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"
  11//config:       default y
  12//config:       depends on LONG_OPTS && !NOMMU
  13//config:       select PLATFORM_LINUX
  14//config:       help
  15//config:         Run program with some namespaces unshared from parent.
  16
  17// depends on LONG_OPTS: it is awkward to exclude code which handles --propagation
  18// and --setgroups based on LONG_OPTS, so instead applet requires LONG_OPTS.
  19// depends on !NOMMU: we need fork()
  20
  21//applet:IF_UNSHARE(APPLET(unshare, BB_DIR_USR_BIN, BB_SUID_DROP))
  22
  23//kbuild:lib-$(CONFIG_UNSHARE) += unshare.o
  24
  25//usage:#define unshare_trivial_usage
  26//usage:       "[OPTIONS] [PROG [ARGS]]"
  27//usage:#define unshare_full_usage "\n"
  28//usage:     "\n        -m, --mount[=FILE]      Unshare mount namespace"
  29//usage:     "\n        -u, --uts[=FILE]        Unshare UTS namespace (hostname etc.)"
  30//usage:     "\n        -i, --ipc[=FILE]        Unshare System V IPC namespace"
  31//usage:     "\n        -n, --net[=FILE]        Unshare network namespace"
  32//usage:     "\n        -p, --pid[=FILE]        Unshare PID namespace"
  33//usage:     "\n        -U, --user[=FILE}       Unshare user namespace"
  34//usage:     "\n        -f, --fork              Fork before execing PROG"
  35//usage:     "\n        -r, --map-root-user     Map current user to root (implies -u)"
  36//usage:     "\n        --mount-proc[=DIR]      Mount /proc filesystem first (implies -m)"
  37//usage:     "\n        --propagation slave|shared|private|unchanged"
  38//usage:     "\n                                Modify mount propagation in mount namespace"
  39//usage:     "\n        --setgroups allow|deny  Control the setgroups syscall in user namespaces"
  40
  41#include <sched.h>
  42#ifndef CLONE_NEWUTS
  43# define CLONE_NEWUTS  0x04000000
  44#endif
  45#ifndef CLONE_NEWIPC
  46# define CLONE_NEWIPC  0x08000000
  47#endif
  48#ifndef CLONE_NEWUSER
  49# define CLONE_NEWUSER 0x10000000
  50#endif
  51#ifndef CLONE_NEWPID
  52# define CLONE_NEWPID  0x20000000
  53#endif
  54#ifndef CLONE_NEWNET
  55# define CLONE_NEWNET  0x40000000
  56#endif
  57
  58#include <sys/mount.h>
  59#ifndef MS_REC
  60# define MS_REC     (1 << 14)
  61#endif
  62#ifndef MS_PRIVATE
  63# define MS_PRIVATE (1 << 18)
  64#endif
  65#ifndef MS_SLAVE
  66# define MS_SLAVE   (1 << 19)
  67#endif
  68#ifndef MS_SHARED
  69# define MS_SHARED  (1 << 20)
  70#endif
  71
  72#include "libbb.h"
  73
  74static void mount_or_die(const char *source, const char *target,
  75                 const char *fstype, unsigned long mountflags)
  76{
  77        if (mount(source, target, fstype, mountflags, NULL)) {
  78                bb_perror_msg_and_die("can't mount %s on %s (flags:0x%lx)",
  79                        source, target, mountflags);
  80                /* fstype is always either NULL or "proc".
  81                 * "proc" is only used to mount /proc.
  82                 * No need to clutter up error message with fstype,
  83                 * it is easily deductible.
  84                 */
  85        }
  86}
  87
  88#define PATH_PROC_SETGROUPS     "/proc/self/setgroups"
  89#define PATH_PROC_UIDMAP        "/proc/self/uid_map"
  90#define PATH_PROC_GIDMAP        "/proc/self/gid_map"
  91
  92struct namespace_descr {
  93        int flag;
  94        const char nsfile4[4];
  95};
  96
  97struct namespace_ctx {
  98        char *path;
  99};
 100
 101enum {
 102        OPT_mount       = 1 << 0,
 103        OPT_uts         = 1 << 1,
 104        OPT_ipc         = 1 << 2,
 105        OPT_net         = 1 << 3,
 106        OPT_pid         = 1 << 4,
 107        OPT_user        = 1 << 5, /* OPT_user, NS_USR_POS, and ns_list[] index must match! */
 108        OPT_fork        = 1 << 6,
 109        OPT_map_root    = 1 << 7,
 110        OPT_mount_proc  = 1 << 8,
 111        OPT_propagation = 1 << 9,
 112        OPT_setgroups   = 1 << 10,
 113};
 114enum {
 115        NS_MNT_POS = 0,
 116        NS_UTS_POS,
 117        NS_IPC_POS,
 118        NS_NET_POS,
 119        NS_PID_POS,
 120        NS_USR_POS, /* OPT_user, NS_USR_POS, and ns_list[] index must match! */
 121        NS_COUNT,
 122};
 123static const struct namespace_descr ns_list[] = {
 124        { CLONE_NEWNS,   "mnt"  },
 125        { CLONE_NEWUTS,  "uts"  },
 126        { CLONE_NEWIPC,  "ipc"  },
 127        { CLONE_NEWNET,  "net"  },
 128        { CLONE_NEWPID,  "pid"  },
 129        { CLONE_NEWUSER, "user" }, /* OPT_user, NS_USR_POS, and ns_list[] index must match! */
 130};
 131
 132/*
 133 * Upstream unshare doesn't support short options for --mount-proc,
 134 * --propagation, --setgroups.
 135 * Optional arguments (namespace mountpoints) exist only for long opts,
 136 * we are forced to use "fake" letters for them.
 137 * '+': stop at first non-option.
 138 */
 139static const char opt_str[] ALIGN1 = "+muinpU""fr""\xfd::""\xfe:""\xff:";
 140static const char unshare_longopts[] ALIGN1 =
 141        "mount\0"               Optional_argument       "\xf0"
 142        "uts\0"                 Optional_argument       "\xf1"
 143        "ipc\0"                 Optional_argument       "\xf2"
 144        "net\0"                 Optional_argument       "\xf3"
 145        "pid\0"                 Optional_argument       "\xf4"
 146        "user\0"                Optional_argument       "\xf5"
 147        "fork\0"                No_argument             "f"
 148        "map-root-user\0"       No_argument             "r"
 149        "mount-proc\0"          Optional_argument       "\xfd"
 150        "propagation\0"         Required_argument       "\xfe"
 151        "setgroups\0"           Required_argument       "\xff"
 152;
 153
 154/* Ugly-looking string reuse trick */
 155#define PRIVATE_STR   "private\0""unchanged\0""shared\0""slave\0"
 156#define PRIVATE_UNCHANGED_SHARED_SLAVE   PRIVATE_STR
 157
 158static unsigned long parse_propagation(const char *prop_str)
 159{
 160        int i = index_in_strings(PRIVATE_UNCHANGED_SHARED_SLAVE, prop_str);
 161        if (i < 0)
 162                bb_error_msg_and_die("unrecognized: --%s=%s", "propagation", prop_str);
 163        if (i == 0)
 164                return MS_REC | MS_PRIVATE;
 165        if (i == 1)
 166                return 0;
 167        if (i == 2)
 168                return MS_REC | MS_SHARED;
 169        return MS_REC | MS_SLAVE;
 170}
 171
 172static void mount_namespaces(pid_t pid, struct namespace_ctx *ns_ctx_list)
 173{
 174        const struct namespace_descr *ns;
 175        struct namespace_ctx *ns_ctx;
 176        int i;
 177
 178        for (i = 0; i < NS_COUNT; i++) {
 179                char nsf[sizeof("/proc/%u/ns/AAAA") + sizeof(int)*3];
 180
 181                ns = &ns_list[i];
 182                ns_ctx = &ns_ctx_list[i];
 183                if (!ns_ctx->path)
 184                        continue;
 185                sprintf(nsf, "/proc/%u/ns/%.4s", (unsigned)pid, ns->nsfile4);
 186                mount_or_die(nsf, ns_ctx->path, NULL, MS_BIND);
 187        }
 188}
 189
 190int unshare_main(int argc, char **argv) MAIN_EXTERNALLY_VISIBLE;
 191int unshare_main(int argc UNUSED_PARAM, char **argv)
 192{
 193        int i;
 194        unsigned int opts;
 195        int unsflags;
 196        uintptr_t need_mount;
 197        const char *proc_mnt_target;
 198        const char *prop_str;
 199        const char *setgrp_str;
 200        unsigned long prop_flags;
 201        uid_t reuid = geteuid();
 202        gid_t regid = getegid();
 203        struct fd_pair fdp;
 204        pid_t child = child; /* for compiler */
 205        struct namespace_ctx ns_ctx_list[NS_COUNT];
 206
 207        memset(ns_ctx_list, 0, sizeof(ns_ctx_list));
 208        proc_mnt_target = "/proc";
 209        prop_str = PRIVATE_STR;
 210        setgrp_str = NULL;
 211
 212        opt_complementary =
 213                "\xf0""m" /* long opts (via their "fake chars") imply short opts */
 214                ":\xf1""u"
 215                ":\xf2""i"
 216                ":\xf3""n"
 217                ":\xf4""p"
 218                ":\xf5""U"
 219                ":ru"      /* --map-root-user or -r implies -u */
 220                ":\xfd""m" /* --mount-proc implies -m */
 221        ;
 222        applet_long_options = unshare_longopts;
 223        opts = getopt32(argv, opt_str,
 224                        &proc_mnt_target, &prop_str, &setgrp_str,
 225                        &ns_ctx_list[NS_MNT_POS].path,
 226                        &ns_ctx_list[NS_UTS_POS].path,
 227                        &ns_ctx_list[NS_IPC_POS].path,
 228                        &ns_ctx_list[NS_NET_POS].path,
 229                        &ns_ctx_list[NS_PID_POS].path,
 230                        &ns_ctx_list[NS_USR_POS].path
 231        );
 232        argv += optind;
 233        //bb_error_msg("opts:0x%x", opts);
 234        //bb_error_msg("mount:%s", ns_ctx_list[NS_MNT_POS].path);
 235        //bb_error_msg("proc_mnt_target:%s", proc_mnt_target);
 236        //bb_error_msg("prop_str:%s", prop_str);
 237        //bb_error_msg("setgrp_str:%s", setgrp_str);
 238        //exit(1);
 239
 240        if (setgrp_str) {
 241                if (strcmp(setgrp_str, "allow") == 0) {
 242                        if (opts & OPT_map_root) {
 243                                bb_error_msg_and_die(
 244                                        "--setgroups=allow and --map-root-user "
 245                                        "are mutually exclusive"
 246                                );
 247                        }
 248                } else {
 249                        /* It's not "allow", must be "deny" */
 250                        if (strcmp(setgrp_str, "deny") != 0)
 251                                bb_error_msg_and_die("unrecognized: --%s=%s",
 252                                        "setgroups", setgrp_str);
 253                }
 254        }
 255
 256        unsflags = 0;
 257        need_mount = 0;
 258        for (i = 0; i < NS_COUNT; i++) {
 259                const struct namespace_descr *ns = &ns_list[i];
 260                struct namespace_ctx *ns_ctx = &ns_ctx_list[i];
 261
 262                if (opts & (1 << i))
 263                        unsflags |= ns->flag;
 264
 265                need_mount |= (uintptr_t)(ns_ctx->path);
 266        }
 267        /* need_mount != 0 if at least one FILE was given */
 268
 269        prop_flags = MS_REC | MS_PRIVATE;
 270        /* Silently ignore --propagation if --mount is not requested. */
 271        if (opts & OPT_mount)
 272                prop_flags = parse_propagation(prop_str);
 273
 274        /*
 275         * Special case: if we were requested to unshare the mount namespace
 276         * AND to make any namespace persistent (by bind mounting it) we need
 277         * to spawn a child process which will wait for the parent to call
 278         * unshare(), then mount parent's namespaces while still in the
 279         * previous namespace.
 280         */
 281        fdp.wr = -1;
 282        if (need_mount && (opts & OPT_mount)) {
 283                /*
 284                 * Can't use getppid() in child, as we can be unsharing the
 285                 * pid namespace.
 286                 */
 287                pid_t ppid = getpid();
 288
 289                xpiped_pair(fdp);
 290
 291                child = xfork();
 292                if (child == 0) {
 293                        /* Child */
 294                        close(fdp.wr);
 295
 296                        /* Wait until parent calls unshare() */
 297                        read(fdp.rd, ns_ctx_list, 1); /* ...using bogus buffer */
 298                        /*close(fdp.rd);*/
 299
 300                        /* Mount parent's unshared namespaces. */
 301                        mount_namespaces(ppid, ns_ctx_list);
 302                        return EXIT_SUCCESS;
 303                }
 304                /* Parent continues */
 305        }
 306
 307        if (unshare(unsflags) != 0)
 308                bb_perror_msg_and_die("unshare(0x%x)", unsflags);
 309
 310        if (fdp.wr >= 0) {
 311                close(fdp.wr); /* Release child */
 312                close(fdp.rd); /* should close fd, to not confuse exec'ed PROG */
 313        }
 314
 315        if (need_mount) {
 316                /* Wait for the child to finish mounting the namespaces. */
 317                if (opts & OPT_mount) {
 318                        int exit_status = wait_for_exitstatus(child);
 319                        if (WIFEXITED(exit_status) &&
 320                            WEXITSTATUS(exit_status) != EXIT_SUCCESS)
 321                                return WEXITSTATUS(exit_status);
 322                } else {
 323                        /*
 324                         * Regular way - we were requested to mount some other
 325                         * namespaces: mount them after the call to unshare().
 326                         */
 327                        mount_namespaces(getpid(), ns_ctx_list);
 328                }
 329        }
 330
 331        /*
 332         * When we're unsharing the pid namespace, it's not the process that
 333         * calls unshare() that is put into the new namespace, but its first
 334         * child. The user may want to use this option to spawn a new process
 335         * that'll become PID 1 in this new namespace.
 336         */
 337        if (opts & OPT_fork) {
 338                xvfork_parent_waits_and_exits();
 339                /* Child continues */
 340        }
 341
 342        if (opts & OPT_map_root) {
 343                char uidmap_buf[sizeof("%u 0 1") + sizeof(int)*3];
 344
 345                /*
 346                 * Since Linux 3.19 unprivileged writing of /proc/self/gid_map
 347                 * has been disabled unless /proc/self/setgroups is written
 348                 * first to permanently disable the ability to call setgroups
 349                 * in that user namespace.
 350                 */
 351                xopen_xwrite_close(PATH_PROC_SETGROUPS, "deny");
 352                sprintf(uidmap_buf, "%u 0 1", (unsigned)reuid);
 353                xopen_xwrite_close(PATH_PROC_UIDMAP, uidmap_buf);
 354                sprintf(uidmap_buf, "%u 0 1", (unsigned)regid);
 355                xopen_xwrite_close(PATH_PROC_GIDMAP, uidmap_buf);
 356        } else
 357        if (setgrp_str) {
 358                /* Write "allow" or "deny" */
 359                xopen_xwrite_close(PATH_PROC_SETGROUPS, setgrp_str);
 360        }
 361
 362        if (opts & OPT_mount) {
 363                mount_or_die("none", "/", NULL, prop_flags);
 364        }
 365
 366        if (opts & OPT_mount_proc) {
 367                /*
 368                 * When creating a new pid namespace, we might want the pid
 369                 * subdirectories in /proc to remain consistent with the new
 370                 * process IDs. Without --mount-proc the pids in /proc would
 371                 * still reflect the old pid namespace. This is why we make
 372                 * /proc private here and then do a fresh mount.
 373                 */
 374                mount_or_die("none", proc_mnt_target, NULL, MS_PRIVATE | MS_REC);
 375                mount_or_die("proc", proc_mnt_target, "proc", MS_NOSUID | MS_NOEXEC | MS_NODEV);
 376        }
 377
 378        exec_prog_or_SHELL(argv);
 379}
 380