#!/usr/bin/perl # # Copyright: 2015 Johannes Schauer # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to # deal in the Software without restriction, including without limitation the # rights to use, copy, modify, merge, publish, distribute, sublicense, and/or # sell copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in # all copies or substantial portions of the Software. # # # This tool tries to be like unshare(1) but with the power of lxc-usernsexec(1) # to map more than one id into the new user namespace by using the programs # newgidmap and newuidmap. Or in other words: This tool tries to be like # lxc-usernsexec(1) but with the power of unshare(1) to unshare more than just # the user and mount namespaces. # # I arbitrarily called it user-unshare because it's like unshare(1) but # doesn't require root (except for the suid root new[ug]idmap programs). # # It is essentially equal to calling: # # $ lxc-usernsexec [opts] -- unshare [opts] -- COMMAND # # Its main points of existence are: # # - as a project for me to learn how unprivileged namespaces work # - written in Perl which means: # - architecture independent (same executable on any architecture) # - easily inspectable by other curious minds # - tons of code comments to let others understand how things work # - no need to install the lxc package in a minimal environment (perl itself # might not be called minimal either but is present in every Debian # installation) # # Further differences to unshare(1): # # - No --setgroups and --map-root-user options. These were only useful when # not more than a single uid and gid was available inside the user # namespace # - CLONE_NEWUSER is always enabled (so there is no --user option) # - persistent namespaces to be used with nsenter cannot be supported because # they require that you have permissions to do `mount --bind` in the host # namespace which requires root privileges # - the --mount-proc=XXX option actually works with XXX being other # directories than /proc # - you are immediately root in the new user namespace and can then switch to # another user with runuser(1) if you so wish # # Further differences to lxc-usernsexec(1): # # - there is no parent that is just waiting for the child to exit and # otherwise wasting pid space (unless you request --fork) # - requires only one pipe instead of two for IPC # - the -m option is now uppercase -M because -m was already the short option # for --mount in unshare(1) # # How it differs from other tools: # # - systemd-nspawn requires to be executed by root and this does not seem # likely to change any time soon: # http://lists.freedesktop.org/archives/systemd-devel/2015-February/028139.html # - linux-user-chroot cheats by being suid root # # # Debian kernels carry a patch named # add-sysctl-to-disallow-unprivileged-CLONE_NEWUSER-by-default.patch # which disables unprivileged usernamespaces by default, to enable it do: # $ echo 1 | sudo tee /proc/sys/kernel/unprivileged_userns_clone > /dev/null # or # $ sysctl -w kernel.unprivileged_userns_clone=1 use warnings; use strict; require 'syscall.ph'; use Getopt::Long; use Pod::Usage; # read the files /etc/subuid and /etc/subgid and return the mapping of user and # group ids, respectively for the given username sub read_subuid_subgid() { my $username = getpwuid $<; my ($subid, $num_subid, $fh, $n); # map the current user and group id to the root user by default my @result = (["u", 0, $<, 1], ["g", 0, $<, 1]); if (-f "/etc/subuid") { open $fh, "<", "/etc/subuid" or die "cannot open /etc/subuid for reading: $!"; while (my $line = <$fh>) { ($n, $subid, $num_subid) = split(/:/, $line, 3); last if ($n eq $username); } close $fh; # all other user ids start after root push @result, ["u", 1, $subid, $num_subid]; } if ( -f "/etc/subgid") { open $fh, "<", "/etc/subgid" or die "cannot open /etc/subgid for reading: $!"; while (my $line = <$fh>) { ($n, $subid, $num_subid) = split(/:/, $line, 3); last if ($n eq $username); } close $fh; # all other group ids start after root push @result, ["g", 1, $subid, $num_subid]; } return @result; } # FIXME: get rid of these constants # from sched.h my $CLONE_NEWNS = 0x20000; my $CLONE_NEWUTS = 0x4000000; my $CLONE_NEWIPC = 0x8000000; my $CLONE_NEWUSER = 0x10000000; my $CLONE_NEWPID = 0x20000000; my $CLONE_NEWNET = 0x40000000; # from sys/mount.h my $MS_NOSUID = 0x2; my $MS_NODEV = 0x4; my $MS_NOEXEC = 0x8; my $MS_BIND = 0x1000; my $MS_REC = 0x4000; my $MS_PRIVATE = 0x40000; my $MS_SLAVE = 0x80000; my $unshare_flags = $CLONE_NEWUSER; my @idmap; my $procmnt = undef; my $fork; Getopt::Long::Configure ("bundling"); GetOptions( 'h|help' => sub { pod2usage(-exitval => 0, -verbose => 2) }, "i|ipc" => sub { $unshare_flags |= $CLONE_NEWIPC }, "m|mount" => sub { $unshare_flags |= $CLONE_NEWNS }, "n|net" => sub { $unshare_flags |= $CLONE_NEWNET }, "p|pid" => sub { $unshare_flags |= $CLONE_NEWPID }, "u|uts" => sub { $unshare_flags |= $CLONE_NEWUTS }, "mount-proc:s" => sub { $fork = 1; # mounting proc requires a forked child or otherwise EPERM $procmnt = $_[1] eq "" ? "/proc" : $_[1]; # /proc is the default $unshare_flags |= $CLONE_NEWNS; }, # implicitly enable --mount "f|fork" => \$fork, "M=s" => sub { # parse -M option and make sure it consists of four parts my @v = split /:/, $_[1], 4; die "invalid format for -m option" if scalar @v != 4; push @idmap, \@v }, ) or pod2usage(-exitval => 2, -verbose => 1); my $origuid = $<; my $origgid = $(; # If the -M option was not given, read /etc/sub[ug]id to get the right # sub[ug]ids for the new[gu]idmap calls later by the child process. # # new[ug]idmap is called with the exact values from /etc/sub[ug]id, # respectively. if (scalar @idmap == 0) { @idmap = read_subuid_subgid; } # Create a pipe for the parent process to signal the child process that it is # done with calling unshare() so that the child can go ahead setting up # uid_map and gid_map. pipe my $rfh, my $wfh; # We have to do this dance with forking a process and then modifying the # parent from the child because: # - new[ug]idmap can only be called on a process id after that process has # unshared the user namespace # - a process looses its capabilities if it performs an execve() with nonzero # user ids see the capabilities(7) man page for details. # - a process that unshared the user namespace by default does not have the # privileges to call new[ug]idmap on itself # # this also works the other way around (the child setting up a user namespace # and being modified from the parent) but that way, the parent would have to # stay around until the child exited (so a pid would be wasted). Additionally, # that variant would require an additional pipe to let the parent signal the # child that it is done with calling new[ug]idmap. The way it is done here, # this signaling can instead be done by wait()-ing for the exit of the child. my $ppid = $$; my $cpid = fork() // die "fork() failed: $!"; if ($cpid == 0) { # child # Close the writing descriptor at our end of the pipe so that we see EOF # when parent closes its descriptor. close $wfh; # Wait for the parent process to finish its unshare() call by waiting for # an EOF. 0 == sysread $rfh, my $c, 1 or die "read() did not receive EOF"; # The program's new[ug]idmap have to be used because they are setuid root. # These privileges are needed to map the ids from /etc/sub[ug]id to the # user namespace set up by the parent. Without these privileges, only the # id of the user itself can be mapped into the new namespace. # # Since new[ug]idmap is setuid root we also don't need to write "deny" to # /proc/$$/setgroups beforehand (this is otherwise required for # unprivileged processes trying to write to /proc/$$/gid_map since kernel # version 3.19 for security reasons) and therefore the parent process # keeps its ability to change its own group here. # # Since /proc/$ppid/[ug]id_map can only be written to once, respectively, # instead of making multiple calls to new[ug]idmap, we assemble a command # line that makes one call each. my $uidmapcmd = ""; my $gidmapcmd = ""; foreach (@idmap) { my ($t, $hostid, $nsid, $range) = @{$_}; if ($t ne "u" and $t ne "g" and $t ne "b") { die "invalid idmap type: $t"; } if ($t eq "u" or $t eq "b") { $uidmapcmd .= " $hostid $nsid $range"; } if ($t eq "g" or $t eq "b") { $gidmapcmd .= " $hostid $nsid $range"; } } if ($uidmapcmd ne "") { 0 == system "newuidmap $ppid $uidmapcmd" or die "newuidmap failed: $!"; } if ($gidmapcmd ne "") { 0 == system "newgidmap $ppid $gidmapcmd" or die "newgidmap failed: $!"; } exit 0; } # parent # After fork()-ing, the parent immediately calls unshare... 0 == syscall &SYS_unshare, $unshare_flags or die "unshare() failed: $!"; # .. and then signals the child process that we are done with the unshare() # call by sending an EOF. close $wfh; # Wait for the child process to finish its setup by waiting for its exit. $cpid == waitpid $cpid, 0 or die "waitpid() failed: $!"; if ($? != 0) { die "child had a non-zero exit status: $?"; } # Currently we are nobody (uid and gid are 65534). So we become root user and # group instead. # # We are using direct syscalls instead of setting $(, $), $< and $> because # then perl would do additional stuff which we don't need or want here, like # checking /proc/sys/kernel/ngroups_max (which might not exist). It would also # also call setgroups() in a way that makes the root user be part of the # group unknown. 0 == syscall &SYS_setgid, 0 or die "setgid failed: $!"; 0 == syscall &SYS_setuid, 0 or die "setuid failed: $!"; 0 == syscall &SYS_setgroups, 0, 0 or die "setgroups failed: $!"; # At this point lxc-usernsexec is checking whether / is mounted MS_SHARED and # if yes, will mount / as MS_SLAVE|MS_REC - not sure why we have to do this # and it seems to work without so we don't do it. #0 == syscall &SYS_mount, 0, my $t = "/", 0, $MS_SLAVE | $MS_REC, 0 or die "mount() failed: $!"; # When the pid namespace is also unshared, then processes expect a master pid # to always be alive within the namespace. To achieve this, we fork() here # instead of exec() to always have one dummy process running as pid 1 inside # the namespace. This is also what the unshare tool does when used with the # --fork option. # # Otherwise, without a pid 1, new processes cannot be forked anymore after pid # 1 finished. if ($fork) { my $cpid = fork() // die "fork() failed: $!"; if ($cpid != 0) { # parent # The parent process will stay alive as pid 1 in this namespace until # the child finishes executing. This is important because pid 1 must # never die or otherwise nothing new can be forked. $cpid == waitpid $cpid, 0 or die "waitpid() failed: $!"; exit $?; } } if (defined $procmnt) { # Directly bind-mounting the system's /proc into $procmnt would be wrong # because we want to restrict proc to the new pid namespace. Thus we need # to mount a new proc. # # FIXME: Funnily, I wasn't able to figure out how to directly mount a new # proc into any other location than the old /proc without getting EINVAL, # therefore, the workaround is to remount the old /proc and then bind mount # that into the new location. # # This operation also fails with: # $ lxc-usernsexec -- unshare --mount-proc=... --fork --pid --mount -- ... # or even with: # $ unshare --mount-proc=... --fork --pid --mount --user -- ... # the authors of unshare(1) should thus be made aware of the workaround # below (or whatever the real fix ends up being): # # The "my $s = ..." and "my $t = ..." constructs are necessary because # string literals cannot be passed as arguments to the syscall function. 0 == syscall &SYS_mount, my $s1 = "none", my $t1 = "/proc", 0, $MS_PRIVATE | $MS_REC, 0 or die "mount() failed: $!"; # FIXME: The following line will fail if --fork wasn't passed. Why??? 0 == syscall &SYS_mount, my $s2 = "proc", my $t2 = "/proc", my $t = "proc", $MS_NOSUID|$MS_NOEXEC|$MS_NODEV, 0 or die "mount() failed: $!"; # Only do a bindmount if proc was required to be mounted elsewhere than # /proc if ($procmnt ne "/proc") { 0 == syscall &SYS_mount, my $s = "/proc", $procmnt, 0, $MS_BIND | $MS_REC, 0 or die "mount failed: $!"; } } ## having this variable set could cause programs looking for unreachable ## machines # see http://bugs.debian.org/780587 #delete $ENV{http_proxy}; # finally, exec our program exec @ARGV or die "exec() failed: $!"; __END__ =head1 NAME user-unshare - unprivileged linux namespaces with uid and gid maps =head1 SYNOPSIS user-unshare [options] [--] command =head1 DESCRIPTION This tool combines the ability of lxc-usernsexec(1) to map ranges of user and group ids into a new user namespace with the ability of unshare(1) to unshare several different kinds of namespaces. The options of this command are nearly a plain copy of lxc-usernsexec(1) and unshare(1). =head1 OPTIONS =over 8 =item B<-h, --help> Print a brief help message and exits. =item B<-M> I The uid map to use in the user namespace. Each map consists of four colon-separate values. First a character 'u', 'g' or 'b' to specify whether this map pertains to user ids, group ids, or both; next the first userid in the user namespace; next the first userid as seen on the host; and finally the number of ids to be mapped. More than one map can be specified. If no map is specified, then by default the full uid and gid ranges granted by /etc/subuid and /etc/subgid will be mapped to the uids and gids starting at 0 in the container. Note that lxc-usernsexec always tries to setuid and setgid to 0 in the namespace. Therefore uid 0 in the namespace must be mapped. =item B<-i, --ipc> Unshare the IPC namespace. =item B<-m, --mount> Unshare the mount namespace. =item B<-n, --net> Unshare the network namespace. =item B<-p, --pid> Unshare the pid namespace. See also the --fork and --mount-proc options. =item B<-u, --uts> Unshare the UTS namespace. =item B<-f, --fork> Fork the specified program as a child process of unshare rather than running it directly. This is useful when creating a new pid namespace. =item B<--mount-proc>I<[=mountpoint]> Just before running the program, mount the proc filesystem at I (default is /proc). This is useful when creating a new pid namespace. It also implies creating a new mount namespace since the /proc mount would otherwise mess up existing programs on the system. The new proc filesystem is explicitly mounted as private (by MS_PRIVATE|MS_REC). =back =head1 EXAMPLES This: lxc-usernsexec -- unshare --mount-proc --fork --ipc --pid --net --mount \ -- sh -c "ls /proc | head && whoami && groups && ip link set lo up && ip addr && /sbin/runuser -u josch whoami" is equivalent to: user-unshare --mount-proc --fork --ipc --pid --net --mount \ -- sh -c "ls /proc | head && whoami && groups && ip link set lo up && ip addr && /sbin/runuser -u josch whoami" Notice that the options are exactly identical. If your user id is 1000, root in a container is mapped to 190000, and you wish to chown a file you own to root in the container, you can use: lxc-usernsexec -m b:0:1000:1 -m b:1:190000:1 -- /bin/chown 1:1 $file which is equivalent to: user-unshare -M b:0:1000:1 -M b:1:190000:1 -- /bin/chown 1:1 $file This maps your userid to root in the user namespace, and 190000 to uid 1. Since root in the user namespace is privileged over all userids mapped into the namespace, you are allowed to change the file ownership, which you could not do on the host using a simple chown. =head1 SEE ALSO unshare(1), lxc-usernsexec(1), user_namespaces(7), newgidmap(1), newuidmap(1) =cut