Hi,
We're moving towards a failsafe migration on DG1, but on DG2+ we'd have
to wedge the GPU when an error occurs during initial clearing or swapin.
But CPU maps don't care whether the GPU is wedged so we need to check
the error status both below and in the TTM fault handler, I guess.
On 8/30/21 2:10 PM, Maarten Lankhorst wrote:
For now, we will only allow async migration when TTM is used,
so the paths we care about are related to TTM.
The mmap path is handled by having the fence in ttm_bo->moving,
when pinning, the binding only becomes available after the moving
fence is signaled, and pinning a cpu map will only work after
the moving fence signals.
This should close all holes where userspace can read a buffer
before it's fully migrated.
Signed-off-by: Maarten Lankhorst <maarten.lankhorst@xxxxxxxxxxxxxxx>
---
drivers/gpu/drm/i915/gem/i915_gem_pages.c | 10 ++++++++++
drivers/gpu/drm/i915/i915_vma.c | 10 ++++++++--
2 files changed, 18 insertions(+), 2 deletions(-)
diff --git a/drivers/gpu/drm/i915/gem/i915_gem_pages.c b/drivers/gpu/drm/i915/gem/i915_gem_pages.c
index 8eb1c3a6fc9c..a044d1829729 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_pages.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_pages.c
@@ -345,6 +345,7 @@ void *i915_gem_object_pin_map(struct drm_i915_gem_object *obj,
enum i915_map_type type)
{
enum i915_map_type has_type;
+ struct dma_fence *moving;
bool pinned;
void *ptr;
int err;
@@ -355,6 +356,15 @@ void *i915_gem_object_pin_map(struct drm_i915_gem_object *obj,
assert_object_held(obj);
+ moving = i915_gem_object_get_moving_fence(obj);
+ if (moving) {
+ err = dma_fence_wait(moving, true);
+ dma_fence_put(moving);
+
+ if (err)
+ return ERR_PTR(err);
+ }
+
Error check on the moving fence?
pinned = !(type & I915_MAP_OVERRIDE);
type &= ~I915_MAP_OVERRIDE;
diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c
index bfba299d905c..320678e9b4e1 100644
--- a/drivers/gpu/drm/i915/i915_vma.c
+++ b/drivers/gpu/drm/i915/i915_vma.c
@@ -398,7 +398,7 @@ int i915_vma_bind(struct i915_vma *vma,
GEM_BUG_ON(!vma->pages);
trace_i915_vma_bind(vma, bind_flags);
- if (work && bind_flags & vma->vm->bind_async_flags) {
+ if (work) {
struct dma_fence *prev;
work->vma = vma;
@@ -1145,6 +1145,7 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
u64 size, u64 alignment, u64 flags)
{
struct i915_vma_work *work = NULL;
+ struct dma_fence *moving = NULL;
intel_wakeref_t wakeref = 0;
unsigned int bound;
int err;
@@ -1168,7 +1169,8 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
if (flags & PIN_GLOBAL)
wakeref = intel_runtime_pm_get(&vma->vm->i915->runtime_pm);
- if (flags & vma->vm->bind_async_flags) {
+ moving = i915_gem_object_get_moving_fence(vma->obj);
+ if (flags & vma->vm->bind_async_flags || moving) {
/* lock VM */
err = i915_vm_lock_objects(vma->vm, ww);
if (err)
@@ -1182,6 +1184,8 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
work->vm = i915_vm_get(vma->vm);
+ dma_fence_work_chain(&work->base, moving);
+
We need to error check the async binding in the dma-fence-work release
callback, and set the VMA_ERROR flag if necessary, at least until we
have the error-safe / gpu wedge stuff. VMA_ERROR is racy, though.
/* Allocate enough page directories to used PTE */
if (vma->vm->allocate_va_range) {
err = i915_vm_alloc_pt_stash(vma->vm,
@@ -1289,6 +1293,8 @@ int i915_vma_pin_ww(struct i915_vma *vma, struct i915_gem_ww_ctx *ww,
err_rpm:
if (wakeref)
intel_runtime_pm_put(&vma->vm->i915->runtime_pm, wakeref);
+ if (moving)
+ dma_fence_put(moving);
i915_vma_put_pages(vma);
return err;
}
