On 1/5/2024 4:22 AM, Jacek Lawrynowicz wrote:
This was not supported properly. A buffer was imported to another VPU
context as a separate buffer object with duplicated sgt.
Both exported and imported buffers could be DMA mapped causing a double
mapping on the same device.
Imported buffer from another VPU context will now have just reference
increased and there will be a single sgt fixing above problem but
buffers still can't be shared among VPU contexts because each context
have its own MMU mapping and ivpu_bo supports only single MMU mapping.
The solution would be to use a mapping list as in panfrost or etnaviv
drivers and it will be implemented in future if required.
Signed-off-by: Jacek Lawrynowicz <jacek.lawrynowicz@xxxxxxxxxxxxxxx>
---
drivers/accel/ivpu/ivpu_gem.c | 44 +++++------------------------------
1 file changed, 6 insertions(+), 38 deletions(-)
diff --git a/drivers/accel/ivpu/ivpu_gem.c b/drivers/accel/ivpu/ivpu_gem.c
index 4de454bfbf91..8cb4d337552e 100644
--- a/drivers/accel/ivpu/ivpu_gem.c
+++ b/drivers/accel/ivpu/ivpu_gem.c
@@ -222,6 +222,12 @@ static int ivpu_bo_open(struct drm_gem_object *obj, struct drm_file *file)
struct ivpu_bo *bo = to_ivpu_bo(obj);
struct ivpu_addr_range *range;
+ if (bo->ctx) {
+ ivpu_warn(vdev, "Can't add BO (vpu_addr 0x%llx) to ctx %u: already in ctx %u\n",
+ bo->vpu_addr, file_priv->ctx.id, bo->ctx->id);
Looks like the vpu_addr is being used as a unique identifier for the BO.
Is that really the best value to use? Seems like if I want to attack
another context, knowing the device address of something that context
owns would be useful information.
