lxcsu

   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 main ():
  46     parser = ArgumentParser()
  47     parser.add_argument("-n", "--nonet",
  48                                         action="store_true", dest="no_netns", default=False,
  49                                         help="Don't enter network namespace")
  50     parser.add_argument("-m", "--nomnt",
  51                                         action="store_true", dest="no_mntns", default=False,
  52                                         help="Don't enter mount namespace")
  53     parser.add_argument("-p", "--nopid",
  54                                         action="store_true", dest="no_pidns", default=False,
  55                                         help="Don't enter pid namespace")
  56     parser.add_argument("-r", "--root",
  57                                         action="store_true", dest="root", default=False,
  58                                         help="Enter as root: be careful")
  59     parser.add_argument("-i","--internal",
  60                                         action="store_true", dest="internal", default=False,
  61                                         help="does *not* prepend '-- -c' to arguments - or invoke lxcsu-internal")
  62     parser.add_argument("-d","--debug",
  63                                         action='store_true', dest='debug', default=False,
  64                                         help="debug option")
  65     parser.add_argument("-s","--nosliceuid",
  66                                         action='store_true', dest="nosliceuid", default=False,
  67                                         help="do not change to slice uid inside of slice")
  68     parser.add_argument("-o","--noslicehome",
  69                                         action='store_true', dest="noslicehome", default=False,
  70                                         help="do not change to slice home directory inside of slice")
  71
  72     if os.path.exists("/etc/lxcsu_default"):
  73         defaults = parser.parse_args(file("/etc/lxcsu_default","r").read().split())
  74         parser.set_defaults(**defaults.__dict__)
  75
  76     parser.add_argument ("slice_name")
  77     parser.add_argument ("command_to_run",nargs="*")
  78
  79     args = parser.parse_args()
  80     slice_name=args.slice_name
  81
  82     # unless we run the symlink 'lxcsu-internal', or we specify the -i option, prepend '--' '-c'
  83     if sys.argv[0].find('internal')>=0: args.internal=True
  84
  85     if len(args.command_to_run)>0 and (args.command_to_run[0] == "/sbin/service"):
  86         # A quick hack to support nodemanager interfaces.py when restarting
  87         # networking in a slice.
  88         args.nosliceuid = True
  89
  90     # plain lxcsu
  91     if not args.internal:
  92         # no command given: enter interactive shell
  93         if not args.command_to_run: args.command_to_run=['/bin/sh']
  94         args.command_to_run = [ '-c' ] + [" ".join(args.command_to_run)]
  95
  96     # support for either setting debug at the top of this file, or on the command-line
  97     if args.debug:
  98         global debug
  99         debug=True
 100
 101     try:
 102         cmd = '/usr/bin/virsh --connect lxc:/// domid %s'%slice_name
 103         # convert to int as a minimal raincheck
 104         driver_pid = int(os.popen(cmd).read().strip())
 105         # locate the pid for the - expected - single child, that would be the init for that VM
 106         #init_pid = int(open("/proc/%s/task/%s/children"%(driver_pid,driver_pid)).read().strip())
 107         init_pid = int(os.popen('pgrep -P %s'%driver_pid).readlines()[0].strip())
 108         # Thierry: I am changing the code below to use child_pid instead of driver_pid
 109         # for the namespace handling features, that I was able to check
 110         # I've left the other ones as they were, i.e. using driver_pid, but I suspect
 111         # they chould be changed as well
 112
 113     except:
 114         print "Domain %s not found"%slice_name
 115         exit(1)
 116
 117     if not driver_pid or not init_pid:
 118         print "Domain %s not started"%slice_name
 119         exit(1)
 120
 121     if debug: print "Found driver_pid",driver_pid,'and init_pid=',init_pid
 122     # xxx probably init_pid here too
 123     arch = getarch('/proc/%s/exe'%driver_pid)
 124
 125     # Set sysctls specific to slice
 126     sysctls = []
 127     sysctl_dir = '/etc/planetlab/vsys-attributes/%s'%slice_name
 128     if (os.access(sysctl_dir,0)):
 129         entries = os.listdir(sysctl_dir)
 130         for e in entries:
 131             prefix = 'vsys_sysctl.'
 132             if (e.startswith(prefix)):
 133                 sysctl_file = '/'.join([sysctl_dir,e])
 134                 sysctl_name = e[len(prefix):]
 135                 sysctl_val = open(sysctl_file).read()
 136                 sysctls.append((sysctl_file, sysctl_name, sysctl_val))
 137
 138     # xxx probably init_pid here too
 139     subdirs = get_cgroup_subdirs_for_pid(driver_pid)
 140     sysfs_root = '/sys/fs/cgroup'
 141
 142     # If the slice is frozen, then we'll get an EBUSY when trying to write to the task
 143     # list for the freezer cgroup. Since the user couldn't do anything anyway, it's best
 144     # in this case to error out the shell. (an alternative would be to un-freeze it,
 145     # add the task, and re-freeze it)
 146     # Enter cgroups
 147     current_cgroup = ''
 148     for subsystem in ['cpuset','memory','blkio','cpuacct','cpuacct,cpu','freezer']:
 149         try:
 150             current_cgroup = subsystem
 151
 152             # There seems to be a bug in the cgroup schema: cpuacct,cpu can become cpu,cpuacct
 153             # We need to handle both
 154             task_path_alt = None
 155             try:
 156                subsystem_comps = subsystem.split(',')
 157                subsystem_comps.reverse()
 158                subsystem_alt = ','.join(subsystem_comps)
 159                tasks_path_alt = [sysfs_root, subsystem_alt, subdirs[subsystem], 'tasks']
 160             except Exception,e:
 161                 pass
 162
 163             tasks_path = [sysfs_root,subsystem,subdirs[subsystem],'tasks']
 164             tasks_path_str = '/'.join(tasks_path)
 165
 166             try:
 167                 f = open(tasks_path_str, 'w')
 168             except:
 169                 tasks_path_alt_str = '/'.join(tasks_path_alt)
 170                 f = open(tasks_path_alt_str, 'w')
 171
 172             f.write(str(os.getpid()))
 173             if (subsystem=='freezer'):
 174                 f.close()
 175
 176         except Exception,e:
 177             if (not subdirs.has_key(subsystem)):
 178                 pass
 179             else:
 180                 if debug: print e
 181                 print "Error assigning cgroup %s (pid=%s) for slice %s"%(current_cgroup,driver_pid, slice_name)
 182                 exit(1)
 183
 184
 185     def chcontext (path):
 186         retcod = setns.chcontext (path)
 187         if retcod != 0:
 188             print 'WARNING - setns(%s)=>%s (ignored)'%(path,retcod)
 189         return retcod
 190
 191     # Use init_pid and not driver_pid to locate reference namespaces
 192     ref_ns = "/proc/%s/ns/"%init_pid
 193
 194     if True:                    chcontext(ref_ns+'uts')
 195     if True:                    chcontext(ref_ns+'ipc')
 196
 197     if (not args.no_pidns):     chcontext(ref_ns+'pid')
 198     if (not args.no_netns):     chcontext(ref_ns+'net')
 199     if (not args.no_mntns):     chcontext(ref_ns+'mnt')
 200
 201     proc_mounted = False
 202     if (not os.access('/proc/self',0)):
 203         proc_mounted = True
 204         setns.proc_mount()
 205
 206     for (sysctl_file, sysctl_name, sysctl_val) in sysctls:
 207         for fn in ["/sbin/sysctl", "/usr/sbin/sysctl", "/bin/sysctl", "/usr/bin/sysctl"]:
 208             if os.path.exists(fn):
 209                 os.system('%s -w %s=%s  >/dev/null 2>&1'%(fn, sysctl_name,sysctl_val))
 210                 break
 211             else:
 212                 print "Error: image does not have a sysctl binary"
 213
 214     # cgroups is not yet LXC-safe, so we need to use the coarse grained access control
 215     # strategy of unmounting the filesystem
 216
 217     umount_result = True
 218     for subsystem in ['cpuset','cpu,cpuacct','memory','devices','freezer','net_cls','blkio','perf_event','systemd']:
 219         fs_path = '/sys/fs/cgroup/%s'%subsystem
 220         if (not umount(fs_path,'-l')):
 221             print 'WARNING - umount failed (ignored) with path=',fs_path
 222             pass
 223             # Leaving these comments for historical reference
 224             #print "Error disabling cgroup access"
 225             #exit(1) - Don't need this because failure here implies failure in the call to umount /sys/fs/cgroup
 226
 227     if (not umount('/sys/fs/cgroup')):
 228         print "Error disabling cgroup access"
 229         exit(1)
 230
 231     fork_pid = os.fork()
 232
 233     def getuid (slicename):
 234         import pwd
 235         try:
 236             return pwd.getpwnam(slicename).pw_uid
 237         except:
 238             return
 239
 240     if (fork_pid == 0):
 241         if (not args.root):
 242             setns.drop_caps()
 243             if (args.nosliceuid):
 244                 # we still want to drop capabilities, but don't want to switch UIDs
 245                 exec_args = [arch,'/bin/sh','--login',]+args.command_to_run
 246             else:
 247                 uid = getuid (slice_name)
 248                 if not uid:
 249                     print "lxcsu could not spot %s in /etc/passwd - exiting"%slice_name
 250                     exit(1)
 251                 exec_args = [arch,'/bin/sh','--uid=%s'%uid,'--login',]+args.command_to_run
 252 # once we can drop f12, it would be nicer to instead go for
 253 # exec_args = [arch,'/usr/sbin/capsh',cap_arg,'--user=%s'%slice_name,'--login',]+args.command_to_run
 254         else:
 255             exec_args = [arch,'/bin/sh','--login']+args.command_to_run
 256
 257         os.environ['SHELL'] = '/bin/sh'
 258         if os.path.exists('/etc/planetlab/lib/bind_public.so'):
 259             os.environ['LD_PRELOAD'] = '/etc/planetlab/lib/bind_public.so'
 260         if not args.noslicehome:
 261             os.environ['HOME'] = '/home/%s'%slice_name
 262             os.chdir("/home/%s"%(slice_name))
 263         if debug: print 'lxcsu:execv:','/usr/bin/setarch',exec_args
 264         os.execv('/usr/bin/setarch',exec_args)
 265     else:
 266         setns.proc_umount()
 267         _,status = os.waitpid(fork_pid,0)
 268         exit(os.WEXITSTATUS(status))
 269
 270 if __name__ == '__main__':
 271         main()