This is documentation for the development version of the project, aka master branch. If you installed Gramine from packages, documentation for the stable version is available at

Introduction to SGX

The Gramine project uses the Intel SGX (Software Guard Extensions) technology to protect software running on untrusted hosts. SGX is a complicated topic, which may be hard to learn, because the documentation is scattered through official/reference documentation, blogposts and academic papers. This page is an attempt to curate a dossier of available reading material.

SGX is an umbrella name of technology that comprises several parts:

  • CPU/platform hardware features: the instruction set, microarchitecture with the PRM memory region and some new MSRs, some new logic in the MMU and so on;

  • the SGX Remote Attestation infrastructure, online services provided by Intel and/or third parties (see DCAP);

  • SDK and assorted software.

SGX is still being developed. The current (March 2024) version of CPU features is referred to as “SGX2” or simply “SGX”. The older instruction set from the original SGX is informally referred to as “SGX1”.

Features which might be considered part of SGX2:

  • EDMM (Enclave Dynamic Memory Management)

  • FLC (Flexible Launch Control; not strictly part of SGX2, but was not part of original SGX hardware either)

  • KSS (Key Separation and Sharing; also not part of SGX2, but was not part of original SGX hardware either)

Around 2022 Intel discontinued SGX support in client CPU cores, and instead introduced it to server cores. The new SGX hardware architecture didn’t change the user-facing ABI, but loosened security guarantees, matching AMD SEV-SNP security model:

  • Merkle tree for memory integrity checking was removed.

  • Hardware RAM MitM attacks are not mitigated anymore: (because of Merkle tree removal)

    • On Icelake server CPUs there’s no integrity protection at all.

    • On Sapphire Rapids server CPUs there’s a 28-bit MAC per each cacheline. It’s possible to bruteforce the MAC or do a replay attack with cacheline granularity (but that still requires a hardware MitM).

  • EPC can now be almost arbitrarily big, significantly improving performance for large workloads.

As of now most of the broadly used server CPUs support SGX2. Only older client CPUs support SGX, so they should not be used in production (because of missing security patches for side-channels).

Introductory reading


Most of the older literature available (especially introduction-level) concerns the original SGX1 only.

Official documentation

Academic research

Installation instructions

See Set up the host environment.

Linux kernel drivers

For historical reasons, there are three SGX drivers currently (March 2024):

SGX terminology

Architectural Enclaves

Architectural Enclaves (AEs) are a set of “system” enclaves concerned with starting and attesting other enclaves. Intel provides reference implementations of these enclaves, though other companies may write their own implementations.

Architectural Enclave Service Manager

The Architectural Enclave Service Manager is responsible for providing SGX applications with access to the Architectural Enclaves. It consists of the Architectural Enclave Service Manager Daemon, which hosts the enclaves, and a component of the SGX SDK, which communicates with the daemon over a Unix socket with the fixed path /var/run/aesmd/aesm.sock.

Asynchronous Enclave Exit

An event caused by an exception occurring during in-enclave execution. CPU saves the current context into SSA, leaves SGX mode and jumps to AEP.

Asynchronous Exit Pointer

An address outside the enclave where CPU will jump in case an exception happens during in-enclave execution.


Attestation is a mechanism to prove the trustworthiness of the SGX enclave to a local or remote party. More specifically, SGX attestation proves that the enclave runs on a real hardware in an up-to-date TEE with the expected initial state. There are two types of the attestation: Local Attestation (between enclaves on the same machines) and Remote Attestation (between enclave and any party, possibly remote).

Data Center Attestation Primitives

A software infrastructure provided by Intel as a reference implementation for the new ECDSA/PCS-based remote attestation. Relies on the Flexible Launch Control hardware feature.

This allows for launching enclaves with Intel’s remote infrastructure only involved in the initial setup. Naturally however, this requires deployment of own infrastructure, so is operationally more complicated. Therefore it is intended for server environments (where you control all the machines).

See also

Orientation Guide


A way to launch enclaves with Intel’s infrastructure, intended for client machines.


A special function call made by non-enclave world into an enclave.


An instance of SGX TEE, residing in a contiguous chunk of usermode address space (ELRANGE) of some process on the system. Application threads may enter and exit the enclave through dedicated CPU instructions. Code running inside an enclave has access to usermode memory of the process which contains it, but not the other way.

Enclave Dynamic Memory Management

A hardware feature of SGX2, allows for dynamic (in enclave runtime) addition and removal of enclave threads and memory, as well as changing memory permissions and type.

Enclave Page Cache

A part of PRM used for caching enclave pages. EPC is only an optimization and its size doesn’t limit possible enclave sizes, though too-small EPC may lead to frequent page swapping and significantly worsen performance.

Enclave Page Cache Map

A part of PRM which holds metadata about EPC pages.

Enhanced Privacy Identification
Enhanced Privacy Identifier

EPID is the attestation protocol originally shipped with SGX. Unlike DCAP, a remote verifier making use of the EPID protocol needs to contact the Intel Attestation Service each time it wishes to attest an enclave.

Contrary to DCAP, EPID may be understood as “opinionated”, with most moving parts fixed and tied to services provided by Intel. This is intended for client enclaves and deprecated for server environments.

EPID attestation can operate in two modes: fully-anonymous (unlinkable) quotes and pseudonymous (linkable) quotes. Unlike fully-anonymous quotes, pseudonymous quotes include an identifier dependent on the identity of the CPU and the developer of the enclave being quoted, which allows determining whether two instances of your enclave are running on the same CPU or not.

If your security model depends on enforcing that the identifiers are different (e.g. because you want to prevent sybil attacks), keep in mind that the enclave host can generate a new identity by performing an epoch reset. The previous identity will then become inaccessible, though.

The attestation mode being used can be chosen by the application enclave, but it must match what was chosen when generating the SPID.

See also


A way to launch enclaves without relying on the Intel’s infrastructure (after initial setup).


An identifier one can obtain from Intel, required to make use of EPID attestation.

Flexible Launch Control

Hardware (CPU) feature that allows substituting Launch Enclave for one not signed by Intel through a change in SGX’s EINIT logic to not require the EINITTOKEN from the Intel-based Launch Enclave. An MSR, which can be locked at boot time, keeps the hash of the public key of the “launching” entity.

With FLC, Launch Enclave can be written by other companies (other than Intel) and must be signed with the key corresponding to the one locked in the MSR (a reference Launch Enclave simply allows all enclaves to run). The MSR can also stay unlocked and then it can be modified at run-time by the VMM or the OS kernel.

Support for FLC can be detected using CPUID instruction, as CPUID.07H:ECX.SGX_LC[bit 30] == 1 (SDM vol. 2A calls this “SGX Launch Control”).

Key Separation and Sharing

A feature that lets developer define additional enclave identity attributes and configuration identifier. Extended enclave identity is defined by the developer on enclave build. Enclave configuration is defined on enclave launch and cannot be modified afterwards.

In addition to the calculated enclave and signer measurements, developer is expected to define a product ID and SVN for her enclaves. These identifiers are part of the SGX Report and are expected to be used in Attestation. They are also used by SGX key derivation to derive different keys per configuration.

KSS adds two more attributes for enclave build and two new ones for enclave launch, which are part of the SGX Report. Additionally, key policy attributes are extended to provide fine-grained control over key derivation.

New build attributes:

  • Extended product ID

  • Family ID

New enclave launch attributes:

  • Config ID

  • Config SVN

This feature was not part of original SGX and therefore is not supported by all SGX-enabled hardware.

Launch Enclave



Local Attestation

In local attestation, the attesting SGX enclave collects attestation evidence in the form of an SGX Report using the EREPORT hardware instruction. This form of attestation is used to send the attestation evidence to a local party (on the same physical machine).

Intel Attestation Service

Internet service provided by Intel for “old” EPID-based remote attestation. Enclaves send SGX quotes to the client/verifier who will forward them to IAS to check their validity.

See also


Provisioning Certification Service, another Internet service provided by Intel.

Memory Encryption Engine




A special function call made by an enclave to the non-enclave world.

SGX Platform Software

Software infrastructure provided by Intel with all special Architectural Enclaves (Provisioning Enclave, Quoting Enclave, Launch Enclave). This mainly refers to the “old” EPID/IAS-based remote attestation.

Processor Reserved Memory

A mostly undocumented region of physical address space reserved by the BIOS for internal use by SGX hardware. Known to contain at least EPC and EPCM.

Provisioning Enclave

One of the Architectural Enclaves of the Intel SGX software infrastructure. It is part of the SGX Platform Software. The Provisioning Enclave is used in EPID based remote attestation. This enclave communicates with the Intel Provisioning Service (IPS) to perform EPID provisioning. The result of this provisioning procedure is the private EPID key securely accessed by the Provisioning Enclave. This procedure happens only during the first deployment of the SGX machine (or, in rare cases, to provision a new EPID key after TCB upgrade). The main user of the Provisioning Enclave is the Quoting Enclave.

Provisioning Certification Enclave

One of the Architectural Enclaves of the Intel SGX software infrastructure. It is part of the SGX Platform Software and DCAP. The Provisioning Certification Enclave is used in DCAP based remote attestation. This enclave communicates with the Intel Provisioning Certification Service (PCS) to perform DCAP provisioning. The result of this provisioning procedure is the DCAP/ECDSA attestation collateral (mainly the X.509 certificate chains rooted in a well-known Intel certificate and Certificate Revocation Lists). This procedure happens during the first deployment of the SGX machine and then periodically to refresh the cached attestation collateral. Typically, to reduce the dependency on PCS, a cloud service provider introduces an intermediate caching service (Provisioning Certification Caching Service, or PCCS) that stores all the attestation collateral obtained from Intel. The main user of the Provisioning Certification Enclave is the Quoting Enclave.

Intel Provisioning Service

Internet service provided by Intel for EPID-based remote attestation. This service provides the corresponding EPID key to the Provisioning Enclave on a remote SGX machine.

Intel Provisioning Certification Service

New internet service provided by Intel for new ECDSA-based remote attestation. Enclave provider creates its own internal Attestation Service where it caches PKI collateral from Intel’s PCS, and the verifier gets the certificate chain from the enclave provider to check validity.

See also


Intel Attestation Service, another Internet service.

Quoting Enclave

One of the Architectural Enclaves of the Intel SGX software infrastructure. It is part of the SGX Platform Software. The Quoting Enclave receives an SGX Report and produces a corresponding SGX Quote. The identity of the Quoting Enclave is publicly known (it signer, its measurement and its attributes) and is vetted by public companies such as Intel (in the form of the certificate chain ending in a publicly known root certificate of the company).

Remote Attestation

For remote attestation, the attesting SGX enclave collects attestation evidence in the form of an SGX Quote using the Quoting Enclave (and the Provisioning Enclave if required). The enclave then may send the collected attestation evidence to the local or remote party, which will verify the evidence and confirm the authenticity and integrity of the attested enclave. After this, the local or remote party trusts the enclave and may establish a secure channel with the enclave and send secrets to it.

Intel SGX Software Development Kit

In the context of SGX, this means a specific piece of software supplied by Intel which helps people write enclaves packed into .so files to be accessible like normal libraries (at least on Linux). Available together with a kernel module and documentation.

SGX Enclave Control Structure



SGX Quote

The SGX quote is the proof of trustworthiness of the enclave and is used during Remote Attestation. The attesting enclave generates the enclave-specific SGX Report, sends the request to the Quoting Enclave using Local Attestation, and the Quoting Enclave returns back the SGX quote with the SGX report embedded in it. The resulting SGX quote contains the enclave’s measurement, attributes and other security-relevant fields, and is tied to the identity of the Quoting Enclave to prove its authenticity. The obtained SGX quote may be later sent to the verifying remote party, which examines the SGX quote and gains trust in the remote enclave.

SGX Report

The SGX report is a data structure that contains the enclave’s measurement, signer identity, attributes and a user-defined 64B string. The SGX report is generated using the EREPORT hardware instruction. It is used during Local Attestation. The SGX report is embedded into the SGX Quote.


The original SGX instruction set, without dynamic resource management.


New SGX instructions and other hardware features that were introduced after the release of the original SGX1 (e.g. EDMM).

Service Provider ID

An identifier provided by Intel, used together with an EPID API key to authenticate to the Intel Attestation Service. You can obtain an SPID through Intel’s Trusted Services Portal.

See EPID for a description of the difference between linkable and unlinkable quotes.

State Save Area



Security Version Number

Each element of the SGX TCB is assigned a Security Version Number (SVN). For the hardware, these SVNs are referred to collectively as CPU_SVN, and for software referred as ISV_SVN. A TCB is considered up to date if all components of the TCB have SVNs greater than or equal to a threshold published by the author of the component.

Trusted Execution Environment

A Trusted Execution Environment (TEE) is an environment where the code executed and the data accessed are isolated and protected in terms of confidentiality (no one has access to the data except the code running inside the TEE) and integrity (no one can change the code and its behavior).

Trusted Computing Base

In context of SGX this has the usual meaning: the set of all components that are critical to security. Any vulnerability in TCB compromises security. Any problem outside TCB is not a vulnerability, i.e. should not compromise security.

In context of Gramine there is also a different meaning (Thread Control Block). Those two should not be confused.

Thread Control Structure