accel/habanalabs: enforce release order of compute device and dma-buf

When user closes the compute device file descriptor without closing a
dma-buf file descriptor, the device will be considered as in use,
leading to hard reset and killing the user process, to ensure the
release of the dma-buf.
Same thing will happen if user first releases the compute device file
and only then the dma-buf.

The implication of this is the duration of hard reset, during which the
device cannot be reacquired.
Moreover, this behavior adds a constraint on a user process to follow
this order of release operations.

To avoid killing the user process and to remove this constraint, enforce
the correct order of release operations inside the driver, by
incrementing the device file refcount for any dma-buf until it is
released.

Signed-off-by: Tomer Tayar <ttayar@habana.ai>
Reviewed-by: Oded Gabbay <ogabbay@kernel.org>
Signed-off-by: Oded Gabbay <ogabbay@kernel.org>
Reviewed-by: Stanislaw Gruszka <stanislaw.gruszka@linux.intel.com>

diff --git a/drivers/accel/habanalabs/common/memory.c b/drivers/accel/habanalabs/common/memory.c
index fd333dd2835eb50a7ff962842538019a9e60b578..7fd8e8f876827e3bad979493ad1d1b6908117d0b 100644 (file)
--- a/drivers/accel/habanalabs/common/memory.c
+++ b/drivers/accel/habanalabs/common/memory.c
@@ -1835,6 +1835,10 @@ static void hl_release_dmabuf(struct dma_buf *dmabuf)
 
        atomic_dec(&ctx->hdev->dmabuf_export_cnt);
        hl_ctx_put(ctx);
+
+       /* Paired with get_file() in export_dmabuf() */
+       fput(ctx->hpriv->filp);
+
        kfree(hl_dmabuf);
 }
 
@@ -1875,6 +1879,12 @@ static int export_dmabuf(struct hl_ctx *ctx,
        hl_ctx_get(hl_dmabuf->ctx);
        atomic_inc(&ctx->hdev->dmabuf_export_cnt);
 
+       /* Get compute device file to enforce release order, such that all exported dma-buf will be
+        * released first and only then the compute device.
+        * Paired with fput() in hl_release_dmabuf().
+        */
+       get_file(ctx->hpriv->filp);
+
        *dmabuf_fd = fd;
 
        return 0;