#define _LARGEFILE64_SOURCE
#define _GNU_SOURCE
+#include "autoconf.h"
+
#include <errno.h>
#include <stdio.h>
#include <fcntl.h>
#include <sys/ptrace.h>
#include <sys/signal.h>
#include <sys/sysinfo.h>
-#include <sys/syscall.h>
#include <sys/resource.h>
+#include <sys/utsname.h>
#include <linux/ptrace.h>
+#if defined(CONFIG_BOX_KERNEL_AMD64) && !defined(CONFIG_BOX_USER_AMD64)
+#include <asm/unistd_32.h>
+#define NATIVE_NR_execve 59 /* 64-bit execve */
+#else
+#include <asm/unistd.h>
+#define NATIVE_NR_execve __NR_execve
+#endif
+
#define NONRET __attribute__((noreturn))
#define UNUSED __attribute__((unused))
#define ARRAY_SIZE(a) (int)(sizeof(a)/sizeof(a[0]))
S(stat) = A_FILENAME,
S(lstat) = A_FILENAME,
S(readlink) = A_FILENAME,
-#ifndef CONFIG_BOX_AMD64
+#ifndef CONFIG_BOX_USER_AMD64
S(oldstat) = A_FILENAME,
S(oldlstat) = A_FILENAME,
S(truncate64) = A_FILENAME,
S(get_thread_area) = A_YES,
S(set_tid_address) = A_YES,
S(exit_group) = A_YES | A_SAMPLE_MEM,
-#ifndef CONFIG_BOX_AMD64
+#ifdef CONFIG_BOX_USER_AMD64
+ S(arch_prctl) = A_YES,
+#else
S(oldfstat) = A_YES,
S(ftruncate64) = A_YES,
S(_llseek) = A_YES,
S(rt_sigqueueinfo) = A_YES | A_LIBERAL,
S(rt_sigsuspend) = A_YES | A_LIBERAL,
S(_sysctl) = A_YES | A_LIBERAL,
-#ifndef CONFIG_BOX_AMD64
+#ifndef CONFIG_BOX_USER_AMD64
S(sigaction) = A_YES | A_LIBERAL,
S(sgetmask) = A_YES | A_LIBERAL,
S(ssetmask) = A_YES | A_LIBERAL,
{ "/proc/meminfo", A_YES },
{ "/proc/self/stat", A_YES },
{ "/proc/self/exe", A_YES }, // Needed by FPC 2.0.x runtime
+ { "/proc/self/maps", A_YES }, // Needed by glibc when it reports arena corruption
};
static struct path_rule *user_path_rules;
/*** Low-level parsing of syscalls ***/
-#ifdef CONFIG_BOX_AMD64
+#ifdef CONFIG_BOX_KERNEL_AMD64
typedef uint64_t arg_t;
#else
typedef uint32_t arg_t;
struct user user;
};
-#ifdef CONFIG_BOX_AMD64
+static int user_mem_fd;
+
+static int read_user_mem(arg_t addr, char *buf, int len)
+{
+ if (!user_mem_fd)
+ {
+ char memname[64];
+ sprintf(memname, "/proc/%d/mem", (int) box_pid);
+ user_mem_fd = open(memname, O_RDONLY);
+ if (user_mem_fd < 0)
+ die("open(%s): %m", memname);
+ }
+ if (lseek64(user_mem_fd, addr, SEEK_SET) < 0)
+ die("lseek64(mem): %m");
+ return read(user_mem_fd, buf, len);
+}
+
+static void close_user_mem(void)
+{
+ if (user_mem_fd)
+ {
+ close(user_mem_fd);
+ user_mem_fd = 0;
+ }
+}
+
+#ifdef CONFIG_BOX_KERNEL_AMD64
static void
-get_syscall_args(struct syscall_args *a)
+get_syscall_args(struct syscall_args *a, int is_exit)
{
if (ptrace(PTRACE_GETREGS, box_pid, NULL, &a->user) < 0)
die("ptrace(PTRACE_GETREGS): %m");
a->sys = a->user.regs.orig_rax;
- a->arg1 = a->user.regs.rdi;
- a->arg2 = a->user.regs.rsi;
- a->arg3 = a->user.regs.rdx;
a->result = a->user.regs.rax;
- // FIXME: Check that it's really a 64-bit syscall
+ /*
+ * CAVEAT: We have to check carefully that this is a real 64-bit syscall.
+ * We test whether the process runs in 64-bit mode, but surprisingly this
+ * is not enough: a 64-bit process can still issue the INT 0x80 instruction
+ * which performs a 32-bit syscall. Currently, the only known way how to
+ * detect this situation is to inspect the instruction code (the kernel
+ * keeps a syscall type flag internally, but it is not accessible from
+ * user space). Hopefully, there is no instruction whose suffix is the
+ * code of the SYSCALL instruction. Sometimes, one would wish the
+ * instruction codes to be unique even when read backwards :)
+ */
+
+ if (is_exit)
+ return;
+
+ int sys_type;
+ uint16_t instr;
+
+ switch (a->user.regs.cs)
+ {
+ case 0x23:
+ // 32-bit CPU mode => only 32-bit syscalls can be issued
+ sys_type = 32;
+ break;
+ case 0x33:
+ // 64-bit CPU mode
+ if (read_user_mem(a->user.regs.rip-2, (char *) &instr, 2) != 2)
+ err("FO: Cannot read syscall instruction");
+ switch (instr)
+ {
+ case 0x050f:
+ break;
+ case 0x80cd:
+ err("FO: Forbidden 32-bit syscall in 64-bit mode");
+ default:
+ err("XX: Unknown syscall instruction %04x", instr);
+ }
+ sys_type = 64;
+ break;
+ default:
+ err("XX: Unknown code segment %04jx", (intmax_t) a->user.regs.cs);
+ }
+
+#ifdef CONFIG_BOX_USER_AMD64
+ if (sys_type != 64)
+ err("FO: Forbidden %d-bit mode syscall", sys_type);
+#else
+ if (sys_type != (exec_seen ? 32 : 64))
+ err("FO: Forbidden %d-bit mode syscall", sys_type);
+#endif
+
+ if (sys_type == 32)
+ {
+ a->arg1 = a->user.regs.rbx;
+ a->arg2 = a->user.regs.rcx;
+ a->arg3 = a->user.regs.rdx;
+ }
+ else
+ {
+ a->arg1 = a->user.regs.rdi;
+ a->arg2 = a->user.regs.rsi;
+ a->arg3 = a->user.regs.rdx;
+ }
}
static void
die("ptrace(PTRACE_SETREGS): %m");
}
+static void
+sanity_check(void)
+{
+}
+
#else
static void
-get_syscall_args(struct syscall_args *a)
+get_syscall_args(struct syscall_args *a, int is_exit UNUSED)
{
if (ptrace(PTRACE_GETREGS, box_pid, NULL, &a->user) < 0)
die("ptrace(PTRACE_GETREGS): %m");
die("ptrace(PTRACE_SETREGS): %m");
}
+static void
+sanity_check(void)
+{
+#if !defined(CONFIG_BOX_ALLOW_INSECURE)
+ struct utsname uts;
+ if (uname(&uts) < 0)
+ die("uname() failed: %m");
+
+ if (!strcmp(uts.machine, "x86_64"))
+ die("Running 32-bit sandbox on 64-bit kernels is inherently unsafe. Please get a 64-bit version.");
+#endif
+}
+
#endif
/*** Syscall checks ***/
valid_filename(arg_t addr)
{
char namebuf[4096], *p, *end;
- static int mem_fd;
if (!file_access)
err("FA: File access forbidden");
if (file_access >= 9)
return;
- if (!mem_fd)
- {
- sprintf(namebuf, "/proc/%d/mem", (int) box_pid);
- mem_fd = open(namebuf, O_RDONLY);
- if (mem_fd < 0)
- die("open(%s): %m", namebuf);
- }
p = end = namebuf;
do
{
l = remains;
if (!l)
err("FA: Access to file with name too long");
- if (lseek64(mem_fd, addr, SEEK_SET) < 0)
- die("lseek64(mem): %m");
- remains = read(mem_fd, end, l);
+ remains = read_user_mem(addr, end, l);
if (remains < 0)
die("read(mem): %m");
if (!remains)
err("FA: Access to file with name out of memory");
- end += l;
- addr += l;
+ end += remains;
+ addr += remains;
}
}
while (*p++);
struct syscall_args a;
static unsigned int sys_tick, last_act;
static arg_t last_sys;
- get_syscall_args(&a);
- arg_t sys = a.sys;
if (++sys_tick & 1) /* Syscall entry */
{
char namebuf[32];
int act;
+
+ get_syscall_args(&a, 0);
+ arg_t sys = a.sys;
msg(">> Syscall %-12s (%08jx,%08jx,%08jx) ", syscall_name(sys, namebuf), (intmax_t) a.arg1, (intmax_t) a.arg2, (intmax_t) a.arg3);
if (!exec_seen)
{
msg("[master] ");
- if (sys == __NR_execve)
- exec_seen = 1;
+ if (sys == NATIVE_NR_execve)
+ {
+ exec_seen = 1;
+ close_user_mem();
+ }
}
else if ((act = valid_syscall(&a)) >= 0)
{
}
else /* Syscall return */
{
- if (sys == ~(arg_t)0)
+ get_syscall_args(&a, 1);
+ if (a.sys == ~(arg_t)0)
{
/* Some syscalls (sigreturn et al.) do not return a value */
if (!(last_act & A_NO_RETVAL))
}
else
{
- if (sys != last_sys)
+ if (a.sys != last_sys)
err("XX: Mismatched syscall entry/exit");
}
if (last_act & A_NO_RETVAL)
if (optind >= argc)
usage();
+ sanity_check();
uid = geteuid();
if (setreuid(uid, uid) < 0)
die("setreuid: %m");