LXR busybox/util-linux/unshare.c

   1/* vi: set sw=4 ts=4: */
   2/*
   3 * Mini unshare implementation for busybox.
   4 *
   5 * Copyright (C) 2016 by Bartosz Golaszewski <bartekgola@gmail.com>
   6 *
   7 * Licensed under GPLv2 or later, see file LICENSE in this source tree.
   8 */
   9//config:config UNSHARE
  10//config:       bool "unshare (7.2 kb)"
  11//config:       default y
  12//config:       depends on !NOMMU
  13//config:       select PLATFORM_LINUX
  14//config:       select LONG_OPTS
  15//config:       help
  16//config:       Run program with some namespaces unshared from parent.
  17
  18// needs LONG_OPTS: it is awkward to exclude code which handles --propagation
  19// and --setgroups based on LONG_OPTS, so instead applet requires LONG_OPTS.
  20// depends on !NOMMU: we need fork()
  21
  22//applet:IF_UNSHARE(APPLET(unshare, BB_DIR_USR_BIN, BB_SUID_DROP))
  23
  24//kbuild:lib-$(CONFIG_UNSHARE) += unshare.o
  25
  26//usage:#define unshare_trivial_usage
  27//usage:       "[OPTIONS] [PROG [ARGS]]"
  28//usage:#define unshare_full_usage "\n"
  29//usage:     "\n        -m,--mount[=FILE]       Unshare mount namespace"
  30//usage:     "\n        -u,--uts[=FILE]         Unshare UTS namespace (hostname etc.)"
  31//usage:     "\n        -i,--ipc[=FILE]         Unshare System V IPC namespace"
  32//usage:     "\n        -n,--net[=FILE]         Unshare network namespace"
  33//usage:     "\n        -p,--pid[=FILE]         Unshare PID namespace"
  34//usage:     "\n        -U,--user[=FILE]        Unshare user namespace"
  35//usage:     "\n        -f,--fork               Fork before execing PROG"
  36//usage:     "\n        -r,--map-root-user      Map current user to root (implies -U)"
  37//usage:     "\n        --mount-proc[=DIR]      Mount /proc filesystem first (implies -m)"
  38//usage:     "\n        --propagation slave|shared|private|unchanged"
  39//usage:     "\n                                Modify mount propagation in mount namespace"
  40//usage:     "\n        --setgroups allow|deny  Control the setgroups syscall in user namespaces"
  41
  42#include <sched.h>
  43#ifndef CLONE_NEWUTS
  44# define CLONE_NEWUTS  0x04000000
  45#endif
  46#ifndef CLONE_NEWIPC
  47# define CLONE_NEWIPC  0x08000000
  48#endif
  49#ifndef CLONE_NEWUSER
  50# define CLONE_NEWUSER 0x10000000
  51#endif
  52#ifndef CLONE_NEWPID
  53# define CLONE_NEWPID  0x20000000
  54#endif
  55#ifndef CLONE_NEWNET
  56# define CLONE_NEWNET  0x40000000
  57#endif
  58
  59#include <sys/mount.h>
  60#ifndef MS_REC
  61# define MS_REC     (1 << 14)
  62#endif
  63#ifndef MS_PRIVATE
  64# define MS_PRIVATE (1 << 18)
  65#endif
  66#ifndef MS_SLAVE
  67# define MS_SLAVE   (1 << 19)
  68#endif
  69#ifndef MS_SHARED
  70# define MS_SHARED  (1 << 20)
  71#endif
  72
  73#include "libbb.h"
  74
  75static void mount_or_die(const char *source, const char *target,
  76                const char *fstype, unsigned long mountflags)
  77{
  78        if (mount(source, target, fstype, mountflags, NULL)) {
  79                bb_perror_msg_and_die("can't mount %s on %s (flags:0x%lx)",
  80                        source, target, mountflags);
  81                /* fstype is always either NULL or "proc".
  82                 * "proc" is only used to mount /proc.
  83                 * No need to clutter up error message with fstype,
  84                 * it is easily deductible.
  85                 */
  86        }
  87}
  88
  89#define PATH_PROC_SETGROUPS     "/proc/self/setgroups"
  90#define PATH_PROC_UIDMAP        "/proc/self/uid_map"
  91#define PATH_PROC_GIDMAP        "/proc/self/gid_map"
  92
  93struct namespace_descr {
  94        int flag;
  95        const char nsfile4[4];
  96};
  97
  98struct namespace_ctx {
  99        char *path;
 100};
 101
 102enum {
 103        OPT_mount       = 1 << 0,
 104        OPT_uts         = 1 << 1,
 105        OPT_ipc         = 1 << 2,
 106        OPT_net         = 1 << 3,
 107        OPT_pid         = 1 << 4,
 108        OPT_user        = 1 << 5, /* OPT_user, NS_USR_POS, and ns_list[] index must match! */
 109        OPT_fork        = 1 << 6,
 110        OPT_map_root    = 1 << 7,
 111        OPT_mount_proc  = 1 << 8,
 112        OPT_propagation = 1 << 9,
 113        OPT_setgroups   = 1 << 10,
 114};
 115enum {
 116        NS_MNT_POS = 0,
 117        NS_UTS_POS,
 118        NS_IPC_POS,
 119        NS_NET_POS,
 120        NS_PID_POS,
 121        NS_USR_POS, /* OPT_user, NS_USR_POS, and ns_list[] index must match! */
 122        NS_COUNT,
 123};
 124static const struct namespace_descr ns_list[] = {
 125        { CLONE_NEWNS,   "mnt"  },
 126        { CLONE_NEWUTS,  "uts"  },
 127        { CLONE_NEWIPC,  "ipc"  },
 128        { CLONE_NEWNET,  "net"  },
 129        { CLONE_NEWPID,  "pid"  },
 130        { CLONE_NEWUSER, "user" }, /* OPT_user, NS_USR_POS, and ns_list[] index must match! */
 131};
 132
 133/*
 134 * Upstream unshare doesn't support short options for --mount-proc,
 135 * --propagation, --setgroups.
 136 * Optional arguments (namespace mountpoints) exist only for long opts,
 137 * we are forced to use "fake" letters for them.
 138 * '+': stop at first non-option.
 139 */
 140#define OPT_STR "+muinpU""fr""\xfd::""\xfe:""\xff:"
 141static const char unshare_longopts[] ALIGN1 =
 142        "mount\0"               Optional_argument       "\xf0"
 143        "uts\0"                 Optional_argument       "\xf1"
 144        "ipc\0"                 Optional_argument       "\xf2"
 145        "net\0"                 Optional_argument       "\xf3"
 146        "pid\0"                 Optional_argument       "\xf4"
 147        "user\0"                Optional_argument       "\xf5"
 148        "fork\0"                No_argument             "f"
 149        "map-root-user\0"       No_argument             "r"
 150        "mount-proc\0"          Optional_argument       "\xfd"
 151        "propagation\0"         Required_argument       "\xfe"
 152        "setgroups\0"           Required_argument       "\xff"
 153;
 154
 155/* Ugly-looking string reuse trick */
 156#define PRIVATE_STR   "private\0""unchanged\0""shared\0""slave\0"
 157#define PRIVATE_UNCHANGED_SHARED_SLAVE   PRIVATE_STR
 158
 159static unsigned long parse_propagation(const char *prop_str)
 160{
 161        int i = index_in_strings(PRIVATE_UNCHANGED_SHARED_SLAVE, prop_str);
 162        if (i < 0)
 163                bb_error_msg_and_die("unrecognized: --%s=%s", "propagation", prop_str);
 164        if (i == 0)
 165                return MS_REC | MS_PRIVATE;
 166        if (i == 1)
 167                return 0;
 168        if (i == 2)
 169                return MS_REC | MS_SHARED;
 170        return MS_REC | MS_SLAVE;
 171}
 172
 173static void mount_namespaces(pid_t pid, struct namespace_ctx *ns_ctx_list)
 174{
 175        const struct namespace_descr *ns;
 176        struct namespace_ctx *ns_ctx;
 177        int i;
 178
 179        for (i = 0; i < NS_COUNT; i++) {
 180                char nsf[sizeof("/proc/%u/ns/AAAA") + sizeof(int)*3];
 181
 182                ns = &ns_list[i];
 183                ns_ctx = &ns_ctx_list[i];
 184                if (!ns_ctx->path)
 185                        continue;
 186                sprintf(nsf, "/proc/%u/ns/%.4s", (unsigned)pid, ns->nsfile4);
 187                mount_or_die(nsf, ns_ctx->path, NULL, MS_BIND);
 188        }
 189}
 190
 191int unshare_main(int argc, char **argv) MAIN_EXTERNALLY_VISIBLE;
 192int unshare_main(int argc UNUSED_PARAM, char **argv)
 193{
 194        int i;
 195        unsigned int opts;
 196        int unsflags;
 197        uintptr_t need_mount;
 198        const char *proc_mnt_target;
 199        const char *prop_str;
 200        const char *setgrp_str;
 201        unsigned long prop_flags;
 202        uid_t reuid = geteuid();
 203        gid_t regid = getegid();
 204        struct fd_pair fdp;
 205        pid_t child = child; /* for compiler */
 206        struct namespace_ctx ns_ctx_list[NS_COUNT];
 207
 208        memset(ns_ctx_list, 0, sizeof(ns_ctx_list));
 209        proc_mnt_target = "/proc";
 210        prop_str = PRIVATE_STR;
 211        setgrp_str = NULL;
 212
 213        opts = getopt32long(argv, "^" OPT_STR "\0"
 214                "\xf0""m" /* long opts (via their "fake chars") imply short opts */
 215                ":\xf1""u"
 216                ":\xf2""i"
 217                ":\xf3""n"
 218                ":\xf4""p"
 219                ":\xf5""U"
 220                ":rU"      /* --map-root-user or -r implies -U */
 221                ":\xfd""m" /* --mount-proc implies -m */
 222                , unshare_longopts,
 223                &proc_mnt_target, &prop_str, &setgrp_str,
 224                &ns_ctx_list[NS_MNT_POS].path,
 225                &ns_ctx_list[NS_UTS_POS].path,
 226                &ns_ctx_list[NS_IPC_POS].path,
 227                &ns_ctx_list[NS_NET_POS].path,
 228                &ns_ctx_list[NS_PID_POS].path,
 229                &ns_ctx_list[NS_USR_POS].path
 230        );
 231        argv += optind;
 232        //bb_error_msg("opts:0x%x", opts);
 233        //bb_error_msg("mount:%s", ns_ctx_list[NS_MNT_POS].path);
 234        //bb_error_msg("proc_mnt_target:%s", proc_mnt_target);
 235        //bb_error_msg("prop_str:%s", prop_str);
 236        //bb_error_msg("setgrp_str:%s", setgrp_str);
 237        //exit(1);
 238
 239        if (setgrp_str) {
 240                if (strcmp(setgrp_str, "allow") == 0) {
 241                        if (opts & OPT_map_root) {
 242                                bb_error_msg_and_die(
 243                                        "--setgroups=allow and --map-root-user "
 244                                        "are mutually exclusive"
 245                                );
 246                        }
 247                } else {
 248                        /* It's not "allow", must be "deny" */
 249                        if (strcmp(setgrp_str, "deny") != 0)
 250                                bb_error_msg_and_die("unrecognized: --%s=%s",
 251                                        "setgroups", setgrp_str);
 252                }
 253        }
 254
 255        unsflags = 0;
 256        need_mount = 0;
 257        for (i = 0; i < NS_COUNT; i++) {
 258                const struct namespace_descr *ns = &ns_list[i];
 259                struct namespace_ctx *ns_ctx = &ns_ctx_list[i];
 260
 261                if (opts & (1 << i))
 262                        unsflags |= ns->flag;
 263
 264                need_mount |= (uintptr_t)(ns_ctx->path);
 265        }
 266        /* need_mount != 0 if at least one FILE was given */
 267
 268        prop_flags = MS_REC | MS_PRIVATE;
 269        /* Silently ignore --propagation if --mount is not requested. */
 270        if (opts & OPT_mount)
 271                prop_flags = parse_propagation(prop_str);
 272
 273        /*
 274         * Special case: if we were requested to unshare the mount namespace
 275         * AND to make any namespace persistent (by bind mounting it) we need
 276         * to spawn a child process which will wait for the parent to call
 277         * unshare(), then mount parent's namespaces while still in the
 278         * previous namespace.
 279         */
 280        fdp.wr = -1;
 281        if (need_mount && (opts & OPT_mount)) {
 282                /*
 283                 * Can't use getppid() in child, as we can be unsharing the
 284                 * pid namespace.
 285                 */
 286                pid_t ppid = getpid();
 287
 288                xpiped_pair(fdp);
 289
 290                child = xfork();
 291                if (child == 0) {
 292                        /* Child */
 293                        close(fdp.wr);
 294
 295                        /* Wait until parent calls unshare() */
 296                        read(fdp.rd, ns_ctx_list, 1); /* ...using bogus buffer */
 297                        /*close(fdp.rd);*/
 298
 299                        /* Mount parent's unshared namespaces. */
 300                        mount_namespaces(ppid, ns_ctx_list);
 301                        return EXIT_SUCCESS;
 302                }
 303                /* Parent continues */
 304        }
 305
 306        if (unshare(unsflags) != 0)
 307                bb_perror_msg_and_die("unshare(0x%x)", unsflags);
 308
 309        if (fdp.wr >= 0) {
 310                close(fdp.wr); /* Release child */
 311                close(fdp.rd); /* should close fd, to not confuse exec'ed PROG */
 312        }
 313
 314        if (need_mount) {
 315                /* Wait for the child to finish mounting the namespaces. */
 316                if (opts & OPT_mount) {
 317                        int exit_status = wait_for_exitstatus(child);
 318                        if (WIFEXITED(exit_status) &&
 319                            WEXITSTATUS(exit_status) != EXIT_SUCCESS)
 320                                return WEXITSTATUS(exit_status);
 321                } else {
 322                        /*
 323                         * Regular way - we were requested to mount some other
 324                         * namespaces: mount them after the call to unshare().
 325                         */
 326                        mount_namespaces(getpid(), ns_ctx_list);
 327                }
 328        }
 329
 330        /*
 331         * When we're unsharing the pid namespace, it's not the process that
 332         * calls unshare() that is put into the new namespace, but its first
 333         * child. The user may want to use this option to spawn a new process
 334         * that'll become PID 1 in this new namespace.
 335         */
 336        if (opts & OPT_fork) {
 337                xvfork_parent_waits_and_exits();
 338                /* Child continues */
 339        }
 340
 341        if (opts & OPT_map_root) {
 342                char uidmap_buf[sizeof("0 %u 1") + sizeof(int)*3];
 343
 344                /*
 345                 * Since Linux 3.19 unprivileged writing of /proc/self/gid_map
 346                 * has been disabled unless /proc/self/setgroups is written
 347                 * first to permanently disable the ability to call setgroups
 348                 * in that user namespace.
 349                 */
 350                xopen_xwrite_close(PATH_PROC_SETGROUPS, "deny");
 351                sprintf(uidmap_buf, "0 %u 1", (unsigned)reuid);
 352                xopen_xwrite_close(PATH_PROC_UIDMAP, uidmap_buf);
 353                sprintf(uidmap_buf, "0 %u 1", (unsigned)regid);
 354                xopen_xwrite_close(PATH_PROC_GIDMAP, uidmap_buf);
 355        } else
 356        if (setgrp_str) {
 357                /* Write "allow" or "deny" */
 358                xopen_xwrite_close(PATH_PROC_SETGROUPS, setgrp_str);
 359        }
 360
 361        if (opts & OPT_mount) {
 362                mount_or_die("none", "/", NULL, prop_flags);
 363        }
 364
 365        if (opts & OPT_mount_proc) {
 366                /*
 367                 * When creating a new pid namespace, we might want the pid
 368                 * subdirectories in /proc to remain consistent with the new
 369                 * process IDs. Without --mount-proc the pids in /proc would
 370                 * still reflect the old pid namespace. This is why we make
 371                 * /proc private here and then do a fresh mount.
 372                 */
 373                mount_or_die("none", proc_mnt_target, NULL, MS_PRIVATE | MS_REC);
 374                mount_or_die("proc", proc_mnt_target, "proc", MS_NOSUID | MS_NOEXEC | MS_NODEV);
 375        }
 376
 377        exec_prog_or_SHELL(argv);
 378}
 379