qemu-seccomp.c

   1 /*
   2  * QEMU seccomp mode 2 support with libseccomp
   3  *
   4  * Copyright IBM, Corp. 2012
   5  *
   6  * Authors:
   7  *  Eduardo Otubo    <eotubo@br.ibm.com>
   8  *
   9  * This work is licensed under the terms of the GNU GPL, version 2.  See
  10  * the COPYING file in the top-level directory.
  11  *
  12  * Contributions after 2012-01-13 are licensed under the terms of the
  13  * GNU GPL, version 2 or (at your option) any later version.
  14  */
  15
  16 #include "qemu/osdep.h"
  17 #include "qapi/error.h"
  18 #include "qemu/config-file.h"
  19 #include "qemu/option.h"
  20 #include "qemu/module.h"
  21 #include <sys/prctl.h>
  22 #include <seccomp.h>
  23 #include "sysemu/seccomp.h"
  24 #include <linux/seccomp.h>
  25
  26 /* For some architectures (notably ARM) cacheflush is not supported until
  27  * libseccomp 2.2.3, but configure enforces that we are using a more recent
  28  * version on those hosts, so it is OK for this check to be less strict.
  29  */
  30 #if SCMP_VER_MAJOR >= 3
  31   #define HAVE_CACHEFLUSH
  32 #elif SCMP_VER_MAJOR == 2 && SCMP_VER_MINOR >= 2
  33   #define HAVE_CACHEFLUSH
  34 #endif
  35
  36 struct QemuSeccompSyscall {
  37     int32_t num;
  38     uint8_t set;
  39     uint8_t narg;
  40     const struct scmp_arg_cmp *arg_cmp;
  41 };
  42
  43 const struct scmp_arg_cmp sched_setscheduler_arg[] = {
  44     /* was SCMP_A1(SCMP_CMP_NE, SCHED_IDLE), but expanded due to GCC 4.x bug */
  45     { .arg = 1, .op = SCMP_CMP_NE, .datum_a = SCHED_IDLE }
  46 };
  47
  48 static const struct QemuSeccompSyscall blacklist[] = {
  49     /* default set of syscalls to blacklist */
  50     { SCMP_SYS(reboot),                 QEMU_SECCOMP_SET_DEFAULT },
  51     { SCMP_SYS(swapon),                 QEMU_SECCOMP_SET_DEFAULT },
  52     { SCMP_SYS(swapoff),                QEMU_SECCOMP_SET_DEFAULT },
  53     { SCMP_SYS(syslog),                 QEMU_SECCOMP_SET_DEFAULT },
  54     { SCMP_SYS(mount),                  QEMU_SECCOMP_SET_DEFAULT },
  55     { SCMP_SYS(umount),                 QEMU_SECCOMP_SET_DEFAULT },
  56     { SCMP_SYS(kexec_load),             QEMU_SECCOMP_SET_DEFAULT },
  57     { SCMP_SYS(afs_syscall),            QEMU_SECCOMP_SET_DEFAULT },
  58     { SCMP_SYS(break),                  QEMU_SECCOMP_SET_DEFAULT },
  59     { SCMP_SYS(ftime),                  QEMU_SECCOMP_SET_DEFAULT },
  60     { SCMP_SYS(getpmsg),                QEMU_SECCOMP_SET_DEFAULT },
  61     { SCMP_SYS(gtty),                   QEMU_SECCOMP_SET_DEFAULT },
  62     { SCMP_SYS(lock),                   QEMU_SECCOMP_SET_DEFAULT },
  63     { SCMP_SYS(mpx),                    QEMU_SECCOMP_SET_DEFAULT },
  64     { SCMP_SYS(prof),                   QEMU_SECCOMP_SET_DEFAULT },
  65     { SCMP_SYS(profil),                 QEMU_SECCOMP_SET_DEFAULT },
  66     { SCMP_SYS(putpmsg),                QEMU_SECCOMP_SET_DEFAULT },
  67     { SCMP_SYS(security),               QEMU_SECCOMP_SET_DEFAULT },
  68     { SCMP_SYS(stty),                   QEMU_SECCOMP_SET_DEFAULT },
  69     { SCMP_SYS(tuxcall),                QEMU_SECCOMP_SET_DEFAULT },
  70     { SCMP_SYS(ulimit),                 QEMU_SECCOMP_SET_DEFAULT },
  71     { SCMP_SYS(vserver),                QEMU_SECCOMP_SET_DEFAULT },
  72     /* obsolete */
  73     { SCMP_SYS(readdir),                QEMU_SECCOMP_SET_OBSOLETE },
  74     { SCMP_SYS(_sysctl),                QEMU_SECCOMP_SET_OBSOLETE },
  75     { SCMP_SYS(bdflush),                QEMU_SECCOMP_SET_OBSOLETE },
  76     { SCMP_SYS(create_module),          QEMU_SECCOMP_SET_OBSOLETE },
  77     { SCMP_SYS(get_kernel_syms),        QEMU_SECCOMP_SET_OBSOLETE },
  78     { SCMP_SYS(query_module),           QEMU_SECCOMP_SET_OBSOLETE },
  79     { SCMP_SYS(sgetmask),               QEMU_SECCOMP_SET_OBSOLETE },
  80     { SCMP_SYS(ssetmask),               QEMU_SECCOMP_SET_OBSOLETE },
  81     { SCMP_SYS(sysfs),                  QEMU_SECCOMP_SET_OBSOLETE },
  82     { SCMP_SYS(uselib),                 QEMU_SECCOMP_SET_OBSOLETE },
  83     { SCMP_SYS(ustat),                  QEMU_SECCOMP_SET_OBSOLETE },
  84     /* privileged */
  85     { SCMP_SYS(setuid),                 QEMU_SECCOMP_SET_PRIVILEGED },
  86     { SCMP_SYS(setgid),                 QEMU_SECCOMP_SET_PRIVILEGED },
  87     { SCMP_SYS(setpgid),                QEMU_SECCOMP_SET_PRIVILEGED },
  88     { SCMP_SYS(setsid),                 QEMU_SECCOMP_SET_PRIVILEGED },
  89     { SCMP_SYS(setreuid),               QEMU_SECCOMP_SET_PRIVILEGED },
  90     { SCMP_SYS(setregid),               QEMU_SECCOMP_SET_PRIVILEGED },
  91     { SCMP_SYS(setresuid),              QEMU_SECCOMP_SET_PRIVILEGED },
  92     { SCMP_SYS(setresgid),              QEMU_SECCOMP_SET_PRIVILEGED },
  93     { SCMP_SYS(setfsuid),               QEMU_SECCOMP_SET_PRIVILEGED },
  94     { SCMP_SYS(setfsgid),               QEMU_SECCOMP_SET_PRIVILEGED },
  95     /* spawn */
  96     { SCMP_SYS(fork),                   QEMU_SECCOMP_SET_SPAWN },
  97     { SCMP_SYS(vfork),                  QEMU_SECCOMP_SET_SPAWN },
  98     { SCMP_SYS(execve),                 QEMU_SECCOMP_SET_SPAWN },
  99     /* resource control */
 100     { SCMP_SYS(getpriority),            QEMU_SECCOMP_SET_RESOURCECTL },
 101     { SCMP_SYS(setpriority),            QEMU_SECCOMP_SET_RESOURCECTL },
 102     { SCMP_SYS(sched_setparam),         QEMU_SECCOMP_SET_RESOURCECTL },
 103     { SCMP_SYS(sched_getparam),         QEMU_SECCOMP_SET_RESOURCECTL },
 104     { SCMP_SYS(sched_setscheduler),     QEMU_SECCOMP_SET_RESOURCECTL,
 105       ARRAY_SIZE(sched_setscheduler_arg), sched_setscheduler_arg },
 106     { SCMP_SYS(sched_getscheduler),     QEMU_SECCOMP_SET_RESOURCECTL },
 107     { SCMP_SYS(sched_setaffinity),      QEMU_SECCOMP_SET_RESOURCECTL },
 108     { SCMP_SYS(sched_getaffinity),      QEMU_SECCOMP_SET_RESOURCECTL },
 109     { SCMP_SYS(sched_get_priority_max), QEMU_SECCOMP_SET_RESOURCECTL },
 110     { SCMP_SYS(sched_get_priority_min), QEMU_SECCOMP_SET_RESOURCECTL },
 111 };
 112
 113 static inline __attribute__((unused)) int
 114 qemu_seccomp(unsigned int operation, unsigned int flags, void *args)
 115 {
 116 #ifdef __NR_seccomp
 117     return syscall(__NR_seccomp, operation, flags, args);
 118 #else
 119     errno = ENOSYS;
 120     return -1;
 121 #endif
 122 }
 123
 124 static uint32_t qemu_seccomp_get_kill_action(void)
 125 {
 126 #if defined(SECCOMP_GET_ACTION_AVAIL) && defined(SCMP_ACT_KILL_PROCESS) && \
 127     defined(SECCOMP_RET_KILL_PROCESS)
 128     {
 129         uint32_t action = SECCOMP_RET_KILL_PROCESS;
 130
 131         if (qemu_seccomp(SECCOMP_GET_ACTION_AVAIL, 0, &action) == 0) {
 132             return SCMP_ACT_KILL_PROCESS;
 133         }
 134     }
 135 #endif
 136
 137     return SCMP_ACT_TRAP;
 138 }
 139
 140
 141 static int seccomp_start(uint32_t seccomp_opts)
 142 {
 143     int rc = 0;
 144     unsigned int i = 0;
 145     scmp_filter_ctx ctx;
 146     uint32_t action = qemu_seccomp_get_kill_action();
 147
 148     ctx = seccomp_init(SCMP_ACT_ALLOW);
 149     if (ctx == NULL) {
 150         rc = -1;
 151         goto seccomp_return;
 152     }
 153
 154     rc = seccomp_attr_set(ctx, SCMP_FLTATR_CTL_TSYNC, 1);
 155     if (rc != 0) {
 156         goto seccomp_return;
 157     }
 158
 159     for (i = 0; i < ARRAY_SIZE(blacklist); i++) {
 160         if (!(seccomp_opts & blacklist[i].set)) {
 161             continue;
 162         }
 163
 164         rc = seccomp_rule_add_array(ctx, action, blacklist[i].num,
 165                                     blacklist[i].narg, blacklist[i].arg_cmp);
 166         if (rc < 0) {
 167             goto seccomp_return;
 168         }
 169     }
 170
 171     rc = seccomp_load(ctx);
 172
 173   seccomp_return:
 174     seccomp_release(ctx);
 175     return rc;
 176 }
 177
 178 #ifdef CONFIG_SECCOMP
 179 int parse_sandbox(void *opaque, QemuOpts *opts, Error **errp)
 180 {
 181     if (qemu_opt_get_bool(opts, "enable", false)) {
 182         uint32_t seccomp_opts = QEMU_SECCOMP_SET_DEFAULT
 183                 | QEMU_SECCOMP_SET_OBSOLETE;
 184         const char *value = NULL;
 185
 186         value = qemu_opt_get(opts, "obsolete");
 187         if (value) {
 188             if (g_str_equal(value, "allow")) {
 189                 seccomp_opts &= ~QEMU_SECCOMP_SET_OBSOLETE;
 190             } else if (g_str_equal(value, "deny")) {
 191                 /* this is the default option, this if is here
 192                  * to provide a little bit of consistency for
 193                  * the command line */
 194             } else {
 195                 error_setg(errp, "invalid argument for obsolete");
 196                 return -1;
 197             }
 198         }
 199
 200         value = qemu_opt_get(opts, "elevateprivileges");
 201         if (value) {
 202             if (g_str_equal(value, "deny")) {
 203                 seccomp_opts |= QEMU_SECCOMP_SET_PRIVILEGED;
 204             } else if (g_str_equal(value, "children")) {
 205                 seccomp_opts |= QEMU_SECCOMP_SET_PRIVILEGED;
 206
 207                 /* calling prctl directly because we're
 208                  * not sure if host has CAP_SYS_ADMIN set*/
 209                 if (prctl(PR_SET_NO_NEW_PRIVS, 1)) {
 210                     error_setg(errp, "failed to set no_new_privs aborting");
 211                     return -1;
 212                 }
 213             } else if (g_str_equal(value, "allow")) {
 214                 /* default value */
 215             } else {
 216                 error_setg(errp, "invalid argument for elevateprivileges");
 217                 return -1;
 218             }
 219         }
 220
 221         value = qemu_opt_get(opts, "spawn");
 222         if (value) {
 223             if (g_str_equal(value, "deny")) {
 224                 seccomp_opts |= QEMU_SECCOMP_SET_SPAWN;
 225             } else if (g_str_equal(value, "allow")) {
 226                 /* default value */
 227             } else {
 228                 error_setg(errp, "invalid argument for spawn");
 229                 return -1;
 230             }
 231         }
 232
 233         value = qemu_opt_get(opts, "resourcecontrol");
 234         if (value) {
 235             if (g_str_equal(value, "deny")) {
 236                 seccomp_opts |= QEMU_SECCOMP_SET_RESOURCECTL;
 237             } else if (g_str_equal(value, "allow")) {
 238                 /* default value */
 239             } else {
 240                 error_setg(errp, "invalid argument for resourcecontrol");
 241                 return -1;
 242             }
 243         }
 244
 245         if (seccomp_start(seccomp_opts) < 0) {
 246             error_setg(errp, "failed to install seccomp syscall filter "
 247                        "in the kernel");
 248             return -1;
 249         }
 250     }
 251
 252     return 0;
 253 }
 254
 255 static QemuOptsList qemu_sandbox_opts = {
 256     .name = "sandbox",
 257     .implied_opt_name = "enable",
 258     .head = QTAILQ_HEAD_INITIALIZER(qemu_sandbox_opts.head),
 259     .desc = {
 260         {
 261             .name = "enable",
 262             .type = QEMU_OPT_BOOL,
 263         },
 264         {
 265             .name = "obsolete",
 266             .type = QEMU_OPT_STRING,
 267         },
 268         {
 269             .name = "elevateprivileges",
 270             .type = QEMU_OPT_STRING,
 271         },
 272         {
 273             .name = "spawn",
 274             .type = QEMU_OPT_STRING,
 275         },
 276         {
 277             .name = "resourcecontrol",
 278             .type = QEMU_OPT_STRING,
 279         },
 280         { /* end of list */ }
 281     },
 282 };
 283
 284 static void seccomp_register(void)
 285 {
 286     bool add = false;
 287
 288     /* FIXME: use seccomp_api_get() >= 2 check when released */
 289
 290 #if defined(SECCOMP_FILTER_FLAG_TSYNC)
 291     int check;
 292
 293     /* check host TSYNC capability, it returns errno == ENOSYS if unavailable */
 294     check = qemu_seccomp(SECCOMP_SET_MODE_FILTER,
 295                          SECCOMP_FILTER_FLAG_TSYNC, NULL);
 296     if (check < 0 && errno == EFAULT) {
 297         add = true;
 298     }
 299 #endif
 300
 301     if (add) {
 302         qemu_add_opts(&qemu_sandbox_opts);
 303     }
 304 }
 305 opts_init(seccomp_register);
 306 #endif