git.onelab.eu Git - lxc-userspace.git/blob

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):

27 output = os.popen('/bin/umount %s 2>&1'%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", dest="nosliceuid", default=False,

51 help="do not change to slice uid inside of slice")

52 parser.add_argument ("slice_name")

53 parser.add_argument ("command_to_run",nargs="*")

54

55 args = parser.parse_args()

56 slice_name=args.slice_name

57

58 # unless we run the symlink 'lxcsu-internal', or we specify the -i option, prepend '--' '-c'

59 if sys.argv[0].find('internal')>=0: args.internal=True

60

61 if len(args.command_to_run)>0 and (args.command_to_run[0] == "/sbin/service"):

62 # A quick hack to support nodemanager interfaces.py when restarting

63 # networking in a slice.

64 args.nosliceuid = True

65

66 # plain lxcsu

67 if not args.internal:

68 # no command given: enter interactive shell

69 if not args.command_to_run: args.command_to_run=['/bin/sh']

70 args.command_to_run = [ '-c' ] + [" ".join(args.command_to_run)]

71

72 # support for either setting debug at the top of this file, or on the command-line

73 if args.debug:

74 global debug

75 debug=True

76

77 try:

78 cmd = 'grep %s /proc/*/cgroup | grep freezer'%slice_name

79 output = os.popen(cmd).readlines()

80 except:

81 print "Error finding slice %s"%slice_name

82 exit(1)

83

84 slice_spec = None

85

86 # provide a default as this is not always properly computed

87 arch = None

88

89 for e in output:

90 try:

91 l = e.rstrip()

92 path = l.split(':')[0]

93 comp = l.rsplit(':')[-1]

94 slice_name_check = comp.rsplit('/')[-1]

95 if debug: print "dealing with >%s<"%slice_name_check

96

97 if (slice_name_check == slice_name):

98 if debug: print "found %s"%slice_name

99 slice_path = path

100 pid = slice_path.split('/')[2]

101 cmdline = open('/proc/%s/cmdline'%pid).read().rstrip('\n\x00')

102 if (cmdline == '/sbin/init') or (cmdline.startswith("init [")):

103 slice_spec = slice_path

104 arch = getarch('/proc/%s/exe'%pid)

105 break

106 except Exception,e:

107 if debug:

108 import traceback

109 print "BEG lxcsu - ignoring exception"

110 traceback.print_exc()

111 print "END lxcsu - ignoring exception"

112 pass

113

114 if (not slice_spec or not pid):

115 print "Not started: %s"%slice_name

116 exit(1)

117

118 if arch is None:

119 arch = 'x86_64'

120

121 # Set sysctls specific to slice

122 sysctls = []

123 sysctl_dir = '/etc/planetlab/vsys-attributes/%s'%slice_name

124 if (os.access(sysctl_dir,0)):

125 entries = os.listdir(sysctl_dir)

126 for e in entries:

127 prefix = 'vsys_sysctl.'

128 if (e.startswith(prefix)):

129 sysctl_file = '/'.join([sysctl_dir,e])

130 sysctl_name = e[len(prefix):]

131 sysctl_val = open(sysctl_file).read()

132 sysctls.append((sysctl_file, sysctl_name, sysctl_val))

133

134 # Enter cgroups

135 try:

136 for subsystem in ['cpuset','memory','blkio']:

137 open('/sys/fs/cgroup/%s/libvirt/lxc/%s/tasks'%(subsystem,slice_name),'w').write(str(os.getpid()))

138

139 except:

140 print "Error assigning resources: %s"%slice_name

141 exit(1)

142

143 try:

144 open('/sys/fs/cgroup/cpuacct/system/libvirtd.service/libvirt/lxc/%s/tasks'%slice_name,'w').write(str(os.getpid()))

145 except:

146 print "Error assigning cpuacct: %s" % slice_name

147 exit(1)

148

149 # If the slice is frozen, then we'll get an EBUSY when trying to write to the task

150 # list for the freezer cgroup. Since the user couldn't do anything anyway, it's best

151 # in this case to error out the shell. (an alternative would be to un-freeze it,

152 # add the task, and re-freeze it)

153 try:

154 f=open('/sys/fs/cgroup/freezer/libvirt/lxc/%s/tasks'%(slice_name),'w')

155 f.write(str(os.getpid()))

156 # note: we need to call f.close() explicitly, or we'll get an exception in

157 # the object destructor, which will not be caught

158 f.close()

159 except:

160 print "Error adding task to freezer cgroup. Slice is probably frozen: %s" % slice_name

161 exit(1)

162

163 setns.chcontext('/proc/%s/ns/uts'%pid)

164 setns.chcontext('/proc/%s/ns/ipc'%pid)

165

166 if (not args.pidns):

167 setns.chcontext('/proc/%s/ns/pid'%pid)

168

169 if (not args.netns):

170 setns.chcontext('/proc/%s/ns/net'%pid)

171

172 if (not args.mntns):

173 setns.chcontext('/proc/%s/ns/mnt'%pid)

174

175 proc_mounted = False

176 if (not os.access('/proc/self',0)):

177 proc_mounted = True

178 setns.proc_mount()

179

180 for (sysctl_file, sysctl_name, sysctl_val) in sysctls:

181 for fn in ["/sbin/sysctl", "/usr/sbin/sysctl", "/bin/sysctl", "/usr/bin/sysctl"]:

182 if os.path.exists(fn):

183 os.system('%s -w %s=%s'%(fn, sysctl_name,sysctl_val))

184 break

185 else:

186 print "Error: image does not have a sysctl binary"

187

188 # cgroups is not yet LXC-safe, so we need to use the coarse grained access control

189 # strategy of unmounting the filesystem

190

191 umount_result = True

192 for subsystem in ['cpuset','cpu,cpuacct','memory','devices','freezer','net_cls','blkio','perf_event']:

193 fs_path = '/sys/fs/cgroup/%s'%subsystem

194 if (not umount(fs_path)):

195 print "Error disabling cgroup access"

196 exit(1)

197

198 if (not umount('/sys/fs/cgroup')):

199 print "Error disabling cgroup access"

200 exit(1)

201

202 pid = os.fork()

203

204 # capsh has a --user option starting with f14

205 # so if only for f12 we need to fake this one

206 #

207 # capsh.c does essentially the following when invoked with --user:

208 # pwd = getpwnam(user); ...

209 # ngroups = MAX_GROUPS;

210 # status = getgrouplist(user, pwd->pw_gid, groups, &ngroups); ...

211 # status = setgroups(ngroups, groups); ...

212 # status = setgid(pwd->pw_gid); ...

213 # status = setuid(pwd->pw_uid); ...

214 # however we cannot simulate that ourselves because if we did in this process then

215 # capsh could not be allowed to mess with caps any more

216

217 def getuid (slicename):

218 import pwd

219 try:

220 return pwd.getpwnam(slicename).pw_uid

221 except:

222 return

223

224 if (pid == 0):

225 cap_arg = '--drop='+drop_capabilities

226

227 if (not args.root):

228 if (args.nosliceuid):

229 # we still want to drop capabilities, but don't want to switch UIDs

230 exec_args = [arch,'/usr/sbin/capsh',cap_arg,'--','--login',]+args.command_to_run

231 else:

232 uid = getuid (slice_name)

233 if not uid:

234 print "lxcsu could not spot %s in /etc/passwd - exiting"%slice_name

235 exit(1)

236 exec_args = [arch,'/usr/sbin/capsh',cap_arg,'--uid=%s'%uid,'--','--login',]+args.command_to_run

237 # once we can drop f12, it would be nicer to instead go for

238 # exec_args = [arch,'/usr/sbin/capsh',cap_arg,'--user=%s'%slice_name,'--','--login',]+args.command_to_run

239 else:

240 exec_args = [arch,'/usr/sbin/capsh','--','--login']+args.command_to_run

241

242 os.environ['SHELL'] = '/bin/sh'

243 os.environ['HOME'] = '/home/%s'%slice_name

244 if os.path.exists('/etc/planetlab/lib/bind_public.so'):

245 os.environ['LD_PRELOAD'] = '/etc/planetlab/lib/bind_public.so'

246 os.chdir("/home/%s"%(slice_name))

247 if debug: print 'lxcsu:execv:','/usr/bin/setarch',exec_args

248 os.execv('/usr/bin/setarch',exec_args)

249 else:

250 setns.proc_umount()

251 _,status = os.waitpid(pid,0)

252 exit(os.WEXITSTATUS(status))

253

254 if __name__ == '__main__':

255 main()