GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE));
+ spin_lock(&obj->vma.lock);
+
if (i915_is_ggtt(vm)) {
if (unlikely(overflows_type(vma->size, u32)))
- goto err_vma;
+ goto err_unlock;
vma->fence_size = i915_gem_fence_size(vm->i915, vma->size,
i915_gem_object_get_tiling(obj),
i915_gem_object_get_stride(obj));
if (unlikely(vma->fence_size < vma->size || /* overflow */
vma->fence_size > vm->total))
- goto err_vma;
+ goto err_unlock;
GEM_BUG_ON(!IS_ALIGNED(vma->fence_size, I915_GTT_MIN_ALIGNMENT));
__set_bit(I915_VMA_GGTT_BIT, __i915_vma_flags(vma));
}
- spin_lock(&obj->vma.lock);
-
rb = NULL;
p = &obj->vma.tree.rb_node;
while (*p) {
return vma;
+err_unlock:
+ spin_unlock(&obj->vma.lock);
err_vma:
i915_vma_free(vma);
return ERR_PTR(-E2BIG);
struct dma_fence_work base;
struct i915_vma *vma;
struct drm_i915_gem_object *pinned;
+ struct i915_sw_dma_fence_cb cb;
enum i915_cache_level cache_level;
unsigned int flags;
};
return vw;
}
+int i915_vma_wait_for_bind(struct i915_vma *vma)
+{
+ int err = 0;
+
+ if (rcu_access_pointer(vma->active.excl.fence)) {
+ struct dma_fence *fence;
+
+ rcu_read_lock();
+ fence = dma_fence_get_rcu_safe(&vma->active.excl.fence);
+ rcu_read_unlock();
+ if (fence) {
+ err = dma_fence_wait(fence, MAX_SCHEDULE_TIMEOUT);
+ dma_fence_put(fence);
+ }
+ }
+
+ return err;
+}
+
/**
* i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
* @vma: VMA to map
trace_i915_vma_bind(vma, bind_flags);
if (work && (bind_flags & ~vma_flags) & vma->vm->bind_async_flags) {
+ struct dma_fence *prev;
+
work->vma = vma;
work->cache_level = cache_level;
work->flags = bind_flags | I915_VMA_ALLOC;
* part of the obj->resv->excl_fence as it only affects
* execution and not content or object's backing store lifetime.
*/
- GEM_BUG_ON(i915_active_has_exclusive(&vma->active));
- i915_active_set_exclusive(&vma->active, &work->base.dma);
+ prev = i915_active_set_exclusive(&vma->active, &work->base.dma);
+ if (prev) {
+ __i915_sw_fence_await_dma_fence(&work->base.chain,
+ prev,
+ &work->cb);
+ dma_fence_put(prev);
+ }
+
work->base.dma.error = 0; /* enable the queue_work() */
if (vma->obj) {
work->pinned = vma->obj;
}
} else {
- GEM_BUG_ON((bind_flags & ~vma_flags) & vma->vm->bind_async_flags);
ret = vma->ops->bind_vma(vma, cache_level, bind_flags);
if (ret)
return ret;
return true;
}
-static void assert_bind_count(const struct drm_i915_gem_object *obj)
-{
- /*
- * Combine the assertion that the object is bound and that we have
- * pinned its pages. But we should never have bound the object
- * more than we have pinned its pages. (For complete accuracy, we
- * assume that no else is pinning the pages, but as a rough assertion
- * that we will not run into problems later, this will do!)
- */
- GEM_BUG_ON(atomic_read(&obj->mm.pages_pin_count) < atomic_read(&obj->bind_count));
-}
-
/**
* i915_vma_insert - finds a slot for the vma in its address space
* @vma: the vma
u64 start, end;
int ret;
- GEM_BUG_ON(i915_vma_is_closed(vma));
GEM_BUG_ON(i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND));
GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color));
- if (vma->obj) {
- struct drm_i915_gem_object *obj = vma->obj;
-
- atomic_inc(&obj->bind_count);
- assert_bind_count(obj);
- }
list_add_tail(&vma->vm_link, &vma->vm->bound_list);
return 0;
* it to be reaped by the shrinker.
*/
list_del(&vma->vm_link);
- if (vma->obj) {
- struct drm_i915_gem_object *obj = vma->obj;
-
- assert_bind_count(obj);
- atomic_dec(&obj->bind_count);
- }
}
static bool try_qad_pin(struct i915_vma *vma, unsigned int flags)
if (flags & PIN_GLOBAL)
wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm);
- /* No more allocations allowed once we hold vm->mutex */
- err = mutex_lock_interruptible(&vma->vm->mutex);
+ /*
+ * Differentiate between user/kernel vma inside the aliasing-ppgtt.
+ *
+ * We conflate the Global GTT with the user's vma when using the
+ * aliasing-ppgtt, but it is still vitally important to try and
+ * keep the use cases distinct. For example, userptr objects are
+ * not allowed inside the Global GTT as that will cause lock
+ * inversions when we have to evict them the mmu_notifier callbacks -
+ * but they are allowed to be part of the user ppGTT which can never
+ * be mapped. As such we try to give the distinct users of the same
+ * mutex, distinct lockclasses [equivalent to how we keep i915_ggtt
+ * and i915_ppgtt separate].
+ *
+ * NB this may cause us to mask real lock inversions -- while the
+ * code is safe today, lockdep may not be able to spot future
+ * transgressions.
+ */
+ err = mutex_lock_interruptible_nested(&vma->vm->mutex,
+ !(flags & PIN_GLOBAL));
if (err)
goto err_fence;
+ /* No more allocations allowed now we hold vm->mutex */
+
+ if (unlikely(i915_vma_is_closed(vma))) {
+ err = -ENOENT;
+ goto err_unlock;
+ }
+
bound = atomic_read(&vma->flags);
if (unlikely(bound & I915_VMA_ERROR)) {
err = -ENOMEM;
mutex_unlock(&vma->vm->mutex);
err_fence:
if (work)
- dma_fence_work_commit(&work->base);
+ dma_fence_work_commit_imm(&work->base);
if (wakeref)
intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref);
err_pages:
do {
err = i915_vma_pin(vma, 0, align, flags | PIN_GLOBAL);
- if (err != -ENOSPC)
+ if (err != -ENOSPC) {
+ if (!err) {
+ err = i915_vma_wait_for_bind(vma);
+ if (err)
+ i915_vma_unpin(vma);
+ }
return err;
+ }
/* Unlike i915_vma_pin, we don't take no for an answer! */
flush_idle_contexts(vm->gt);
void i915_vma_parked(struct intel_gt *gt)
{
struct i915_vma *vma, *next;
+ LIST_HEAD(closed);
spin_lock_irq(>->closed_lock);
list_for_each_entry_safe(vma, next, >->closed_vma, closed_link) {
if (!kref_get_unless_zero(&obj->base.refcount))
continue;
- if (i915_vm_tryopen(vm)) {
- list_del_init(&vma->closed_link);
- } else {
+ if (!i915_vm_tryopen(vm)) {
i915_gem_object_put(obj);
- obj = NULL;
+ continue;
}
- spin_unlock_irq(>->closed_lock);
+ list_move(&vma->closed_link, &closed);
+ }
+ spin_unlock_irq(>->closed_lock);
- if (obj) {
- __i915_vma_put(vma);
- i915_gem_object_put(obj);
- }
+ /* As the GT is held idle, no vma can be reopened as we destroy them */
+ list_for_each_entry_safe(vma, next, &closed, closed_link) {
+ struct drm_i915_gem_object *obj = vma->obj;
+ struct i915_address_space *vm = vma->vm;
- i915_vm_close(vm);
+ INIT_LIST_HEAD(&vma->closed_link);
+ __i915_vma_put(vma);
- /* Restart after dropping lock */
- spin_lock_irq(>->closed_lock);
- next = list_first_entry(>->closed_vma,
- typeof(*next), closed_link);
+ i915_gem_object_put(obj);
+ i915_vm_close(vm);
}
- spin_unlock_irq(>->closed_lock);
}
static void __i915_vma_iounmap(struct i915_vma *vma)
GEM_BUG_ON(!i915_vma_is_pinned(vma));
/* Wait for the vma to be bound before we start! */
- err = i915_request_await_active(rq, &vma->active);
+ err = i915_request_await_active(rq, &vma->active,
+ I915_ACTIVE_AWAIT_EXCL);
if (err)
return err;
dma_resv_add_shared_fence(vma->resv, &rq->fence);
obj->write_domain = 0;
}
+
+ if (flags & EXEC_OBJECT_NEEDS_FENCE && vma->fence)
+ i915_active_add_request(&vma->fence->active, rq);
+
obj->read_domains |= I915_GEM_GPU_DOMAINS;
obj->mm.dirty = true;
lockdep_assert_held(&vma->vm->mutex);
- /*
- * First wait upon any activity as retiring the request may
- * have side-effects such as unpinning or even unbinding this vma.
- *
- * XXX Actually waiting under the vm->mutex is a hinderance and
- * should be pipelined wherever possible. In cases where that is
- * unavoidable, we should lift the wait to before the mutex.
- */
- ret = i915_vma_sync(vma);
- if (ret)
- return ret;
-
if (i915_vma_is_pinned(vma)) {
vma_print_allocator(vma, "is pinned");
return -EAGAIN;
GEM_BUG_ON(i915_vma_is_active(vma));
if (i915_vma_is_map_and_fenceable(vma)) {
+ /* Force a pagefault for domain tracking on next user access */
+ i915_vma_revoke_mmap(vma);
+
/*
* Check that we have flushed all writes through the GGTT
* before the unbind, other due to non-strict nature of those
* indirect writes they may end up referencing the GGTT PTE
* after the unbind.
+ *
+ * Note that we may be concurrently poking at the GGTT_WRITE
+ * bit from set-domain, as we mark all GGTT vma associated
+ * with an object. We know this is for another vma, as we
+ * are currently unbinding this one -- so if this vma will be
+ * reused, it will be refaulted and have its dirty bit set
+ * before the next write.
*/
i915_vma_flush_writes(vma);
- GEM_BUG_ON(i915_vma_has_ggtt_write(vma));
/* release the fence reg _after_ flushing */
- ret = i915_vma_revoke_fence(vma);
- if (ret)
- return ret;
-
- /* Force a pagefault for domain tracking on next user access */
- i915_vma_revoke_mmap(vma);
+ i915_vma_revoke_fence(vma);
__i915_vma_iounmap(vma);
clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
trace_i915_vma_unbind(vma);
vma->ops->unbind_vma(vma);
}
- atomic_and(~(I915_VMA_BIND_MASK | I915_VMA_ERROR), &vma->flags);
+ atomic_and(~(I915_VMA_BIND_MASK | I915_VMA_ERROR | I915_VMA_GGTT_WRITE),
+ &vma->flags);
i915_vma_detach(vma);
vma_unbind_pages(vma);
if (!drm_mm_node_allocated(&vma->node))
return 0;
+ /* Optimistic wait before taking the mutex */
+ err = i915_vma_sync(vma);
+ if (err)
+ goto out_rpm;
+
+ if (i915_vma_is_pinned(vma)) {
+ vma_print_allocator(vma, "is pinned");
+ return -EAGAIN;
+ }
+
if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))
/* XXX not always required: nop_clear_range */
wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm);
- err = mutex_lock_interruptible(&vm->mutex);
+ err = mutex_lock_interruptible_nested(&vma->vm->mutex, !wakeref);
if (err)
- return err;
+ goto out_rpm;
err = __i915_vma_unbind(vma);
mutex_unlock(&vm->mutex);
+out_rpm:
if (wakeref)
intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref);
-
return err;
}