slicesu

   1 #!/usr/bin/python
   2
   3 import sys
   4 import os
   5 import setns
   6 import pdb
   7
   8 from argparse import ArgumentParser
   9
  10 # can set to True here, but also use the -d option
  11 debug = False
  12
  13 def getarch(f):
  14         output = os.popen('readelf -h %s 2>&1'%f).readlines()
  15         classlines = [x for x in output if ('Class' in x.split(':')[0])]
  16         line = classlines[0]
  17         c = line.split(':')[1]
  18         if ('ELF64' in c):
  19                 return 'x86_64'
  20         elif ('ELF32' in c):
  21                 return 'i686'
  22         else:
  23                 raise Exception('Could not determine architecture')
  24
  25 def get_cgroup_subdirs_for_pid(pid):
  26         cgroup_info_file = '/proc/%s/cgroup'%pid
  27         cgroup_lines = open(cgroup_info_file).read().splitlines()
  28
  29         subdirs = {}
  30         for line in cgroup_lines:
  31                 try:
  32                         _, cgroup_name, subdir = line.split(':')
  33                         subdirs[cgroup_name] = subdir
  34                 except Exception, e:
  35                         print "Error reading cgroup info: %s"%str(e)
  36                         pass
  37
  38         return subdirs
  39
  40
  41 def umount(fs_dir, opts=''):
  42         output = os.popen('/bin/umount %s %s 2>&1'%(opts, fs_dir)).read()
  43         return ('device is busy' not in output)
  44
  45 def kvmsu(slice_name, args):
  46         # Find MAC address of virt interface
  47         ifstr = os.popen('/usr/bin/virsh domiflist %s'%slice_name).read().strip()
  48         mac_address = ifstr.split()[-1]
  49
  50         # Find IP address of virt interface
  51         ipaddr_str = os.popen('arp -an | grep %s'%'78:2b:cb:2d:f0:60').read().strip()
  52         ip_addr_part = ipaddr_str.split()[1]
  53         if (ip_addr_part[0]=='(' and ip_addr_part[-1]==')'):
  54                 ip_addr = ip_addr_part[1:len(ip_addr_part)-1]
  55
  56         os.execv('/usr/bin/ssh',['/usr/bin/ssh','-o','StrictHostKeyChecking=no','root@%s'%ip_addr,'"$@"'])
  57
  58
  59 def lxcsu(slice_name, args):
  60         # plain lxcsu
  61         if not args.internal:
  62                 # no command given: enter interactive shell
  63                 if not args.command_to_run: args.command_to_run=['/bin/sh']
  64                 args.command_to_run = [ '-c' ] + [" ".join(args.command_to_run)]
  65
  66         # support for either setting debug at the top of this file, or on the command-line
  67         if args.debug:
  68                 global debug
  69                 debug=True
  70
  71         try:
  72                 cmd = '/usr/bin/virsh --connect lxc:/// domid %s'%slice_name
  73                 # convert to int as a minimal raincheck
  74                 driver_pid = int(os.popen(cmd).read().strip())
  75                 # locate the pid for the - expected - single child, that would be the init for that VM
  76                 #init_pid = int(open("/proc/%s/task/%s/children"%(driver_pid,driver_pid)).read().strip())
  77                 init_pid = int(os.popen('pgrep -P %s'%driver_pid).readlines()[0].strip())
  78                 # Thierry: I am changing the code below to use child_pid instead of driver_pid
  79                 # for the namespace handling features, that I was able to check
  80                 # I've left the other ones as they were, i.e. using driver_pid, but I suspect
  81                 # they chould be changed as well
  82
  83         except:
  84                 print "Domain %s not found"%slice_name
  85                 exit(1)
  86
  87         if not driver_pid or not init_pid:
  88                 print "Domain %s not started"%slice_name
  89                 exit(1)
  90
  91         if debug: print "Found driver_pid",driver_pid,'and init_pid=',init_pid
  92         # xxx probably init_pid here too
  93         arch = getarch('/proc/%s/exe'%driver_pid)
  94
  95         # Set sysctls specific to slice
  96         sysctls = []
  97         sysctl_dir = '/etc/planetlab/vsys-attributes/%s'%slice_name
  98         if (os.access(sysctl_dir,0)):
  99                 entries = os.listdir(sysctl_dir)
 100                 for e in entries:
 101                         prefix = 'vsys_sysctl.'
 102                         if (e.startswith(prefix)):
 103                                 sysctl_file = '/'.join([sysctl_dir,e])
 104                                 sysctl_name = e[len(prefix):]
 105                                 sysctl_val = open(sysctl_file).read()
 106                                 sysctls.append((sysctl_file, sysctl_name, sysctl_val))
 107
 108         # xxx probably init_pid here too
 109         subdirs = get_cgroup_subdirs_for_pid(driver_pid)
 110         sysfs_root = '/sys/fs/cgroup'
 111
 112         # If the slice is frozen, then we'll get an EBUSY when trying to write to the task
 113         # list for the freezer cgroup. Since the user couldn't do anything anyway, it's best
 114         # in this case to error out the shell. (an alternative would be to un-freeze it,
 115         # add the task, and re-freeze it)
 116         # Enter cgroups
 117         current_cgroup = ''
 118         for subsystem in ['cpuset','memory','blkio','cpuacct','cpuacct,cpu','freezer']:
 119                 try:
 120                         current_cgroup = subsystem
 121
 122                         # There seems to be a bug in the cgroup schema: cpuacct,cpu can become cpu,cpuacct
 123                         # We need to handle both
 124                         task_path_alt = None
 125                         try:
 126                            subsystem_comps = subsystem.split(',')
 127                            subsystem_comps.reverse()
 128                            subsystem_alt = ','.join(subsystem_comps)
 129                            tasks_path_alt = [sysfs_root, subsystem_alt, subdirs[subsystem], 'tasks']
 130                         except Exception,e:
 131                                 pass
 132
 133                         tasks_path = [sysfs_root,subsystem,subdirs[subsystem],'tasks']
 134                         tasks_path_str = '/'.join(tasks_path)
 135
 136                         try:
 137                                 f = open(tasks_path_str, 'w')
 138                         except:
 139                                 tasks_path_alt_str = '/'.join(tasks_path_alt)
 140                                 f = open(tasks_path_alt_str, 'w')
 141
 142                         f.write(str(os.getpid()))
 143                         if (subsystem=='freezer'):
 144                                 f.close()
 145
 146                 except Exception,e:
 147                         if (not subdirs.has_key(subsystem)):
 148                                 pass
 149                         else:
 150                                 if debug: print e
 151                                 print "Error assigning cgroup %s (pid=%s) for slice %s"%(current_cgroup,driver_pid, slice_name)
 152                                 exit(1)
 153
 154
 155         def chcontext (path):
 156                 retcod = setns.chcontext (path)
 157                 if retcod != 0:
 158                         print 'WARNING - setns(%s)=>%s (ignored)'%(path,retcod)
 159                 return retcod
 160
 161         # Use init_pid and not driver_pid to locate reference namespaces
 162         ref_ns = "/proc/%s/ns/"%init_pid
 163
 164         if True:                                        chcontext(ref_ns+'uts')
 165         if True:                                        chcontext(ref_ns+'ipc')
 166
 167         if (not args.no_pidns):         chcontext(ref_ns+'pid')
 168         if (not args.no_netns):         chcontext(ref_ns+'net')
 169         if (not args.no_mntns):         chcontext(ref_ns+'mnt')
 170
 171         proc_mounted = False
 172         if (not os.access('/proc/self',0)):
 173                 proc_mounted = True
 174                 setns.proc_mount()
 175
 176         for (sysctl_file, sysctl_name, sysctl_val) in sysctls:
 177                 for fn in ["/sbin/sysctl", "/usr/sbin/sysctl", "/bin/sysctl", "/usr/bin/sysctl"]:
 178                         if os.path.exists(fn):
 179                                 os.system('%s -w %s=%s  >/dev/null 2>&1'%(fn, sysctl_name,sysctl_val))
 180                                 break
 181                         else:
 182                                 print "Error: image does not have a sysctl binary"
 183
 184         # cgroups is not yet LXC-safe, so we need to use the coarse grained access control
 185         # strategy of unmounting the filesystem
 186
 187         umount_result = True
 188         for subsystem in ['cpuset','cpu,cpuacct','memory','devices','freezer','net_cls','blkio','perf_event','systemd']:
 189                 fs_path = '/sys/fs/cgroup/%s'%subsystem
 190                 if (not umount(fs_path,'-l')):
 191                         print 'WARNING - umount failed (ignored) with path=',fs_path
 192                         pass
 193                         # Leaving these comments for historical reference
 194                         #print "Error disabling cgroup access"
 195                         #exit(1) - Don't need this because failure here implies failure in the call to umount /sys/fs/cgroup
 196
 197         if (not umount('/sys/fs/cgroup')):
 198                 print "Error disabling cgroup access"
 199                 exit(1)
 200
 201         fork_pid = os.fork()
 202
 203         def getuid (slicename):
 204                 import pwd
 205                 try:
 206                         return pwd.getpwnam(slicename).pw_uid
 207                 except:
 208                         return
 209
 210         if (fork_pid == 0):
 211                 if (not args.root):
 212                         setns.drop_caps()
 213                         if (args.nosliceuid):
 214                                 # we still want to drop capabilities, but don't want to switch UIDs
 215                                 exec_args = [arch,'/bin/sh','--login',]+args.command_to_run
 216                         else:
 217                                                                 # let's keep this check even though we don't use the uid
 218                                                                 # as a way of checking the existence of the slice account
 219                                 uid = getuid (slice_name)
 220                                 if not uid:
 221                                         print "lxcsu could not spot %s in /etc/passwd - exiting"%slice_name
 222                                         exit(1)
 223                                 exec_args = [arch,'/usr/bin/sudo','-u',slice_name,'/bin/sh','--login',]+args.command_to_run
 224 # once we can drop f12, it would be nicer to instead go for
 225 # exec_args = [arch,'/usr/sbin/capsh',cap_arg,'--user=%s'%slice_name,'--login',]+args.command_to_run
 226                 else:
 227                         exec_args = [arch,'/bin/sh','--login']+args.command_to_run
 228
 229                 os.environ['SHELL'] = '/bin/sh'
 230                 if os.path.exists('/etc/planetlab/lib/bind_public.so'):
 231                         os.environ['LD_PRELOAD'] = '/etc/planetlab/lib/bind_public.so'
 232                 if not args.noslicehome:
 233                         os.environ['HOME'] = '/home/%s'%slice_name
 234                         os.chdir("/home/%s"%(slice_name))
 235                 if debug: print 'lxcsu:execv:','/usr/bin/setarch',exec_args
 236                 os.execv('/usr/bin/setarch',exec_args)
 237         else:
 238                 setns.proc_umount()
 239                 _,status = os.waitpid(fork_pid,0)
 240                 exit(os.WEXITSTATUS(status))
 241
 242 def main ():
 243         parser = ArgumentParser()
 244         parser.add_argument("-n", "--nonet",
 245                                                                                 action="store_true", dest="no_netns", default=False,
 246                                                                                 help="Don't enter network namespace")
 247         parser.add_argument("-m", "--nomnt",
 248                                                                                 action="store_true", dest="no_mntns", default=False,
 249                                                                                 help="Don't enter mount namespace")
 250         parser.add_argument("-p", "--nopid",
 251                                                                                 action="store_true", dest="no_pidns", default=False,
 252                                                                                 help="Don't enter pid namespace")
 253         parser.add_argument("-r", "--root",
 254                                                                                 action="store_true", dest="root", default=False,
 255                                                                                 help="Enter as root: be careful")
 256         parser.add_argument("-i","--internal",
 257                                                                                 action="store_true", dest="internal", default=False,
 258                                                                                 help="does *not* prepend '-- -c' to arguments - or invoke lxcsu-internal")
 259         parser.add_argument("-d","--debug",
 260                                                                                 action='store_true', dest='debug', default=False,
 261                                                                                 help="debug option")
 262         parser.add_argument("-s","--nosliceuid",
 263                                                                                 action='store_true', dest="nosliceuid", default=False,
 264                                                                                 help="do not change to slice uid inside of slice")
 265         parser.add_argument("-o","--noslicehome",
 266                                                                                 action='store_true', dest="noslicehome", default=False,
 267                                                                                 help="do not change to slice home directory inside of slice")
 268
 269         if os.path.exists("/etc/lxcsu_default"):
 270                 defaults = parser.parse_args(file("/etc/lxcsu_default","r").read().split())
 271                 parser.set_defaults(**defaults.__dict__)
 272
 273         parser.add_argument ("slice_name")
 274         parser.add_argument ("command_to_run",nargs="*")
 275
 276         args = parser.parse_args()
 277         slice_name=args.slice_name
 278
 279         # unless we run the symlink 'lxcsu-internal', or we specify the -i option, prepend '--' '-c'
 280         if sys.argv[0].find('internal')>=0: args.internal=True
 281
 282         if len(args.command_to_run)>0 and (args.command_to_run[0] == "/sbin/service"):
 283                 # A quick hack to support nodemanager interfaces.py when restarting
 284                 # networking in a slice.
 285                 args.nosliceuid = True
 286
 287         try:
 288                 cmd = '/usr/bin/virsh dominfo %s'%slice_name
 289                 info = os.popen(cmd).read()
 290                 d = {}
 291
 292                 ostype = None
 293                 for line in info.split('\n'):
 294                         pair = line.split(':')
 295                         d[pair[0].strip()] = pair[1].strip()
 296                         if (pair[0].strip()=='OS Type'):
 297                                 ostype = pair[1].strip()
 298                                 break
 299         except:
 300                 print "Domain %s not found"%slice_name
 301
 302         is_kvm = ostype=='hvm'
 303
 304         if (is_kvm):
 305                 kvmsu(slice_name, args)
 306         else:
 307                 lxcsu(slice_name, args)
 308
 309 if __name__ == '__main__':
 310                 main()