LibSandbox: Add reusable sandbox building blocks

Move the Linux no_new_privs, Landlock, and seccomp policy plumbing
into LibSandbox so individual services can describe the privileges they
need without copying the BPF and kernel feature detection machinery.

Keep ImageDecoder's sandbox policy service-local by composing the new
building blocks in SandboxLinux.cpp. This preserves the existing syscall
allowlist while making the policy easier to audit and reuse.
This commit is contained in:
Andreas Kling 2026-06-09 11:00:24 +02:00 committed by Andreas Kling
parent e83836b928
commit ba87a32626
8 changed files with 937 additions and 486 deletions

View file

@ -12,6 +12,7 @@ add_subdirectory(LibJS)
add_subdirectory(LibMain)
add_subdirectory(LibRegex)
add_subdirectory(LibRequests)
add_subdirectory(LibSandbox)
add_subdirectory(LibSync)
add_subdirectory(LibSyntax)
add_subdirectory(LibTLS)

View file

@ -0,0 +1,11 @@
set(SOURCES
Sandbox.cpp
)
if (LINUX)
list(APPEND SOURCES
Seccomp.cpp
)
endif()
ladybird_lib(LibSandbox sandbox)

View file

@ -0,0 +1,102 @@
/*
* Copyright (c) 2026-present, the Ladybird developers.
*
* SPDX-License-Identifier: BSD-2-Clause
*/
#include <LibSandbox/Sandbox.h>
#if defined(AK_OS_LINUX)
# include <AK/ScopeGuard.h>
# include <errno.h>
# include <linux/landlock.h>
# include <sys/prctl.h>
# include <sys/syscall.h>
# include <unistd.h>
#endif
#if defined(__GLIBC__)
# include <malloc.h>
#endif
namespace Sandbox {
ErrorOr<void> install_no_new_privileges()
{
#if defined(AK_OS_LINUX)
if (prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0) < 0)
return Error::from_syscall("prctl(PR_SET_NO_NEW_PRIVS)"sv, errno);
#endif
return {};
}
ErrorOr<void> configure_runtime()
{
#if defined(AK_OS_LINUX) && defined(__GLIBC__) && defined(M_ARENA_MAX) && !defined(HAS_ADDRESS_SANITIZER)
if (mallopt(M_ARENA_MAX, 4) == 0)
return Error::from_string_literal("mallopt(M_ARENA_MAX) failed");
#endif
return {};
}
ErrorOr<void> restrict_filesystem_with_landlock()
{
#if defined(AK_OS_LINUX) && defined(__NR_landlock_create_ruleset) && defined(__NR_landlock_restrict_self)
auto landlock_abi = syscall(__NR_landlock_create_ruleset, nullptr, 0, LANDLOCK_CREATE_RULESET_VERSION);
if (landlock_abi < 0) {
if (errno == ENOSYS || errno == EOPNOTSUPP || errno == EINVAL)
return {};
return Error::from_syscall("landlock_create_ruleset(LANDLOCK_CREATE_RULESET_VERSION)"sv, errno);
}
if (landlock_abi == 0)
return {};
landlock_ruleset_attr ruleset_attributes {};
ruleset_attributes.handled_access_fs = LANDLOCK_ACCESS_FS_EXECUTE
| LANDLOCK_ACCESS_FS_WRITE_FILE
| LANDLOCK_ACCESS_FS_READ_FILE
| LANDLOCK_ACCESS_FS_READ_DIR
| LANDLOCK_ACCESS_FS_REMOVE_DIR
| LANDLOCK_ACCESS_FS_REMOVE_FILE
| LANDLOCK_ACCESS_FS_MAKE_CHAR
| LANDLOCK_ACCESS_FS_MAKE_DIR
| LANDLOCK_ACCESS_FS_MAKE_REG
| LANDLOCK_ACCESS_FS_MAKE_SOCK
| LANDLOCK_ACCESS_FS_MAKE_FIFO
| LANDLOCK_ACCESS_FS_MAKE_BLOCK
| LANDLOCK_ACCESS_FS_MAKE_SYM;
# ifdef LANDLOCK_ACCESS_FS_REFER
if (landlock_abi >= 2)
ruleset_attributes.handled_access_fs |= LANDLOCK_ACCESS_FS_REFER;
# endif
# ifdef LANDLOCK_ACCESS_FS_TRUNCATE
if (landlock_abi >= 3)
ruleset_attributes.handled_access_fs |= LANDLOCK_ACCESS_FS_TRUNCATE;
# endif
# if defined(LANDLOCK_ACCESS_NET_BIND_TCP) && defined(LANDLOCK_ACCESS_NET_CONNECT_TCP)
if (landlock_abi >= 4)
ruleset_attributes.handled_access_net = LANDLOCK_ACCESS_NET_BIND_TCP | LANDLOCK_ACCESS_NET_CONNECT_TCP;
auto ruleset_attributes_size = landlock_abi >= 4
? sizeof(ruleset_attributes)
: offsetof(landlock_ruleset_attr, handled_access_net);
# else
auto ruleset_attributes_size = sizeof(ruleset_attributes);
# endif
auto ruleset_fd = syscall(__NR_landlock_create_ruleset, &ruleset_attributes, ruleset_attributes_size, 0);
if (ruleset_fd < 0)
return Error::from_syscall("landlock_create_ruleset"sv, errno);
ArmedScopeGuard close_ruleset_fd = [&] {
close(static_cast<int>(ruleset_fd));
};
if (syscall(__NR_landlock_restrict_self, ruleset_fd, 0) < 0)
return Error::from_syscall("landlock_restrict_self"sv, errno);
#endif
return {};
}
}

View file

@ -0,0 +1,17 @@
/*
* Copyright (c) 2026-present, the Ladybird developers.
*
* SPDX-License-Identifier: BSD-2-Clause
*/
#pragma once
#include <AK/Error.h>
namespace Sandbox {
[[nodiscard]] ErrorOr<void> install_no_new_privileges();
[[nodiscard]] ErrorOr<void> configure_runtime();
[[nodiscard]] ErrorOr<void> restrict_filesystem_with_landlock();
}

View file

@ -0,0 +1,750 @@
/*
* Copyright (c) 2026-present, the Ladybird developers.
*
* SPDX-License-Identifier: BSD-2-Clause
*/
#include <LibSandbox/Seccomp.h>
#include <errno.h>
#include <fcntl.h>
#include <linux/audit.h>
#include <linux/sched.h>
#include <linux/seccomp.h>
#include <stddef.h>
#include <sys/ioctl.h>
#include <sys/mman.h>
#include <sys/prctl.h>
#include <sys/syscall.h>
#include <sys/ucontext.h>
#include <unistd.h>
namespace Sandbox {
namespace {
#if defined(__x86_64__)
static constexpr u32 audit_architecture = AUDIT_ARCH_X86_64;
#elif defined(__aarch64__)
static constexpr u32 audit_architecture = AUDIT_ARCH_AARCH64;
#elif defined(__riscv) && __riscv_xlen == 64
static constexpr u32 audit_architecture = AUDIT_ARCH_RISCV64;
#else
# error "Add 64-bit seccomp audit architecture for this Linux architecture"
#endif
static constexpr u32 thread_clone_required_flags = CLONE_VM | CLONE_FS | CLONE_FILES | CLONE_SIGHAND | CLONE_THREAD;
static constexpr u32 thread_clone_allowed_flags = thread_clone_required_flags
| CLONE_SYSVSEM
| CLONE_SETTLS
| CLONE_PARENT_SETTID
| CLONE_CHILD_CLEARTID
| CLONE_DETACHED
| CLONE_CHILD_SETTID;
#ifdef O_LARGEFILE
static constexpr unsigned read_only_open_flags = O_CLOEXEC | O_LARGEFILE;
#else
static constexpr unsigned read_only_open_flags = O_CLOEXEC;
#endif
#define SECCOMP_LOAD_SYSCALL_NR BPF_STMT(BPF_LD | BPF_W | BPF_ABS, static_cast<unsigned int>(offsetof(seccomp_data, nr)))
#define SECCOMP_LOAD_ARCHITECTURE BPF_STMT(BPF_LD | BPF_W | BPF_ABS, static_cast<unsigned int>(offsetof(seccomp_data, arch)))
#define SECCOMP_LOAD_ARGUMENT(index) BPF_STMT(BPF_LD | BPF_W | BPF_ABS, static_cast<unsigned int>(offsetof(seccomp_data, args[(index)])))
#define SECCOMP_ALLOW BPF_STMT(BPF_RET | BPF_K, SECCOMP_RET_ALLOW)
#define SECCOMP_TRAP BPF_STMT(BPF_RET | BPF_K, SECCOMP_RET_TRAP)
#define SECCOMP_ERRNO(error) BPF_STMT(BPF_RET | BPF_K, SECCOMP_RET_ERRNO | ((error) & SECCOMP_RET_DATA))
#define SECCOMP_ALLOW_NOTHING BPF_STMT(BPF_ALU | BPF_ADD | BPF_K, 0)
#define SECCOMP_APPEND_ALLOW_SYSCALL(policy, name) \
do { \
(policy).append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, __NR_##name, 0, 1)); \
(policy).append(SECCOMP_ALLOW); \
} while (0)
#define SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(policy, name) \
IF_DEFINED_##name(SECCOMP_APPEND_ALLOW_SYSCALL(policy, name), (policy).append(SECCOMP_ALLOW_NOTHING))
#define IF_DEFINED_accept4(if_defined, if_not_defined) if_defined
#define IF_DEFINED_brk(if_defined, if_not_defined) if_defined
#define IF_DEFINED_clock_getres(if_defined, if_not_defined) if_defined
#define IF_DEFINED_clock_gettime(if_defined, if_not_defined) if_defined
#define IF_DEFINED_clock_nanosleep(if_defined, if_not_defined) if_defined
#define IF_DEFINED_clone(if_defined, if_not_defined) if_defined
#define IF_DEFINED_clone3(if_defined, if_not_defined) if_defined
#define IF_DEFINED_close(if_defined, if_not_defined) if_defined
#define IF_DEFINED_dup(if_defined, if_not_defined) if_defined
#define IF_DEFINED_dup3(if_defined, if_not_defined) if_defined
#define IF_DEFINED_exit(if_defined, if_not_defined) if_defined
#define IF_DEFINED_exit_group(if_defined, if_not_defined) if_defined
#define IF_DEFINED_fcntl(if_defined, if_not_defined) if_defined
#define IF_DEFINED_fcntl64(if_defined, if_not_defined) if_defined
#define IF_DEFINED_fstat(if_defined, if_not_defined) if_defined
#define IF_DEFINED_ftruncate(if_defined, if_not_defined) if_defined
#define IF_DEFINED_futex(if_defined, if_not_defined) if_defined
#define IF_DEFINED_futex_time64(if_defined, if_not_defined) if_defined
#define IF_DEFINED_getcpu(if_defined, if_not_defined) if_defined
#define IF_DEFINED_getegid(if_defined, if_not_defined) if_defined
#define IF_DEFINED_geteuid(if_defined, if_not_defined) if_defined
#define IF_DEFINED_getgid(if_defined, if_not_defined) if_defined
#define IF_DEFINED_getpeername(if_defined, if_not_defined) if_defined
#define IF_DEFINED_getpid(if_defined, if_not_defined) if_defined
#define IF_DEFINED_getrandom(if_defined, if_not_defined) if_defined
#define IF_DEFINED_getrlimit(if_defined, if_not_defined) if_defined
#define IF_DEFINED_getrusage(if_defined, if_not_defined) if_defined
#define IF_DEFINED_getsockname(if_defined, if_not_defined) if_defined
#define IF_DEFINED_getsockopt(if_defined, if_not_defined) if_defined
#define IF_DEFINED_gettid(if_defined, if_not_defined) if_defined
#define IF_DEFINED_gettimeofday(if_defined, if_not_defined) if_defined
#define IF_DEFINED_getuid(if_defined, if_not_defined) if_defined
#define IF_DEFINED_ioctl(if_defined, if_not_defined) if_defined
#define IF_DEFINED_lseek(if_defined, if_not_defined) if_defined
#define IF_DEFINED_madvise(if_defined, if_not_defined) if_defined
#define IF_DEFINED_membarrier(if_defined, if_not_defined) if_defined
#define IF_DEFINED_memfd_create(if_defined, if_not_defined) if_defined
#define IF_DEFINED_mmap(if_defined, if_not_defined) if_defined
#define IF_DEFINED_mmap2(if_defined, if_not_defined) if_defined
#define IF_DEFINED_mprotect(if_defined, if_not_defined) if_defined
#define IF_DEFINED_mremap(if_defined, if_not_defined) if_defined
#define IF_DEFINED_munmap(if_defined, if_not_defined) if_defined
#define IF_DEFINED_nanosleep(if_defined, if_not_defined) if_defined
#define IF_DEFINED_pipe2(if_defined, if_not_defined) if_defined
#define IF_DEFINED_poll(if_defined, if_not_defined) if_defined
#define IF_DEFINED_ppoll(if_defined, if_not_defined) if_defined
#define IF_DEFINED_prctl(if_defined, if_not_defined) if_defined
#define IF_DEFINED_prlimit64(if_defined, if_not_defined) if_defined
#define IF_DEFINED_pselect6(if_defined, if_not_defined) if_defined
#define IF_DEFINED_pwrite64(if_defined, if_not_defined) if_defined
#define IF_DEFINED_read(if_defined, if_not_defined) if_defined
#define IF_DEFINED_recvfrom(if_defined, if_not_defined) if_defined
#define IF_DEFINED_recvmmsg(if_defined, if_not_defined) if_defined
#define IF_DEFINED_recvmsg(if_defined, if_not_defined) if_defined
#define IF_DEFINED_restart_syscall(if_defined, if_not_defined) if_defined
#define IF_DEFINED_rseq(if_defined, if_not_defined) if_defined
#define IF_DEFINED_rt_sigaction(if_defined, if_not_defined) if_defined
#define IF_DEFINED_rt_sigprocmask(if_defined, if_not_defined) if_defined
#define IF_DEFINED_rt_sigreturn(if_defined, if_not_defined) if_defined
#define IF_DEFINED_sched_getaffinity(if_defined, if_not_defined) if_defined
#define IF_DEFINED_sched_yield(if_defined, if_not_defined) if_defined
#define IF_DEFINED_sendmmsg(if_defined, if_not_defined) if_defined
#define IF_DEFINED_sendmsg(if_defined, if_not_defined) if_defined
#define IF_DEFINED_sendto(if_defined, if_not_defined) if_defined
#define IF_DEFINED_set_robust_list(if_defined, if_not_defined) if_defined
#define IF_DEFINED_set_tid_address(if_defined, if_not_defined) if_defined
#define IF_DEFINED_setsockopt(if_defined, if_not_defined) if_defined
#define IF_DEFINED_sigaltstack(if_defined, if_not_defined) if_defined
#define IF_DEFINED_socketpair(if_defined, if_not_defined) if_defined
#define IF_DEFINED_sysinfo(if_defined, if_not_defined) if_defined
#define IF_DEFINED_tgkill(if_defined, if_not_defined) if_defined
#define IF_DEFINED_umask(if_defined, if_not_defined) if_defined
#define IF_DEFINED_uname(if_defined, if_not_defined) if_defined
#define IF_DEFINED_write(if_defined, if_not_defined) if_defined
#ifndef __NR_accept4
# undef IF_DEFINED_accept4
# define IF_DEFINED_accept4(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_clock_getres
# undef IF_DEFINED_clock_getres
# define IF_DEFINED_clock_getres(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_clock_gettime
# undef IF_DEFINED_clock_gettime
# define IF_DEFINED_clock_gettime(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_clock_nanosleep
# undef IF_DEFINED_clock_nanosleep
# define IF_DEFINED_clock_nanosleep(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_clone3
# undef IF_DEFINED_clone3
# define IF_DEFINED_clone3(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_dup3
# undef IF_DEFINED_dup3
# define IF_DEFINED_dup3(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_fcntl64
# undef IF_DEFINED_fcntl64
# define IF_DEFINED_fcntl64(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_fstat
# undef IF_DEFINED_fstat
# define IF_DEFINED_fstat(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_futex_time64
# undef IF_DEFINED_futex_time64
# define IF_DEFINED_futex_time64(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_getcpu
# undef IF_DEFINED_getcpu
# define IF_DEFINED_getcpu(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_getrandom
# undef IF_DEFINED_getrandom
# define IF_DEFINED_getrandom(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_membarrier
# undef IF_DEFINED_membarrier
# define IF_DEFINED_membarrier(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_memfd_create
# undef IF_DEFINED_memfd_create
# define IF_DEFINED_memfd_create(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_mmap2
# undef IF_DEFINED_mmap2
# define IF_DEFINED_mmap2(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_pipe2
# undef IF_DEFINED_pipe2
# define IF_DEFINED_pipe2(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_poll
# undef IF_DEFINED_poll
# define IF_DEFINED_poll(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_ppoll
# undef IF_DEFINED_ppoll
# define IF_DEFINED_ppoll(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_prlimit64
# undef IF_DEFINED_prlimit64
# define IF_DEFINED_prlimit64(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_pselect6
# undef IF_DEFINED_pselect6
# define IF_DEFINED_pselect6(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_pwrite64
# undef IF_DEFINED_pwrite64
# define IF_DEFINED_pwrite64(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_recvmmsg
# undef IF_DEFINED_recvmmsg
# define IF_DEFINED_recvmmsg(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_restart_syscall
# undef IF_DEFINED_restart_syscall
# define IF_DEFINED_restart_syscall(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_rseq
# undef IF_DEFINED_rseq
# define IF_DEFINED_rseq(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_sendmmsg
# undef IF_DEFINED_sendmmsg
# define IF_DEFINED_sendmmsg(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_sigaltstack
# undef IF_DEFINED_sigaltstack
# define IF_DEFINED_sigaltstack(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_sysinfo
# undef IF_DEFINED_sysinfo
# define IF_DEFINED_sysinfo(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_umask
# undef IF_DEFINED_umask
# define IF_DEFINED_umask(if_defined, if_not_defined) if_not_defined
#endif
static size_t safe_string_length(char const* string)
{
size_t length = 0;
while (string[length] != '\0')
++length;
return length;
}
static void write_bytes(char const* string, size_t length)
{
while (length > 0) {
auto written = write(STDERR_FILENO, string, length);
if (written <= 0)
return;
string += written;
length -= static_cast<size_t>(written);
}
}
static void write_string(char const* string)
{
write_bytes(string, safe_string_length(string));
}
static void write_unsigned(u64 value)
{
char buffer[20];
size_t index = sizeof(buffer);
do {
buffer[--index] = static_cast<char>('0' + (value % 10));
value /= 10;
} while (value > 0);
write_bytes(buffer + index, sizeof(buffer) - index);
}
static void write_hex(u64 value)
{
constexpr char const* digits = "0123456789abcdef";
char buffer[16];
size_t index = sizeof(buffer);
do {
buffer[--index] = digits[value & 0xf];
value >>= 4;
} while (value > 0);
write_string("0x");
write_bytes(buffer + index, sizeof(buffer) - index);
}
#define CASE_SYSCALL_NAME(name) \
case __NR_##name: \
return #name
static char const* syscall_name(long syscall_number)
{
switch (syscall_number) {
#ifdef __NR_accept
CASE_SYSCALL_NAME(accept);
#endif
#ifdef __NR_accept4
CASE_SYSCALL_NAME(accept4);
#endif
#ifdef __NR_access
CASE_SYSCALL_NAME(access);
#endif
#ifdef __NR_bind
CASE_SYSCALL_NAME(bind);
#endif
#ifdef __NR_clone
CASE_SYSCALL_NAME(clone);
#endif
#ifdef __NR_clone3
CASE_SYSCALL_NAME(clone3);
#endif
#ifdef __NR_connect
CASE_SYSCALL_NAME(connect);
#endif
#ifdef __NR_execve
CASE_SYSCALL_NAME(execve);
#endif
#ifdef __NR_execveat
CASE_SYSCALL_NAME(execveat);
#endif
#ifdef __NR_fcntl
CASE_SYSCALL_NAME(fcntl);
#endif
#ifdef __NR_fcntl64
CASE_SYSCALL_NAME(fcntl64);
#endif
#ifdef __NR_fstat
CASE_SYSCALL_NAME(fstat);
#endif
#ifdef __NR_futex
CASE_SYSCALL_NAME(futex);
#endif
#ifdef __NR_getdents64
CASE_SYSCALL_NAME(getdents64);
#endif
#ifdef __NR_ioctl
CASE_SYSCALL_NAME(ioctl);
#endif
#ifdef __NR_link
CASE_SYSCALL_NAME(link);
#endif
#ifdef __NR_linkat
CASE_SYSCALL_NAME(linkat);
#endif
#ifdef __NR_memfd_create
CASE_SYSCALL_NAME(memfd_create);
#endif
#ifdef __NR_mkdir
CASE_SYSCALL_NAME(mkdir);
#endif
#ifdef __NR_mkdirat
CASE_SYSCALL_NAME(mkdirat);
#endif
#ifdef __NR_mmap
CASE_SYSCALL_NAME(mmap);
#endif
#ifdef __NR_mmap2
CASE_SYSCALL_NAME(mmap2);
#endif
#ifdef __NR_mprotect
CASE_SYSCALL_NAME(mprotect);
#endif
#ifdef __NR_newfstatat
CASE_SYSCALL_NAME(newfstatat);
#endif
#ifdef __NR_open
CASE_SYSCALL_NAME(open);
#endif
#ifdef __NR_openat
CASE_SYSCALL_NAME(openat);
#endif
#ifdef __NR_prctl
CASE_SYSCALL_NAME(prctl);
#endif
#ifdef __NR_readlink
CASE_SYSCALL_NAME(readlink);
#endif
#ifdef __NR_readlinkat
CASE_SYSCALL_NAME(readlinkat);
#endif
#ifdef __NR_rename
CASE_SYSCALL_NAME(rename);
#endif
#ifdef __NR_renameat
CASE_SYSCALL_NAME(renameat);
#endif
#ifdef __NR_renameat2
CASE_SYSCALL_NAME(renameat2);
#endif
#ifdef __NR_rmdir
CASE_SYSCALL_NAME(rmdir);
#endif
#ifdef __NR_sched_setaffinity
CASE_SYSCALL_NAME(sched_setaffinity);
#endif
#ifdef __NR_sched_setparam
CASE_SYSCALL_NAME(sched_setparam);
#endif
#ifdef __NR_sched_setscheduler
CASE_SYSCALL_NAME(sched_setscheduler);
#endif
#ifdef __NR_seccomp
CASE_SYSCALL_NAME(seccomp);
#endif
#ifdef __NR_socket
CASE_SYSCALL_NAME(socket);
#endif
#ifdef __NR_statx
CASE_SYSCALL_NAME(statx);
#endif
#ifdef __NR_unlink
CASE_SYSCALL_NAME(unlink);
#endif
#ifdef __NR_unlinkat
CASE_SYSCALL_NAME(unlinkat);
#endif
default:
return "unknown";
}
}
#undef CASE_SYSCALL_NAME
static void handle_sigsys(int, siginfo_t* info, void* context)
{
write_string("Sandbox violation: disallowed syscall ");
write_string(syscall_name(info->si_syscall));
write_string(" (");
write_unsigned(static_cast<u64>(info->si_syscall));
write_string("), architecture ");
write_hex(info->si_arch);
write_string(", instruction pointer ");
write_hex(reinterpret_cast<FlatPtr>(info->si_call_addr));
#if defined(__x86_64__)
auto* machine_context = &static_cast<ucontext_t*>(context)->uc_mcontext;
write_string(", args [");
write_hex(static_cast<u64>(machine_context->gregs[REG_RDI]));
write_string(", ");
write_hex(static_cast<u64>(machine_context->gregs[REG_RSI]));
write_string(", ");
write_hex(static_cast<u64>(machine_context->gregs[REG_RDX]));
write_string(", ");
write_hex(static_cast<u64>(machine_context->gregs[REG_R10]));
write_string(", ");
write_hex(static_cast<u64>(machine_context->gregs[REG_R8]));
write_string(", ");
write_hex(static_cast<u64>(machine_context->gregs[REG_R9]));
write_string("]");
#elif defined(__aarch64__)
auto* machine_context = &static_cast<ucontext_t*>(context)->uc_mcontext;
write_string(", args [");
for (size_t i = 0; i < 6; ++i) {
if (i != 0)
write_string(", ");
write_hex(static_cast<u64>(machine_context->regs[i]));
}
write_string("]");
#endif
write_string("\n");
struct sigaction action {};
action.sa_handler = SIG_DFL;
sigemptyset(&action.sa_mask);
sigaction(SIGSYS, &action, nullptr);
raise(SIGSYS);
_exit(128 + SIGSYS);
}
static ErrorOr<void> install_sigsys_handler()
{
struct sigaction action {};
action.sa_sigaction = handle_sigsys;
action.sa_flags = SA_SIGINFO;
if (sigemptyset(&action.sa_mask) < 0)
return Error::from_syscall("sigemptyset"sv, errno);
if (sigaction(SIGSYS, &action, nullptr) < 0)
return Error::from_syscall("sigaction(SIGSYS)"sv, errno);
return {};
}
}
SeccompPolicy::SeccompPolicy()
{
append_architecture_check();
append_load_syscall_number();
}
void SeccompPolicy::append(sock_filter filter)
{
MUST(m_filter.try_append(filter));
}
void SeccompPolicy::append_architecture_check()
{
append(SECCOMP_LOAD_ARCHITECTURE);
append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, audit_architecture, 1, 0));
append(SECCOMP_TRAP);
}
void SeccompPolicy::append_load_syscall_number()
{
append(SECCOMP_LOAD_SYSCALL_NR);
}
void SeccompPolicy::append_kill()
{
append(SECCOMP_TRAP);
}
void SeccompPolicy::deny_readonly_filesystem_probes()
{
#ifdef __NR_open
append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, __NR_open, 0, 5));
append(SECCOMP_LOAD_ARGUMENT(1));
append(BPF_STMT(BPF_ALU | BPF_AND | BPF_K, ~read_only_open_flags));
append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, 0, 0, 1));
append(SECCOMP_ERRNO(EACCES));
append(SECCOMP_LOAD_SYSCALL_NR);
#endif
#ifdef __NR_openat
append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, __NR_openat, 0, 5));
append(SECCOMP_LOAD_ARGUMENT(2));
append(BPF_STMT(BPF_ALU | BPF_AND | BPF_K, ~read_only_open_flags));
append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, 0, 0, 1));
append(SECCOMP_ERRNO(EACCES));
append(SECCOMP_LOAD_SYSCALL_NR);
#endif
}
void SeccompPolicy::allow_file_descriptor_operations()
{
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, read);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, write);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, close);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, fstat);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, dup);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, dup3);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, pipe2);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, ftruncate);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, lseek);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, pwrite64);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, memfd_create);
append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, __NR_fcntl, 0, 5));
append(SECCOMP_LOAD_ARGUMENT(1));
append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, F_GETFD, 0, 1));
append(SECCOMP_ALLOW);
append(SECCOMP_LOAD_SYSCALL_NR);
append(BPF_STMT(BPF_ALU | BPF_ADD | BPF_K, 0));
append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, __NR_fcntl, 0, 5));
append(SECCOMP_LOAD_ARGUMENT(1));
append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, F_SETFD, 0, 1));
append(SECCOMP_ALLOW);
append(SECCOMP_LOAD_SYSCALL_NR);
append(BPF_STMT(BPF_ALU | BPF_ADD | BPF_K, 0));
append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, __NR_fcntl, 0, 5));
append(SECCOMP_LOAD_ARGUMENT(1));
append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, F_GETFL, 0, 1));
append(SECCOMP_ALLOW);
append(SECCOMP_LOAD_SYSCALL_NR);
append(BPF_STMT(BPF_ALU | BPF_ADD | BPF_K, 0));
append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, __NR_fcntl, 0, 5));
append(SECCOMP_LOAD_ARGUMENT(1));
append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, F_SETFL, 0, 1));
append(SECCOMP_ALLOW);
append(SECCOMP_LOAD_SYSCALL_NR);
append(BPF_STMT(BPF_ALU | BPF_ADD | BPF_K, 0));
append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, __NR_ioctl, 0, 5));
append(SECCOMP_LOAD_ARGUMENT(1));
append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, FIONBIO, 0, 1));
append(SECCOMP_ALLOW);
append(SECCOMP_LOAD_SYSCALL_NR);
append(BPF_STMT(BPF_ALU | BPF_ADD | BPF_K, 0));
}
void SeccompPolicy::allow_ipc()
{
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, poll);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, ppoll);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, pselect6);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, recvmsg);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, recvfrom);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, recvmmsg);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, sendmsg);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, sendto);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, sendmmsg);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, socketpair);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, getsockopt);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, setsockopt);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, getsockname);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, getpeername);
}
void SeccompPolicy::allow_memory_without_executable_mappings()
{
append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, __NR_mmap, 0, 5));
append(SECCOMP_LOAD_ARGUMENT(2));
append(BPF_STMT(BPF_ALU | BPF_AND | BPF_K, PROT_EXEC));
append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, 0, 0, 1));
append(SECCOMP_ALLOW);
append(SECCOMP_LOAD_SYSCALL_NR);
#ifdef __NR_mmap2
append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, __NR_mmap2, 0, 5));
append(SECCOMP_LOAD_ARGUMENT(2));
append(BPF_STMT(BPF_ALU | BPF_AND | BPF_K, PROT_EXEC));
append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, 0, 0, 1));
append(SECCOMP_ALLOW);
append(SECCOMP_LOAD_SYSCALL_NR);
#endif
append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, __NR_mprotect, 0, 5));
append(SECCOMP_LOAD_ARGUMENT(2));
append(BPF_STMT(BPF_ALU | BPF_AND | BPF_K, PROT_EXEC));
append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, 0, 0, 1));
append(SECCOMP_ALLOW);
append(SECCOMP_LOAD_SYSCALL_NR);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, mremap);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, munmap);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, madvise);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, brk);
}
void SeccompPolicy::allow_threads()
{
#ifdef __NR_clone
append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, __NR_clone, 0, 8));
append(SECCOMP_LOAD_ARGUMENT(0));
append(BPF_STMT(BPF_ALU | BPF_AND | BPF_K, thread_clone_required_flags));
append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, thread_clone_required_flags, 0, 4));
append(SECCOMP_LOAD_ARGUMENT(0));
append(BPF_STMT(BPF_ALU | BPF_AND | BPF_K, ~thread_clone_allowed_flags));
append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, 0, 0, 1));
append(SECCOMP_ALLOW);
append(SECCOMP_LOAD_SYSCALL_NR);
#endif
#ifdef __NR_clone3
append(BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, __NR_clone3, 0, 1));
append(SECCOMP_ERRNO(ENOSYS));
#endif
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, futex);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, futex_time64);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, set_tid_address);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, set_robust_list);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, rseq);
}
void SeccompPolicy::allow_signals()
{
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, rt_sigaction);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, rt_sigprocmask);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, rt_sigreturn);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, sigaltstack);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, tgkill);
}
void SeccompPolicy::allow_clocks()
{
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, clock_gettime);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, clock_getres);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, clock_nanosleep);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, nanosleep);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, restart_syscall);
}
void SeccompPolicy::allow_process_metadata()
{
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, getpid);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, gettid);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, getuid);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, geteuid);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, getgid);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, getegid);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, getrandom);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, gettimeofday);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, getrlimit);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, prlimit64);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, getrusage);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, sched_getaffinity);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, sched_yield);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, getcpu);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, membarrier);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, sysinfo);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, uname);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, umask);
}
void SeccompPolicy::allow_common_runtime()
{
allow_memory_without_executable_mappings();
allow_threads();
allow_signals();
allow_clocks();
allow_process_metadata();
allow_prctl();
allow_exit();
}
void SeccompPolicy::allow_prctl()
{
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, prctl);
}
void SeccompPolicy::allow_exit()
{
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, exit);
SECCOMP_APPEND_ALLOW_SYSCALL_IF_DEFINED(*this, exit_group);
}
ErrorOr<void> SeccompPolicy::install()
{
TRY(install_sigsys_handler());
append_kill();
sock_fprog program {
.len = static_cast<unsigned short>(m_filter.size()),
.filter = m_filter.data(),
};
#ifdef __NR_seccomp
if (syscall(__NR_seccomp, SECCOMP_SET_MODE_FILTER, SECCOMP_FILTER_FLAG_TSYNC, &program) < 0)
return Error::from_syscall("seccomp(SECCOMP_SET_MODE_FILTER)"sv, errno);
#else
if (prctl(PR_SET_SECCOMP, SECCOMP_MODE_FILTER, &program) < 0)
return Error::from_syscall("prctl(PR_SET_SECCOMP)"sv, errno);
#endif
return {};
}
}

View file

@ -0,0 +1,42 @@
/*
* Copyright (c) 2026-present, the Ladybird developers.
*
* SPDX-License-Identifier: BSD-2-Clause
*/
#pragma once
#include <AK/Error.h>
#include <AK/Vector.h>
#include <linux/filter.h>
namespace Sandbox {
class SeccompPolicy {
public:
SeccompPolicy();
void deny_readonly_filesystem_probes();
void allow_file_descriptor_operations();
void allow_ipc();
void allow_memory_without_executable_mappings();
void allow_threads();
void allow_signals();
void allow_clocks();
void allow_process_metadata();
void allow_common_runtime();
void allow_prctl();
void allow_exit();
[[nodiscard]] ErrorOr<void> install();
private:
void append(sock_filter);
void append_architecture_check();
void append_load_syscall_number();
void append_kill();
Vector<sock_filter> m_filter;
};
}

View file

@ -31,7 +31,7 @@ endif()
target_include_directories(imagedecoderservice PRIVATE ${CMAKE_CURRENT_BINARY_DIR}/../..)
target_include_directories(imagedecoderservice PRIVATE ${LADYBIRD_SOURCE_DIR}/Services/)
target_link_libraries(ImageDecoder PRIVATE imagedecoderservice LibCore LibMain LibThreading)
target_link_libraries(ImageDecoder PRIVATE imagedecoderservice LibCore LibMain LibSandbox LibThreading)
target_link_libraries(imagedecoderservice PRIVATE LibCore LibGfx LibIPC LibImageDecoderClient LibMain LibSync LibThreading)
if (WIN32)

View file

@ -4,497 +4,25 @@
* SPDX-License-Identifier: BSD-2-Clause
*/
#include <AK/Platform.h>
#include <AK/ScopeGuard.h>
#include <ImageDecoder/Sandbox.h>
#include <errno.h>
#include <fcntl.h>
#include <linux/audit.h>
#include <linux/filter.h>
#include <linux/landlock.h>
#include <linux/sched.h>
#include <linux/seccomp.h>
#include <stddef.h>
#include <sys/ioctl.h>
#include <sys/mman.h>
#include <sys/prctl.h>
#include <sys/socket.h>
#include <sys/syscall.h>
#include <unistd.h>
#if defined(__GLIBC__)
# include <malloc.h>
#endif
#include <LibSandbox/Sandbox.h>
#include <LibSandbox/Seccomp.h>
namespace ImageDecoder {
namespace {
#if defined(__x86_64__)
static constexpr u32 audit_architecture = AUDIT_ARCH_X86_64;
#elif defined(__aarch64__)
static constexpr u32 audit_architecture = AUDIT_ARCH_AARCH64;
#elif defined(__riscv) && __riscv_xlen == 64
static constexpr u32 audit_architecture = AUDIT_ARCH_RISCV64;
#else
# error "Add 64-bit seccomp audit architecture for this Linux architecture"
#endif
static constexpr u32 thread_clone_required_flags = CLONE_VM | CLONE_FS | CLONE_FILES | CLONE_SIGHAND | CLONE_THREAD;
static constexpr u32 thread_clone_allowed_flags = thread_clone_required_flags
| CLONE_SYSVSEM
| CLONE_SETTLS
| CLONE_PARENT_SETTID
| CLONE_CHILD_CLEARTID
| CLONE_DETACHED
| CLONE_CHILD_SETTID;
#define SECCOMP_LOAD_SYSCALL_NR BPF_STMT(BPF_LD | BPF_W | BPF_ABS, static_cast<unsigned int>(offsetof(seccomp_data, nr)))
#define SECCOMP_LOAD_ARCHITECTURE BPF_STMT(BPF_LD | BPF_W | BPF_ABS, static_cast<unsigned int>(offsetof(seccomp_data, arch)))
#define SECCOMP_LOAD_ARGUMENT(index) BPF_STMT(BPF_LD | BPF_W | BPF_ABS, static_cast<unsigned int>(offsetof(seccomp_data, args[(index)])))
#define SECCOMP_ALLOW BPF_STMT(BPF_RET | BPF_K, SECCOMP_RET_ALLOW)
#define SECCOMP_KILL BPF_STMT(BPF_RET | BPF_K, SECCOMP_RET_KILL_PROCESS)
#define SECCOMP_ERRNO(error) BPF_STMT(BPF_RET | BPF_K, SECCOMP_RET_ERRNO | ((error) & SECCOMP_RET_DATA))
#define SECCOMP_ALLOW_SYSCALL(name) \
BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, __NR_##name, 0, 1), SECCOMP_ALLOW
#define SECCOMP_ALLOW_NOTHING BPF_STMT(BPF_ALU | BPF_ADD | BPF_K, 0)
#define SECCOMP_ALLOW_SYSCALL_IF_DEFINED(name) \
IF_DEFINED_##name(SECCOMP_ALLOW_SYSCALL(name), SECCOMP_ALLOW_NOTHING)
#define SECCOMP_DENY_READONLY_OPEN_SYSCALL(name, flags_argument_index) \
BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, __NR_##name, 0, 5), \
SECCOMP_LOAD_ARGUMENT(flags_argument_index), BPF_STMT(BPF_ALU | BPF_AND | BPF_K, ~read_only_open_flags), \
BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, 0, 0, 1), SECCOMP_ERRNO(EACCES), SECCOMP_LOAD_SYSCALL_NR
#define SECCOMP_ALLOW_THREAD_CLONE_SYSCALL(name) \
BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, __NR_##name, 0, 8), \
SECCOMP_LOAD_ARGUMENT(0), BPF_STMT(BPF_ALU | BPF_AND | BPF_K, thread_clone_required_flags), \
BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, thread_clone_required_flags, 0, 4), SECCOMP_LOAD_ARGUMENT(0), \
BPF_STMT(BPF_ALU | BPF_AND | BPF_K, ~thread_clone_allowed_flags), \
BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, 0, 0, 1), SECCOMP_ALLOW, SECCOMP_LOAD_SYSCALL_NR
#ifdef O_LARGEFILE
static constexpr unsigned read_only_open_flags = O_CLOEXEC | O_LARGEFILE;
#else
static constexpr unsigned read_only_open_flags = O_CLOEXEC;
#endif
#define IF_DEFINED_accept4(if_defined, if_not_defined) if_defined
#define IF_DEFINED_brk(if_defined, if_not_defined) if_defined
#define IF_DEFINED_clock_getres(if_defined, if_not_defined) if_defined
#define IF_DEFINED_clock_gettime(if_defined, if_not_defined) if_defined
#define IF_DEFINED_clock_nanosleep(if_defined, if_not_defined) if_defined
#define IF_DEFINED_clone(if_defined, if_not_defined) if_defined
#define IF_DEFINED_clone3(if_defined, if_not_defined) if_defined
#define IF_DEFINED_close(if_defined, if_not_defined) if_defined
#define IF_DEFINED_dup(if_defined, if_not_defined) if_defined
#define IF_DEFINED_dup3(if_defined, if_not_defined) if_defined
#define IF_DEFINED_exit(if_defined, if_not_defined) if_defined
#define IF_DEFINED_exit_group(if_defined, if_not_defined) if_defined
#define IF_DEFINED_fcntl(if_defined, if_not_defined) if_defined
#define IF_DEFINED_fcntl64(if_defined, if_not_defined) if_defined
#define IF_DEFINED_fstat(if_defined, if_not_defined) if_defined
#define IF_DEFINED_ftruncate(if_defined, if_not_defined) if_defined
#define IF_DEFINED_futex(if_defined, if_not_defined) if_defined
#define IF_DEFINED_futex_time64(if_defined, if_not_defined) if_defined
#define IF_DEFINED_getcpu(if_defined, if_not_defined) if_defined
#define IF_DEFINED_getegid(if_defined, if_not_defined) if_defined
#define IF_DEFINED_geteuid(if_defined, if_not_defined) if_defined
#define IF_DEFINED_getgid(if_defined, if_not_defined) if_defined
#define IF_DEFINED_getpeername(if_defined, if_not_defined) if_defined
#define IF_DEFINED_getpid(if_defined, if_not_defined) if_defined
#define IF_DEFINED_getrandom(if_defined, if_not_defined) if_defined
#define IF_DEFINED_getrlimit(if_defined, if_not_defined) if_defined
#define IF_DEFINED_getrusage(if_defined, if_not_defined) if_defined
#define IF_DEFINED_getsockname(if_defined, if_not_defined) if_defined
#define IF_DEFINED_getsockopt(if_defined, if_not_defined) if_defined
#define IF_DEFINED_gettid(if_defined, if_not_defined) if_defined
#define IF_DEFINED_gettimeofday(if_defined, if_not_defined) if_defined
#define IF_DEFINED_getuid(if_defined, if_not_defined) if_defined
#define IF_DEFINED_ioctl(if_defined, if_not_defined) if_defined
#define IF_DEFINED_lseek(if_defined, if_not_defined) if_defined
#define IF_DEFINED_madvise(if_defined, if_not_defined) if_defined
#define IF_DEFINED_membarrier(if_defined, if_not_defined) if_defined
#define IF_DEFINED_memfd_create(if_defined, if_not_defined) if_defined
#define IF_DEFINED_mmap(if_defined, if_not_defined) if_defined
#define IF_DEFINED_mmap2(if_defined, if_not_defined) if_defined
#define IF_DEFINED_mprotect(if_defined, if_not_defined) if_defined
#define IF_DEFINED_mremap(if_defined, if_not_defined) if_defined
#define IF_DEFINED_munmap(if_defined, if_not_defined) if_defined
#define IF_DEFINED_nanosleep(if_defined, if_not_defined) if_defined
#define IF_DEFINED_pipe2(if_defined, if_not_defined) if_defined
#define IF_DEFINED_poll(if_defined, if_not_defined) if_defined
#define IF_DEFINED_ppoll(if_defined, if_not_defined) if_defined
#define IF_DEFINED_prctl(if_defined, if_not_defined) if_defined
#define IF_DEFINED_prlimit64(if_defined, if_not_defined) if_defined
#define IF_DEFINED_pselect6(if_defined, if_not_defined) if_defined
#define IF_DEFINED_pwrite64(if_defined, if_not_defined) if_defined
#define IF_DEFINED_read(if_defined, if_not_defined) if_defined
#define IF_DEFINED_recvfrom(if_defined, if_not_defined) if_defined
#define IF_DEFINED_recvmmsg(if_defined, if_not_defined) if_defined
#define IF_DEFINED_recvmsg(if_defined, if_not_defined) if_defined
#define IF_DEFINED_restart_syscall(if_defined, if_not_defined) if_defined
#define IF_DEFINED_rseq(if_defined, if_not_defined) if_defined
#define IF_DEFINED_rt_sigaction(if_defined, if_not_defined) if_defined
#define IF_DEFINED_rt_sigprocmask(if_defined, if_not_defined) if_defined
#define IF_DEFINED_rt_sigreturn(if_defined, if_not_defined) if_defined
#define IF_DEFINED_sched_getaffinity(if_defined, if_not_defined) if_defined
#define IF_DEFINED_sched_yield(if_defined, if_not_defined) if_defined
#define IF_DEFINED_sendmmsg(if_defined, if_not_defined) if_defined
#define IF_DEFINED_sendmsg(if_defined, if_not_defined) if_defined
#define IF_DEFINED_sendto(if_defined, if_not_defined) if_defined
#define IF_DEFINED_set_robust_list(if_defined, if_not_defined) if_defined
#define IF_DEFINED_set_tid_address(if_defined, if_not_defined) if_defined
#define IF_DEFINED_setsockopt(if_defined, if_not_defined) if_defined
#define IF_DEFINED_sigaltstack(if_defined, if_not_defined) if_defined
#define IF_DEFINED_socketpair(if_defined, if_not_defined) if_defined
#define IF_DEFINED_sysinfo(if_defined, if_not_defined) if_defined
#define IF_DEFINED_tgkill(if_defined, if_not_defined) if_defined
#define IF_DEFINED_umask(if_defined, if_not_defined) if_defined
#define IF_DEFINED_uname(if_defined, if_not_defined) if_defined
#define IF_DEFINED_write(if_defined, if_not_defined) if_defined
#ifndef __NR_accept4
# undef IF_DEFINED_accept4
# define IF_DEFINED_accept4(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_clock_getres
# undef IF_DEFINED_clock_getres
# define IF_DEFINED_clock_getres(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_clock_gettime
# undef IF_DEFINED_clock_gettime
# define IF_DEFINED_clock_gettime(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_clock_nanosleep
# undef IF_DEFINED_clock_nanosleep
# define IF_DEFINED_clock_nanosleep(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_clone3
# undef IF_DEFINED_clone3
# define IF_DEFINED_clone3(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_dup3
# undef IF_DEFINED_dup3
# define IF_DEFINED_dup3(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_fcntl64
# undef IF_DEFINED_fcntl64
# define IF_DEFINED_fcntl64(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_fstat
# undef IF_DEFINED_fstat
# define IF_DEFINED_fstat(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_futex_time64
# undef IF_DEFINED_futex_time64
# define IF_DEFINED_futex_time64(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_getcpu
# undef IF_DEFINED_getcpu
# define IF_DEFINED_getcpu(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_getrandom
# undef IF_DEFINED_getrandom
# define IF_DEFINED_getrandom(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_membarrier
# undef IF_DEFINED_membarrier
# define IF_DEFINED_membarrier(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_memfd_create
# undef IF_DEFINED_memfd_create
# define IF_DEFINED_memfd_create(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_mmap2
# undef IF_DEFINED_mmap2
# define IF_DEFINED_mmap2(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_pipe2
# undef IF_DEFINED_pipe2
# define IF_DEFINED_pipe2(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_poll
# undef IF_DEFINED_poll
# define IF_DEFINED_poll(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_ppoll
# undef IF_DEFINED_ppoll
# define IF_DEFINED_ppoll(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_prlimit64
# undef IF_DEFINED_prlimit64
# define IF_DEFINED_prlimit64(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_pselect6
# undef IF_DEFINED_pselect6
# define IF_DEFINED_pselect6(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_pwrite64
# undef IF_DEFINED_pwrite64
# define IF_DEFINED_pwrite64(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_recvmmsg
# undef IF_DEFINED_recvmmsg
# define IF_DEFINED_recvmmsg(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_restart_syscall
# undef IF_DEFINED_restart_syscall
# define IF_DEFINED_restart_syscall(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_rseq
# undef IF_DEFINED_rseq
# define IF_DEFINED_rseq(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_sendmmsg
# undef IF_DEFINED_sendmmsg
# define IF_DEFINED_sendmmsg(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_sigaltstack
# undef IF_DEFINED_sigaltstack
# define IF_DEFINED_sigaltstack(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_sysinfo
# undef IF_DEFINED_sysinfo
# define IF_DEFINED_sysinfo(if_defined, if_not_defined) if_not_defined
#endif
#ifndef __NR_umask
# undef IF_DEFINED_umask
# define IF_DEFINED_umask(if_defined, if_not_defined) if_not_defined
#endif
#define SECCOMP_ALLOW_MMAP_WITHOUT_EXEC(name, protection_argument) \
BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, __NR_##name, 0, 5), \
SECCOMP_LOAD_ARGUMENT(protection_argument), \
BPF_STMT(BPF_ALU | BPF_AND | BPF_K, PROT_EXEC), \
BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, 0, 0, 1), \
SECCOMP_ALLOW, \
SECCOMP_LOAD_SYSCALL_NR
#define SECCOMP_ALLOW_IOCTL_COMMAND(command) \
BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, __NR_ioctl, 0, 5), \
SECCOMP_LOAD_ARGUMENT(1), \
BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, command, 0, 1), \
SECCOMP_ALLOW, \
SECCOMP_LOAD_SYSCALL_NR, \
BPF_STMT(BPF_ALU | BPF_ADD | BPF_K, 0)
#define SECCOMP_ALLOW_FCNTL_COMMAND(command) \
BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, __NR_fcntl, 0, 5), \
SECCOMP_LOAD_ARGUMENT(1), \
BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, command, 0, 1), \
SECCOMP_ALLOW, \
SECCOMP_LOAD_SYSCALL_NR, \
BPF_STMT(BPF_ALU | BPF_ADD | BPF_K, 0)
ErrorOr<void> install_no_new_privileges()
{
if (prctl(PR_SET_NO_NEW_PRIVS, 1, 0, 0, 0) < 0)
return Error::from_syscall("prctl(PR_SET_NO_NEW_PRIVS)"sv, errno);
return {};
}
ErrorOr<void> configure_runtime_for_sandbox()
{
#if defined(__GLIBC__) && defined(M_ARENA_MAX) && !defined(HAS_ADDRESS_SANITIZER)
if (mallopt(M_ARENA_MAX, 4) == 0)
return Error::from_string_literal("mallopt(M_ARENA_MAX) failed");
#endif
return {};
}
ErrorOr<void> restrict_filesystem_with_landlock()
{
#if defined(__NR_landlock_create_ruleset) && defined(__NR_landlock_restrict_self)
auto landlock_abi = syscall(__NR_landlock_create_ruleset, nullptr, 0, LANDLOCK_CREATE_RULESET_VERSION);
if (landlock_abi < 0) {
if (errno == ENOSYS || errno == EOPNOTSUPP || errno == EINVAL)
return {};
return Error::from_syscall("landlock_create_ruleset(LANDLOCK_CREATE_RULESET_VERSION)"sv, errno);
}
if (landlock_abi == 0)
return {};
landlock_ruleset_attr ruleset_attributes {};
ruleset_attributes.handled_access_fs = LANDLOCK_ACCESS_FS_EXECUTE
| LANDLOCK_ACCESS_FS_WRITE_FILE
| LANDLOCK_ACCESS_FS_READ_FILE
| LANDLOCK_ACCESS_FS_READ_DIR
| LANDLOCK_ACCESS_FS_REMOVE_DIR
| LANDLOCK_ACCESS_FS_REMOVE_FILE
| LANDLOCK_ACCESS_FS_MAKE_CHAR
| LANDLOCK_ACCESS_FS_MAKE_DIR
| LANDLOCK_ACCESS_FS_MAKE_REG
| LANDLOCK_ACCESS_FS_MAKE_SOCK
| LANDLOCK_ACCESS_FS_MAKE_FIFO
| LANDLOCK_ACCESS_FS_MAKE_BLOCK
| LANDLOCK_ACCESS_FS_MAKE_SYM;
# ifdef LANDLOCK_ACCESS_FS_REFER
if (landlock_abi >= 2)
ruleset_attributes.handled_access_fs |= LANDLOCK_ACCESS_FS_REFER;
# endif
# ifdef LANDLOCK_ACCESS_FS_TRUNCATE
if (landlock_abi >= 3)
ruleset_attributes.handled_access_fs |= LANDLOCK_ACCESS_FS_TRUNCATE;
# endif
# if defined(LANDLOCK_ACCESS_NET_BIND_TCP) && defined(LANDLOCK_ACCESS_NET_CONNECT_TCP)
if (landlock_abi >= 4)
ruleset_attributes.handled_access_net = LANDLOCK_ACCESS_NET_BIND_TCP | LANDLOCK_ACCESS_NET_CONNECT_TCP;
auto ruleset_attributes_size = landlock_abi >= 4
? sizeof(ruleset_attributes)
: offsetof(landlock_ruleset_attr, handled_access_net);
# else
auto ruleset_attributes_size = sizeof(ruleset_attributes);
# endif
auto ruleset_fd = syscall(__NR_landlock_create_ruleset, &ruleset_attributes, ruleset_attributes_size, 0);
if (ruleset_fd < 0)
return Error::from_syscall("landlock_create_ruleset"sv, errno);
ArmedScopeGuard close_ruleset_fd = [&] {
close(static_cast<int>(ruleset_fd));
};
if (syscall(__NR_landlock_restrict_self, ruleset_fd, 0) < 0)
return Error::from_syscall("landlock_restrict_self"sv, errno);
#endif
return {};
}
ErrorOr<void> install_seccomp_filter()
{
sock_filter filter[] = {
SECCOMP_LOAD_ARCHITECTURE,
BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, audit_architecture, 1, 0),
SECCOMP_KILL,
SECCOMP_LOAD_SYSCALL_NR,
// Deny read-only runtime probes without granting filesystem access.
#ifdef __NR_open
SECCOMP_DENY_READONLY_OPEN_SYSCALL(open, 1),
#endif
#ifdef __NR_openat
SECCOMP_DENY_READONLY_OPEN_SYSCALL(openat, 2),
#endif
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(read),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(write),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(close),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(poll),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(ppoll),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(pselect6),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(recvmsg),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(recvfrom),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(recvmmsg),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(sendmsg),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(sendto),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(sendmmsg),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(socketpair),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(getsockopt),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(setsockopt),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(getsockname),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(getpeername),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(fstat),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(dup),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(dup3),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(pipe2),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(ftruncate),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(lseek),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(pwrite64),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(memfd_create),
SECCOMP_ALLOW_MMAP_WITHOUT_EXEC(mmap, 2),
#ifdef __NR_mmap2
SECCOMP_ALLOW_MMAP_WITHOUT_EXEC(mmap2, 2),
#endif
SECCOMP_ALLOW_MMAP_WITHOUT_EXEC(mprotect, 2),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(mremap),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(munmap),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(madvise),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(brk),
SECCOMP_ALLOW_FCNTL_COMMAND(F_GETFD),
SECCOMP_ALLOW_FCNTL_COMMAND(F_SETFD),
SECCOMP_ALLOW_FCNTL_COMMAND(F_GETFL),
SECCOMP_ALLOW_FCNTL_COMMAND(F_SETFL),
SECCOMP_ALLOW_IOCTL_COMMAND(FIONBIO),
#ifdef __NR_clone
SECCOMP_ALLOW_THREAD_CLONE_SYSCALL(clone),
#endif
#ifdef __NR_clone3
BPF_JUMP(BPF_JMP | BPF_JEQ | BPF_K, __NR_clone3, 0, 1),
SECCOMP_ERRNO(ENOSYS),
#endif
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(futex),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(futex_time64),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(set_tid_address),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(set_robust_list),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(rseq),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(rt_sigaction),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(rt_sigprocmask),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(rt_sigreturn),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(sigaltstack),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(tgkill),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(clock_gettime),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(clock_getres),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(clock_nanosleep),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(nanosleep),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(restart_syscall),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(getpid),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(gettid),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(getuid),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(geteuid),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(getgid),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(getegid),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(getrandom),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(gettimeofday),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(getrlimit),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(prlimit64),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(getrusage),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(sched_getaffinity),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(sched_yield),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(getcpu),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(membarrier),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(sysinfo),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(uname),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(umask),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(prctl),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(exit),
SECCOMP_ALLOW_SYSCALL_IF_DEFINED(exit_group),
SECCOMP_KILL,
};
sock_fprog program {
.len = static_cast<unsigned short>(sizeof(filter) / sizeof(filter[0])),
.filter = filter,
};
#ifdef __NR_seccomp
if (syscall(__NR_seccomp, SECCOMP_SET_MODE_FILTER, SECCOMP_FILTER_FLAG_TSYNC, &program) < 0)
return Error::from_syscall("seccomp(SECCOMP_SET_MODE_FILTER)"sv, errno);
#else
if (prctl(PR_SET_SECCOMP, SECCOMP_MODE_FILTER, &program) < 0)
return Error::from_syscall("prctl(PR_SET_SECCOMP)"sv, errno);
#endif
return {};
}
}
ErrorOr<void> apply_sandbox()
{
TRY(install_no_new_privileges());
TRY(configure_runtime_for_sandbox());
TRY(restrict_filesystem_with_landlock());
TRY(install_seccomp_filter());
TRY(Sandbox::install_no_new_privileges());
TRY(Sandbox::configure_runtime());
TRY(Sandbox::restrict_filesystem_with_landlock());
Sandbox::SeccompPolicy policy;
policy.deny_readonly_filesystem_probes();
policy.allow_file_descriptor_operations();
policy.allow_ipc();
policy.allow_common_runtime();
TRY(policy.install());
return {};
}