summaryrefslogtreecommitdiff
path: root/drivers/gpu/drm/i915/i915_vma.c
diff options
context:
space:
mode:
Diffstat (limited to 'drivers/gpu/drm/i915/i915_vma.c')
-rw-r--r--drivers/gpu/drm/i915/i915_vma.c1102
1 files changed, 817 insertions, 285 deletions
diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c
index 29a858c53bdd..2c0a63664e13 100644
--- a/drivers/gpu/drm/i915/i915_vma.c
+++ b/drivers/gpu/drm/i915/i915_vma.c
@@ -23,20 +23,40 @@
*/
#include <linux/sched/mm.h>
+#include <linux/dma-fence-array.h>
+
#include <drm/drm_gem.h>
+#include <drm/drm_print.h>
+#include "display/intel_fb.h"
#include "display/intel_frontbuffer.h"
-
#include "gem/i915_gem_lmem.h"
+#include "gem/i915_gem_object_frontbuffer.h"
+#include "gem/i915_gem_tiling.h"
#include "gt/intel_engine.h"
#include "gt/intel_engine_heartbeat.h"
#include "gt/intel_gt.h"
+#include "gt/intel_gt_pm.h"
#include "gt/intel_gt_requests.h"
+#include "gt/intel_tlb.h"
#include "i915_drv.h"
+#include "i915_gem_evict.h"
#include "i915_sw_fence_work.h"
#include "i915_trace.h"
#include "i915_vma.h"
+#include "i915_vma_resource.h"
+
+static inline void assert_vma_held_evict(const struct i915_vma *vma)
+{
+ /*
+ * We may be forced to unbind when the vm is dead, to clean it up.
+ * This is the only exception to the requirement of the object lock
+ * being held.
+ */
+ if (kref_read(&vma->vm->ref))
+ assert_object_held_shared(vma->obj);
+}
static struct kmem_cache *slab_vmas;
@@ -59,14 +79,16 @@ static void vma_print_allocator(struct i915_vma *vma, const char *reason)
char buf[512];
if (!vma->node.stack) {
- DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: unknown owner\n",
- vma->node.start, vma->node.size, reason);
+ drm_dbg(vma->obj->base.dev,
+ "vma.node [%08llx + %08llx] %s: unknown owner\n",
+ vma->node.start, vma->node.size, reason);
return;
}
stack_depot_snprint(vma->node.stack, buf, sizeof(buf), 0);
- DRM_DEBUG_DRIVER("vma.node [%08llx + %08llx] %s: inserted at %s\n",
- vma->node.start, vma->node.size, reason, buf);
+ drm_dbg(vma->obj->base.dev,
+ "vma.node [%08llx + %08llx] %s: inserted at %s\n",
+ vma->node.start, vma->node.size, reason, buf);
}
#else
@@ -84,22 +106,53 @@ static inline struct i915_vma *active_to_vma(struct i915_active *ref)
static int __i915_vma_active(struct i915_active *ref)
{
- return i915_vma_tryget(active_to_vma(ref)) ? 0 : -ENOENT;
+ struct i915_vma *vma = active_to_vma(ref);
+
+ if (!i915_vma_tryget(vma))
+ return -ENOENT;
+
+ /*
+ * Exclude global GTT VMA from holding a GT wakeref
+ * while active, otherwise GPU never goes idle.
+ */
+ if (!i915_vma_is_ggtt(vma)) {
+ /*
+ * Since we and our _retire() counterpart can be
+ * called asynchronously, storing a wakeref tracking
+ * handle inside struct i915_vma is not safe, and
+ * there is no other good place for that. Hence,
+ * use untracked variants of intel_gt_pm_get/put().
+ */
+ intel_gt_pm_get_untracked(vma->vm->gt);
+ }
+
+ return 0;
}
static void __i915_vma_retire(struct i915_active *ref)
{
- i915_vma_put(active_to_vma(ref));
+ struct i915_vma *vma = active_to_vma(ref);
+
+ if (!i915_vma_is_ggtt(vma)) {
+ /*
+ * Since we can be called from atomic contexts,
+ * use an async variant of intel_gt_pm_put().
+ */
+ intel_gt_pm_put_async_untracked(vma->vm->gt);
+ }
+
+ i915_vma_put(vma);
}
static struct i915_vma *
vma_create(struct drm_i915_gem_object *obj,
struct i915_address_space *vm,
- const struct i915_ggtt_view *view)
+ const struct i915_gtt_view *view)
{
struct i915_vma *pos = ERR_PTR(-E2BIG);
struct i915_vma *vma;
struct rb_node *rb, **p;
+ int err;
/* The aliasing_ppgtt should never be used directly! */
GEM_BUG_ON(vm == &vm->gt->ggtt->alias->vm);
@@ -108,8 +161,6 @@ vma_create(struct drm_i915_gem_object *obj,
if (vma == NULL)
return ERR_PTR(-ENOMEM);
- kref_init(&vma->ref);
- vma->vm = i915_vm_get(vm);
vma->ops = &vm->vma_ops;
vma->obj = obj;
vma->size = obj->base.size;
@@ -125,10 +176,12 @@ vma_create(struct drm_i915_gem_object *obj,
}
INIT_LIST_HEAD(&vma->closed_link);
+ INIT_LIST_HEAD(&vma->obj_link);
+ RB_CLEAR_NODE(&vma->obj_node);
- if (view && view->type != I915_GGTT_VIEW_NORMAL) {
- vma->ggtt_view = *view;
- if (view->type == I915_GGTT_VIEW_PARTIAL) {
+ if (view && view->type != I915_GTT_VIEW_NORMAL) {
+ vma->gtt_view = *view;
+ if (view->type == I915_GTT_VIEW_PARTIAL) {
GEM_BUG_ON(range_overflows_t(u64,
view->partial.offset,
view->partial.size,
@@ -136,10 +189,10 @@ vma_create(struct drm_i915_gem_object *obj,
vma->size = view->partial.size;
vma->size <<= PAGE_SHIFT;
GEM_BUG_ON(vma->size > obj->base.size);
- } else if (view->type == I915_GGTT_VIEW_ROTATED) {
+ } else if (view->type == I915_GTT_VIEW_ROTATED) {
vma->size = intel_rotation_info_size(&view->rotated);
vma->size <<= PAGE_SHIFT;
- } else if (view->type == I915_GGTT_VIEW_REMAPPED) {
+ } else if (view->type == I915_GTT_VIEW_REMAPPED) {
vma->size = intel_remapped_info_size(&view->remapped);
vma->size <<= PAGE_SHIFT;
}
@@ -150,8 +203,16 @@ vma_create(struct drm_i915_gem_object *obj,
GEM_BUG_ON(!IS_ALIGNED(vma->size, I915_GTT_PAGE_SIZE));
- spin_lock(&obj->vma.lock);
+ err = mutex_lock_interruptible(&vm->mutex);
+ if (err) {
+ pos = ERR_PTR(err);
+ goto err_vma;
+ }
+ vma->vm = vm;
+ list_add_tail(&vma->vm_link, &vm->unbound_list);
+
+ spin_lock(&obj->vma.lock);
if (i915_is_ggtt(vm)) {
if (unlikely(overflows_type(vma->size, u32)))
goto err_unlock;
@@ -209,13 +270,15 @@ vma_create(struct drm_i915_gem_object *obj,
list_add_tail(&vma->obj_link, &obj->vma.list);
spin_unlock(&obj->vma.lock);
+ mutex_unlock(&vm->mutex);
return vma;
err_unlock:
spin_unlock(&obj->vma.lock);
+ list_del_init(&vma->vm_link);
+ mutex_unlock(&vm->mutex);
err_vma:
- i915_vm_put(vm);
i915_vma_free(vma);
return pos;
}
@@ -223,7 +286,7 @@ err_vma:
static struct i915_vma *
i915_vma_lookup(struct drm_i915_gem_object *obj,
struct i915_address_space *vm,
- const struct i915_ggtt_view *view)
+ const struct i915_gtt_view *view)
{
struct rb_node *rb;
@@ -261,12 +324,12 @@ i915_vma_lookup(struct drm_i915_gem_object *obj,
struct i915_vma *
i915_vma_instance(struct drm_i915_gem_object *obj,
struct i915_address_space *vm,
- const struct i915_ggtt_view *view)
+ const struct i915_gtt_view *view)
{
struct i915_vma *vma;
GEM_BUG_ON(view && !i915_is_ggtt_or_dpt(vm));
- GEM_BUG_ON(!atomic_read(&vm->open));
+ GEM_BUG_ON(!kref_read(&vm->ref));
spin_lock(&obj->vma.lock);
vma = i915_vma_lookup(obj, vm, view);
@@ -284,33 +347,41 @@ struct i915_vma_work {
struct dma_fence_work base;
struct i915_address_space *vm;
struct i915_vm_pt_stash stash;
- struct i915_vma *vma;
- struct drm_i915_gem_object *pinned;
+ struct i915_vma_resource *vma_res;
+ struct drm_i915_gem_object *obj;
struct i915_sw_dma_fence_cb cb;
- enum i915_cache_level cache_level;
+ unsigned int pat_index;
unsigned int flags;
};
static void __vma_bind(struct dma_fence_work *work)
{
struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
- struct i915_vma *vma = vw->vma;
+ struct i915_vma_resource *vma_res = vw->vma_res;
- vma->ops->bind_vma(vw->vm, &vw->stash,
- vma, vw->cache_level, vw->flags);
+ /*
+ * We are about the bind the object, which must mean we have already
+ * signaled the work to potentially clear/move the pages underneath. If
+ * something went wrong at that stage then the object should have
+ * unknown_state set, in which case we need to skip the bind.
+ */
+ if (i915_gem_object_has_unknown_state(vw->obj))
+ return;
+
+ vma_res->ops->bind_vma(vma_res->vm, &vw->stash,
+ vma_res, vw->pat_index, vw->flags);
}
static void __vma_release(struct dma_fence_work *work)
{
struct i915_vma_work *vw = container_of(work, typeof(*vw), base);
- if (vw->pinned) {
- __i915_gem_object_unpin_pages(vw->pinned);
- i915_gem_object_put(vw->pinned);
- }
+ if (vw->obj)
+ i915_gem_object_put(vw->obj);
i915_vm_free_pt_stash(vw->vm, &vw->stash);
- i915_vm_put(vw->vm);
+ if (vw->vma_res)
+ i915_vma_resource_put(vw->vma_res);
}
static const struct dma_fence_work_ops bind_ops = {
@@ -374,36 +445,57 @@ static int i915_vma_verify_bind_complete(struct i915_vma *vma)
#define i915_vma_verify_bind_complete(_vma) 0
#endif
+I915_SELFTEST_EXPORT void
+i915_vma_resource_init_from_vma(struct i915_vma_resource *vma_res,
+ struct i915_vma *vma)
+{
+ struct drm_i915_gem_object *obj = vma->obj;
+
+ i915_vma_resource_init(vma_res, vma->vm, vma->pages, &vma->page_sizes,
+ obj->mm.rsgt, i915_gem_object_is_readonly(obj),
+ i915_gem_object_is_lmem(obj), obj->mm.region,
+ vma->ops, vma->private, __i915_vma_offset(vma),
+ __i915_vma_size(vma), vma->size, vma->guard);
+}
+
/**
* i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space.
* @vma: VMA to map
- * @cache_level: mapping cache level
+ * @pat_index: PAT index to set in PTE
* @flags: flags like global or local mapping
* @work: preallocated worker for allocating and binding the PTE
+ * @vma_res: pointer to a preallocated vma resource. The resource is either
+ * consumed or freed.
*
* DMA addresses are taken from the scatter-gather table of this object (or of
* this VMA in case of non-default GGTT views) and PTE entries set up.
* Note that DMA addresses are also the only part of the SG table we care about.
*/
int i915_vma_bind(struct i915_vma *vma,
- enum i915_cache_level cache_level,
+ unsigned int pat_index,
u32 flags,
- struct i915_vma_work *work)
+ struct i915_vma_work *work,
+ struct i915_vma_resource *vma_res)
{
u32 bind_flags;
u32 vma_flags;
+ int ret;
lockdep_assert_held(&vma->vm->mutex);
GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
- GEM_BUG_ON(vma->size > vma->node.size);
+ GEM_BUG_ON(vma->size > i915_vma_size(vma));
if (GEM_DEBUG_WARN_ON(range_overflows(vma->node.start,
vma->node.size,
- vma->vm->total)))
+ vma->vm->total))) {
+ i915_vma_resource_free(vma_res);
return -ENODEV;
+ }
- if (GEM_DEBUG_WARN_ON(!flags))
+ if (GEM_DEBUG_WARN_ON(!flags)) {
+ i915_vma_resource_free(vma_res);
return -EINVAL;
+ }
bind_flags = flags;
bind_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
@@ -412,17 +504,45 @@ int i915_vma_bind(struct i915_vma *vma,
vma_flags &= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
bind_flags &= ~vma_flags;
- if (bind_flags == 0)
+ if (bind_flags == 0) {
+ i915_vma_resource_free(vma_res);
return 0;
+ }
GEM_BUG_ON(!atomic_read(&vma->pages_count));
+ /* Wait for or await async unbinds touching our range */
+ if (work && bind_flags & vma->vm->bind_async_flags)
+ ret = i915_vma_resource_bind_dep_await(vma->vm,
+ &work->base.chain,
+ vma->node.start,
+ vma->node.size,
+ true,
+ GFP_NOWAIT |
+ __GFP_RETRY_MAYFAIL |
+ __GFP_NOWARN);
+ else
+ ret = i915_vma_resource_bind_dep_sync(vma->vm, vma->node.start,
+ vma->node.size, true);
+ if (ret) {
+ i915_vma_resource_free(vma_res);
+ return ret;
+ }
+
+ if (vma->resource || !vma_res) {
+ /* Rebinding with an additional I915_VMA_*_BIND */
+ GEM_WARN_ON(!vma_flags);
+ i915_vma_resource_free(vma_res);
+ } else {
+ i915_vma_resource_init_from_vma(vma_res, vma);
+ vma->resource = vma_res;
+ }
trace_i915_vma_bind(vma, bind_flags);
if (work && bind_flags & vma->vm->bind_async_flags) {
struct dma_fence *prev;
- work->vma = vma;
- work->cache_level = cache_level;
+ work->vma_res = i915_vma_resource_get(vma->resource);
+ work->pat_index = pat_index;
work->flags = bind_flags;
/*
@@ -443,18 +563,17 @@ int i915_vma_bind(struct i915_vma *vma,
}
work->base.dma.error = 0; /* enable the queue_work() */
-
- __i915_gem_object_pin_pages(vma->obj);
- work->pinned = i915_gem_object_get(vma->obj);
+ work->obj = i915_gem_object_get(vma->obj);
} else {
- if (vma->obj) {
- int ret;
+ ret = i915_gem_object_wait_moving_fence(vma->obj, true);
+ if (ret) {
+ i915_vma_resource_free(vma->resource);
+ vma->resource = NULL;
- ret = i915_gem_object_wait_moving_fence(vma->obj, true);
- if (ret)
- return ret;
+ return ret;
}
- vma->ops->bind_vma(vma->vm, NULL, vma, cache_level, bind_flags);
+ vma->ops->bind_vma(vma->vm, NULL, vma->resource, pat_index,
+ bind_flags);
}
atomic_or(bind_flags, &vma->flags);
@@ -466,12 +585,8 @@ void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
void __iomem *ptr;
int err;
- if (!i915_gem_object_is_lmem(vma->obj)) {
- if (GEM_WARN_ON(!i915_vma_is_map_and_fenceable(vma))) {
- err = -ENODEV;
- goto err;
- }
- }
+ if (WARN_ON_ONCE(vma->obj->flags & I915_BO_ALLOC_GPU_ONLY))
+ return IOMEM_ERR_PTR(-EINVAL);
GEM_BUG_ON(!i915_vma_is_ggtt(vma));
GEM_BUG_ON(!i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND));
@@ -485,20 +600,33 @@ void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
* of pages, that way we can also drop the
* I915_BO_ALLOC_CONTIGUOUS when allocating the object.
*/
- if (i915_gem_object_is_lmem(vma->obj))
+ if (i915_gem_object_is_lmem(vma->obj)) {
ptr = i915_gem_object_lmem_io_map(vma->obj, 0,
vma->obj->base.size);
- else
+ } else if (i915_vma_is_map_and_fenceable(vma)) {
ptr = io_mapping_map_wc(&i915_vm_to_ggtt(vma->vm)->iomap,
- vma->node.start,
- vma->node.size);
+ i915_vma_offset(vma),
+ i915_vma_size(vma));
+ } else {
+ ptr = (void __iomem *)
+ i915_gem_object_pin_map(vma->obj, I915_MAP_WC);
+ if (IS_ERR(ptr)) {
+ err = PTR_ERR(ptr);
+ goto err;
+ }
+ ptr = page_pack_bits(ptr, 1);
+ }
+
if (ptr == NULL) {
err = -ENOMEM;
goto err;
}
if (unlikely(cmpxchg(&vma->iomap, NULL, ptr))) {
- io_mapping_unmap(ptr);
+ if (page_unmask_bits(ptr))
+ __i915_gem_object_release_map(vma->obj);
+ else
+ io_mapping_unmap(ptr);
ptr = vma->iomap;
}
}
@@ -512,12 +640,12 @@ void __iomem *i915_vma_pin_iomap(struct i915_vma *vma)
i915_vma_set_ggtt_write(vma);
/* NB Access through the GTT requires the device to be awake. */
- return ptr;
+ return page_mask_bits(ptr);
err_unpin:
__i915_vma_unpin(vma);
err:
- return IO_ERR_PTR(err);
+ return IOMEM_ERR_PTR(err);
}
void i915_vma_flush_writes(struct i915_vma *vma)
@@ -530,6 +658,8 @@ void i915_vma_unpin_iomap(struct i915_vma *vma)
{
GEM_BUG_ON(vma->iomap == NULL);
+ /* XXX We keep the mapping until __i915_vma_unbind()/evict() */
+
i915_vma_flush_writes(vma);
i915_vma_unpin_fence(vma);
@@ -565,22 +695,26 @@ bool i915_vma_misplaced(const struct i915_vma *vma,
if (test_bit(I915_VMA_ERROR_BIT, __i915_vma_flags(vma)))
return true;
- if (vma->node.size < size)
+ if (i915_vma_size(vma) < size)
return true;
GEM_BUG_ON(alignment && !is_power_of_2(alignment));
- if (alignment && !IS_ALIGNED(vma->node.start, alignment))
+ if (alignment && !IS_ALIGNED(i915_vma_offset(vma), alignment))
return true;
if (flags & PIN_MAPPABLE && !i915_vma_is_map_and_fenceable(vma))
return true;
if (flags & PIN_OFFSET_BIAS &&
- vma->node.start < (flags & PIN_OFFSET_MASK))
+ i915_vma_offset(vma) < (flags & PIN_OFFSET_MASK))
return true;
if (flags & PIN_OFFSET_FIXED &&
- vma->node.start != (flags & PIN_OFFSET_MASK))
+ i915_vma_offset(vma) != (flags & PIN_OFFSET_MASK))
+ return true;
+
+ if (flags & PIN_OFFSET_GUARD &&
+ vma->guard < (flags & PIN_OFFSET_MASK))
return true;
return false;
@@ -593,10 +727,11 @@ void __i915_vma_set_map_and_fenceable(struct i915_vma *vma)
GEM_BUG_ON(!i915_vma_is_ggtt(vma));
GEM_BUG_ON(!vma->fence_size);
- fenceable = (vma->node.size >= vma->fence_size &&
- IS_ALIGNED(vma->node.start, vma->fence_alignment));
+ fenceable = (i915_vma_size(vma) >= vma->fence_size &&
+ IS_ALIGNED(i915_vma_offset(vma), vma->fence_alignment));
- mappable = vma->node.start + vma->fence_size <= i915_vm_to_ggtt(vma->vm)->mappable_end;
+ mappable = i915_ggtt_offset(vma) + vma->fence_size <=
+ i915_vm_to_ggtt(vma->vm)->mappable_end;
if (mappable && fenceable)
set_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
@@ -639,29 +774,32 @@ bool i915_gem_valid_gtt_space(struct i915_vma *vma, unsigned long color)
/**
* i915_vma_insert - finds a slot for the vma in its address space
* @vma: the vma
+ * @ww: An optional struct i915_gem_ww_ctx
* @size: requested size in bytes (can be larger than the VMA)
* @alignment: required alignment
* @flags: mask of PIN_* flags to use
*
* First we try to allocate some free space that meets the requirements for
- * the VMA. Failiing that, if the flags permit, it will evict an old VMA,
- * preferrably the oldest idle entry to make room for the new VMA.
+ * the VMA. Failing that, if the flags permit, it will evict an old VMA,
+ * preferably the oldest idle entry to make room for the new VMA.
*
* Returns:
* 0 on success, negative error code otherwise.
*/
static int
-i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
+i915_vma_insert(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
+ u64 size, u64 alignment, u64 flags)
{
- unsigned long color;
+ unsigned long color, guard;
u64 start, end;
int ret;
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(hweight64(flags & (PIN_OFFSET_GUARD | PIN_OFFSET_FIXED | PIN_OFFSET_BIAS)) > 1);
size = max(size, vma->size);
- alignment = max(alignment, vma->display_alignment);
+ alignment = max_t(typeof(alignment), alignment, vma->display_alignment);
if (flags & PIN_MAPPABLE) {
size = max_t(typeof(size), size, vma->fence_size);
alignment = max_t(typeof(alignment),
@@ -672,6 +810,18 @@ i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
GEM_BUG_ON(!IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
GEM_BUG_ON(!is_power_of_2(alignment));
+ guard = vma->guard; /* retain guard across rebinds */
+ if (flags & PIN_OFFSET_GUARD) {
+ GEM_BUG_ON(overflows_type(flags & PIN_OFFSET_MASK, u32));
+ guard = max_t(u32, guard, flags & PIN_OFFSET_MASK);
+ }
+ /*
+ * As we align the node upon insertion, but the hardware gets
+ * node.start + guard, the easiest way to make that work is
+ * to make the guard a multiple of the alignment size.
+ */
+ guard = ALIGN(guard, alignment);
+
start = flags & PIN_OFFSET_BIAS ? flags & PIN_OFFSET_MASK : 0;
GEM_BUG_ON(!IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
@@ -682,43 +832,58 @@ i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
end = min_t(u64, end, (1ULL << 32) - I915_GTT_PAGE_SIZE);
GEM_BUG_ON(!IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
- /* If binding the object/GGTT view requires more space than the entire
+ alignment = max(alignment, i915_vm_obj_min_alignment(vma->vm, vma->obj));
+
+ /*
+ * If binding the object/GGTT view requires more space than the entire
* aperture has, reject it early before evicting everything in a vain
* attempt to find space.
*/
- if (size > end) {
- DRM_DEBUG("Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
- size, flags & PIN_MAPPABLE ? "mappable" : "total",
- end);
+ if (size > end - 2 * guard) {
+ drm_dbg(vma->obj->base.dev,
+ "Attempting to bind an object larger than the aperture: request=%llu > %s aperture=%llu\n",
+ size, flags & PIN_MAPPABLE ? "mappable" : "total", end);
return -ENOSPC;
}
color = 0;
+
if (i915_vm_has_cache_coloring(vma->vm))
- color = vma->obj->cache_level;
+ color = vma->obj->pat_index;
if (flags & PIN_OFFSET_FIXED) {
u64 offset = flags & PIN_OFFSET_MASK;
if (!IS_ALIGNED(offset, alignment) ||
range_overflows(offset, size, end))
return -EINVAL;
+ /*
+ * The caller knows not of the guard added by others and
+ * requests for the offset of the start of its buffer
+ * to be fixed, which may not be the same as the position
+ * of the vma->node due to the guard pages.
+ */
+ if (offset < guard || offset + size > end - guard)
+ return -ENOSPC;
- ret = i915_gem_gtt_reserve(vma->vm, &vma->node,
- size, offset, color,
- flags);
+ ret = i915_gem_gtt_reserve(vma->vm, ww, &vma->node,
+ size + 2 * guard,
+ offset - guard,
+ color, flags);
if (ret)
return ret;
} else {
+ size += 2 * guard;
/*
* We only support huge gtt pages through the 48b PPGTT,
* however we also don't want to force any alignment for
* objects which need to be tightly packed into the low 32bits.
*
* Note that we assume that GGTT are limited to 4GiB for the
- * forseeable future. See also i915_ggtt_offset().
+ * foreseeable future. See also i915_ggtt_offset().
*/
if (upper_32_bits(end - 1) &&
- vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
+ vma->page_sizes.sg > I915_GTT_PAGE_SIZE &&
+ !HAS_64K_PAGES(vma->vm->i915)) {
/*
* We can't mix 64K and 4K PTEs in the same page-table
* (2M block), and so to avoid the ugliness and
@@ -742,7 +907,7 @@ i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
size = round_up(size, I915_GTT_PAGE_SIZE_2M);
}
- ret = i915_gem_gtt_insert(vma->vm, &vma->node,
+ ret = i915_gem_gtt_insert(vma->vm, ww, &vma->node,
size, alignment, color,
start, end, flags);
if (ret)
@@ -754,7 +919,8 @@ i915_vma_insert(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
GEM_BUG_ON(!drm_mm_node_allocated(&vma->node));
GEM_BUG_ON(!i915_gem_valid_gtt_space(vma, color));
- list_add_tail(&vma->vm_link, &vma->vm->bound_list);
+ list_move_tail(&vma->vm_link, &vma->vm->bound_list);
+ vma->guard = guard;
return 0;
}
@@ -770,15 +936,23 @@ i915_vma_detach(struct i915_vma *vma)
* vma, we can drop its hold on the backing storage and allow
* it to be reaped by the shrinker.
*/
- list_del(&vma->vm_link);
+ list_move_tail(&vma->vm_link, &vma->vm->unbound_list);
}
static bool try_qad_pin(struct i915_vma *vma, unsigned int flags)
{
unsigned int bound;
- bool pinned = true;
bound = atomic_read(&vma->flags);
+
+ if (flags & PIN_VALIDATE) {
+ flags &= I915_VMA_BIND_MASK;
+
+ return (flags & bound) == flags;
+ }
+
+ /* with the lock mandatory for unbind, we don't race here */
+ flags &= I915_VMA_BIND_MASK;
do {
if (unlikely(flags & ~bound))
return false;
@@ -786,34 +960,10 @@ static bool try_qad_pin(struct i915_vma *vma, unsigned int flags)
if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR)))
return false;
- if (!(bound & I915_VMA_PIN_MASK))
- goto unpinned;
-
GEM_BUG_ON(((bound + 1) & I915_VMA_PIN_MASK) == 0);
} while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
return true;
-
-unpinned:
- /*
- * If pin_count==0, but we are bound, check under the lock to avoid
- * racing with a concurrent i915_vma_unbind().
- */
- mutex_lock(&vma->vm->mutex);
- do {
- if (unlikely(bound & (I915_VMA_OVERFLOW | I915_VMA_ERROR))) {
- pinned = false;
- break;
- }
-
- if (unlikely(flags & ~bound)) {
- pinned = false;
- break;
- }
- } while (!atomic_try_cmpxchg(&vma->flags, &bound, bound + 1));
- mutex_unlock(&vma->vm->mutex);
-
- return pinned;
}
static struct scatterlist *
@@ -823,7 +973,7 @@ rotate_pages(struct drm_i915_gem_object *obj, unsigned int offset,
struct sg_table *st, struct scatterlist *sg)
{
unsigned int column, row;
- unsigned int src_idx;
+ pgoff_t src_idx;
for (column = 0; column < width; column++) {
unsigned int left;
@@ -853,7 +1003,7 @@ rotate_pages(struct drm_i915_gem_object *obj, unsigned int offset,
/*
* The DE ignores the PTEs for the padding tiles, the sg entry
- * here is just a conenience to indicate how many padding PTEs
+ * here is just a convenience to indicate how many padding PTEs
* to insert at this spot.
*/
sg_set_page(sg, NULL, left, 0);
@@ -909,30 +1059,39 @@ err_st_alloc:
}
static struct scatterlist *
-remap_pages(struct drm_i915_gem_object *obj,
- unsigned int offset, unsigned int alignment_pad,
- unsigned int width, unsigned int height,
- unsigned int src_stride, unsigned int dst_stride,
- struct sg_table *st, struct scatterlist *sg)
+add_padding_pages(unsigned int count,
+ struct sg_table *st, struct scatterlist *sg)
+{
+ st->nents++;
+
+ /*
+ * The DE ignores the PTEs for the padding tiles, the sg entry
+ * here is just a convenience to indicate how many padding PTEs
+ * to insert at this spot.
+ */
+ sg_set_page(sg, NULL, count * I915_GTT_PAGE_SIZE, 0);
+ sg_dma_address(sg) = 0;
+ sg_dma_len(sg) = count * I915_GTT_PAGE_SIZE;
+ sg = sg_next(sg);
+
+ return sg;
+}
+
+static struct scatterlist *
+remap_tiled_color_plane_pages(struct drm_i915_gem_object *obj,
+ unsigned long offset, unsigned int alignment_pad,
+ unsigned int width, unsigned int height,
+ unsigned int src_stride, unsigned int dst_stride,
+ struct sg_table *st, struct scatterlist *sg,
+ unsigned int *gtt_offset)
{
unsigned int row;
if (!width || !height)
return sg;
- if (alignment_pad) {
- st->nents++;
-
- /*
- * The DE ignores the PTEs for the padding tiles, the sg entry
- * here is just a convenience to indicate how many padding PTEs
- * to insert at this spot.
- */
- sg_set_page(sg, NULL, alignment_pad * 4096, 0);
- sg_dma_address(sg) = 0;
- sg_dma_len(sg) = alignment_pad * 4096;
- sg = sg_next(sg);
- }
+ if (alignment_pad)
+ sg = add_padding_pages(alignment_pad, st, sg);
for (row = 0; row < height; row++) {
unsigned int left = width * I915_GTT_PAGE_SIZE;
@@ -969,18 +1128,98 @@ remap_pages(struct drm_i915_gem_object *obj,
if (!left)
continue;
+ sg = add_padding_pages(left >> PAGE_SHIFT, st, sg);
+ }
+
+ *gtt_offset += alignment_pad + dst_stride * height;
+
+ return sg;
+}
+
+static struct scatterlist *
+remap_contiguous_pages(struct drm_i915_gem_object *obj,
+ pgoff_t obj_offset,
+ unsigned int count,
+ struct sg_table *st, struct scatterlist *sg)
+{
+ struct scatterlist *iter;
+ unsigned int offset;
+
+ iter = i915_gem_object_get_sg_dma(obj, obj_offset, &offset);
+ GEM_BUG_ON(!iter);
+
+ do {
+ unsigned int len;
+
+ len = min(sg_dma_len(iter) - (offset << PAGE_SHIFT),
+ count << PAGE_SHIFT);
+ sg_set_page(sg, NULL, len, 0);
+ sg_dma_address(sg) =
+ sg_dma_address(iter) + (offset << PAGE_SHIFT);
+ sg_dma_len(sg) = len;
+
st->nents++;
+ count -= len >> PAGE_SHIFT;
+ if (count == 0)
+ return sg;
- /*
- * The DE ignores the PTEs for the padding tiles, the sg entry
- * here is just a conenience to indicate how many padding PTEs
- * to insert at this spot.
- */
- sg_set_page(sg, NULL, left, 0);
- sg_dma_address(sg) = 0;
- sg_dma_len(sg) = left;
- sg = sg_next(sg);
- }
+ sg = __sg_next(sg);
+ iter = __sg_next(iter);
+ offset = 0;
+ } while (1);
+}
+
+static struct scatterlist *
+remap_linear_color_plane_pages(struct drm_i915_gem_object *obj,
+ pgoff_t obj_offset, unsigned int alignment_pad,
+ unsigned int size,
+ struct sg_table *st, struct scatterlist *sg,
+ unsigned int *gtt_offset)
+{
+ if (!size)
+ return sg;
+
+ if (alignment_pad)
+ sg = add_padding_pages(alignment_pad, st, sg);
+
+ sg = remap_contiguous_pages(obj, obj_offset, size, st, sg);
+ sg = sg_next(sg);
+
+ *gtt_offset += alignment_pad + size;
+
+ return sg;
+}
+
+static struct scatterlist *
+remap_color_plane_pages(const struct intel_remapped_info *rem_info,
+ struct drm_i915_gem_object *obj,
+ int color_plane,
+ struct sg_table *st, struct scatterlist *sg,
+ unsigned int *gtt_offset)
+{
+ unsigned int alignment_pad = 0;
+
+ if (rem_info->plane_alignment)
+ alignment_pad = ALIGN(*gtt_offset, rem_info->plane_alignment) - *gtt_offset;
+
+ if (rem_info->plane[color_plane].linear)
+ sg = remap_linear_color_plane_pages(obj,
+ rem_info->plane[color_plane].offset,
+ alignment_pad,
+ rem_info->plane[color_plane].size,
+ st, sg,
+ gtt_offset);
+
+ else
+ sg = remap_tiled_color_plane_pages(obj,
+ rem_info->plane[color_plane].offset,
+ alignment_pad,
+ rem_info->plane[color_plane].width,
+ rem_info->plane[color_plane].height,
+ rem_info->plane[color_plane].src_stride,
+ rem_info->plane[color_plane].dst_stride,
+ st, sg,
+ gtt_offset);
return sg;
}
@@ -1009,21 +1248,8 @@ intel_remap_pages(struct intel_remapped_info *rem_info,
st->nents = 0;
sg = st->sgl;
- for (i = 0 ; i < ARRAY_SIZE(rem_info->plane); i++) {
- unsigned int alignment_pad = 0;
-
- if (rem_info->plane_alignment)
- alignment_pad = ALIGN(gtt_offset, rem_info->plane_alignment) - gtt_offset;
-
- sg = remap_pages(obj,
- rem_info->plane[i].offset, alignment_pad,
- rem_info->plane[i].width, rem_info->plane[i].height,
- rem_info->plane[i].src_stride, rem_info->plane[i].dst_stride,
- st, sg);
-
- gtt_offset += alignment_pad +
- rem_info->plane[i].dst_stride * rem_info->plane[i].height;
- }
+ for (i = 0 ; i < ARRAY_SIZE(rem_info->plane); i++)
+ sg = remap_color_plane_pages(rem_info, obj, i, st, sg, &gtt_offset);
i915_sg_trim(st);
@@ -1041,13 +1267,12 @@ err_st_alloc:
}
static noinline struct sg_table *
-intel_partial_pages(const struct i915_ggtt_view *view,
+intel_partial_pages(const struct i915_gtt_view *view,
struct drm_i915_gem_object *obj)
{
struct sg_table *st;
- struct scatterlist *sg, *iter;
+ struct scatterlist *sg;
unsigned int count = view->partial.size;
- unsigned int offset;
int ret = -ENOMEM;
st = kmalloc(sizeof(*st), GFP_KERNEL);
@@ -1058,34 +1283,14 @@ intel_partial_pages(const struct i915_ggtt_view *view,
if (ret)
goto err_sg_alloc;
- iter = i915_gem_object_get_sg_dma(obj, view->partial.offset, &offset);
- GEM_BUG_ON(!iter);
-
- sg = st->sgl;
st->nents = 0;
- do {
- unsigned int len;
-
- len = min(sg_dma_len(iter) - (offset << PAGE_SHIFT),
- count << PAGE_SHIFT);
- sg_set_page(sg, NULL, len, 0);
- sg_dma_address(sg) =
- sg_dma_address(iter) + (offset << PAGE_SHIFT);
- sg_dma_len(sg) = len;
- st->nents++;
- count -= len >> PAGE_SHIFT;
- if (count == 0) {
- sg_mark_end(sg);
- i915_sg_trim(st); /* Drop any unused tail entries. */
+ sg = remap_contiguous_pages(obj, view->partial.offset, count, st, st->sgl);
- return st;
- }
+ sg_mark_end(sg);
+ i915_sg_trim(st); /* Drop any unused tail entries. */
- sg = __sg_next(sg);
- iter = __sg_next(iter);
- offset = 0;
- } while (1);
+ return st;
err_sg_alloc:
kfree(st);
@@ -1097,7 +1302,6 @@ static int
__i915_vma_get_pages(struct i915_vma *vma)
{
struct sg_table *pages;
- int ret;
/*
* The vma->pages are only valid within the lifespan of the borrowed
@@ -1107,41 +1311,39 @@ __i915_vma_get_pages(struct i915_vma *vma)
*/
GEM_BUG_ON(!i915_gem_object_has_pinned_pages(vma->obj));
- switch (vma->ggtt_view.type) {
+ switch (vma->gtt_view.type) {
default:
- GEM_BUG_ON(vma->ggtt_view.type);
+ GEM_BUG_ON(vma->gtt_view.type);
fallthrough;
- case I915_GGTT_VIEW_NORMAL:
+ case I915_GTT_VIEW_NORMAL:
pages = vma->obj->mm.pages;
break;
- case I915_GGTT_VIEW_ROTATED:
+ case I915_GTT_VIEW_ROTATED:
pages =
- intel_rotate_pages(&vma->ggtt_view.rotated, vma->obj);
+ intel_rotate_pages(&vma->gtt_view.rotated, vma->obj);
break;
- case I915_GGTT_VIEW_REMAPPED:
+ case I915_GTT_VIEW_REMAPPED:
pages =
- intel_remap_pages(&vma->ggtt_view.remapped, vma->obj);
+ intel_remap_pages(&vma->gtt_view.remapped, vma->obj);
break;
- case I915_GGTT_VIEW_PARTIAL:
- pages = intel_partial_pages(&vma->ggtt_view, vma->obj);
+ case I915_GTT_VIEW_PARTIAL:
+ pages = intel_partial_pages(&vma->gtt_view, vma->obj);
break;
}
- ret = 0;
if (IS_ERR(pages)) {
- ret = PTR_ERR(pages);
- pages = NULL;
drm_err(&vma->vm->i915->drm,
- "Failed to get pages for VMA view type %u (%d)!\n",
- vma->ggtt_view.type, ret);
+ "Failed to get pages for VMA view type %u (%ld)!\n",
+ vma->gtt_view.type, PTR_ERR(pages));
+ return PTR_ERR(pages);
}
vma->pages = pages;
- return ret;
+ return 0;
}
I915_SELFTEST_EXPORT int i915_vma_get_pages(struct i915_vma *vma)
@@ -1170,28 +1372,38 @@ err_unpin:
return err;
}
+void vma_invalidate_tlb(struct i915_address_space *vm, u32 *tlb)
+{
+ struct intel_gt *gt;
+ int id;
+
+ if (!tlb)
+ return;
+
+ /*
+ * Before we release the pages that were bound by this vma, we
+ * must invalidate all the TLBs that may still have a reference
+ * back to our physical address. It only needs to be done once,
+ * so after updating the PTE to point away from the pages, record
+ * the most recent TLB invalidation seqno, and if we have not yet
+ * flushed the TLBs upon release, perform a full invalidation.
+ */
+ for_each_gt(gt, vm->i915, id)
+ WRITE_ONCE(tlb[id],
+ intel_gt_next_invalidate_tlb_full(gt));
+}
+
static void __vma_put_pages(struct i915_vma *vma, unsigned int count)
{
/* We allocate under vma_get_pages, so beware the shrinker */
- struct sg_table *pages = READ_ONCE(vma->pages);
-
GEM_BUG_ON(atomic_read(&vma->pages_count) < count);
if (atomic_sub_return(count, &vma->pages_count) == 0) {
- /*
- * The atomic_sub_return is a read barrier for the READ_ONCE of
- * vma->pages above.
- *
- * READ_ONCE is safe because this is either called from the same
- * function (i915_vma_pin_ww), or guarded by vma->vm->mutex.
- *
- * TODO: We're leaving vma->pages dangling, until vma->obj->resv
- * lock is required.
- */
- if (pages != vma->obj->mm.pages) {
- sg_free_table(pages);
- kfree(pages);
+ if (vma->pages != vma->obj->mm.pages) {
+ sg_free_table(vma->pages);
+ kfree(vma->pages);
}
+ vma->pages = NULL;
i915_gem_object_unpin_pages(vma->obj);
}
@@ -1224,7 +1436,8 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
{
struct i915_vma_work *work = NULL;
struct dma_fence *moving = NULL;
- intel_wakeref_t wakeref = 0;
+ struct i915_vma_resource *vma_res = NULL;
+ intel_wakeref_t wakeref;
unsigned int bound;
int err;
@@ -1237,18 +1450,23 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
GEM_BUG_ON(!(flags & (PIN_USER | PIN_GLOBAL)));
/* First try and grab the pin without rebinding the vma */
- if (try_qad_pin(vma, flags & I915_VMA_BIND_MASK))
+ if (try_qad_pin(vma, flags))
return 0;
err = i915_vma_get_pages(vma);
if (err)
return err;
- if (flags & PIN_GLOBAL)
- wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm);
+ /*
+ * In case of a global GTT, we must hold a runtime-pm wakeref
+ * while global PTEs are updated. In other cases, we hold
+ * the rpm reference while the VMA is active. Since runtime
+ * resume may require allocations, which are forbidden inside
+ * vm->mutex, get the first rpm wakeref outside of the mutex.
+ */
+ wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm);
- moving = vma->obj ? i915_gem_object_get_moving_fence(vma->obj) : NULL;
- if (flags & vma->vm->bind_async_flags || moving) {
+ if (flags & vma->vm->bind_async_flags) {
/* lock VM */
err = i915_vm_lock_objects(vma->vm, ww);
if (err)
@@ -1260,7 +1478,11 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
goto err_rpm;
}
- work->vm = i915_vm_get(vma->vm);
+ work->vm = vma->vm;
+
+ err = i915_gem_object_get_moving_fence(vma->obj, &moving);
+ if (err)
+ goto err_rpm;
dma_fence_work_chain(&work->base, moving);
@@ -1278,6 +1500,12 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
}
}
+ vma_res = i915_vma_resource_alloc();
+ if (IS_ERR(vma_res)) {
+ err = PTR_ERR(vma_res);
+ goto err_fence;
+ }
+
/*
* Differentiate between user/kernel vma inside the aliasing-ppgtt.
*
@@ -1298,7 +1526,7 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
err = mutex_lock_interruptible_nested(&vma->vm->mutex,
!(flags & PIN_GLOBAL));
if (err)
- goto err_fence;
+ goto err_vma_res;
/* No more allocations allowed now we hold vm->mutex */
@@ -1319,7 +1547,8 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
}
if (unlikely(!(flags & ~bound & I915_VMA_BIND_MASK))) {
- __i915_vma_pin(vma);
+ if (!(flags & PIN_VALIDATE))
+ __i915_vma_pin(vma);
goto err_unlock;
}
@@ -1328,7 +1557,7 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
goto err_unlock;
if (!(bound & I915_VMA_BIND_MASK)) {
- err = i915_vma_insert(vma, size, alignment, flags);
+ err = i915_vma_insert(vma, ww, size, alignment, flags);
if (err)
goto err_active;
@@ -1338,8 +1567,9 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
GEM_BUG_ON(!vma->pages);
err = i915_vma_bind(vma,
- vma->obj->cache_level,
- flags, work);
+ vma->obj->pat_index,
+ flags, work, vma_res);
+ vma_res = NULL;
if (err)
goto err_remove;
@@ -1348,8 +1578,10 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
atomic_add(I915_VMA_PAGES_ACTIVE, &vma->pages_count);
list_move_tail(&vma->vm_link, &vma->vm->bound_list);
- __i915_vma_pin(vma);
- GEM_BUG_ON(!i915_vma_is_pinned(vma));
+ if (!(flags & PIN_VALIDATE)) {
+ __i915_vma_pin(vma);
+ GEM_BUG_ON(!i915_vma_is_pinned(vma));
+ }
GEM_BUG_ON(!i915_vma_is_bound(vma, flags));
GEM_BUG_ON(i915_vma_misplaced(vma, size, alignment, flags));
@@ -1362,12 +1594,25 @@ err_active:
i915_active_release(&vma->active);
err_unlock:
mutex_unlock(&vma->vm->mutex);
+err_vma_res:
+ i915_vma_resource_free(vma_res);
err_fence:
- if (work)
- dma_fence_work_commit_imm(&work->base);
+ if (work) {
+ /*
+ * When pinning VMA to GGTT on CHV or BXT with VTD enabled,
+ * commit VMA binding asynchronously to avoid risk of lock
+ * inversion among reservation_ww locks held here and
+ * cpu_hotplug_lock acquired from stop_machine(), which we
+ * wrap around GGTT updates when running in those environments.
+ */
+ if (i915_vma_is_ggtt(vma) &&
+ intel_vm_no_concurrent_access_wa(vma->vm->i915))
+ dma_fence_work_commit(&work->base);
+ else
+ dma_fence_work_commit_imm(&work->base);
+ }
err_rpm:
- if (wakeref)
- intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref);
+ intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref);
if (moving)
dma_fence_put(moving);
@@ -1376,6 +1621,26 @@ err_rpm:
return err;
}
+int i915_vma_pin(struct i915_vma *vma, u64 size, u64 alignment, u64 flags)
+{
+ struct i915_gem_ww_ctx ww;
+ int err;
+
+ i915_gem_ww_ctx_init(&ww, true);
+retry:
+ err = i915_gem_object_lock(vma->obj, &ww);
+ if (!err)
+ err = i915_vma_pin_ww(vma, &ww, size, alignment, flags);
+ if (err == -EDEADLK) {
+ err = i915_gem_ww_ctx_backoff(&ww);
+ if (!err)
+ goto retry;
+ }
+ i915_gem_ww_ctx_fini(&ww);
+
+ return err;
+}
+
static void flush_idle_contexts(struct intel_gt *gt)
{
struct intel_engine_cs *engine;
@@ -1391,6 +1656,8 @@ static int __i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
u32 align, unsigned int flags)
{
struct i915_address_space *vm = vma->vm;
+ struct intel_gt *gt;
+ struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
int err;
do {
@@ -1406,9 +1673,15 @@ static int __i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
}
/* Unlike i915_vma_pin, we don't take no for an answer! */
- flush_idle_contexts(vm->gt);
+ list_for_each_entry(gt, &ggtt->gt_list, ggtt_link)
+ flush_idle_contexts(gt);
if (mutex_lock_interruptible(&vm->mutex) == 0) {
- i915_gem_evict_vm(vm);
+ /*
+ * We pass NULL ww here, as we don't want to unbind
+ * locked objects when called from execbuf when pinning
+ * is removed. This would probably regress badly.
+ */
+ i915_gem_evict_vm(vm, NULL, NULL);
mutex_unlock(&vm->mutex);
}
} while (1);
@@ -1425,9 +1698,7 @@ int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
if (ww)
return __i915_ggtt_pin(vma, ww, align, flags);
-#ifdef CONFIG_LOCKDEP
- WARN_ON(dma_resv_held(vma->obj->base.resv));
-#endif
+ lockdep_assert_not_held(&vma->obj->base.resv->lock.base);
for_i915_gem_ww(&_ww, err, true) {
err = i915_gem_object_lock(vma->obj, &_ww);
@@ -1438,6 +1709,26 @@ int i915_ggtt_pin(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
return err;
}
+/**
+ * i915_ggtt_clear_scanout - Clear scanout flag for all objects ggtt vmas
+ * @obj: i915 GEM object
+ * This function clears scanout flags for objects ggtt vmas. These flags are set
+ * when object is pinned for display use and this function to clear them all is
+ * targeted to be called by frontbuffer tracking code when the frontbuffer is
+ * about to be released.
+ */
+void i915_ggtt_clear_scanout(struct drm_i915_gem_object *obj)
+{
+ struct i915_vma *vma;
+
+ spin_lock(&obj->vma.lock);
+ for_each_ggtt_vma(vma, obj) {
+ i915_vma_clear_scanout(vma);
+ vma->display_alignment = I915_GTT_MIN_ALIGNMENT;
+ }
+ spin_unlock(&obj->vma.lock);
+}
+
static void __vma_close(struct i915_vma *vma, struct intel_gt *gt)
{
/*
@@ -1475,46 +1766,107 @@ void i915_vma_close(struct i915_vma *vma)
static void __i915_vma_remove_closed(struct i915_vma *vma)
{
- struct intel_gt *gt = vma->vm->gt;
-
- spin_lock_irq(&gt->closed_lock);
list_del_init(&vma->closed_link);
- spin_unlock_irq(&gt->closed_lock);
}
void i915_vma_reopen(struct i915_vma *vma)
{
+ struct intel_gt *gt = vma->vm->gt;
+
+ spin_lock_irq(&gt->closed_lock);
if (i915_vma_is_closed(vma))
__i915_vma_remove_closed(vma);
+ spin_unlock_irq(&gt->closed_lock);
+}
+
+static void force_unbind(struct i915_vma *vma)
+{
+ if (!drm_mm_node_allocated(&vma->node))
+ return;
+
+ atomic_and(~I915_VMA_PIN_MASK, &vma->flags);
+ WARN_ON(__i915_vma_unbind(vma));
+ GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
}
-void i915_vma_release(struct kref *ref)
+static void release_references(struct i915_vma *vma, struct intel_gt *gt,
+ bool vm_ddestroy)
{
- struct i915_vma *vma = container_of(ref, typeof(*vma), ref);
struct drm_i915_gem_object *obj = vma->obj;
- if (drm_mm_node_allocated(&vma->node)) {
- mutex_lock(&vma->vm->mutex);
- atomic_and(~I915_VMA_PIN_MASK, &vma->flags);
- WARN_ON(__i915_vma_unbind(vma));
- mutex_unlock(&vma->vm->mutex);
- GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
- }
GEM_BUG_ON(i915_vma_is_active(vma));
spin_lock(&obj->vma.lock);
list_del(&vma->obj_link);
if (!RB_EMPTY_NODE(&vma->obj_node))
rb_erase(&vma->obj_node, &obj->vma.tree);
+
spin_unlock(&obj->vma.lock);
+ spin_lock_irq(&gt->closed_lock);
__i915_vma_remove_closed(vma);
- i915_vm_put(vma->vm);
+ spin_unlock_irq(&gt->closed_lock);
+
+ if (vm_ddestroy)
+ i915_vm_resv_put(vma->vm);
i915_active_fini(&vma->active);
+ GEM_WARN_ON(vma->resource);
i915_vma_free(vma);
}
+/*
+ * i915_vma_destroy_locked - Remove all weak reference to the vma and put
+ * the initial reference.
+ *
+ * This function should be called when it's decided the vma isn't needed
+ * anymore. The caller must assure that it doesn't race with another lookup
+ * plus destroy, typically by taking an appropriate reference.
+ *
+ * Current callsites are
+ * - __i915_gem_object_pages_fini()
+ * - __i915_vm_close() - Blocks the above function by taking a reference on
+ * the object.
+ * - __i915_vma_parked() - Blocks the above functions by taking a reference
+ * on the vm and a reference on the object. Also takes the object lock so
+ * destruction from __i915_vma_parked() can be blocked by holding the
+ * object lock. Since the object lock is only allowed from within i915 with
+ * an object refcount, holding the object lock also implicitly blocks the
+ * vma freeing from __i915_gem_object_pages_fini().
+ *
+ * Because of locks taken during destruction, a vma is also guaranteed to
+ * stay alive while the following locks are held if it was looked up while
+ * holding one of the locks:
+ * - vm->mutex
+ * - obj->vma.lock
+ * - gt->closed_lock
+ */
+void i915_vma_destroy_locked(struct i915_vma *vma)
+{
+ lockdep_assert_held(&vma->vm->mutex);
+
+ force_unbind(vma);
+ list_del_init(&vma->vm_link);
+ release_references(vma, vma->vm->gt, false);
+}
+
+void i915_vma_destroy(struct i915_vma *vma)
+{
+ struct intel_gt *gt;
+ bool vm_ddestroy;
+
+ mutex_lock(&vma->vm->mutex);
+ force_unbind(vma);
+ list_del_init(&vma->vm_link);
+ vm_ddestroy = vma->vm_ddestroy;
+ vma->vm_ddestroy = false;
+
+ /* vma->vm may be freed when releasing vma->vm->mutex. */
+ gt = vma->vm->gt;
+ mutex_unlock(&vma->vm->mutex);
+ release_references(vma, gt, vm_ddestroy);
+}
+
void i915_vma_parked(struct intel_gt *gt)
{
struct i915_vma *vma, *next;
@@ -1530,7 +1882,7 @@ void i915_vma_parked(struct intel_gt *gt)
if (!kref_get_unless_zero(&obj->base.refcount))
continue;
- if (!i915_vm_tryopen(vm)) {
+ if (!i915_vm_tryget(vm)) {
i915_gem_object_put(obj);
continue;
}
@@ -1544,11 +1896,19 @@ void i915_vma_parked(struct intel_gt *gt)
struct drm_i915_gem_object *obj = vma->obj;
struct i915_address_space *vm = vma->vm;
- INIT_LIST_HEAD(&vma->closed_link);
- __i915_vma_put(vma);
+ if (i915_gem_object_trylock(obj, NULL)) {
+ INIT_LIST_HEAD(&vma->closed_link);
+ i915_vma_destroy(vma);
+ i915_gem_object_unlock(obj);
+ } else {
+ /* back you go.. */
+ spin_lock_irq(&gt->closed_lock);
+ list_add(&vma->closed_link, &gt->closed_vma);
+ spin_unlock_irq(&gt->closed_lock);
+ }
i915_gem_object_put(obj);
- i915_vm_close(vm);
+ i915_vm_put(vm);
}
}
@@ -1559,7 +1919,10 @@ static void __i915_vma_iounmap(struct i915_vma *vma)
if (vma->iomap == NULL)
return;
- io_mapping_unmap(vma->iomap);
+ if (page_unmask_bits(vma->iomap))
+ __i915_gem_object_release_map(vma->obj);
+ else
+ io_mapping_unmap(vma->iomap);
vma->iomap = NULL;
}
@@ -1575,7 +1938,7 @@ void i915_vma_revoke_mmap(struct i915_vma *vma)
GEM_BUG_ON(!vma->obj->userfault_count);
node = &vma->mmo->vma_node;
- vma_offset = vma->ggtt_view.partial.offset << PAGE_SHIFT;
+ vma_offset = vma->gtt_view.partial.offset << PAGE_SHIFT;
unmap_mapping_range(vma->vm->i915->drm.anon_inode->i_mapping,
drm_vma_node_offset_addr(node) + vma_offset,
vma->size,
@@ -1596,8 +1959,6 @@ static int __i915_vma_move_to_active(struct i915_vma *vma, struct i915_request *
{
int err;
- GEM_BUG_ON(!i915_vma_is_pinned(vma));
-
/* Wait for the vma to be bound before we start! */
err = __i915_request_await_bind(rq, vma);
if (err)
@@ -1616,36 +1977,59 @@ int _i915_vma_move_to_active(struct i915_vma *vma,
assert_object_held(obj);
+ GEM_BUG_ON(!vma->pages);
+
+ if (!(flags & __EXEC_OBJECT_NO_REQUEST_AWAIT)) {
+ err = i915_request_await_object(rq, vma->obj, flags & EXEC_OBJECT_WRITE);
+ if (unlikely(err))
+ return err;
+ }
err = __i915_vma_move_to_active(vma, rq);
if (unlikely(err))
return err;
+ /*
+ * Reserve fences slot early to prevent an allocation after preparing
+ * the workload and associating fences with dma_resv.
+ */
+ if (fence && !(flags & __EXEC_OBJECT_NO_RESERVE)) {
+ struct dma_fence *curr;
+ int idx;
+
+ dma_fence_array_for_each(curr, idx, fence)
+ ;
+ err = dma_resv_reserve_fences(vma->obj->base.resv, idx);
+ if (unlikely(err))
+ return err;
+ }
+
if (flags & EXEC_OBJECT_WRITE) {
- struct intel_frontbuffer *front;
+ struct i915_frontbuffer *front;
- front = __intel_frontbuffer_get(obj);
+ front = i915_gem_object_frontbuffer_lookup(obj);
if (unlikely(front)) {
- if (intel_frontbuffer_invalidate(front, ORIGIN_CS))
+ if (intel_frontbuffer_invalidate(&front->base, ORIGIN_CS))
i915_active_add_request(&front->write, rq);
- intel_frontbuffer_put(front);
+ i915_gem_object_frontbuffer_put(front);
}
+ }
- if (fence) {
- dma_resv_add_excl_fence(vma->obj->base.resv, fence);
+ if (fence) {
+ struct dma_fence *curr;
+ enum dma_resv_usage usage;
+ int idx;
+
+ if (flags & EXEC_OBJECT_WRITE) {
+ usage = DMA_RESV_USAGE_WRITE;
obj->write_domain = I915_GEM_DOMAIN_RENDER;
obj->read_domains = 0;
- }
- } else {
- if (!(flags & __EXEC_OBJECT_NO_RESERVE)) {
- err = dma_resv_reserve_shared(vma->obj->base.resv, 1);
- if (unlikely(err))
- return err;
- }
-
- if (fence) {
- dma_resv_add_shared_fence(vma->obj->base.resv, fence);
+ } else {
+ usage = DMA_RESV_USAGE_READ;
obj->write_domain = 0;
}
+
+ dma_fence_array_for_each(curr, idx, fence)
+ dma_resv_add_fence(vma->obj->base.resv, curr, usage);
}
if (flags & EXEC_OBJECT_NEEDS_FENCE && vma->fence)
@@ -1658,9 +2042,13 @@ int _i915_vma_move_to_active(struct i915_vma *vma,
return 0;
}
-void __i915_vma_evict(struct i915_vma *vma)
+struct dma_fence *__i915_vma_evict(struct i915_vma *vma, bool async)
{
+ struct i915_vma_resource *vma_res = vma->resource;
+ struct dma_fence *unbind_fence;
+
GEM_BUG_ON(i915_vma_is_pinned(vma));
+ assert_vma_held_evict(vma);
if (i915_vma_is_map_and_fenceable(vma)) {
/* Force a pagefault for domain tracking on next user access */
@@ -1684,21 +2072,54 @@ void __i915_vma_evict(struct i915_vma *vma)
/* release the fence reg _after_ flushing */
i915_vma_revoke_fence(vma);
- __i915_vma_iounmap(vma);
clear_bit(I915_VMA_CAN_FENCE_BIT, __i915_vma_flags(vma));
}
+
+ __i915_vma_iounmap(vma);
+
GEM_BUG_ON(vma->fence);
GEM_BUG_ON(i915_vma_has_userfault(vma));
- if (likely(atomic_read(&vma->vm->open))) {
- trace_i915_vma_unbind(vma);
- vma->ops->unbind_vma(vma->vm, vma);
- }
+ /* Object backend must be async capable. */
+ GEM_WARN_ON(async && !vma->resource->bi.pages_rsgt);
+
+ /* If vm is not open, unbind is a nop. */
+ vma_res->needs_wakeref = i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND) &&
+ kref_read(&vma->vm->ref);
+ vma_res->skip_pte_rewrite = !kref_read(&vma->vm->ref) ||
+ vma->vm->skip_pte_rewrite;
+ trace_i915_vma_unbind(vma);
+
+ if (async)
+ unbind_fence = i915_vma_resource_unbind(vma_res,
+ vma->obj->mm.tlb);
+ else
+ unbind_fence = i915_vma_resource_unbind(vma_res, NULL);
+
+ vma->resource = NULL;
+
atomic_and(~(I915_VMA_BIND_MASK | I915_VMA_ERROR | I915_VMA_GGTT_WRITE),
&vma->flags);
i915_vma_detach(vma);
+
+ if (!async) {
+ if (unbind_fence) {
+ dma_fence_wait(unbind_fence, false);
+ dma_fence_put(unbind_fence);
+ unbind_fence = NULL;
+ }
+ vma_invalidate_tlb(vma->vm, vma->obj->mm.tlb);
+ }
+
+ /*
+ * Binding itself may not have completed until the unbind fence signals,
+ * so don't drop the pages until that happens, unless the resource is
+ * async_capable.
+ */
+
vma_unbind_pages(vma);
+ return unbind_fence;
}
int __i915_vma_unbind(struct i915_vma *vma)
@@ -1706,6 +2127,7 @@ int __i915_vma_unbind(struct i915_vma *vma)
int ret;
lockdep_assert_held(&vma->vm->mutex);
+ assert_vma_held_evict(vma);
if (!drm_mm_node_allocated(&vma->node))
return 0;
@@ -1725,18 +2147,55 @@ int __i915_vma_unbind(struct i915_vma *vma)
return ret;
GEM_BUG_ON(i915_vma_is_active(vma));
- __i915_vma_evict(vma);
+ __i915_vma_evict(vma, false);
drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */
return 0;
}
+static struct dma_fence *__i915_vma_unbind_async(struct i915_vma *vma)
+{
+ struct dma_fence *fence;
+
+ lockdep_assert_held(&vma->vm->mutex);
+
+ if (!drm_mm_node_allocated(&vma->node))
+ return NULL;
+
+ if (i915_vma_is_pinned(vma) ||
+ &vma->obj->mm.rsgt->table != vma->resource->bi.pages)
+ return ERR_PTR(-EAGAIN);
+
+ /*
+ * We probably need to replace this with awaiting the fences of the
+ * object's dma_resv when the vma active goes away. When doing that
+ * we need to be careful to not add the vma_resource unbind fence
+ * immediately to the object's dma_resv, because then unbinding
+ * the next vma from the object, in case there are many, will
+ * actually await the unbinding of the previous vmas, which is
+ * undesirable.
+ */
+ if (i915_sw_fence_await_active(&vma->resource->chain, &vma->active,
+ I915_ACTIVE_AWAIT_EXCL |
+ I915_ACTIVE_AWAIT_ACTIVE) < 0) {
+ return ERR_PTR(-EBUSY);
+ }
+
+ fence = __i915_vma_evict(vma, true);
+
+ drm_mm_remove_node(&vma->node); /* pairs with i915_vma_release() */
+
+ return fence;
+}
+
int i915_vma_unbind(struct i915_vma *vma)
{
struct i915_address_space *vm = vma->vm;
- intel_wakeref_t wakeref = 0;
+ intel_wakeref_t wakeref = NULL;
int err;
+ assert_object_held_shared(vma->obj);
+
/* Optimistic wait before taking the mutex */
err = i915_vma_sync(vma);
if (err)
@@ -1767,6 +2226,79 @@ out_rpm:
return err;
}
+int i915_vma_unbind_async(struct i915_vma *vma, bool trylock_vm)
+{
+ struct drm_i915_gem_object *obj = vma->obj;
+ struct i915_address_space *vm = vma->vm;
+ intel_wakeref_t wakeref = NULL;
+ struct dma_fence *fence;
+ int err;
+
+ /*
+ * We need the dma-resv lock since we add the
+ * unbind fence to the dma-resv object.
+ */
+ assert_object_held(obj);
+
+ if (!drm_mm_node_allocated(&vma->node))
+ return 0;
+
+ if (i915_vma_is_pinned(vma)) {
+ vma_print_allocator(vma, "is pinned");
+ return -EAGAIN;
+ }
+
+ if (!obj->mm.rsgt)
+ return -EBUSY;
+
+ err = dma_resv_reserve_fences(obj->base.resv, 2);
+ if (err)
+ return -EBUSY;
+
+ /*
+ * It would be great if we could grab this wakeref from the
+ * async unbind work if needed, but we can't because it uses
+ * kmalloc and it's in the dma-fence signalling critical path.
+ */
+ if (i915_vma_is_bound(vma, I915_VMA_GLOBAL_BIND))
+ wakeref = intel_runtime_pm_get(&vm->i915->runtime_pm);
+
+ if (trylock_vm && !mutex_trylock(&vm->mutex)) {
+ err = -EBUSY;
+ goto out_rpm;
+ } else if (!trylock_vm) {
+ err = mutex_lock_interruptible_nested(&vm->mutex, !wakeref);
+ if (err)
+ goto out_rpm;
+ }
+
+ fence = __i915_vma_unbind_async(vma);
+ mutex_unlock(&vm->mutex);
+ if (IS_ERR_OR_NULL(fence)) {
+ err = PTR_ERR_OR_ZERO(fence);
+ goto out_rpm;
+ }
+
+ dma_resv_add_fence(obj->base.resv, fence, DMA_RESV_USAGE_READ);
+ dma_fence_put(fence);
+
+out_rpm:
+ if (wakeref)
+ intel_runtime_pm_put(&vm->i915->runtime_pm, wakeref);
+ return err;
+}
+
+int i915_vma_unbind_unlocked(struct i915_vma *vma)
+{
+ int err;
+
+ i915_gem_object_lock(vma->obj, NULL);
+ err = i915_vma_unbind(vma);
+ i915_gem_object_unlock(vma->obj);
+
+ return err;
+}
+
struct i915_vma *i915_vma_make_unshrinkable(struct i915_vma *vma)
{
i915_gem_object_make_unshrinkable(vma->obj);
generated by cgit (git 2.34.1) at 2025-12-26 03:51:41 +0000