lxcsu

   1 #!/usr/bin/python
   2
   3 import sys
   4 import os
   5 import setns
   6
   7 from argparse import ArgumentParser
   8
   9 drop_capabilities='cap_sys_admin,cap_sys_boot,cap_sys_module'
  10
  11 # can set to True here, but also use the -d option
  12 debug = False
  13
  14 def getarch(f):
  15         output = os.popen('readelf -h %s 2>&1'%f).readlines()
  16         classlines = [x for x in output if ('Class' in x.split(':')[0])]
  17         line = classlines[0]
  18         c = line.split(':')[1]
  19         if ('ELF64' in c):
  20                 return 'x86_64'
  21         elif ('ELF32' in c):
  22                 return 'i686'
  23         else:
  24                 raise Exception('Could not determine architecture')
  25
  26 def umount(fs_dir, opts=''):
  27         output = os.popen('/bin/umount %s %s 2>&1'%(opts, fs_dir)).read()
  28         return ('device is busy' not in output)
  29
  30 def main ():
  31         parser = ArgumentParser()
  32         parser.add_argument("-n", "--nonet",
  33                                                 action="store_true", dest="netns", default=False,
  34                                                 help="Don't enter network namespace")
  35         parser.add_argument("-m", "--nomnt",
  36                                                 action="store_true", dest="mntns", default=False,
  37                                                 help="Don't enter mount namespace")
  38         parser.add_argument("-p", "--nopid",
  39                                                 action="store_true", dest="pidns", default=False,
  40                                                 help="Don't enter pid namespace")
  41         parser.add_argument("-r", "--root",
  42                                                 action="store_true", dest="root", default=False,
  43                                                 help="Enter as root: be careful")
  44         parser.add_argument("-i","--internal",
  45                                                 action="store_true", dest="internal", default=False,
  46                                                 help="does *not* prepend '-- -c' to arguments - or invoke lxcsu-internal")
  47         parser.add_argument("-d","--debug",
  48                                                 action='store_true', dest='debug', default=False,
  49                                                 help="debug option")
  50         parser.add_argument("-s","--nosliceuid",
  51                                                 action='store_true', dest="nosliceuid", default=False,
  52                                                 help="do not change to slice uid inside of slice")
  53         parser.add_argument("-o","--noslicehome",
  54                                                 action='store_true', dest="noslicehome", default=False,
  55                                                 help="do not change to slice home directory inside of slice")
  56
  57         if os.path.exists("/etc/lxcsu_default"):
  58                 defaults = parser.parse_args(file("/etc/lxcsu_default","r").read().split())
  59                 parser.set_defaults(**defaults.__dict__)
  60
  61         parser.add_argument ("slice_name")
  62         parser.add_argument ("command_to_run",nargs="*")
  63
  64         args = parser.parse_args()
  65         slice_name=args.slice_name
  66
  67         # unless we run the symlink 'lxcsu-internal', or we specify the -i option, prepend '--' '-c'
  68         if sys.argv[0].find('internal')>=0: args.internal=True
  69
  70         if len(args.command_to_run)>0 and (args.command_to_run[0] == "/sbin/service"):
  71                 # A quick hack to support nodemanager interfaces.py when restarting
  72                 # networking in a slice.
  73                 args.nosliceuid = True
  74
  75         # plain lxcsu
  76         if not args.internal:
  77                 # no command given: enter interactive shell
  78                 if not args.command_to_run: args.command_to_run=['/bin/sh']
  79                 args.command_to_run = [ '-c' ] + [" ".join(args.command_to_run)]
  80
  81         # support for either setting debug at the top of this file, or on the command-line
  82         if args.debug:
  83                 global debug
  84                 debug=True
  85
  86         try:
  87                 cmd = '/usr/bin/virsh --connect lxc:/// domid %s'%slice_name
  88                 pidnum = int(os.popen(cmd).read().rstrip())
  89         except:
  90                 print "Error finding slice %s"%slice_name
  91                 exit(1)
  92
  93         pid = '%s'%pidnum
  94         cmdline = open('/proc/%s/cmdline'%pidnum).read().rstrip('\n\x00')
  95         arch = getarch('/proc/%s/exe'%pid)
  96
  97         if (not pid):
  98                 print "Not started: %s"%slice_name
  99                 exit(1)
 100
 101         if arch is None:
 102                 arch = 'x86_64'
 103
 104         # Set sysctls specific to slice
 105         sysctls = []
 106         sysctl_dir = '/etc/planetlab/vsys-attributes/%s'%slice_name
 107         if (os.access(sysctl_dir,0)):
 108                 entries = os.listdir(sysctl_dir)
 109                 for e in entries:
 110                         prefix = 'vsys_sysctl.'
 111                         if (e.startswith(prefix)):
 112                                 sysctl_file = '/'.join([sysctl_dir,e])
 113                                 sysctl_name = e[len(prefix):]
 114                                 sysctl_val = open(sysctl_file).read()
 115                                 sysctls.append((sysctl_file, sysctl_name, sysctl_val))
 116
 117         # Enter cgroups
 118         try:
 119                 for subsystem in ['cpuset','memory','blkio']:
 120                         open('/sys/fs/cgroup/%s/libvirt/lxc/%s/tasks'%(subsystem,slice_name),'w').write(str(os.getpid()))
 121
 122         except:
 123                 print "Error assigning resources: %s"%slice_name
 124                 exit(1)
 125
 126         try:
 127                 open('/sys/fs/cgroup/cpuacct/system/libvirtd.service/libvirt/lxc/%s/tasks'%slice_name,'w').write(str(os.getpid()))
 128         except:
 129                 print "Error assigning cpuacct: %s" % slice_name
 130                 exit(1)
 131
 132         # If the slice is frozen, then we'll get an EBUSY when trying to write to the task
 133         # list for the freezer cgroup. Since the user couldn't do anything anyway, it's best
 134         # in this case to error out the shell. (an alternative would be to un-freeze it,
 135         # add the task, and re-freeze it)
 136         try:
 137                 f=open('/sys/fs/cgroup/freezer/libvirt/lxc/%s/tasks'%(slice_name),'w')
 138                 f.write(str(os.getpid()))
 139                 # note: we need to call f.close() explicitly, or we'll get an exception in
 140                 # the object destructor, which will not be caught
 141                 f.close()
 142         except:
 143                 print "Error adding task to freezer cgroup. Slice is probably frozen: %s" % slice_name
 144                 exit(1)
 145
 146         setns.chcontext('/proc/%s/ns/uts'%pid)
 147         setns.chcontext('/proc/%s/ns/ipc'%pid)
 148
 149         if (not args.pidns):
 150                 setns.chcontext('/proc/%s/ns/pid'%pid)
 151
 152         if (not args.netns):
 153                 setns.chcontext('/proc/%s/ns/net'%pid)
 154
 155         if (not args.mntns):
 156                 setns.chcontext('/proc/%s/ns/mnt'%pid)
 157
 158         proc_mounted = False
 159         if (not os.access('/proc/self',0)):
 160                 proc_mounted = True
 161                 setns.proc_mount()
 162
 163         for (sysctl_file, sysctl_name, sysctl_val) in sysctls:
 164                                 for fn in ["/sbin/sysctl", "/usr/sbin/sysctl", "/bin/sysctl", "/usr/bin/sysctl"]:
 165                                         if os.path.exists(fn):
 166                                                 os.system('%s -w %s=%s  >/dev/null 2>&1'%(fn, sysctl_name,sysctl_val))
 167                                                 break
 168                                 else:
 169                                         print "Error: image does not have a sysctl binary"
 170
 171         # cgroups is not yet LXC-safe, so we need to use the coarse grained access control
 172         # strategy of unmounting the filesystem
 173
 174         umount_result = True
 175         for subsystem in ['cpuset','cpu,cpuacct','memory','devices','freezer','net_cls','blkio','perf_event','systemd']:
 176                 fs_path = '/sys/fs/cgroup/%s'%subsystem
 177                 if (not umount(fs_path,'-l')):
 178                         pass
 179                         # Leaving these comments for historical reference
 180                         #print "Error disabling cgroup access"
 181                         #exit(1) - Don't need this because failure here implies failure in the call to umount /sys/fs/cgroup
 182
 183         if (not umount('/sys/fs/cgroup')):
 184                 print "Error disabling cgroup access"
 185                 exit(1)
 186
 187         pid = os.fork()
 188
 189         # capsh has a --user option starting with f14
 190         # so if only for f12 we need to fake this one
 191         #
 192         # capsh.c does essentially the following when invoked with --user:
 193         #               pwd = getpwnam(user); ...
 194         #               ngroups = MAX_GROUPS;
 195         #               status = getgrouplist(user, pwd->pw_gid, groups, &ngroups); ...
 196         #               status = setgroups(ngroups, groups); ...
 197         #               status = setgid(pwd->pw_gid); ...
 198         #               status = setuid(pwd->pw_uid); ...
 199         # however we cannot simulate that ourselves because if we did in this process then
 200         # capsh could not be allowed to mess with caps any more
 201
 202         def getuid (slicename):
 203                 import pwd
 204                 try:
 205                         return pwd.getpwnam(slicename).pw_uid
 206                 except:
 207                         return
 208
 209         if (pid == 0):
 210                 cap_arg = '--drop='+drop_capabilities
 211
 212                 if (not args.root):
 213                         if (args.nosliceuid):
 214                                 # we still want to drop capabilities, but don't want to switch UIDs
 215                                 exec_args = [arch,'/usr/sbin/capsh',cap_arg,'--','--login',]+args.command_to_run
 216                         else:
 217                                 uid = getuid (slice_name)
 218                                 if not uid:
 219                                         print "lxcsu could not spot %s in /etc/passwd - exiting"%slice_name
 220                                         exit(1)
 221                                 exec_args = [arch,'/usr/sbin/capsh',cap_arg,'--uid=%s'%uid,'--','--login',]+args.command_to_run
 222 # once we can drop f12, it would be nicer to instead go for
 223 #                                exec_args = [arch,'/usr/sbin/capsh',cap_arg,'--user=%s'%slice_name,'--','--login',]+args.command_to_run
 224                 else:
 225                         exec_args = [arch,'/usr/sbin/capsh','--','--login']+args.command_to_run
 226
 227                 os.environ['SHELL'] = '/bin/sh'
 228                 if os.path.exists('/etc/planetlab/lib/bind_public.so'):
 229                         os.environ['LD_PRELOAD'] = '/etc/planetlab/lib/bind_public.so'
 230                 if not args.noslicehome:
 231                         os.environ['HOME'] = '/home/%s'%slice_name
 232                         os.chdir("/home/%s"%(slice_name))
 233                 if debug: print 'lxcsu:execv:','/usr/bin/setarch',exec_args
 234                 os.execv('/usr/bin/setarch',exec_args)
 235         else:
 236                 setns.proc_umount()
 237                 _,status = os.waitpid(pid,0)
 238                 exit(os.WEXITSTATUS(status))
 239
 240 if __name__ == '__main__':
 241         main()