Re: [Intel-gfx] [PATCH v7 20/20] drm/i915/vm_bind: Async vm_unbind support

[Niranjana Vishwanathapura]

On Tue, Nov 15, 2022 at 04:58:42PM +0100, Andi Shyti wrote:

Hi Niranjana,

On Sat, Nov 12, 2022 at 11:57:32PM -0800, Niranjana Vishwanathapura wrote:

Asynchronously unbind the vma upon vm_unbind call.
Fall back to synchronous unbind if backend doesn't support
async unbind or if async unbind fails.

No need for vm_unbind out fence support as i915 will internally
handle all sequencing and user need not try to sequence any
operation with the unbind completion.

v2: use i915_vma_destroy_async in vm_unbind ioctl

Signed-off-by: Niranjana Vishwanathapura <niranjana.vishwanathap...@intel.com>
---
 .../drm/i915/gem/i915_gem_vm_bind_object.c    |  2 +-
 drivers/gpu/drm/i915/i915_vma.c               | 51 +++++++++++++++++--
 drivers/gpu/drm/i915/i915_vma.h               |  1 +
 include/uapi/drm/i915_drm.h                   |  3 +-
 4 files changed, 51 insertions(+), 6 deletions(-)

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_vm_bind_object.c 
b/drivers/gpu/drm/i915/gem/i915_gem_vm_bind_object.c
index d87d1210365b..36651b447966 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_vm_bind_object.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_vm_bind_object.c
@@ -210,7 +210,7 @@ static int i915_gem_vm_unbind_vma(struct i915_address_space 
*vm,
         */
        obj = vma->obj;
        i915_gem_object_lock(obj, NULL);
-       i915_vma_destroy(vma);
+       i915_vma_destroy_async(vma);
        i915_gem_object_unlock(obj);

        i915_gem_object_put(obj);
diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c
index 7cf77c67d755..483d25f2425c 100644
--- a/drivers/gpu/drm/i915/i915_vma.c
+++ b/drivers/gpu/drm/i915/i915_vma.c
@@ -42,6 +42,8 @@
 #include "i915_vma.h"
 #include "i915_vma_resource.h"

+static struct dma_fence *__i915_vma_unbind_async(struct i915_vma *vma);
+
 static inline void assert_vma_held_evict(const struct i915_vma *vma)
 {
        /*
@@ -1713,7 +1715,7 @@ void i915_vma_reopen(struct i915_vma *vma)
        spin_unlock_irq(&gt->closed_lock);
 }

-static void force_unbind(struct i915_vma *vma)
+static void force_unbind(struct i915_vma *vma, bool async)

I still like the defines on this, they look cleaner, but it's a
matter of taste.

Ok, will change.

Niranjana

Reviewed-by: Andi Shyti <andi.sh...@linux.intel.com>

Andi