Caution

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 https://gramine.readthedocs.io/en/stable/.

Manifest syntax

A manifest file is an application-specific configuration text file that specifies the environment and resources for running an application inside Gramine. A manifest file contains key-value pairs (as well as more complicated table and array objects) in the TOML syntax. For the details of the TOML syntax, see the official documentation.

A typical string entry looks like this:

[Key][.Key][.Key] = "[Value]"

A typical integer entry looks similar to the above but without double quotes:

[Key][.Key][.Key] = [Value]

Comments can be inlined in a manifest by starting them with a hash sign (# comment...).

There is also a preprocessor available: gramine-manifest, which renders manifests from Jinja templates.

Common syntax

Log level

loader.log_level = "[none|error|warning|debug|trace|all]"
(Default: "error")

loader.log_file = "[PATH]"

This configures Gramine’s debug log. The log_level option specifies what messages to enable (e.g. loader.log_level = "debug" will enable all messages of type error, warning and debug). By default, the messages are printed to the standard error. If log_file is specified, the messages will be appended to that file.

Gramine outputs log messages of the following types:

error: A serious error preventing Gramine from operating properly (for example, error initializing one of the components).
warning: A non-fatal issue. Might mean that application is requesting something unsupported or poorly emulated.
debug: Detailed information about Gramine’s operation and internals.
trace: More detailed information, such as all system calls requested by the application. Might contain a lot of noise.

Warning

Only error log level is suitable for production. Other levels may leak sensitive data.

Loader entrypoint

loader.entrypoint.uri = "[URI]"
(Default: "<path to libsysdb.so>")

loader.entrypoint.sha256 = "[HASH]"

loader.entrypoint.uri specifies the LibOS component that Gramine will load and run before loading the first executable of the user application. Note: currently, there is only one LibOS implementation: libsysdb.so, and there is no need to specify this option explicitly.

Additionally, the loader entrypoint is cryptographically hashed at build time. At startup, Gramine verifies that the entrypoint’s hash matches what is stored in the manifest in loader.entrypoint.sha256. The signer tool will automatically generate the hash of this file and add it to the SGX-specific manifest (.manifest.sgx). The manifest writer may also explicitly specify the hash; in this case, hashing of the file will be skipped by the signer tool and the value in loader.entrypoint.sha256 will be used instead.

Note that the loader (the PAL binary) loads the LibOS binary specified in loader.entrypoint and passes control to this binary. Next, the LibOS binary loads the actual executable (the user application) specified in libos.entrypoint. Also note that, in contrast to libos.entrypoint, the loader.entrypoint.uri option specifies a PAL URI (with the file: prefix).

LibOS Entrypoint

libos.entrypoint = "[PATH]"

This specifies the first executable of the user application which is to be started when spawning a Gramine instance from this manifest file. Needs to be a path inside Gramine pointing to a mounted file. Relative paths will be interpreted as starting from the current working directory (i.e. from / by default, or fs.start_dir if specified).

The recommended usage is to provide an absolute path of the executable. This executable must also be in the Gramine’s filesystem, either directly mounted or within a directory that is mounted. For example, if one wishes to execute the Python 3.8 interpreter, one would specify it as the entrypoint and mount the Python executable at the expected path within the Gramine filesystem:

libos.entrypoint = "/usr/bin/python3.8"

fs.mounts = [
  { path = "/usr/bin/python3.8", uri = "file:/usr/bin/python3.8" },
  # Or, if using a binary from your local directory:
  # { path = "/usr/bin/python3.8", uri = "file:python3.8" },
]

Command-line arguments

loader.insecure__use_cmdline_argv = true

or

loader.argv = ["arg0", "arg1", "arg2", ...]

or

loader.argv_src_file = "file:file_with_serialized_argv"

If you want your application to use commandline arguments, you must choose one of the three mutually exclusive options:

set loader.insecure__use_cmdline_argv (insecure in almost all cases),
put commandline arguments into loader.argv array (note that the first argument is typically the program name and the actual arguments start with the second array item; see this link for an explanation),
point loader.argv_src_file to a file containing output of gramine-argv-serializer.

If none of the above arguments-handling manifest options is specified in the manifest, the application will get argv = [ <libos.entrypoint value> ].

loader.argv_src_file is intended to point to either a trusted file or an encrypted file. The former allows to securely hardcode arguments, the latter allows the arguments to be provided at runtime from an external (trusted) source.

Note

Pointing to an encrypted file is currently not supported, due to the fact that encryption key provisioning currently happens after setting up arguments.

Domain names configuration

sys.enable_extra_runtime_domain_names_conf = [true|false]
(Default: false)

This option will generate the following extra configuration:

Hostname (obtained by apps via nodename field in uname syscall), set to the host’s hostname at initialization.
Pseudo-file /etc/resolv.conf, with keywords:
- nameserver
- search
- options [edns0] [inet6] [rotate] [use-vc]
Unsupported keywords and malformed lines from /etc/resolv.conf are ignored.

The functionality is achieved by taking the host’s configuration via various APIs and reading the host’s configuration files. In the case of Linux PAL, most information comes from the host’s /etc. The gathered information is used to create /etc files inside Gramine’s filesystem, or change Gramine process configuration. For security-enforcing modes (such as SGX), Gramine additionally sanitizes the information gathered from the host. Invalid host’s configuration is reported as an error (e.g. invalid hostname, or invalid IPv4 address in nameserver keyword).

Note that Gramine supports only a subset of the configuration. Refer to the list of supported keywords.

This option takes precedence over fs.mounts. This means that etc files provided via fs.mounts will be overridden with the ones added via this option.

Environment variables

loader.insecure__use_host_env = [true|false]

By default, environment variables from the host will not be passed to the app. This can be overridden by the option above, but most applications and runtime libraries trust their environment variables and are completely insecure when these are attacker-controlled. For example, an attacker can execute an additional dynamic library by specifying LD_PRELOAD variable.

To securely set up the execution environment for an app you should use one or both of the following options:

loader.env.[ENVIRON] = "[VALUE]"
or
loader.env.[ENVIRON] = { value = "[VALUE]" }
or
loader.env.[ENVIRON] = { passthrough = true }

loader.env_src_file = "file:file_with_serialized_envs"

loader.env.[ENVIRON] adds/overwrites/passes a single environment variable and can be used multiple times to specify more than one variable. To add/overwrite the environment variable, specify a TOML string ("[VALUE]") or a TOML table with the key-value pair { value = "[VALUE]" }. To pass the environment variable from the host, specify a TOML table with the key-value pair { passthrough = true }. If you specify a variable, it needs to either have a value or be a passthrough.

loader.env_src_file allows to specify a URI to a file containing serialized environment, which can be generated using gramine-argv-serializer. This option is intended to point to either a trusted file or an encrypted file. The former allows to securely hardcode environments (in a more flexible way than loader.env.[ENVIRON] option), the latter allows the environments to be provided at runtime from an external (trusted) source.

Note

Pointing to an encrypted file is currently not supported, due to the fact that encryption key provisioning currently happens after setting up environment variables.

If the same variable is set in both, then loader.env.[ENVIRON] takes precedence. It is prohibited to specify both value and passthrough keys for the same environment variable. If manifest option insecure__use_host_env is specified, then passthrough = true manifest options have no effect (they are “consumed” by insecure__use_host_env).

Note

It is tempting to try to passthrough all environment variables using insecure__use_host_env and then disallow some of them using passthrough = false. However, this deny list approach is intentionally prohibited because it’s inherently insecure (doesn’t provide any real security). Gramine loudly fails if passthrough = false manifest options are set.

User ID and Group ID

loader.uid = [NUM]
loader.gid = [NUM]
(Default: 0)

This specifies the initial, Gramine emulated user/group ID and effective user/group ID. It must be non-negative. By default, Gramine emulates the user/group ID and effective user/group ID as the root user (uid = gid = 0).

Disabling ASLR

loader.insecure__disable_aslr = [true|false]
(Default: false)

This specifies whether to disable Address Space Layout Randomization (ASLR). Since disabling ASLR worsens security of the application, ASLR is enabled by default.

Check invalid pointers

libos.check_invalid_pointers = [true|false]
(Default: true)

This specifies whether to enable checks of invalid pointers on syscall invocations. In particular, when this manifest option is set to true, Gramine’s LibOS will return an EFAULT error code if a user-supplied buffer points to an invalid memory region. Setting this manifest option to false may improve performance for certain workloads but may also generate SIGSEGV/SIGBUS exceptions for some applications that specifically use invalid pointers (though this is not expected for most real-world applications).

Limit on open file descriptors

sys.fds.limit = [NUM]
(default: 900)

This specifies the maximum number of open file descriptors in the Gramine process. More specifically, this option sets the RLIMIT_NOFILE resource limit: it specifies a value one greater than the maximum file descriptor number that can be opened by the Gramine process.

Stack size

sys.stack.size = "[SIZE]"
(default: "256K")

This specifies the stack size of each thread in each Gramine process. The default value is determined by the library OS. Units like K (KiB), M (MiB), and G (GiB) can be appended to the values for convenience. For example, sys.stack.size = "1M" indicates a 1 MiB stack size.

Program break (brk) size

sys.brk.max_size = "[SIZE]"
(default: "256K")

This specifies the maximal program break (brk) size in each Gramine process. The default value of the program break size is determined by the library OS. Units like K (KiB), M (MiB), and G (GiB) can be appended to the values for convenience. For example, sys.brk.max_size = "1M" indicates a 1 MiB brk size.

Allowing host-based insecure eventfd

sys.insecure__allow_eventfd = [true|false]
(Default: false)

By default, Gramine implements eventfd in a secure but restricted way: currently this secure implementation only works when eventfd usage is confined to a single process (note that application may still be multi-process and spawn child processes, but eventfds created in parent will be invalid in children).

However, sometimes it is acceptable for applications to use host-based insecure eventfd implementation. This implementation works without the above-mentioned restriction in multi-process applications. Use sys.insecure__allow_eventfd manifest syntax to switch to this insecure implementation.

Note

sys.insecure__allow_eventfd is pass-through and thus potentially insecure in e.g. SGX environments. It is the responsibility of the app developer to analyze the app usage of eventfd, with security implications in mind.

External SIGTERM injection

sys.enable_sigterm_injection = [true|false]
(Default: false)

This specifies whether to allow for a one-time injection of SIGTERM signal into Gramine. Could be useful to handle graceful shutdown. Be careful! In SGX environment, the untrusted host could inject that signal in an arbitrary moment. Examine what your application’s SIGTERM handler does and whether it poses any security threat.

Disallowing subprocesses (fork)

sys.disallow_subprocesses = [true|false]
(Default: false)

This specifies whether to block applications from creating child processes (e.g. via fork() or clone() system calls). The intuition is that many applications have fallbacks when they fail to spawn a child process (e.g. Python). Could be useful in SGX environments: child processes consume EPC memory which is a limited resource.

Note

This option is not a security feature - Gramine by-design is only a one-way sandbox, which doesn’t protect the host from the enclave. Don’t use this option if you want to somehow mitigate running untrusted enclaves. Instead, to achieve this, you need to run the whole Gramine inside a proper security sandbox.

Mocking syscalls

sys.debug__mock_syscalls = [
  { name = "syscall_name1", return = 0 },   # no-op syscall
  { name = "syscall_name2", return = -38 }, # denied syscall (ENOSYS)
]

This syntax specifies the system calls that are mocked when executed in Gramine (i.e. they return a specified value without any other side effects). If return field is skipped, then the default value is 0 (no-op).

Be warned that returning a success (e.g. 0) but skipping the possible side effects of the syscall may introduce bugs to the application, if the application expects these side effects (e.g. mocking the futex syscall may lead to silent introduction of race conditions or hangs).

As one example, to skip sched_yield syscall, specify:

sys.debug__mock_syscalls = [
  { name = "sched_yield", return = 0 },
]

As another example, to disallow eventfd completely, specify:

sys.debug__mock_syscalls = [
  { name = "eventfd",  return = -38 },
  { name = "eventfd2", return = -38 },
]

Note

This option is not a replacement for sys.disallow_subprocesses (see above). This is because the clone() syscall has two usages: (1) it is used to spawn subprocesses by Glibc and many other libraries and runtimes and (2) it is also used to create threads in the same process. The sys.disallow_subprocesses manifest option disables only the first usage, whereas sys.debug__mock_syscalls = [ { name = "clone", ... } ] disables both usages.

Warning

This option is not a security feature. Its rationale is improving performance (the example of sched_yield), mocking syscalls currently not implemented in Gramine, and limiting syscalls exposed to the app.

Warning

If used incorrectly, this option may break syscall semantics and make your application unsound! This option is for advanced users only and for experimenting/debugging.

Root FS mount point

fs.root.type = "[chroot|...]"
fs.root.uri  = "[URI]"

This syntax specifies the root filesystem to be mounted inside the library OS. Both parameters are optional. If not specified, then Gramine mounts the current working directory as the root.

FS mount points

fs.mounts = [
  { type = "[chroot|...]", path = "[PATH]", uri = "[URI]" },
  { type = "[chroot|...]", path = "[PATH]", uri = "[URI]" },
]

Or, as separate sections:

[[fs.mounts]]
type = "[chroot|...]"
path = "[PATH]"
uri  = "[URI]"

[[fs.mounts]]
type = "[chroot|...]"
path = "[PATH]"
uri  = "[URI]"

This syntax specifies how filesystems are mounted inside the library OS. For dynamically linked binaries, usually at least one chroot mount point is required in the manifest (the mount point of linked libraries). The filesystems will be mounted in the order in which they appear in the manifest.

Note

Keep in mind that TOML does not allow trailing commas in inline tables: { path = "...", uri = "...", } is a syntax error.

The type parameter specifies the mount point type. If omitted, it defaults to "chroot". The path parameter must be an absolute path (i.e., must begin with /).

Gramine currently supports the following types of mount points:

chroot (default): Host-backed files. All host files and sub-directories found under [URI] are forwarded to the Gramine instance and placed under [PATH]. For example, with a host-level path specified as uri = "file:/one/path/" and forwarded to Gramine via path = "/another/path", a host-level file /one/path/file is visible to graminized application as /another/path/file. This concept is similar to FreeBSD’s chroot and to Docker’s named volumes. Files under chroot mount points support mmap and fork/clone.
encrypted: Host-backed encrypted files. See Encrypted files for more information.
untrusted_shm: Untrusted shared memory files. See Untrusted shared memory for more information.
tmpfs: Temporary in-memory-only files. These files are not backed by host-level files. The tmpfs files are created under [PATH] (this path is empty on Gramine instance startup) and are destroyed when a Gramine instance terminates. The [URI] parameter is always ignored and can be omitted.

tmpfs is especially useful in trusted environments (like Intel SGX) for securely storing temporary files. This concept is similar to Linux’s tmpfs. Currently there is a limitation that each process has its own, non-shared tmpfs (i.e., processes don’t see each other’s files).

Start (current working) directory

fs.start_dir = "[URI]"

This syntax specifies the start (current working) directory. If not specified, then Gramine sets the root directory as the start directory (see fs.root).

Allowed IOCTLs

sys.ioctl_structs.[identifier] = [memory-layout-format]

sys.allowed_ioctls = [
  { request_code = [NUM], struct = "[identifier-of-ioctl-struct]" },
]

By default, Gramine disables all device-backed and socket IOCTLs. This syntax allows to explicitly allow a set of IOCTLs on devices (devices must be explicitly mounted via fs.mounts manifest syntax) and sockets (e.g. for SIOCGIFCONF and SIOCGIFHWADDR). Only IOCTLs with the request_code argument found among the manifest-listed IOCTLs are allowed to pass-through to the host. Each IOCTL entry may also contain a reference to an IOCTL struct in the struct field, in case the third IOCTL argument is intended to be translated by Gramine.

Available IOCTL structs are described via sys.ioctl_structs. Each IOCTL struct describes the memory layout of the third argument to the ioctl system call (typically a pointer to a complex nested object passed to the device). Description of the memory layout is required for a deep copy of the IOCTL struct. Here we use the term memory region to denote a separate contiguous region of memory and the term sub-region of a memory region to denote a part of the memory region that has properties different from other sub-regions in the same memory region (e.g., should it be copied in or out of Gramine memory, is it a pointer to another memory region, etc.). For example, a C struct can be considered one memory region, and fields of this C struct can be considered sub-regions of this memory region.

We also use the terms Gramine memory and host memory. This distinction is currently relevant only for SGX environments: Gramine memory means in-SGX-enclave memory and host memory means outside-of-SGX-enclave untrusted memory.

Memory layout of the IOCTL struct is described using the TOML syntax of inline arrays (for each new separate memory region) and inline tables (for each sub-region in one memory region). Each sub-region is described via the following keys:

name is an optional name for this sub-region; mainly used to find length-specifying sub-regions and nested memory regions.
alignment is an optional alignment of the memory region; may be specified only in the first sub-region of a memory region (all other sub-regions are contiguous with the first sub-region, so specifying their alignment doesn’t make sense). The default value is 1.
size is the size of this sub-region. The size field may be a string with the name of another sub-region that contains the size value or an integer with the constant size measured in unit units (default unit is 1 byte; also see below). For example, size = "strlen" denotes a size field that will be calculated dynamically during IOCTL execution based on the sub-region named strlen, whereas size = 16 denotes a sub-region of size 16B. Note that ptr sub-regions must not specify the size field.
unit is an optional unit of measurement for size. It is 1 byte by default. Unit of measurement must be a constant integer. For example, size = "strlen" and unit = 2 denote a wide-char string (where each character is 2B long) of a dynamically specified length.
adjustment is an optional integer adjustment for size (always specified in bytes). It is 0 bytes by default. This field must be a constant (possibly negative) integer. For example, size = 6, unit = 2 and adjustment = -8 results in a total size of 4B.
direction = "none" | "out" | "in" | "inout" is an optional direction of copy for this sub-region (from the app point of view). For example, direction = "out" denotes a sub-region to be copied out of Gramine memory to host memory, i.e., this sub-region is an input to the host device. The default value is none which is useful for e.g. padding of structs. This field must be ommitted if the ptr field is specified for this sub-region (pointer sub-regions contain the pointer value which will be unconditionally rewired to point to host memory).
ptr = inlined-memory-region or ptr = "another-ioctl-struct" specifies a pointer to another, nested memory region. This field is required when describing complex IOCTL structs. Such pointer memory region always has the implicit size of 8B, and the pointer value is always rewired by Gramine to the memory region in host memory (containing a corresponding nested memory region). If ptr is specified together with array_len, it describes an array of pointers to these memory regions. (In other words, ptr is an array of pointers to memory regions with array_len = 1 by default.) This may be recursive with the NULL value being a guard, which allows describing linked lists.
array_len is an optional number of items in the ptr array. This field cannot be specified with non-ptr sub-regions. The default value is 1.
onlyif = "simple boolean expression" allows to condition the sub-region based on a boolean expression. The sub-region is taken into account only if the expression evaluates to true. The only currently supported formats of expressions are token1 == token2 and token1 != token2, where token1 and token2 may be constant non-negative integers or sub-region names (referenced sub-regions must contain a native-endian integer value of size 1..8B).

Consider this simple C snippet:

struct ioctl_read {
    size_t buf_size;                /* copied from Gramine to device */
    char* buf;                      /* copied from device to Gramine */
} __attribute__((aligned(0x1000))); /* alignment just for illustration */

This translates into the following manifest syntax:

sys.ioctl_structs.ioctl_read = [
    {
        name      = "buf_size",
        size      = 8,
        direction = "out",
        alignment = 0x1000
    },
    {
        ptr = [
            {
                size      = "buf_size",
                direction = "in"
            }
        ]
    }
]

The above example specifies a root struct (first memory region) that consists of two sub-regions: the first one contains an 8-byte size value, the second one is an 8-byte pointer value. This pointer points to another memory region in Gramine memory that contains a single sub-region of size buf_size. This nested sub-region is copied from the device into the Gramine memory.

IOCTLs that use the above struct in a third argument are defined like this:

sys.allowed_ioctls = [
  { request_code = 0x12345678, struct = "ioctl_read" },
  { request_code = 0x87654321, struct = "ioctl_read" },
]

If the IOCTL’s third argument should be passed directly as-is (or unused at all), then the struct key must be an empty string or not exist at all:

sys.allowed_ioctls = [
  { request_code = 0x43218765, struct = "" },
  { request_code = 0x87654321 },
]

Note

IOCTLs for device communication are pass-through and thus potentially insecure by themselves in e.g. SGX environments:

IOCTL arguments are passed as-is from the app to the untrusted host, which may lead to leaks of enclave data.
Untrusted host can change IOCTL arguments as it wishes when passing them from Gramine to the device and back.

It is the responsibility of the app developer to correctly use IOCTLs, with security implications in mind. In most cases, IOCTL arguments should be encrypted or integrity-protected with a key pre-shared between Gramine and the device.

Experimental flock (BSD-style locks) support

sys.experimental__enable_flock = [true|false]
(Default: false)

This syntax enables the flock system call in Gramine.

Warning

This syscall is still under development and may contain security vulnerabilities. This is temporary; the syscall will be enabled by default in the future after thorough validation and this syntax will be removed then.

SGX syntax

If Gramine is not running with SGX, the SGX-specific syntax is ignored. All keys in the SGX-specific syntax are optional.

Debug/production enclave

sgx.debug = [true|false]
(Default: false)

This syntax specifies whether the enclave can be debugged. Set it to true for a debug enclave and to false for a production enclave.

EDMM

sgx.edmm_enable = [true|false]
(Default: false)

This setting enables the EDMM feature (after-enclave-creation memory management). If set to true, Gramine will refuse to start on CPUs which do not support EDMM feature.

When this feature is enabled, Gramine does not add heap pages (uninitialized memory) to the enclave at creation time. Instead, memory is added to the enclave on demand (note that for mappings requested with MAP_NORESERVE, the enclave pages are lazily committed on page fault events). This can greatly reduce startup time for bigger enclaves, reduce the EPC usage (as only actually allocated memory is used) and allow for changing memory permissions (without this Gramine allocates all dynamic memory as RWX). Unfortunately it can negatively impact performance, as adding a page to the enclave at runtime is a more expensive operation than adding the page before enclave creation (because it involves more enclave exits and syscalls).

When this feature is enabled, it is not necessary to specify sgx.enclave_size (Gramine will automatically set it to 1TB which should be enough for any application). However if sgx.enclave_size is specified, this explicit value will take precedence.

Note

Support for EDMM first appeared in Linux 6.0.

Enclave size

sgx.enclave_size = "[SIZE]"
(default: "256M" without EDMM, "1024G" with EDMM)

This syntax specifies the size of the enclave set during enclave creation time if EDMM is not enabled (sgx.edmm_enable = false) or the maximal size that the enclave can grow to if EDMM is enabled (sgx.edmm_enable = true).

The PAL and library OS code/data count towards this size value, as well as the application memory itself: application’s code, stack, heap, loaded application libraries, etc. The application cannot allocate memory that exceeds this limit.

Be careful when setting the enclave size to large values: when creating enclaves which do not have the EDMM feature enabled, Gramine not only reserves sgx.enclave_size bytes of virtual address space but also commits them to the backing store (EPC, RAM and/or swap file). For example, if sgx.enclave_size = "4G", then 4GB of EPC/RAM will be immediately allocated to back the enclave memory (recall that EPC is the SGX-protected part of RAM). Thus, if your system has 4GB of backing store or less, then the host Linux kernel will fail to start the SGX enclave and will typically print the Killed message. If you encounter this situation, you can try the following:

If possible, decrease sgx.enclave_size to a value less than the amount of RAM. For example, if you have 4GB of RAM, set sgx.enclave_size = "2G".
Switch to a system that has more RAM. For example, if you must use sgx.enclave_size = "4G", move to a system with at least 5GB of RAM.
If the above options are ruled out, then increase the swap file size (recall that the swap file is a space on hard disk used as a virtual “extension” to real RAM). For example, if you have 4GB of RAM and you must use sgx.enclave_size = "4G", then create the swap file of size 1GB. Note that as soon as the SGX application starts using the swap file, its performance degrades significantly!

Also, be careful with multi-process SGX applications: each new child process runs in its own SGX enclave and thus requires an additional sgx.enclave_size amount of RAM. For example, if you run bash -c ls and your manifest contains sgx.enclave_size = "4G", then two SGX enclaves (bash and ls processes) will consume 8GB of RAM in total. If there is less than 8GB of RAM (+ swap file) on your system, such bash -c ls SGX workload will fail. Note this does not apply to the enclaves with EDMM enabled, where memory is not reserved upfront and is allocated on demand.

Number of threads

sgx.max_threads = [NUM]
(Default: 4)

If EDMM is not enabled (sgx.edmm_enable = false), then this syntax specifies the maximum number of threads that can be created inside the enclave (recall that SGX v1 requires a predetermined maximum number of thread slots). The application cannot have more threads than this limit at a time (however, it is possible to create new threads after old threads are destroyed).

If EDMM is enabled (sgx.edmm_enable = true), then this syntax specifies the number of pre-allocated thread slots (must be at least 1). However, the maximum number of threads can exceed this limit during enclave execution, by dynamically allocating new thread slots.

Note that Gramine uses several helper threads internally:

The IPC thread to facilitate inter-process communication. This thread is always spawned at Gramine startup. Its activity depends on the communication patterns among Gramine processes; if there is only one Gramine process, the IPC thread always sleeps.
The Async thread to implement timers and other asynchronous events/notifications. This thread is spawned on demand. It terminates itself if there are no pending events/notifications.
The TLS-handshake thread on pipes creation. This thread is spawned on demand, each time a new pipe is created. It terminates itself immediately after the TLS handshake is performed.

Given these internal threads, sgx.max_threads should be set to at least 4 even for single-threaded applications (to accommodate for the main thread, the IPC thread, the Async thread and one TLS-handshake thread).

Note

This option will be renamed after non-EDMM platform support is dropped.

Number of RPC threads (Exitless feature)

sgx.insecure__rpc_thread_num = [NUM]
(Default: 0)

This syntax specifies the number of RPC threads that are created outside of the enclave. RPC threads are helper threads that run in untrusted mode alongside enclave threads. RPC threads issue system calls on behalf of enclave threads. This allows “exitless” design when application threads never leave the enclave (except for a few syscalls where there is no benefit, e.g., nanosleep()).

If the user specifies 0 or omits this directive, then no RPC threads are created, and all system calls perform an enclave exit (“normal” execution).

Note that the number of created RPC threads should match the maximum number of simultaneous enclave threads. If there are more RPC threads, then CPU time is wasted. If there are less RPC threads, some enclave threads may starve, especially if there are many blocking system calls by other enclave threads.

The Exitless feature may be detrimental for performance. It trades slow OCALLs/ECALLs for fast shared-memory communication at the cost of occupying more CPU cores and burning more CPU cycles. For example, a single-threaded Redis instance on Linux becomes 5-threaded on Gramine with Exitless. Thus, Exitless may negatively impact throughput but may improve latency.

This feature is currently marked as insecure, because it reads and writes to untrusted memory in potentially insecure manner - susceptible to CVE-2022-21233 (INTEL-SA-00657) and CVE-2022-21166 (INTEL-SA-00615) respectively.

SGX EXINFO

sgx.use_exinfo = [true|false]
(Default: false)

sgx.insecure__allow_memfaults_without_exinfo = [true|false]
(Default: false)

If sgx.use_exinfo is set, user application can retrieve faulting address in signal handler in case of page/general protection faults (#PF and #GP). Otherwise (set to false), the behavior depends on sgx.insecure__allow_memfaults_without_exinfo:

If sgx.insecure__allow_memfaults_without_exinfo is unset (default), then Gramine terminates with an error, to prevent a possible attack.
Otherwise the exception is allowed and Gramine forwards it to the application, and the faulting address is provided as 0.

The default value for sgx.use_exinfo is false because some frameworks/runtimes could otherwise print the callstack and variables/registers on exceptions, potentially leaking data.

Note

The option sgx.insecure__allow_memfaults_without_exinfo is provided only to allow debugging/testing on old CPUs that do not support the EXINFO feature. Without EXINFO support, a malicious host may attack the application by injecting a memory fault. This option is thus insecure and must not be used in production environments! It will be removed in near future.

Optional CPU features (AVX, AVX512, AMX, MPX, PKRU)

sgx.cpu_features.avx    = "[unspecified|disabled|required]"
sgx.cpu_features.avx512 = "[unspecified|disabled|required]"
sgx.cpu_features.amx    = "[unspecified|disabled|required]"
(Default: "unspecified")

sgx.cpu_features.mpx    = "[disabled|required]"
sgx.cpu_features.pkru   = "[disabled|required]"
(Default: "disabled")

For the SGX threat model, Gramine divides CPU features in two subsets: not-security-hardening features (currently AVX, AVX512 and AMX) and security-hardening features (MPX and PKRU).

The "unspecified" syntax applies only to not-security-hardening features. It means that the enclave initialization will succeed regardless of whether the CPU feature is available on the platform or not. The CPU features will be enabled in the enclave if they are available on the platform. Security-hardening features cannot be marked as "unspecified" because this would lead to different security properties of the application on different platforms.

The "disabled" syntax disables the CPU feature inside the enclave even if this CPU feature is available on the platform. This may improve enclave performance because this CPU feature will not be saved and restored during enclave entry/exit. This syntax is provided to improve performance of applications that are known to not rely on certain CPU features. Be aware that if the application relies on some disabled CPU features, the application will fail with SIGILL (“illegal instruction”). For example, if the application is built with AVX support, and AVX is disabled in the manifest, the application will crash.

The "required" syntax ensures that the CPU feature is available and enabled for the enclave. If such option is set in the manifest but the CPU feature is unavailable on the platform, enclave initialization will fail.

When in doubt, it is recommended to keep the default values for these features. In this case, Gramine auto-detects the corresponding not-security-hardening CPU features on the platform and enables them if available, and disables security-hardening CPU features. This allows the SGX enclave to be executed on the widest range of platforms.

At the technical level, the mapping from CPU features to SGX fields is as follows:

"unspecified": SIGSTRUCT.ATTRIBUTEMASK[feature] = 0. The untrusted loader of Gramine sets SECS.ATTRIBUTES[feature] = 0 if it can’t detect the feature on the platform, and SECS.ATTRIBUTES[feature] = 1 otherwise.
"required": SIGSTRUCT.ATTRIBUTEMASK[feature] = 1 and SIGSTRUCT.ATTRIBUTES[feature] = 1.
"disabled": SIGSTRUCT.ATTRIBUTEMASK[feature] = 1 and SIGSTRUCT.ATTRIBUTES[feature] = 0.

ISV Product ID and SVN

sgx.isvprodid = [NUM]
sgx.isvsvn    = [NUM]
(Default: 0)

This syntax specifies the ISV Product ID and SVN to be added to the enclave signature.

Attribute masks for SGX sealing key derivation

sgx.seal_key.flags_mask = "[8-byte hex value]"  (default: "0xffffffffffffffff")
sgx.seal_key.xfrm_mask  = "[8-byte hex value]"  (default: "0xfffffffffff9ff1b")
sgx.seal_key.misc_mask  = "[4-byte hex value]"  (default: "0xffffffff")

This syntax specifies masks used to generate the SGX sealing key. These masks correspond to the following SGX KEYREQUEST struct fields:

flags_mask: KEYREQUEST.ATTRIBUTESMASK.FLAGS
xfrm_mask: KEYREQUEST.ATTRIBUTESMASK.XFRM
misc_mask: KEYREQUEST.MISCMASK

Most users do not need to set these masks. Only advanced users with knowledge of SGX sealing should use these masks. In particular, these masks allow to specify a subset of enclave/machine attributes to be used in sealing key derivation. Moreover, these masks themselves are used in sealing key derivation.

Allowed files

sgx.allowed_files = [
  "[URI]",
  "[URI]",
]

This syntax specifies the files/directories (with the file: prefix) and devices (with the dev: prefix) that are allowed to be created or opened in the enclave unconditionally. In other words, allowed files, directories and devices can be opened for reading/writing and can be created if they do not exist already. Allowed files are not cryptographically hashed and are thus not protected.

Warning

It is insecure to allow files containing code or critical information; developers must not allow files blindly! Instead, use trusted or encrypted files.

Similarly, communication with allowed devices is pass-through and thus potentially insecure by itself. It is the responsibility of the app developer to correctly communicate with devices, with security implications in mind.

Trusted files

# entries can be strings
sgx.trusted_files = [
  "[URI]",
  "[URI]",
]

# entries can also be tables
[[sgx.trusted_files]]
uri = "[URI]"
sha256 = "[HASH]"

This syntax specifies the files to be cryptographically hashed at build time; at runtime, these files may only be accessed by the app if the files’ hashes match what is stored in the manifest. This implies that trusted files can be only opened for reading (not for writing) and cannot be created if they do not exist already. The signer tool will automatically generate hashes of these files and add them to the SGX-specific manifest (.manifest.sgx). The manifest writer may also specify the hash for a file using the TOML-table syntax, in the field sha256; in this case, hashing of the file will be skipped by the signer tool and the value in sha256 field will be used instead.

Marking files as trusted is especially useful for shared libraries: a trusted library cannot be silently replaced by a malicious host because the hash verification will fail.

Encrypted files

fs.mounts = [
  { type = "encrypted", path = "[PATH]", uri = "[URI]", key_name = "[KEY_NAME]" },
]

fs.insecure__keys.[KEY_NAME] = "[32-character hex value]"

This syntax allows mounting files that are encrypted on disk and transparently decrypted when accessed by Gramine or by application running inside Gramine. Encrypted files guarantee data confidentiality and integrity (tamper resistance), as well as file swap protection (an encrypted file can only be accessed when in a specific host path).

Encrypted files were previously known as protected files, and some Gramine tools might still use the old name.

URI can be a file or a directory. If a directory is mounted, all existing files/directories within it are recursively treated as encrypted (and are expected to be encrypted in the PF format). New files created in an encrypted mount are also automatically treated as encrypted.

Warning

The current implementation assumes that type = "encrypted" mounts do not overlap on host, i.e. there are no host files reachable through more than one type = "encrypted" mount. Otherwise, changes made to such files might not be correctly persisted by Gramine.

Note that path size of an encrypted file is limited to 512 bytes and filename size is limited to 260 bytes.

The key_name mount parameter specifies the name of the encryption key. If omitted, it will default to "default". This feature can be used to mount different files or directories with different encryption keys.

fs.insecure__keys.[KEY_NAME] can be used to specify the encryption keys directly in manifest. This option must be used only for debugging purposes.

Warning

sgx.insecure__keys.[KEY_NAME] hard-codes the key in the manifest. This option is thus insecure and must not be used in production environments! Typically, you want to provision the encryption keys using SGX local/remote attestation, thus you should not specify any sgx.insecure__keys.[KEY_NAME] manifest options at all. Instead, use the Secret Provisioning interface (see Attestation and Secret Provisioning).

Key names beginning with underscore (_) denote special keys provided by Gramine:

"_sgx_mrenclave" (SGX only) is the SGX sealing key based on the MRENCLAVE identity of the enclave. This is useful to allow only the same enclave (on the same platform) to unseal files, i.e., this key is not accessible to any other software other than the specific enclave on the specific platform.
"_sgx_mrsigner" (SGX only) is the SGX sealing key based on the MRSIGNER identity of the enclave. This is useful to allow all enclaves signed with the same key (and on the same platform) to unseal files, i.e., this key is not accessible to any other software other than the specific set of same-MRSIGNER enclaves on the specific platform.

Warning

The same key must not be used for the encrypted-files mount and for the application’s own crypto operations. Such “double” use of the same key may lead to compromise of the key. For example, specifying an FS mount via {type = "encrypted", ..., key_name = "_sgx_mrenclave"} in the manifest and using the same key obtained via /dev/attestation/keys/_sgx_mrenclave in the application is insecure. If you need to derive encryption keys from such a “doubly-used” key, you must apply a KDF.

Untrusted shared memory

fs.mounts = [
  { type = "untrusted_shm", path = "[PATH]", uri = "[URI]" },
]

This syntax allows mounting shared memory objects that are accessible by both the application running inside Gramine and by other host software/hardware (host OS, other host processes, devices connected to the host). In Gramine, untrusted shared memory applies to files which must be mapped into application address space with the MAP_SHARED flag.

URI can be a file or a directory (with a dev: prefix). If a directory is mounted, all files under this directory are treated as shared memory objects (but sub-directories are inaccessible for security reasons). New files created in a shared memory mount are also automatically treated as shared memory objects. Creating sub-directories in a shared memory mount is not allowed, for security reasons. Files in a shared memory mount (or the mounted directory itself) need to be explicitly listed as allowed_files to be accessed. See Allowed files for more information.

Typically, you should mount the directory /dev/shm/ (it is used for sharing data between processes and devices) and allow specific files in it. When this directory is mounted, the Gramine application may create the files – called “shared memory objects” in POSIX – under this directory (for example, this is how shm_open() Glibc function works). It is not recommended to allow a directory unless the application creates shared memory objects with unpredictable names. Allowing a directory creates a risk of exposing unexpected data to the host.

Note

Adding shared memory mounts is insecure by itself in SGX environments:

All data put in shared memory reside outside of the SGX enclave.

Typically applications do not encrypt the data put in shared memory, which may lead to leaks of enclave data.

Untrusted host can modify data in shared memory as it wishes, so applications could see or operate on maliciously modified data.

It is the responsibility of the app developer to correctly use shared memory, with security implications in mind. In most cases, data in shared memory should be preliminarily encrypted or integrity-protected by the user app with a key pre-shared between all participating processes (and possibly devices that use this shared memory).

File check policy

sgx.file_check_policy = "[strict|allow_all_but_log]"
(Default: "strict")

This syntax specifies the file check policy, determining the behavior of authentication when opening files or devices. By default, only files explicitly listed as trusted_files and files or devices explicitly listed as allowed_files are allowed for access.

If the file check policy is allow_all_but_log, all files and devices other than trusted and allowed are allowed for access, and Gramine emits a warning message for every such file/device. Effectively, this policy operates on all unknown files and devices as if they were listed as allowed_files. (However, this policy still does not allow writing/creating files specified as trusted.) This policy is a convenient way to determine the set of files that the ported application uses.

Attestation and quotes

sgx.remote_attestation = "[none|epid|dcap]"
(Default: "none")

sgx.ra_client_linkable = [true|false]
sgx.ra_client_spid     = "[HEX]"
(Only for EPID based attestation)

This syntax specifies the parameters for remote attestation. By default, it is not enabled.

For EPID based attestation, remote_attestation must be set to epid. In addition, ra_client_linkable and ra_client_spid must be filled with your registered Intel SGX EPID Attestation Service credentials (linkable/unlinkable mode and SPID of the client respectively).

For DCAP based attestation, remote_attestation must be set to dcap. ra_client_spid and ra_client_linkable are ignored.

Pre-heating enclave

sgx.preheat_enclave = [true|false]
(Default: false)

When enabled, this option instructs Gramine to pre-fault all heap pages during initialization. This has a negative impact on the total run time, but shifts the EPC page faults cost to the initialization phase, which can be useful in a scenario where a server starts and receives connections / work packages only after some time. It also makes the later run time and latency much more predictable.

Please note that using this option makes sense only when the EPC is large enough to hold the whole heap area.

This option is invalid (i.e. must be false) if specified together with sgx.edmm_enable, as there are no heap pages to pre-fault.

Enabling SGX enclave stats

sgx.enable_stats = [true|false]
(Default: false)

This syntax specifies whether to enable SGX enclave-specific statistics:

TCS.FLAGS.DBGOPTIN flag. This flag is set in all enclave threads and enables certain debug and profiling features with enclaves, including breakpoints, performance counters, Intel PT, etc.
Printing the stats on SGX-specific events. Currently supported stats are: number of EENTERs (corresponds to ECALLs plus returns from OCALLs), number of EEXITs (corresponds to OCALLs plus returns from ECALLs) and number of AEXs (corresponds to interrupts/exceptions/signals during enclave execution). Prints overall stats at the end of enclave execution.
Printing the SGX enclave loading time at startup. The enclave loading time includes creating the enclave, adding enclave pages, measuring them and initializing the enclave.

For this option to take effect, Gramine must be compiled with --buildtype=debug or --buildtype=debugoptimized. Otherwise (if built in release mode), Gramine will exit with an error.

Warning

This option is insecure and cannot be used with production enclaves (sgx.debug = false). If a production enclave is started with this option set, Gramine will fail initialization of the enclave.

SGX profiling

sgx.profile.enable = ["none"|"main"|"all"]
(Default: "none")

This syntax specifies whether to enable SGX profiling. Gramine must be compiled with --buildtype=debug or --buildtype=debugoptimized for this option to work (the latter is advised). In addition, the manifest must contain sgx.debug = true.

If this option is set to main, the main process will collect IP samples and save them as sgx-perf.data. If it is set to all, all processes will collect samples and save them to sgx-perf-<PID>.data.

The saved files can be viewed with the perf tool, e.g. perf report -i sgx-perf.data.

See SGX profiling for more information.

Warning

This option is insecure and cannot be used with production enclaves (sgx.debug = false). If a production enclave is started with this option set, Gramine will fail initialization of the enclave.

sgx.profile.mode = ["aex"|"ocall_inner"|"ocall_outer"]
(Default: "aex")

Specifies what events to record:

aex: Records enclave state during asynchronous enclave exit (AEX). Use this to check where the CPU time is spent in the enclave.
ocall_inner: Records enclave state during OCALL.
ocall_outer: Records the outer OCALL function, i.e., what OCALL handlers are going to be executed. Does not include stack information (cannot be used with sgx.profile.with_stack = true).

See also OCALL profiling for more detailed advice regarding the OCALL modes.

sgx.profile.with_stack = [true|false]
(Default: false)

This syntax specifies whether to include stack information with the profiling data. This will enable perf report to show call chains. However, it will make the output file much bigger, and slow down the process.

sgx.profile.frequency = [INTEGER]
(Default: 50)

This syntax specifies approximate frequency at which profiling samples are taken (in samples per second). Lower values will mean less accurate results, but also lower overhead.

Note that the accuracy is limited by how often the process is interrupted by Linux scheduler: the effective maximum is 250 samples per second.

Note

This option applies only to aex mode. In the ocall_* modes, currently all samples are taken.

SGX profiling with Intel VTune Profiler

sgx.vtune_profile = [true|false]
(Default: false)

This syntax specifies whether to enable SGX profiling with Intel VTune Profiler. Gramine must be compiled with --buildtype=debug or --buildtype=debugoptimized for this option to work (the latter is advised). In addition, the application manifest must also contain sgx.debug = true.

Note

The manifest options sgx.vtune_profile and sgx.profile.* can work independently.

See Profiling SGX hotspots with Intel VTune Profiler for more information.

Deprecated options

Loader entrypoint (deprecated syntax)

loader.entrypoint = "[URI]"

This syntax was renamed to loader.entrypoint.uri.