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_action(int set)
 125 {
 126     switch (set) {
 127     case QEMU_SECCOMP_SET_DEFAULT:
 128     case QEMU_SECCOMP_SET_OBSOLETE:
 129     case QEMU_SECCOMP_SET_PRIVILEGED:
 130     case QEMU_SECCOMP_SET_SPAWN: {
 131 #if defined(SECCOMP_GET_ACTION_AVAIL) && defined(SCMP_ACT_KILL_PROCESS) && \
 132     defined(SECCOMP_RET_KILL_PROCESS)
 133         static int kill_process = -1;
 134         if (kill_process == -1) {
 135             uint32_t action = SECCOMP_RET_KILL_PROCESS;
 136
 137             if (qemu_seccomp(SECCOMP_GET_ACTION_AVAIL, 0, &action) == 0) {
 138                 kill_process = 1;
 139             }
 140             kill_process = 0;
 141         }
 142         if (kill_process == 1) {
 143             return SCMP_ACT_KILL_PROCESS;
 144         }
 145 #endif
 146         return SCMP_ACT_TRAP;
 147     }
 148
 149     case QEMU_SECCOMP_SET_RESOURCECTL:
 150         return SCMP_ACT_ERRNO(EPERM);
 151
 152     default:
 153         g_assert_not_reached();
 154     }
 155 }
 156
 157
 158 static int seccomp_start(uint32_t seccomp_opts, Error **errp)
 159 {
 160     int rc = -1;
 161     unsigned int i = 0;
 162     scmp_filter_ctx ctx;
 163
 164     ctx = seccomp_init(SCMP_ACT_ALLOW);
 165     if (ctx == NULL) {
 166         error_setg(errp, "failed to initialize seccomp context");
 167         goto seccomp_return;
 168     }
 169
 170     rc = seccomp_attr_set(ctx, SCMP_FLTATR_CTL_TSYNC, 1);
 171     if (rc != 0) {
 172         error_setg_errno(errp, -rc,
 173                          "failed to set seccomp thread synchronization");
 174         goto seccomp_return;
 175     }
 176
 177     for (i = 0; i < ARRAY_SIZE(blacklist); i++) {
 178         uint32_t action;
 179         if (!(seccomp_opts & blacklist[i].set)) {
 180             continue;
 181         }
 182
 183         action = qemu_seccomp_get_action(blacklist[i].set);
 184         rc = seccomp_rule_add_array(ctx, action, blacklist[i].num,
 185                                     blacklist[i].narg, blacklist[i].arg_cmp);
 186         if (rc < 0) {
 187             error_setg_errno(errp, -rc,
 188                              "failed to add seccomp blacklist rules");
 189             goto seccomp_return;
 190         }
 191     }
 192
 193     rc = seccomp_load(ctx);
 194     if (rc < 0) {
 195         error_setg_errno(errp, -rc,
 196                          "failed to load seccomp syscall filter in kernel");
 197     }
 198
 199   seccomp_return:
 200     seccomp_release(ctx);
 201     return rc < 0 ? -1 : 0;
 202 }
 203
 204 #ifdef CONFIG_SECCOMP
 205 int parse_sandbox(void *opaque, QemuOpts *opts, Error **errp)
 206 {
 207     if (qemu_opt_get_bool(opts, "enable", false)) {
 208         uint32_t seccomp_opts = QEMU_SECCOMP_SET_DEFAULT
 209                 | QEMU_SECCOMP_SET_OBSOLETE;
 210         const char *value = NULL;
 211
 212         value = qemu_opt_get(opts, "obsolete");
 213         if (value) {
 214             if (g_str_equal(value, "allow")) {
 215                 seccomp_opts &= ~QEMU_SECCOMP_SET_OBSOLETE;
 216             } else if (g_str_equal(value, "deny")) {
 217                 /* this is the default option, this if is here
 218                  * to provide a little bit of consistency for
 219                  * the command line */
 220             } else {
 221                 error_setg(errp, "invalid argument for obsolete");
 222                 return -1;
 223             }
 224         }
 225
 226         value = qemu_opt_get(opts, "elevateprivileges");
 227         if (value) {
 228             if (g_str_equal(value, "deny")) {
 229                 seccomp_opts |= QEMU_SECCOMP_SET_PRIVILEGED;
 230             } else if (g_str_equal(value, "children")) {
 231                 seccomp_opts |= QEMU_SECCOMP_SET_PRIVILEGED;
 232
 233                 /* calling prctl directly because we're
 234                  * not sure if host has CAP_SYS_ADMIN set*/
 235                 if (prctl(PR_SET_NO_NEW_PRIVS, 1)) {
 236                     error_setg(errp, "failed to set no_new_privs aborting");
 237                     return -1;
 238                 }
 239             } else if (g_str_equal(value, "allow")) {
 240                 /* default value */
 241             } else {
 242                 error_setg(errp, "invalid argument for elevateprivileges");
 243                 return -1;
 244             }
 245         }
 246
 247         value = qemu_opt_get(opts, "spawn");
 248         if (value) {
 249             if (g_str_equal(value, "deny")) {
 250                 seccomp_opts |= QEMU_SECCOMP_SET_SPAWN;
 251             } else if (g_str_equal(value, "allow")) {
 252                 /* default value */
 253             } else {
 254                 error_setg(errp, "invalid argument for spawn");
 255                 return -1;
 256             }
 257         }
 258
 259         value = qemu_opt_get(opts, "resourcecontrol");
 260         if (value) {
 261             if (g_str_equal(value, "deny")) {
 262                 seccomp_opts |= QEMU_SECCOMP_SET_RESOURCECTL;
 263             } else if (g_str_equal(value, "allow")) {
 264                 /* default value */
 265             } else {
 266                 error_setg(errp, "invalid argument for resourcecontrol");
 267                 return -1;
 268             }
 269         }
 270
 271         if (seccomp_start(seccomp_opts, errp) < 0) {
 272             return -1;
 273         }
 274     }
 275
 276     return 0;
 277 }
 278
 279 static QemuOptsList qemu_sandbox_opts = {
 280     .name = "sandbox",
 281     .implied_opt_name = "enable",
 282     .head = QTAILQ_HEAD_INITIALIZER(qemu_sandbox_opts.head),
 283     .desc = {
 284         {
 285             .name = "enable",
 286             .type = QEMU_OPT_BOOL,
 287         },
 288         {
 289             .name = "obsolete",
 290             .type = QEMU_OPT_STRING,
 291         },
 292         {
 293             .name = "elevateprivileges",
 294             .type = QEMU_OPT_STRING,
 295         },
 296         {
 297             .name = "spawn",
 298             .type = QEMU_OPT_STRING,
 299         },
 300         {
 301             .name = "resourcecontrol",
 302             .type = QEMU_OPT_STRING,
 303         },
 304         { /* end of list */ }
 305     },
 306 };
 307
 308 static void seccomp_register(void)
 309 {
 310     bool add = false;
 311
 312     /* FIXME: use seccomp_api_get() >= 2 check when released */
 313
 314 #if defined(SECCOMP_FILTER_FLAG_TSYNC)
 315     int check;
 316
 317     /* check host TSYNC capability, it returns errno == ENOSYS if unavailable */
 318     check = qemu_seccomp(SECCOMP_SET_MODE_FILTER,
 319                          SECCOMP_FILTER_FLAG_TSYNC, NULL);
 320     if (check < 0 && errno == EFAULT) {
 321         add = true;
 322     }
 323 #endif
 324
 325     if (add) {
 326         qemu_add_opts(&qemu_sandbox_opts);
 327     }
 328 }
 329 opts_init(seccomp_register);
 330 #endif