A lightweight WebAssembly runtime that is fast, secure, and standards-compliant
wasi-nn: check input tensor data length in pytorch and openvino (#14959)
netliomax25-codeffff2e2db24a
diff --git a/crates/wasi-nn/src/backend/openvino.rs b/crates/wasi-nn/src/backend/openvino.rsindex ef6fc2b38a..1b5b7c0bbd 100644--- a/crates/wasi-nn/src/backend/openvino.rs+++ b/crates/wasi-nn/src/backend/openvino.rs@@ -100,6 +100,13 @@ impl BackendExecutionContext for OpenvinoExecutionContext {let shape = Shape::new(&dimensions)?;let mut new_tensor = OvTensor::new(precision, &shape)?;let buffer = new_tensor.get_raw_data_mut()?;+ if buffer.len() != tensor.data.len() {+ return Err(BackendError::BackendAccess(wasmtime::format_err!(+ "input tensor has {} bytes of data but its dimensions and type need {} bytes",+ tensor.data.len(),+ buffer.len()+ )));+ }buffer.copy_from_slice(&tensor.data);// Assign the tensor to the request.match id {@@ -128,6 +135,13 @@ impl BackendExecutionContext for OpenvinoExecutionContext {let shape = Shape::new(&dimensions)?;let mut new_tensor = OvTensor::new(precision, &shape)?;let buffer = new_tensor.get_raw_data_mut()?;+ if buffer.len() != input.tensor.data.len() {+ return Err(BackendError::BackendAccess(wasmtime::format_err!(+ "input tensor has {} bytes of data but its dimensions and type need {} bytes",+ input.tensor.data.len(),+ buffer.len()+ )));+ }buffer.copy_from_slice(&input.tensor.data);self.0.set_tensor(&input.name, &new_tensor)?;diff --git a/crates/wasi-nn/src/backend/pytorch.rs b/crates/wasi-nn/src/backend/pytorch.rsindex 1cdf2e1b8c..90f4dc7925 100644--- a/crates/wasi-nn/src/backend/pytorch.rs+++ b/crates/wasi-nn/src/backend/pytorch.rs@@ -105,14 +105,7 @@ struct PytorchExecutionContext {/// See [PR #77](https://github.com/WebAssembly/wasi-nn/pull/77), at which point multiple named parameters for `Tensor` inputs is planned to be supported in pytorch backend.impl BackendExecutionContext for PytorchExecutionContext {fn set_input(&mut self, id: Id, input_tensor: &Tensor) -> Result<(), BackendError> {- let kind = input_tensor.ty.try_into()?;- let dimensions = input_tensor- .dimensions- .iter()- .map(|&dim| dim as i64)- .collect::<Vec<_>>();- let tensor = TchTensor::from_data_size(&input_tensor.data, &dimensions, kind)- .to_device(map_execution_target_to_string(self.target));+ let tensor = to_tch_tensor(input_tensor, self.target)?;match id {Id::Index(i) => {// Check if id_type is already set and if it matches the current id type@@ -165,16 +158,7 @@ impl BackendExecutionContext for PytorchExecutionContext {self.inputs.clear();self.id_type = None;for input in &inputs {- let kind = input.tensor.ty.try_into()?;- let dimensions = input- .tensor- .dimensions- .iter()- .map(|&dim| dim as i64)- .collect::<Vec<_>>();-- let tensor = TchTensor::from_data_size(&input.tensor.data, &dimensions, kind)- .to_device(map_execution_target_to_string(self.target));+ let tensor = to_tch_tensor(&input.tensor, self.target)?;self.inputs.push(Some(tensor));// Set ID type to Name since we're using named tensors@@ -250,6 +234,42 @@ impl BackendExecutionContext for PytorchExecutionContext {}}+/// Converts a wasi-nn [`Tensor`] into a PyTorch tensor on `target`.+///+/// `TchTensor::from_data_size` copies as many bytes as the dimensions and+/// element type call for without looking at the length of the data, so check+/// that the guest-provided data is exactly that size first.+fn to_tch_tensor(tensor: &Tensor, target: ExecutionTarget) -> Result<TchTensor, BackendError> {+ let kind: Kind = tensor.ty.try_into()?;+ let expected = tensor+ .dimensions+ .iter()+ .try_fold(kind.elt_size_in_bytes(), |count, &dim| {+ count.checked_mul(usize::try_from(dim).ok()?)+ })+ .ok_or_else(|| {+ BackendError::BackendAccess(wasmtime::format_err!(+ "tensor dimensions are too large: {:?}",+ tensor.dimensions+ ))+ })?;+ if tensor.data.len() != expected {+ return Err(BackendError::BackendAccess(wasmtime::format_err!(+ "input tensor has {} bytes of data but dimensions {:?} of type {:?} need {expected} bytes",+ tensor.data.len(),+ tensor.dimensions,+ tensor.ty+ )));+ }+ let dimensions = tensor+ .dimensions+ .iter()+ .map(|&dim| dim as i64)+ .collect::<Vec<_>>();+ Ok(TchTensor::from_data_size(&tensor.data, &dimensions, kind)+ .to_device(map_execution_target_to_string(target)))+}+fn map_execution_target_to_string(target: ExecutionTarget) -> Device {match target {ExecutionTarget::Cpu => Device::Cpu,@@ -313,3 +333,23 @@ impl From<TchError> for BackendError {BackendError::BackendAccess(wasmtime::Error::new(e))}}++#[cfg(test)]+mod tests {+ use super::*;++ #[test]+ fn to_tch_tensor_checks_data_length() {+ let exact = Tensor::new(vec![2, 3], TensorType::Fp32, vec![0; 24]);+ assert!(to_tch_tensor(&exact, ExecutionTarget::Cpu).is_ok());++ let short = Tensor::new(vec![4096], TensorType::Fp32, vec![0; 4]);+ assert!(to_tch_tensor(&short, ExecutionTarget::Cpu).is_err());++ let long = Tensor::new(vec![2], TensorType::Fp16, vec![0; 8]);+ assert!(to_tch_tensor(&long, ExecutionTarget::Cpu).is_err());++ let huge = Tensor::new(vec![u32::MAX; 4], TensorType::I64, vec![]);+ assert!(to_tch_tensor(&huge, ExecutionTarget::Cpu).is_err());+ }+}