_control *args = data;
827 struct i915_address_space *vm;
834 if (args->extensions)
841 err = mutex_lock_interruptible(&file_priv->vm_idr_lock);
845 vm = idr_remove(&file_priv->vm_idr, id);
847 mutex_unlock(&file_priv->vm_idr_lock);
855 struct context_barrier_task {
856 struct i915_active base;
857 void (*task)(void *data);
861 static void cb_retire(struct i915_active *base)
863 struct context_barrier_task *cb = container_of(base, typeof(*cb), base);
868 i915_active_fini(&cb->base);
872 I915_SELFTEST_DECLARE(static intel_engine_mask_t context_barrier_inject_fault);
873 static int context_barrier_task(struct i915_gem_context *ctx,
874 intel_engine_mask_t engines,
875 bool (*skip)(struct intel_context *ce, void *data),
876 int (*emit)(struct i915_request *rq, void *data),
877 void (*task)(void *data),
880 struct drm_i915_private *i915 = ctx->i915;
881 struct context_barrier_task *cb;
882 struct i915_gem_engines_iter it;
883 struct intel_context *ce;
886 lockdep_assert_held(&i915->drm.struct_mutex);
889 cb = kmalloc(sizeof(*cb), GFP_KERNEL);
893 i915_active_init(i915, &cb->base, NULL, cb_retire);
894 err = i915_active_acquire(&cb->base);
900 for_each_gem_engine(ce, i915_gem_context_lock_engines(ctx), it) {
901 struct i915_request *rq;
903 if (I915_SELFTEST_ONLY(context_barrier_inject_fault &
909 if (!(ce->engine->mask & engines))
912 if (skip && skip(ce, data))
915 rq = intel_context_create_request(ce);
923 err = emit(rq, data);
925 err = i915_active_ref(&cb->base, rq->fence.context, rq);
927 i915_request_add(rq);
931 i915_gem_context_unlock_engines(ctx);
933 cb->task = err ? NULL : task; /* caller needs to unwind instead */
936 i915_active_release(&cb->base);
941 static int get_ppgtt(struct drm_i915_file_private *file_priv,
942 struct i915_gem_context *ctx,
943 struct drm_i915_gem_context_param *args)
945 struct i915_address_space *vm;
951 /* XXX rcu acquire? */
952 ret = mutex_lock_interruptible(&ctx->i915->drm.struct_mutex);
956 vm = i915_vm_get(ctx->vm);
957 mutex_unlock(&ctx->i915->drm.struct_mutex);
959 ret = mutex_lock_interruptible(&file_priv->vm_idr_lock);
963 ret = idr_alloc(&file_priv->vm_idr, vm, 0, 0, GFP_KERNEL);
975 mutex_unlock(&file_priv->vm_idr_lock);
981 static void set_ppgtt_barrier(void *data)
983 struct i915_address_space *old = data;
985 if (INTEL_GEN(old->i915) < 8)
986 gen6_ppgtt_unpin_all(i915_vm_to_ppgtt(old));
991 static int emit_ppgtt_update(struct i915_request *rq, void *data)
993 struct i915_address_space *vm = rq->hw_context->vm;
994 struct intel_engine_cs *engine = rq->engine;
995 u32 base = engine->mmio_base;
999 if (i915_vm_is_4lvl(vm)) {
1000 struct i915_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1001 const dma_addr_t pd_daddr = px_dma(ppgtt->pd);
1003 cs = intel_ring_begin(rq, 6);
1007 *cs++ = MI_LOAD_REGISTER_IMM(2);
1009 *cs++ = i915_mmio_reg_offset(GEN8_RING_PDP_UDW(base, 0));
1010 *cs++ = upper_32_bits(pd_daddr);
1011 *cs++ = i915_mmio_reg_offset(GEN8_RING_PDP_LDW(base, 0));
1012 *cs++ = lower_32_bits(pd_daddr);
1015 intel_ring_advance(rq, cs);
1016 } else if (HAS_LOGICAL_RING_CONTEXTS(engine->i915)) {
1017 struct i915_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1019 cs = intel_ring_begin(rq, 4 * GEN8_3LVL_PDPES + 2);
1023 *cs++ = MI_LOAD_REGISTER_IMM(2 * GEN8_3LVL_PDPES);
1024 for (i = GEN8_3LVL_PDPES; i--; ) {
1025 const dma_addr_t pd_daddr = i915_page_dir_dma_addr(ppgtt, i);
1027 *cs++ = i915_mmio_reg_offset(GEN8_RING_PDP_UDW(base, i));
1028 *cs++ = upper_32_bits(pd_daddr);
1029 *cs++ = i915_mmio_reg_offset(GEN8_RING_PDP_LDW(base, i));
1030 *cs++ = lower_32_bits(pd_daddr);
1033 intel_ring_advance(rq, cs);
1035 /* ppGTT is not part of the legacy context image */
1036 gen6_ppgtt_pin(i915_vm_to_ppgtt(vm));
1042 static bool skip_ppgtt_update(struct intel_context *ce, void *data)
1044 if (HAS_LOGICAL_RING_CONTEXTS(ce->engine->i915))
1047 return !atomic_read(&ce->pin_count);
1050 static int set_ppgtt(struct drm_i915_file_private *file_priv,
1051 struct i915_gem_context *ctx,
1052 struct drm_i915_gem_context_param *args)
1054 struct i915_address_space *vm, *old;
1063 if (upper_32_bits(args->value))
1066 err = mutex_lock_interruptible(&file_priv->vm_idr_lock);
1070 vm = idr_find(&file_priv->vm_idr, args->value);
1073 mutex_unlock(&file_priv->vm_idr_lock);
1077 err = mutex_lock_interruptible(&ctx->i915->drm.struct_mutex);
1084 /* Teardown the existing obj:vma cache, it will have to be rebuilt. */
1085 mutex_lock(&ctx->mutex);
1087 mutex_unlock(&ctx->mutex);
1089 old = __set_ppgtt(ctx, vm);
1092 * We need to flush any requests using the current ppgtt before
1093 * we release it as the requests do not hold a reference themselves,
1094 * only indirectly through the context.
1096 err = context_barrier_task(ctx, ALL_ENGINES,
1102 i915_vm_put(__set_ppgtt(ctx, old));
1107 mutex_unlock(&ctx->i915->drm.struct_mutex);
1114 static int gen8_emit_rpcs_config(struct i915_request *rq,
1115 struct intel_context *ce,
1116 struct intel_sseu sseu)
1121 cs = intel_ring_begin(rq, 4);
1125 offset = i915_ggtt_offset(ce->state) +
1126 LRC_STATE_PN * PAGE_SIZE +
1127 (CTX_R_PWR_CLK_STATE + 1) * 4;
1129 *cs++ = MI_STORE_DWORD_IMM_GEN4 | MI_USE_GGTT;
1130 *cs++ = lower_32_bits(offset);
1131 *cs++ = upper_32_bits(offset);
1132 *cs++ = intel_sseu_make_rpcs(rq->i915, &sseu);
1134 intel_ring_advance(rq, cs);
1140 gen8_modify_rpcs(struct intel_context *ce, struct intel_sseu sseu)
1142 struct i915_request *rq;
1145 lockdep_assert_held(&ce->pin_mutex);
1148 * If the context is not idle, we have to submit an ordered request to
1149 * modify its context image via the kernel context (writing to our own
1150 * image, or into the registers directory, does not stick). Pristine
1151 * and idle contexts will be configured on pinning.
1153 if (!intel_context_is_pinned(ce))
1156 rq = i915_request_create(ce->engine->kernel_context);
1160 /* Serialise with the remote context */
1161 ret = intel_context_prepare_remote_request(ce, rq);
1163 ret = gen8_emit_rpcs_config(rq, ce, sseu);
1165 i915_request_add(rq);
1170 __intel_context_reconfigure_sseu(struct intel_context *ce,
1171 struct intel_sseu sseu)
1175 GEM_BUG_ON(INTEL_GEN(ce->engine->i915) < 8);
1177 ret = intel_context_lock_pinned(ce);
1181 /* Nothing to do if unmodified. */
1182 if (!memcmp(&ce->sseu, &sseu, sizeof(sseu)))
1185 ret = gen8_modify_rpcs(ce, sseu);
1190 intel_context_unlock_pinned(ce);
1195 intel_context_reconfigure_sseu(struct intel_context *ce, struct intel_sseu sseu)
1197 struct drm_i915_private *i915 = ce->engine->i915;
1200 ret = mutex_lock_interruptible(&i915->drm.struct_mutex);
1204 ret = __intel_context_reconfigure_sseu(ce, sseu);
1206 mutex_unlock(&i915->drm.struct_mutex);
1212 user_to_context_sseu(struct drm_i915_private *i915,
1213 const struct drm_i915_gem_context_param_sseu *user,
1214 struct intel_sseu *context)
1216 const struct sseu_dev_info *device = &RUNTIME_INFO(i915)->sseu;
1218 /* No zeros in any field. */
1219 if (!user->slice_mask || !user->subslice_mask ||
1220 !user->min_eus_per_subslice || !user->max_eus_per_subslice)
1224 if (user->max_eus_per_subslice < user->min_eus_per_subslice)
1228 * Some future proofing on the types since the uAPI is wider than the
1229 * current internal implementation.
1231 if (overflows_type(user->slice_mask, context->slice_mask) ||
1232 overflows_type(user->subslice_mask, context->subslice_mask) ||
1233 overflows_type(user->min_eus_per_subslice,
1234 context->min_eus_per_subslice) ||
1235 overflows_type(user->max_eus_per_subslice,
1236 context->max_eus_per_subslice))
1239 /* Check validity against hardware. */
1240 if (user->slice_mask & ~device->slice_mask)
1243 if (user->subslice_mask & ~device->subslice_mask[0])
1246 if (user->max_eus_per_subslice > device->max_eus_per_subslice)
1249 context->slice_mask = user->slice_mask;
1250 context->subslice_mask = user->subslice_mask;
1251 context->min_eus_per_subslice = user->min_eus_per_subslice;
1252 context->max_eus_per_subslice = user->max_eus_per_subslice;
1254 /* Part specific restrictions. */
1255 if (IS_GEN(i915, 11)) {
1256 unsigned int hw_s = hweight8(device->slice_mask);
1257 unsigned int hw_ss_per_s = hweight8(device->subslice_mask[0]);
1258 unsigned int req_s = hweight8(context->slice_mask);
1259 unsigned int req_ss = hweight8(context->subslice_mask);
1262 * Only full subslice enablement is possible if more than one
1263 * slice is turned on.
1265 if (req_s > 1 && req_ss != hw_ss_per_s)
1269 * If more than four (SScount bitfield limit) subslices are
1270 * requested then the number has to be even.
1272 if (req_ss > 4 && (req_ss & 1))
1276 * If only one slice is enabled and subslice count is below the
1277 * device full enablement, it must be at most half of the all
1278 * available subslices.
1280 if (req_s == 1 && req_ss < hw_ss_per_s &&
1281 req_ss > (hw_ss_per_s / 2))
1284 /* ABI restriction - VME use case only. */
1286 /* All slices or one slice only. */
1287 if (req_s != 1 && req_s != hw_s)
1291 * Half subslices or full enablement only when one slice is
1295 (req_ss != hw_ss_per_s && req_ss != (hw_ss_per_s / 2)))
1298 /* No EU configuration changes. */
1299 if ((user->min_eus_per_subslice !=
1300 device->max_eus_per_subslice) ||
1301 (user->max_eus_per_subslice !=
1302 device->max_eus_per_subslice))
1309 static int set_sseu(struct i915_gem_context *ctx,
1310 struct drm_i915_gem_context_param *args)
1312 struct drm_i915_private *i915 = ctx->i915;
1313 struct drm_i915_gem_context_param_sseu user_sseu;
1314 struct intel_context *ce;
1315 struct intel_sseu sseu;
1316 unsigned long lookup;
1319 if (args->size < sizeof(user_sseu))
1322 if (!IS_GEN(i915, 11))
1325 if (copy_from_user(&user_sseu, u64_to_user_ptr(args->value),
1332 if (user_sseu.flags & ~(I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX))
1336 if (user_sseu.flags & I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX)
1337 lookup |= LOOKUP_USER_INDEX;
1339 ce = lookup_user_engine(ctx, lookup, &user_sseu.engine);
1343 /* Only render engine supports RPCS configuration. */
1344 if (ce->engine->class != RENDER_CLASS) {
1349 ret = user_to_context_sseu(i915, &user_sseu, &sseu);
1353 ret = intel_context_reconfigure_sseu(ce, sseu);
1357 args->size = sizeof(user_sseu);
1360 intel_context_put(ce);
1364 struct set_engines {
1365 struct i915_gem_context *ctx;
1366 struct i915_gem_engines *engines;
1370 set_engines__load_balance(struct i915_user_extension __user *base, void *data)
1372 struct i915_context_engines_load_balance __user *ext =
1373 container_of_user(base, typeof(*ext), base);
1374 const struct set_engines *set = data;
1375 struct intel_engine_cs *stack[16];
1376 struct intel_engine_cs **siblings;
1377 struct intel_context *ce;
1378 u16 num_siblings, idx;
1382 if (!HAS_EXECLISTS(set->ctx->i915))
1385 if (USES_GUC_SUBMISSION(set->ctx->i915))
1386 return -ENODEV; /* not implement yet */
1388 if (get_user(idx, &ext->engine_index))
1391 if (idx >= set->engines->num_engines) {
1392 DRM_DEBUG("Invalid placement value, %d >= %d\n",
1393 idx, set->engines->num_engines);
1397 idx = array_index_nospec(idx, set->engines->num_engines);
1398 if (set->engines->engines[idx]) {
1399 DRM_DEBUG("Invalid placement[%d], already occupied\n", idx);
1403 if (get_user(num_siblings, &ext->num_siblings))
1406 err = check_user_mbz(&ext->flags);
1410 err = check_user_mbz(&ext->mbz64);
1415 if (num_siblings > ARRAY_SIZE(stack)) {
1416 siblings = kmalloc_array(num_siblings,
1423 for (n = 0; n < num_siblings; n++) {
1424 struct i915_engine_class_instance ci;
1426 if (copy_from_user(&ci, &ext->engines[n], sizeof(ci))) {
1431 siblings[n] = intel_engine_lookup_user(set->ctx->i915,
1433 ci.engine_instance);
1435 DRM_DEBUG("Invalid sibling[%d]: { class:%d, inst:%d }\n",
1436 n, ci.engine_class, ci.engine_instance);
1442 ce = intel_execlists_create_virtual(set->ctx, siblings, n);
1448 if (cmpxchg(&set->engines->engines[idx], NULL, ce)) {
1449 intel_context_put(ce);
1455 if (siblings != stack)
1462 set_engines__bond(struct i915_user_extension __user *base, void *data)
1464 struct i915_context_engines_bond __user *ext =
1465 container_of_user(base, typeof(*ext), base);
1466 const struct set_engines *set = data;
1467 struct i915_engine_class_instance ci;
1468 struct intel_engine_cs *virtual;
1469 struct intel_engine_cs *master;
1473 if (get_user(idx, &ext->virtual_index))
1476 if (idx >= set->engines->num_engines) {
1477 DRM_DEBUG("Invalid index for virtual engine: %d >= %d\n",
1478 idx, set->engines->num_engines);
1482 idx = array_index_nospec(idx, set->engines->num_engines);
1483 if (!set->engines->engines[idx]) {
1484 DRM_DEBUG("Invalid engine at %d\n", idx);
1487 virtual = set->engines->engines[idx]->engine;
1489 err = check_user_mbz(&ext->flags);
1493 for (n = 0; n < ARRAY_SIZE(ext->mbz64); n++) {
1494 err = check_user_mbz(&ext->mbz64[n]);
1499 if (copy_from_user(&ci, &ext->master, sizeof(ci)))
1502 master = intel_engine_lookup_user(set->ctx->i915,
1503 ci.engine_class, ci.engine_instance);
1505 DRM_DEBUG("Unrecognised master engine: { class:%u, instance:%u }\n",
1506 ci.engine_class, ci.engine_instance);
1510 if (get_user(num_bonds, &ext->num_bonds))
1513 for (n = 0; n < num_bonds; n++) {
1514 struct intel_engine_cs *bond;
1516 if (copy_from_user(&ci, &ext->engines[n], sizeof(ci)))
1519 bond = intel_engine_lookup_user(set->ctx->i915,
1521 ci.engine_instance);
1523 DRM_DEBUG("Unrecognised engine[%d] for bonding: { class:%d, instance: %d }\n",
1524 n, ci.engine_class, ci.engine_instance);
1529 * A non-virtual engine has no siblings to choose between; and
1530 * a submit fence will always be directed to the one engine.
1532 if (intel_engine_is_virtual(virtual)) {
1533 err = intel_virtual_engine_attach_bond(virtual,
1544 static const i915_user_extension_fn set_engines__extensions[] = {
1545 [I915_CONTEXT_ENGINES_EXT_LOAD_BALANCE] = set_engines__load_balance,
1546 [I915_CONTEXT_ENGINES_EXT_BOND] = set_engines__bond,
1550 set_engines(struct i915_gem_context *ctx,
1551 const struct drm_i915_gem_context_param *args)
1553 struct i915_context_param_engines __user *user =
1554 u64_to_user_ptr(args->value);
1555 struct set_engines set = { .ctx = ctx };
1556 unsigned int num_engines, n;
1560 if (!args->size) { /* switch back to legacy user_ring_map */
1561 if (!i915_gem_context_user_engines(ctx))
1564 set.engines = default_engines(ctx);
1565 if (IS_ERR(set.engines))
1566 return PTR_ERR(set.engines);
1571 BUILD_BUG_ON(!IS_ALIGNED(sizeof(*user), sizeof(*user->engines)));
1572 if (args->size < sizeof(*user) ||
1573 !IS_ALIGNED(args->size, sizeof(*user->engines))) {
1574 DRM_DEBUG("Invalid size for engine array: %d\n",
1580 * Note that I915_EXEC_RING_MASK limits execbuf to only using the
1581 * first 64 engines defined here.
1583 num_engines = (args->size - sizeof(*user)) / sizeof(*user->engines);
1585 set.engines = kmalloc(struct_size(set.engines, engines, num_engines),
1590 init_rcu_head(&set.engines->rcu);
1591 for (n = 0; n < num_engines; n++) {
1592 struct i915_engine_class_instance ci;
1593 struct intel_engine_cs *engine;
1595 if (copy_from_user(&ci, &user->engines[n], sizeof(ci))) {
1596 __free_engines(set.engines, n);
1600 if (ci.engine_class == (u16)I915_ENGINE_CLASS_INVALID &&
1601 ci.engine_instance == (u16)I915_ENGINE_CLASS_INVALID_NONE) {
1602 set.engines->engines[n] = NULL;
1606 engine = intel_engine_lookup_user(ctx->i915,
1608 ci.engine_instance);
1610 DRM_DEBUG("Invalid engine[%d]: { class:%d, instance:%d }\n",
1611 n, ci.engine_class, ci.engine_instance);
1612 __free_engines(set.engines, n);
1616 set.engines->engines[n] = intel_context_create(ctx, engine);
1617 if (!set.engines->engines[n]) {
1618 __free_engines(set.engines, n);
1622 set.engines->num_engines = num_engines;
1625 if (!get_user(extensions, &user->extensions))
1626 err = i915_user_extensions(u64_to_user_ptr(extensions),
1627 set_engines__extensions,
1628 ARRAY_SIZE(set_engines__extensions),
1631 free_engines(set.engines);
1636 mutex_lock(&ctx->engines_mutex);
1638 i915_gem_context_set_user_engines(ctx);
1640 i915_gem_context_clear_user_engines(ctx);
1641 rcu_swap_protected(ctx->engines, set.engines, 1);
1642 mutex_unlock(&ctx->engines_mutex);
1644 call_rcu(&set.engines->rcu, free_engines_rcu);
1649 static struct i915_gem_engines *
1650 __copy_engines(struct i915_gem_engines *e)
1652 struct i915_gem_engines *copy;
1655 copy = kmalloc(struct_size(e, engines, e->num_engines), GFP_KERNEL);
1657 return ERR_PTR(-ENOMEM);
1659 init_rcu_head(©->rcu);
1660 for (n = 0; n < e->num_engines; n++) {
1662 copy->engines[n] = intel_context_get(e->engines[n]);
1664 copy->engines[n] = NULL;
1666 copy->num_engines = n;
1672 get_engines(struct i915_gem_context *ctx,
1673 struct drm_i915_gem_context_param *args)
1675 struct i915_context_param_engines __user *user;
1676 struct i915_gem_engines *e;
1677 size_t n, count, size;
1680 err = mutex_lock_interruptible(&ctx->engines_mutex);
1685 if (i915_gem_context_user_engines(ctx))
1686 e = __copy_engines(i915_gem_context_engines(ctx));
1687 mutex_unlock(&ctx->engines_mutex);
1688 if (IS_ERR_OR_NULL(e)) {
1690 return PTR_ERR_OR_ZERO(e);
1693 count = e->num_engines;
1695 /* Be paranoid in case we have an impedance mismatch */
1696 if (!check_struct_size(user, engines, count, &size)) {
1700 if (overflows_type(size, args->size)) {
1710 if (args->size < size) {
1715 user = u64_to_user_ptr(args->value);
1716 if (!access_ok(user, size)) {
1721 if (put_user(0, &user->extensions)) {
1726 for (n = 0; n < count; n++) {
1727 struct i915_engine_class_instance ci = {
1728 .engine_class = I915_ENGINE_CLASS_INVALID,
1729 .engine_instance = I915_ENGINE_CLASS_INVALID_NONE,
1732 if (e->engines[n]) {
1733 ci.engine_class = e->engines[n]->engine->uabi_class;
1734 ci.engine_instance = e->engines[n]->engine->uabi_instance;
1737 if (copy_to_user(&user->engines[n], &ci, sizeof(ci))) {
1750 static int ctx_setparam(struct drm_i915_file_private *fpriv,
1751 struct i915_gem_context *ctx,
1752 struct drm_i915_gem_context_param *args)
1756 switch (args->param) {
1757 case I915_CONTEXT_PARAM_NO_ZEROMAP:
1760 else if (args->value)
1761 set_bit(UCONTEXT_NO_ZEROMAP, &ctx->user_flags);
1763 clear_bit(UCONTEXT_NO_ZEROMAP, &ctx->user_flags);
1766 case I915_CONTEXT_PARAM_NO_ERROR_CAPTURE:
1769 else if (args->value)
1770 i915_gem_context_set_no_error_capture(ctx);
1772 i915_gem_context_clear_no_error_capture(ctx);
1775 case I915_CONTEXT_PARAM_BANNABLE:
1778 else if (!capable(CAP_SYS_ADMIN) && !args->value)
1780 else if (args->value)
1781 i915_gem_context_set_bannable(ctx);
1783 i915_gem_context_clear_bannable(ctx);
1786 case I915_CONTEXT_PARAM_RECOVERABLE:
1789 else if (args->value)
1790 i915_gem_context_set_recoverable(ctx);
1792 i915_gem_context_clear_recoverable(ctx);
1795 case I915_CONTEXT_PARAM_PRIORITY:
1797 s64 priority = args->value;
1801 else if (!(ctx->i915->caps.scheduler & I915_SCHEDULER_CAP_PRIORITY))
1803 else if (priority > I915_CONTEXT_MAX_USER_PRIORITY ||
1804 priority < I915_CONTEXT_MIN_USER_PRIORITY)
1806 else if (priority > I915_CONTEXT_DEFAULT_PRIORITY &&
1807 !capable(CAP_SYS_NICE))
1810 ctx->sched.priority =
1811 I915_USER_PRIORITY(priority);
1815 case I915_CONTEXT_PARAM_SSEU:
1816 ret = set_sseu(ctx, args);
1819 case I915_CONTEXT_PARAM_VM:
1820 ret = set_ppgtt(fpriv, ctx, args);
1823 case I915_CONTEXT_PARAM_ENGINES:
1824 ret = set_engines(ctx, args);
1827 case I915_CONTEXT_PARAM_BAN_PERIOD:
1837 struct i915_gem_context *ctx;
1838 struct drm_i915_file_private *fpriv;
1841 static int create_setparam(struct i915_user_extension __user *ext, void *data)
1843 struct drm_i915_gem_context_create_ext_setparam local;
1844 const struct create_ext *arg = data;
1846 if (copy_from_user(&local, ext, sizeof(local)))
1849 if (local.param.ctx_id)
1852 return ctx_setparam(arg->fpriv, arg->ctx, &local.param);
1855 static int clone_engines(struct i915_gem_context *dst,
1856 struct i915_gem_context *src)
1858 struct i915_gem_engines *e = i915_gem_context_lock_engines(src);
1859 struct i915_gem_engines *clone;
1863 clone = kmalloc(struct_size(e, engines, e->num_engines), GFP_KERNEL);
1867 init_rcu_head(&clone->rcu);
1868 for (n = 0; n < e->num_engines; n++) {
1869 struct intel_engine_cs *engine;
1871 if (!e->engines[n]) {
1872 clone->engines[n] = NULL;
1875 engine = e->engines[n]->engine;
1878 * Virtual engines are singletons; they can only exist
1879 * inside a single context, because they embed their
1880 * HW context... As each virtual context implies a single
1881 * timeline (each engine can only dequeue a single request
1882 * at any time), it would be surprising for two contexts
1883 * to use the same engine. So let's create a copy of
1884 * the virtual engine instead.
1886 if (intel_engine_is_virtual(engine))
1888 intel_execlists_clone_virtual(dst, engine);
1890 clone->engines[n] = intel_context_create(dst, engine);
1891 if (IS_ERR_OR_NULL(clone->engines[n])) {
1892 __free_engines(clone, n);
1896 clone->num_engines = n;
1898 user_engines = i915_gem_context_user_engines(src);
1899 i915_gem_context_unlock_engines(src);
1901 free_engines(dst->engines);
1902 RCU_INIT_POINTER(dst->engines, clone);
1904 i915_gem_context_set_user_engines(dst);
1906 i915_gem_context_clear_user_engines(dst);
1910 i915_gem_context_unlock_engines(src);
1914 static int clone_flags(struct i915_gem_context *dst,
1915 struct i915_gem_context *src)
1917 dst->user_flags = src->user_flags;
1921 static int clone_schedattr(struct i915_gem_context *dst,
1922 struct i915_gem_context *src)
1924 dst->sched = src->sched;
1928 static int clone_sseu(struct i915_gem_context *dst,
1929 struct i915_gem_context *src)
1931 struct i915_gem_engines *e = i915_gem_context_lock_engines(src);
1932 struct i915_gem_engines *clone;
1936 clone = dst->engines; /* no locking required; sole access */
1937 if (e->num_engines != clone->num_engines) {
1942 for (n = 0; n < e->num_engines; n++) {
1943 struct intel_context *ce = e->engines[n];
1945 if (clone->engines[n]->engine->class != ce->engine->class) {
1946 /* Must have compatible engine maps! */
1951 /* serialises with set_sseu */
1952 err = intel_context_lock_pinned(ce);
1956 clone->engines[n]->sseu = ce->sseu;
1957 intel_context_unlock_pinned(ce);
1962 i915_gem_context_unlock_engines(src);
1966 static int clone_timeline(struct i915_gem_context *dst,
1967 struct i915_gem_context *src)
1969 if (src->timeline) {
1970 GEM_BUG_ON(src->timeline == dst->timeline);
1973 intel_timeline_put(dst->timeline);
1974 dst->timeline = intel_timeline_get(src->timeline);
1980 static int clone_vm(struct i915_gem_context *dst,
1981 struct i915_gem_context *src)
1983 struct i915_address_space *vm;
1987 vm = READ_ONCE(src->vm);
1991 if (!kref_get_unless_zero(&vm->ref))
1995 * This ppgtt may have be reallocated between
1996 * the read and the kref, and reassigned to a third
1997 * context. In order to avoid inadvertent sharing
1998 * of this ppgtt with that third context (and not
1999 * src), we have to confirm that we have the same
2000 * ppgtt after passing through the strong memory
2001 * barrier implied by a successful
2002 * kref_get_unless_zero().
2004 * Once we have acquired the current ppgtt of src,
2005 * we no longer care if it is released from src, as
2006 * it cannot be reallocated elsewhere.
2009 if (vm == READ_ONCE(src->vm))
2017 __assign_ppgtt(dst, vm);
2024 static int create_clone(struct i915_user_extension __user *ext, void *data)
2026 static int (* const fn[])(struct i915_gem_context *dst,
2027 struct i915_gem_context *src) = {
2028 #define MAP(x, y) [ilog2(I915_CONTEXT_CLONE_##x)] = y
2029 MAP(ENGINES, clone_engines),
2030 MAP(FLAGS, clone_flags),
2031 MAP(SCHEDATTR, clone_schedattr),
2032 MAP(SSEU, clone_sseu),
2033 MAP(TIMELINE, clone_timeline),
2037 struct drm_i915_gem_context_create_ext_clone local;
2038 const struct create_ext *arg = data;
2039 struct i915_gem_context *dst = arg->ctx;
2040 struct i915_gem_context *src;
2043 if (copy_from_user(&local, ext, sizeof(local)))
2046 BUILD_BUG_ON(GENMASK(BITS_PER_TYPE(local.flags) - 1, ARRAY_SIZE(fn)) !=
2047 I915_CONTEXT_CLONE_UNKNOWN);
2049 if (local.flags & I915_CONTEXT_CLONE_UNKNOWN)
2056 src = __i915_gem_context_lookup_rcu(arg->fpriv, local.clone_id);
2061 GEM_BUG_ON(src == dst);
2063 for (bit = 0; bit < ARRAY_SIZE(fn); bit++) {
2064 if (!(local.flags & BIT(bit)))
2067 err = fn[bit](dst, src);
2075 static const i915_user_extension_fn create_extensions[] = {
2076 [I915_CONTEXT_CREATE_EXT_SETPARAM] = create_setparam,
2077 [I915_CONTEXT_CREATE_EXT_CLONE] = create_clone,
2080 static bool client_is_banned(struct drm_i915_file_private *file_priv)
2082 return atomic_read(&file_priv->ban_score) >= I915_CLIENT_SCORE_BANNED;
2085 int i915_gem_context_create_ioctl(struct drm_device *dev, void *data,
2086 struct drm_file *file)
2088 struct drm_i915_private *i915 = to_i915(dev);
2089 struct drm_i915_gem_context_create_ext *args = data;
2090 struct create_ext ext_data;
2093 if (!DRIVER_CAPS(i915)->has_logical_contexts)
2096 if (args->flags & I915_CONTEXT_CREATE_FLAGS_UNKNOWN)
2099 ret = intel_gt_terminally_wedged(&i915->gt);
2103 ext_data.fpriv = file->driver_priv;
2104 if (client_is_banned(ext_data.fpriv)) {
2105 DRM_DEBUG("client %s[%d] banned from creating ctx\n",
2107 pid_nr(get_task_pid(current, PIDTYPE_PID)));
2111 ret = i915_mutex_lock_interruptible(dev);
2115 ext_data.ctx = i915_gem_create_context(i915, args->flags);
2116 mutex_unlock(&dev->struct_mutex);
2117 if (IS_ERR(ext_data.ctx))
2118 return PTR_ERR(ext_data.ctx);
2120 if (args->flags & I915_CONTEXT_CREATE_FLAGS_USE_EXTENSIONS) {
2121 ret = i915_user_extensions(u64_to_user_ptr(args->extensions),
2123 ARRAY_SIZE(create_extensions),
2129 ret = gem_context_register(ext_data.ctx, ext_data.fpriv);
2134 DRM_DEBUG("HW context %d created\n", args->ctx_id);
2139 context_close(ext_data.ctx);
2143 int i915_gem_context_destroy_ioctl(struct drm_device *dev, void *data,
2144 struct drm_file *file)
2146 struct drm_i915_gem_context_destroy *args = data;
2147 struct drm_i915_file_private *file_priv = file->driver_priv;
2148 struct i915_gem_context *ctx;
2156 if (mutex_lock_interruptible(&file_priv->context_idr_lock))
2159 ctx = idr_remove(&file_priv->context_idr, args->ctx_id);
2160 mutex_unlock(&file_priv->context_idr_lock);
2168 static int get_sseu(struct i915_gem_context *ctx,
2169 struct drm_i915_gem_context_param *args)
2171 struct drm_i915_gem_context_param_sseu user_sseu;
2172 struct intel_context *ce;
2173 unsigned long lookup;
2176 if (args->size == 0)
2178 else if (args->size < sizeof(user_sseu))
2181 if (copy_from_user(&user_sseu, u64_to_user_ptr(args->value),
2188 if (user_sseu.flags & ~(I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX))
2192 if (user_sseu.flags & I915_CONTEXT_SSEU_FLAG_ENGINE_INDEX)
2193 lookup |= LOOKUP_USER_INDEX;
2195 ce = lookup_user_engine(ctx, lookup, &user_sseu.engine);
2199 err = intel_context_lock_pinned(ce); /* serialises with set_sseu */
2201 intel_context_put(ce);
2205 user_sseu.slice_mask = ce->sseu.slice_mask;
2206 user_sseu.subslice_mask = ce->sseu.subslice_mask;
2207 user_sseu.min_eus_per_subslice = ce->sseu.min_eus_per_subslice;
2208 user_sseu.max_eus_per_subslice = ce->sseu.max_eus_per_subslice;
2210 intel_context_unlock_pinned(ce);
2211 intel_context_put(ce);
2213 if (copy_to_user(u64_to_user_ptr(args->value), &user_sseu,
2218 args->size = sizeof(user_sseu);
2223 int i915_gem_context_getparam_ioctl(struct drm_device *dev, void *data,
2224 struct drm_file *file)
2226 struct drm_i915_file_private *file_priv = file->driver_priv;
2227 struct drm_i915_gem_context_param *args = data;
2228 struct i915_gem_context *ctx;
2231 ctx = i915_gem_context_lookup(file_priv, args->ctx_id);
2235 switch (args->param) {
2236 case I915_CONTEXT_PARAM_NO_ZEROMAP:
2238 args->value = test_bit(UCONTEXT_NO_ZEROMAP, &ctx->user_flags);
2241 case I915_CONTEXT_PARAM_GTT_SIZE:
2244 args->value = ctx->vm->total;
2245 else if (to_i915(dev)->ggtt.alias)
2246 args->value = to_i915(dev)->ggtt.alias->vm.total;
2248 args->value = to_i915(dev)->ggtt.vm.total;
2251 case I915_CONTEXT_PARAM_NO_ERROR_CAPTURE:
2253 args->value = i915_gem_context_no_error_capture(ctx);
2256 case I915_CONTEXT_PARAM_BANNABLE:
2258 args->value = i915_gem_context_is_bannable(ctx);
2261 case I915_CONTEXT_PARAM_RECOVERABLE:
2263 args->value = i915_gem_context_is_recoverable(ctx);
2266 case I915_CONTEXT_PARAM_PRIORITY:
2268 args->value = ctx->sched.priority >> I915_USER_PRIORITY_SHIFT;
2271 case I915_CONTEXT_PARAM_SSEU:
2272 ret = get_sseu(ctx, args);
2275 case I915_CONTEXT_PARAM_VM:
2276 ret = get_ppgtt(file_priv, ctx, args);
2279 case I915_CONTEXT_PARAM_ENGINES:
2280 ret = get_engines(ctx, args);
2283 case I915_CONTEXT_PARAM_BAN_PERIOD:
2289 i915_gem_context_put(ctx);
2293 int i915_gem_context_setparam_ioctl(struct drm_device *dev, void *data,
2294 struct drm_file *file)
2296 struct drm_i915_file_private *file_priv = file->driver_priv;
2297 struct drm_i915_gem_context_param *args = data;
2298 struct i915_gem_context *ctx;
2301 ctx = i915_gem_context_lookup(file_priv, args->ctx_id);
2305 ret = ctx_setparam(file_priv, ctx, args);
2307 i915_gem_context_put(ctx);
2311 int i915_gem_context_reset_stats_ioctl(struct drm_device *dev,
2312 void *data, struct drm_file *file)
2314 struct drm_i915_private *dev_priv = to_i915(dev);
2315 struct drm_i915_reset_stats *args = data;
2316 struct i915_gem_context *ctx;
2319 if (args->flags || args->pad)
2324 ctx = __i915_gem_context_lookup_rcu(file->driver_priv, args->ctx_id);
2329 * We opt for unserialised reads here. This may result in tearing
2330 * in the extremely unlikely event of a GPU hang on this context
2331 * as we are querying them. If we need that extra layer of protection,
2332 * we should wrap the hangstats with a seqlock.
2335 if (capable(CAP_SYS_ADMIN))
2336 args->reset_count = i915_reset_count(&dev_priv->gpu_error);
2338 args->reset_count = 0;
2340 args->batch_active = atomic_read(&ctx->guilty_count);
2341 args->batch_pending = atomic_read(&ctx->active_count);
2349 int __i915_gem_context_pin_hw_id(struct i915_gem_context *ctx)
2351 struct drm_i915_private *i915 = ctx->i915;
2354 mutex_lock(&i915->contexts.mutex);
2356 GEM_BUG_ON(i915_gem_context_is_closed(ctx));
2358 if (list_empty(&ctx->hw_id_link)) {
2359 GEM_BUG_ON(atomic_read(&ctx->hw_id_pin_count));
2361 err = assign_hw_id(i915, &ctx->hw_id);
2365 list_add_tail(&ctx->hw_id_link, &i915->contexts.hw_id_list);
2368 GEM_BUG_ON(atomic_read(&ctx->hw_id_pin_count) == ~0u);
2369 atomic_inc(&ctx->hw_id_pin_count);
2372 mutex_unlock(&i915->contexts.mutex);
2376 /* GEM context-engines iterator: for_each_gem_engine() */
2377 struct intel_context *
2378 i915_gem_engines_iter_next(struct i915_gem_engines_iter *it)
2380 const struct i915_gem_engines *e = it->engines;
2381 struct intel_context *ctx;
2384 if (it->idx >= e->num_engines)
2387 ctx = e->engines[it->idx++];
2393 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
2394 #include "selftests/mock_context.c"
2395 #include "selftests/i915_gem_context.c"
2398 static void i915_global_gem_context_shrink(void)
2400 kmem_cache_shrink(global.slab_luts);
2403 static void i915_global_gem_context_exit(void)
2405 kmem_cache_destroy(global.slab_luts);
2408 static struct i915_global_gem_context global = { {
2409 .shrink = i915_global_gem_context_shrink,
2410 .exit = i915_global_gem_context_exit,
2413 int __init i915_global_gem_context_init(void)
2415 global.slab_luts = KMEM_CACHE(i915_lut_handle, 0);
2416 if (!global.slab_luts)
2419 i915_global_register(&global.base);