1.. SPDX-License-Identifier: GPL-2.0 2 3============== 4Nitro Enclaves 5============== 6 7Overview 8======== 9 10Nitro Enclaves (NE) is a new Amazon Elastic Compute Cloud (EC2) capability 11that allows customers to carve out isolated compute environments within EC2 12instances [1]. 13 14For example, an application that processes sensitive data and runs in a VM, 15can be separated from other applications running in the same VM. This 16application then runs in a separate VM than the primary VM, namely an enclave. 17It runs alongside the VM that spawned it. This setup matches low latency 18applications needs. 19 20The current supported architectures for the NE kernel driver, available in the 21upstream Linux kernel, are x86 and ARM64. 22 23The resources that are allocated for the enclave, such as memory and CPUs, are 24carved out of the primary VM. Each enclave is mapped to a process running in the 25primary VM, that communicates with the NE kernel driver via an ioctl interface. 26 27In this sense, there are two components: 28 291. An enclave abstraction process - a user space process running in the primary 30VM guest that uses the provided ioctl interface of the NE driver to spawn an 31enclave VM (that's 2 below). 32 33There is a NE emulated PCI device exposed to the primary VM. The driver for this 34new PCI device is included in the NE driver. 35 36The ioctl logic is mapped to PCI device commands e.g. the NE_START_ENCLAVE ioctl 37maps to an enclave start PCI command. The PCI device commands are then 38translated into actions taken on the hypervisor side; that's the Nitro 39hypervisor running on the host where the primary VM is running. The Nitro 40hypervisor is based on core KVM technology. 41 422. The enclave itself - a VM running on the same host as the primary VM that 43spawned it. Memory and CPUs are carved out of the primary VM and are dedicated 44for the enclave VM. An enclave does not have persistent storage attached. 45 46The memory regions carved out of the primary VM and given to an enclave need to 47be aligned 2 MiB / 1 GiB physically contiguous memory regions (or multiple of 48this size e.g. 8 MiB). The memory can be allocated e.g. by using hugetlbfs from 49user space [2][3][7]. The memory size for an enclave needs to be at least 5064 MiB. The enclave memory and CPUs need to be from the same NUMA node. 51 52An enclave runs on dedicated cores. CPU 0 and its CPU siblings need to remain 53available for the primary VM. A CPU pool has to be set for NE purposes by an 54user with admin capability. See the cpu list section from the kernel 55documentation [4] for how a CPU pool format looks. 56 57An enclave communicates with the primary VM via a local communication channel, 58using virtio-vsock [5]. The primary VM has virtio-pci vsock emulated device, 59while the enclave VM has a virtio-mmio vsock emulated device. The vsock device 60uses eventfd for signaling. The enclave VM sees the usual interfaces - local 61APIC and IOAPIC - to get interrupts from virtio-vsock device. The virtio-mmio 62device is placed in memory below the typical 4 GiB. 63 64The application that runs in the enclave needs to be packaged in an enclave 65image together with the OS ( e.g. kernel, ramdisk, init ) that will run in the 66enclave VM. The enclave VM has its own kernel and follows the standard Linux 67boot protocol [6][8]. 68 69The kernel bzImage, the kernel command line, the ramdisk(s) are part of the 70Enclave Image Format (EIF); plus an EIF header including metadata such as magic 71number, eif version, image size and CRC. 72 73Hash values are computed for the entire enclave image (EIF), the kernel and 74ramdisk(s). That's used, for example, to check that the enclave image that is 75loaded in the enclave VM is the one that was intended to be run. 76 77These crypto measurements are included in a signed attestation document 78generated by the Nitro Hypervisor and further used to prove the identity of the 79enclave; KMS is an example of service that NE is integrated with and that checks 80the attestation doc. 81 82The enclave image (EIF) is loaded in the enclave memory at offset 8 MiB. The 83init process in the enclave connects to the vsock CID of the primary VM and a 84predefined port - 9000 - to send a heartbeat value - 0xb7. This mechanism is 85used to check in the primary VM that the enclave has booted. The CID of the 86primary VM is 3. 87 88If the enclave VM crashes or gracefully exits, an interrupt event is received by 89the NE driver. This event is sent further to the user space enclave process 90running in the primary VM via a poll notification mechanism. Then the user space 91enclave process can exit. 92 93[1] https://aws.amazon.com/ec2/nitro/nitro-enclaves/ 94[2] https://www.kernel.org/doc/html/latest/admin-guide/mm/hugetlbpage.html 95[3] https://lwn.net/Articles/807108/ 96[4] https://www.kernel.org/doc/html/latest/admin-guide/kernel-parameters.html 97[5] https://man7.org/linux/man-pages/man7/vsock.7.html 98[6] https://www.kernel.org/doc/html/latest/x86/boot.html 99[7] https://www.kernel.org/doc/html/latest/arm64/hugetlbpage.html 100[8] https://www.kernel.org/doc/html/latest/arm64/booting.html 101