slicesu

   1 #!/usr/bin/python
   2
   3 import sys
   4 import os
   5 import setns
   6 import pdb
   7 import re
   8
   9 from argparse import ArgumentParser
  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 get_cgroup_subdirs_for_pid(pid):
  27         cgroup_info_file = '/proc/%s/cgroup'%pid
  28         cgroup_lines = open(cgroup_info_file).read().splitlines()
  29
  30         subdirs = {}
  31         for line in cgroup_lines:
  32                 try:
  33                         _, cgroup_name, subdir = line.split(':')
  34                         subdirs[cgroup_name] = subdir
  35                 except Exception, e:
  36                         print "Error reading cgroup info: %s"%str(e)
  37                         pass
  38
  39         return subdirs
  40
  41
  42 def umount(fs_dir, opts=''):
  43         output = os.popen('/bin/umount %s %s 2>&1'%(opts, fs_dir)).read()
  44         return ('device is busy' not in output)
  45
  46 def kvmsu(slice_name, args):
  47     mac_addresses = []
  48
  49     # Find MAC addresses of virt interfaces
  50     p = os.popen("virsh domiflist %s"%slice_name)
  51     for line in p:
  52         m = re.search(r'([0-9A-F]{2}[:-]){5}([0-9A-F]{2})', line, re.I)
  53         if m:
  54             mac_addresses.append(m.group())
  55
  56     # Find IP address of a reachable virt interface
  57     # Here we're taking the one attached to br-nat, but this is specific to OpenCloud
  58     ip_addr = None
  59     try:
  60         f = open('/var/lib/dnsmasq/br-nat.hosts', 'r')
  61         for line in f:
  62             for mac in mac_addresses:
  63                 m = re.search("^%s,[0-9.]+$"%mac, line, re.I)
  64                 if m:
  65                     ip_addr = m.group().split(',')[1]
  66                     break
  67         f.close()
  68     except:
  69         pass
  70
  71     if ip_addr:
  72         # Need to login to a Ubuntu Cloud image as 'ubuntu' user
  73         # Not sure how to make this more general
  74             os.execv('/usr/bin/ssh',
  75                     ['/usr/bin/ssh','-o','StrictHostKeyChecking=no','ubuntu@%s'%ip_addr]
  76                     + args.command_to_run)
  77     else:
  78         # What should happen here?
  79         print "Cannot map instance to IP address"
  80         exit(1)
  81
  82
  83
  84 def lxcsu(slice_name, args):
  85         # plain lxcsu
  86         if not args.internal:
  87                 # no command given: enter interactive shell
  88                 if not args.command_to_run: args.command_to_run=['/bin/sh']
  89                 args.command_to_run = [ '-c' ] + [" ".join(args.command_to_run)]
  90
  91         # support for either setting debug at the top of this file, or on the command-line
  92         if args.debug:
  93                 global debug
  94                 debug=True
  95
  96         try:
  97                 cmd = '/usr/bin/virsh --connect lxc:/// domid %s'%slice_name
  98                 # convert to int as a minimal raincheck
  99                 driver_pid = int(os.popen(cmd).read().strip())
 100                 # locate the pid for the - expected - single child, that would be the init for that VM
 101                 #init_pid = int(open("/proc/%s/task/%s/children"%(driver_pid,driver_pid)).read().strip())
 102                 init_pid = int(os.popen('pgrep -P %s'%driver_pid).readlines()[0].strip())
 103                 # Thierry: I am changing the code below to use child_pid instead of driver_pid
 104                 # for the namespace handling features, that I was able to check
 105                 # I've left the other ones as they were, i.e. using driver_pid, but I suspect
 106                 # they chould be changed as well
 107
 108         except:
 109                 print "Domain %s not found"%slice_name
 110                 exit(1)
 111
 112         if not driver_pid or not init_pid:
 113                 print "Domain %s not started"%slice_name
 114                 exit(1)
 115
 116         if debug: print "Found driver_pid",driver_pid,'and init_pid=',init_pid
 117         # xxx probably init_pid here too
 118         arch = getarch('/proc/%s/exe'%driver_pid)
 119
 120         # Set sysctls specific to slice
 121         sysctls = []
 122         sysctl_dir = '/etc/planetlab/vsys-attributes/%s'%slice_name
 123         if (os.access(sysctl_dir,0)):
 124                 entries = os.listdir(sysctl_dir)
 125                 for e in entries:
 126                         prefix = 'vsys_sysctl.'
 127                         if (e.startswith(prefix)):
 128                                 sysctl_file = '/'.join([sysctl_dir,e])
 129                                 sysctl_name = e[len(prefix):]
 130                                 sysctl_val = open(sysctl_file).read()
 131                                 sysctls.append((sysctl_file, sysctl_name, sysctl_val))
 132
 133         # xxx probably init_pid here too
 134         subdirs = get_cgroup_subdirs_for_pid(driver_pid)
 135         sysfs_root = '/sys/fs/cgroup'
 136
 137         # If the slice is frozen, then we'll get an EBUSY when trying to write to the task
 138         # list for the freezer cgroup. Since the user couldn't do anything anyway, it's best
 139         # in this case to error out the shell. (an alternative would be to un-freeze it,
 140         # add the task, and re-freeze it)
 141         # Enter cgroups
 142         current_cgroup = ''
 143         for subsystem in ['cpuset','memory','blkio','cpuacct','cpuacct,cpu','freezer']:
 144                 try:
 145                         current_cgroup = subsystem
 146
 147                         # There seems to be a bug in the cgroup schema: cpuacct,cpu can become cpu,cpuacct
 148                         # We need to handle both
 149                         task_path_alt = None
 150                         try:
 151                            subsystem_comps = subsystem.split(',')
 152                            subsystem_comps.reverse()
 153                            subsystem_alt = ','.join(subsystem_comps)
 154                            tasks_path_alt = [sysfs_root, subsystem_alt, subdirs[subsystem], 'tasks']
 155                         except Exception,e:
 156                                 pass
 157
 158                         tasks_path = [sysfs_root,subsystem,subdirs[subsystem],'tasks']
 159                         tasks_path_str = '/'.join(tasks_path)
 160
 161                         try:
 162                                 f = open(tasks_path_str, 'w')
 163                         except:
 164                                 tasks_path_alt_str = '/'.join(tasks_path_alt)
 165                                 f = open(tasks_path_alt_str, 'w')
 166
 167                         f.write(str(os.getpid()))
 168                         if (subsystem=='freezer'):
 169                                 f.close()
 170
 171                 except Exception,e:
 172                         if (not subdirs.has_key(subsystem)):
 173                                 pass
 174                         else:
 175                                 if debug: print e
 176                                 print "Error assigning cgroup %s (pid=%s) for slice %s"%(current_cgroup,driver_pid, slice_name)
 177                                 exit(1)
 178
 179
 180         def chcontext (path):
 181                 retcod = setns.chcontext (path)
 182                 if retcod != 0:
 183                         print 'WARNING - setns(%s)=>%s (ignored)'%(path,retcod)
 184                 return retcod
 185
 186         # Use init_pid and not driver_pid to locate reference namespaces
 187         ref_ns = "/proc/%s/ns/"%init_pid
 188
 189         if True:                                        chcontext(ref_ns+'uts')
 190         if True:                                        chcontext(ref_ns+'ipc')
 191
 192         if (not args.no_pidns):         chcontext(ref_ns+'pid')
 193         if (not args.no_netns):         chcontext(ref_ns+'net')
 194         if (not args.no_mntns):         chcontext(ref_ns+'mnt')
 195
 196         proc_mounted = False
 197         if (not os.access('/proc/self',0)):
 198                 proc_mounted = True
 199                 setns.proc_mount()
 200
 201         for (sysctl_file, sysctl_name, sysctl_val) in sysctls:
 202                 for fn in ["/sbin/sysctl", "/usr/sbin/sysctl", "/bin/sysctl", "/usr/bin/sysctl"]:
 203                         if os.path.exists(fn):
 204                                 os.system('%s -w %s=%s  >/dev/null 2>&1'%(fn, sysctl_name,sysctl_val))
 205                                 break
 206                         else:
 207                                 print "Error: image does not have a sysctl binary"
 208
 209         # cgroups is not yet LXC-safe, so we need to use the coarse grained access control
 210         # strategy of unmounting the filesystem
 211
 212         umount_result = True
 213         for subsystem in ['cpuset','cpu,cpuacct','memory','devices','freezer','net_cls','blkio','perf_event','systemd']:
 214                 fs_path = '/sys/fs/cgroup/%s'%subsystem
 215                 if (not umount(fs_path,'-l')):
 216                         print 'WARNING - umount failed (ignored) with path=',fs_path
 217                         pass
 218                         # Leaving these comments for historical reference
 219                         #print "Error disabling cgroup access"
 220                         #exit(1) - Don't need this because failure here implies failure in the call to umount /sys/fs/cgroup
 221
 222         if (not umount('/sys/fs/cgroup')):
 223                 print "Error disabling cgroup access"
 224                 exit(1)
 225
 226         fork_pid = os.fork()
 227
 228         def getuid (slicename):
 229                 import pwd
 230                 try:
 231                         return pwd.getpwnam(slicename).pw_uid
 232                 except:
 233                         return
 234
 235         if (fork_pid == 0):
 236                 if (not args.root):
 237                         setns.drop_caps()
 238                         if (args.nosliceuid):
 239                                 # we still want to drop capabilities, but don't want to switch UIDs
 240                                 exec_args = [arch,'/bin/sh','--login',]+args.command_to_run
 241                         else:
 242                                                                 # let's keep this check even though we don't use the uid
 243                                                                 # as a way of checking the existence of the slice account
 244                                 uid = getuid (slice_name)
 245                                 if not uid:
 246                                         print "lxcsu could not spot %s in /etc/passwd - exiting"%slice_name
 247                                         exit(1)
 248                                 exec_args = [arch,'/usr/bin/sudo','-u',slice_name,'/bin/sh','--login',]+args.command_to_run
 249 # once we can drop f12, it would be nicer to instead go for
 250 # exec_args = [arch,'/usr/sbin/capsh',cap_arg,'--user=%s'%slice_name,'--login',]+args.command_to_run
 251                 else:
 252                         exec_args = [arch,'/bin/sh','--login']+args.command_to_run
 253
 254                 os.environ['SHELL'] = '/bin/sh'
 255                 if os.path.exists('/etc/planetlab/lib/bind_public.so'):
 256                         os.environ['LD_PRELOAD'] = '/etc/planetlab/lib/bind_public.so'
 257                 if not args.noslicehome:
 258                         os.environ['HOME'] = '/home/%s'%slice_name
 259                         os.chdir("/home/%s"%(slice_name))
 260                 if debug: print 'lxcsu:execv:','/usr/bin/setarch',exec_args
 261                 os.execv('/usr/bin/setarch',exec_args)
 262         else:
 263                 setns.proc_umount()
 264                 _,status = os.waitpid(fork_pid,0)
 265                 exit(os.WEXITSTATUS(status))
 266
 267 def main ():
 268         parser = ArgumentParser()
 269         parser.add_argument("-n", "--nonet",
 270                                                                                 action="store_true", dest="no_netns", default=False,
 271                                                                                 help="Don't enter network namespace")
 272         parser.add_argument("-m", "--nomnt",
 273                                                                                 action="store_true", dest="no_mntns", default=False,
 274                                                                                 help="Don't enter mount namespace")
 275         parser.add_argument("-p", "--nopid",
 276                                                                                 action="store_true", dest="no_pidns", default=False,
 277                                                                                 help="Don't enter pid namespace")
 278         parser.add_argument("-r", "--root",
 279                                                                                 action="store_true", dest="root", default=False,
 280                                                                                 help="Enter as root: be careful")
 281         parser.add_argument("-i","--internal",
 282                                                                                 action="store_true", dest="internal", default=False,
 283                                                                                 help="does *not* prepend '-- -c' to arguments - or invoke lxcsu-internal")
 284         parser.add_argument("-d","--debug",
 285                                                                                 action='store_true', dest='debug', default=False,
 286                                                                                 help="debug option")
 287         parser.add_argument("-s","--nosliceuid",
 288                                                                                 action='store_true', dest="nosliceuid", default=False,
 289                                                                                 help="do not change to slice uid inside of slice")
 290         parser.add_argument("-o","--noslicehome",
 291                                                                                 action='store_true', dest="noslicehome", default=False,
 292                                                                                 help="do not change to slice home directory inside of slice")
 293
 294         if os.path.exists("/etc/lxcsu_default"):
 295                 defaults = parser.parse_args(file("/etc/lxcsu_default","r").read().split())
 296                 parser.set_defaults(**defaults.__dict__)
 297
 298         parser.add_argument ("slice_name")
 299         parser.add_argument ("command_to_run",nargs="*")
 300
 301         args = parser.parse_args()
 302         slice_name=args.slice_name
 303
 304         # unless we run the symlink 'lxcsu-internal', or we specify the -i option, prepend '--' '-c'
 305         if sys.argv[0].find('internal')>=0: args.internal=True
 306
 307         if len(args.command_to_run)>0 and (args.command_to_run[0] == "/sbin/service"):
 308                 # A quick hack to support nodemanager interfaces.py when restarting
 309                 # networking in a slice.
 310                 args.nosliceuid = True
 311
 312         try:
 313                 cmd = '/usr/bin/virsh dominfo %s'%slice_name
 314                 info = os.popen(cmd).read()
 315                 d = {}
 316
 317                 ostype = None
 318                 for line in info.split('\n'):
 319                         pair = line.split(':')
 320                         d[pair[0].strip()] = pair[1].strip()
 321                         if (pair[0].strip()=='OS Type'):
 322                                 ostype = pair[1].strip()
 323                                 break
 324         except:
 325                 print "Domain %s not found"%slice_name
 326
 327         is_kvm = ostype=='hvm'
 328
 329         if (is_kvm):
 330                 kvmsu(slice_name, args)
 331         else:
 332                 lxcsu(slice_name, args)
 333
 334 if __name__ == '__main__':
 335                 main()