softmmu/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 denylist[] = {
  49     /* default set of syscalls that should get blocked */
  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(setpriority),            QEMU_SECCOMP_SET_RESOURCECTL },
 101     { SCMP_SYS(sched_setparam),         QEMU_SECCOMP_SET_RESOURCECTL },
 102     { SCMP_SYS(sched_setscheduler),     QEMU_SECCOMP_SET_RESOURCECTL,
 103       ARRAY_SIZE(sched_setscheduler_arg), sched_setscheduler_arg },
 104     { SCMP_SYS(sched_setaffinity),      QEMU_SECCOMP_SET_RESOURCECTL },
 105 };
 106
 107 static inline __attribute__((unused)) int
 108 qemu_seccomp(unsigned int operation, unsigned int flags, void *args)
 109 {
 110 #ifdef __NR_seccomp
 111     return syscall(__NR_seccomp, operation, flags, args);
 112 #else
 113     errno = ENOSYS;
 114     return -1;
 115 #endif
 116 }
 117
 118 static uint32_t qemu_seccomp_get_action(int set)
 119 {
 120     switch (set) {
 121     case QEMU_SECCOMP_SET_DEFAULT:
 122     case QEMU_SECCOMP_SET_OBSOLETE:
 123     case QEMU_SECCOMP_SET_PRIVILEGED:
 124     case QEMU_SECCOMP_SET_SPAWN: {
 125 #if defined(SECCOMP_GET_ACTION_AVAIL) && defined(SCMP_ACT_KILL_PROCESS) && \
 126     defined(SECCOMP_RET_KILL_PROCESS)
 127         static int kill_process = -1;
 128         if (kill_process == -1) {
 129             uint32_t action = SECCOMP_RET_KILL_PROCESS;
 130
 131             if (qemu_seccomp(SECCOMP_GET_ACTION_AVAIL, 0, &action) == 0) {
 132                 kill_process = 1;
 133             } else {
 134                 kill_process = 0;
 135             }
 136         }
 137         if (kill_process == 1) {
 138             return SCMP_ACT_KILL_PROCESS;
 139         }
 140 #endif
 141         return SCMP_ACT_TRAP;
 142     }
 143
 144     case QEMU_SECCOMP_SET_RESOURCECTL:
 145         return SCMP_ACT_ERRNO(EPERM);
 146
 147     default:
 148         g_assert_not_reached();
 149     }
 150 }
 151
 152
 153 static int seccomp_start(uint32_t seccomp_opts, Error **errp)
 154 {
 155     int rc = -1;
 156     unsigned int i = 0;
 157     scmp_filter_ctx ctx;
 158
 159     ctx = seccomp_init(SCMP_ACT_ALLOW);
 160     if (ctx == NULL) {
 161         error_setg(errp, "failed to initialize seccomp context");
 162         goto seccomp_return;
 163     }
 164
 165     rc = seccomp_attr_set(ctx, SCMP_FLTATR_CTL_TSYNC, 1);
 166     if (rc != 0) {
 167         error_setg_errno(errp, -rc,
 168                          "failed to set seccomp thread synchronization");
 169         goto seccomp_return;
 170     }
 171
 172     for (i = 0; i < ARRAY_SIZE(denylist); i++) {
 173         uint32_t action;
 174         if (!(seccomp_opts & denylist[i].set)) {
 175             continue;
 176         }
 177
 178         action = qemu_seccomp_get_action(denylist[i].set);
 179         rc = seccomp_rule_add_array(ctx, action, denylist[i].num,
 180                                     denylist[i].narg, denylist[i].arg_cmp);
 181         if (rc < 0) {
 182             error_setg_errno(errp, -rc,
 183                              "failed to add seccomp denylist rules");
 184             goto seccomp_return;
 185         }
 186     }
 187
 188     rc = seccomp_load(ctx);
 189     if (rc < 0) {
 190         error_setg_errno(errp, -rc,
 191                          "failed to load seccomp syscall filter in kernel");
 192     }
 193
 194   seccomp_return:
 195     seccomp_release(ctx);
 196     return rc < 0 ? -1 : 0;
 197 }
 198
 199 int parse_sandbox(void *opaque, QemuOpts *opts, Error **errp)
 200 {
 201     if (qemu_opt_get_bool(opts, "enable", false)) {
 202         uint32_t seccomp_opts = QEMU_SECCOMP_SET_DEFAULT
 203                 | QEMU_SECCOMP_SET_OBSOLETE;
 204         const char *value = NULL;
 205
 206         value = qemu_opt_get(opts, "obsolete");
 207         if (value) {
 208             if (g_str_equal(value, "allow")) {
 209                 seccomp_opts &= ~QEMU_SECCOMP_SET_OBSOLETE;
 210             } else if (g_str_equal(value, "deny")) {
 211                 /* this is the default option, this if is here
 212                  * to provide a little bit of consistency for
 213                  * the command line */
 214             } else {
 215                 error_setg(errp, "invalid argument for obsolete");
 216                 return -1;
 217             }
 218         }
 219
 220         value = qemu_opt_get(opts, "elevateprivileges");
 221         if (value) {
 222             if (g_str_equal(value, "deny")) {
 223                 seccomp_opts |= QEMU_SECCOMP_SET_PRIVILEGED;
 224             } else if (g_str_equal(value, "children")) {
 225                 seccomp_opts |= QEMU_SECCOMP_SET_PRIVILEGED;
 226
 227                 /* calling prctl directly because we're
 228                  * not sure if host has CAP_SYS_ADMIN set*/
 229                 if (prctl(PR_SET_NO_NEW_PRIVS, 1)) {
 230                     error_setg(errp, "failed to set no_new_privs aborting");
 231                     return -1;
 232                 }
 233             } else if (g_str_equal(value, "allow")) {
 234                 /* default value */
 235             } else {
 236                 error_setg(errp, "invalid argument for elevateprivileges");
 237                 return -1;
 238             }
 239         }
 240
 241         value = qemu_opt_get(opts, "spawn");
 242         if (value) {
 243             if (g_str_equal(value, "deny")) {
 244                 seccomp_opts |= QEMU_SECCOMP_SET_SPAWN;
 245             } else if (g_str_equal(value, "allow")) {
 246                 /* default value */
 247             } else {
 248                 error_setg(errp, "invalid argument for spawn");
 249                 return -1;
 250             }
 251         }
 252
 253         value = qemu_opt_get(opts, "resourcecontrol");
 254         if (value) {
 255             if (g_str_equal(value, "deny")) {
 256                 seccomp_opts |= QEMU_SECCOMP_SET_RESOURCECTL;
 257             } else if (g_str_equal(value, "allow")) {
 258                 /* default value */
 259             } else {
 260                 error_setg(errp, "invalid argument for resourcecontrol");
 261                 return -1;
 262             }
 263         }
 264
 265         if (seccomp_start(seccomp_opts, errp) < 0) {
 266             return -1;
 267         }
 268     }
 269
 270     return 0;
 271 }
 272
 273 static QemuOptsList qemu_sandbox_opts = {
 274     .name = "sandbox",
 275     .implied_opt_name = "enable",
 276     .head = QTAILQ_HEAD_INITIALIZER(qemu_sandbox_opts.head),
 277     .desc = {
 278         {
 279             .name = "enable",
 280             .type = QEMU_OPT_BOOL,
 281         },
 282         {
 283             .name = "obsolete",
 284             .type = QEMU_OPT_STRING,
 285         },
 286         {
 287             .name = "elevateprivileges",
 288             .type = QEMU_OPT_STRING,
 289         },
 290         {
 291             .name = "spawn",
 292             .type = QEMU_OPT_STRING,
 293         },
 294         {
 295             .name = "resourcecontrol",
 296             .type = QEMU_OPT_STRING,
 297         },
 298         { /* end of list */ }
 299     },
 300 };
 301
 302 static void seccomp_register(void)
 303 {
 304     bool add = false;
 305
 306     /* FIXME: use seccomp_api_get() >= 2 check when released */
 307
 308 #if defined(SECCOMP_FILTER_FLAG_TSYNC)
 309     int check;
 310
 311     /* check host TSYNC capability, it returns errno == ENOSYS if unavailable */
 312     check = qemu_seccomp(SECCOMP_SET_MODE_FILTER,
 313                          SECCOMP_FILTER_FLAG_TSYNC, NULL);
 314     if (check < 0 && errno == EFAULT) {
 315         add = true;
 316     }
 317 #endif
 318
 319     if (add) {
 320         qemu_add_opts(&qemu_sandbox_opts);
 321     }
 322 }
 323 opts_init(seccomp_register);