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 ("slice_name")

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

52

53 args = parser.parse_args()

54 slice_name=args.slice_name

55

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

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

58

59 # plain lxcsu

60 if not args.internal:

61 # no command given: enter interactive shell

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

63 args.command_to_run = [ '-c' ] + args.command_to_run

64

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

66 if args.debug:

67 global debug

68 debug=True

69

70 try:

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

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

73 except:

74 print "Error finding slice %s"%slice_name

75 exit(1)

76

77 slice_spec = None

78

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

80 arch = None

81

82 for e in output:

83 try:

84 l = e.rstrip()

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

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

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

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

89

90 if (slice_name_check == slice_name):

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

92 slice_path = path

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

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

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

96 slice_spec = slice_path

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

98 break

99 except Exception,e:

100 if debug:

101 import traceback

102 print "BEG lxcsu - ignoring exception"

103 traceback.print_exc()

104 print "END lxcsu - ignoring exception"

105 pass

106

107 if (not slice_spec or not pid):

108 print "Not started: %s"%slice_name

109 exit(1)

110

111 if arch is None:

112 arch = 'x86_64'

113

114 # Set sysctls specific to slice

115 sysctls = []

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

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

118 entries = os.listdir(sysctl_dir)

119 for e in entries:

120 prefix = 'vsys_sysctl.'

121 if (e.startswith(prefix)):

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

123 sysctl_name = e[len(prefix):]

124 sysctl_val = open(sysctl_file).read()

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

126

127 # Enter cgroups

128 try:

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

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

131

132 except:

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

134 exit(1)

135

136 try:

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

138 except:

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

140 exit(1)

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 try:

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

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

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

150 # the object destructor, which will not be caught

151 f.close()

152 except:

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

154 exit(1)

155

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

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

158

159 if (not args.pidns):

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

161

162 if (not args.netns):

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

164

165 if (not args.mntns):

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

167

168 proc_mounted = False

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

170 proc_mounted = True

171 setns.proc_mount()

172

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

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

175 if os.path.exists(fn):

176 os.system("mount -o remount,rw none /proc/sys")

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

178 os.system("mount -o remount,ro none /proc/sys")

179 break

180 else:

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

182

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

184 # strategy of unmounting the filesystem

185

186 umount_result = True

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

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

189 if (not umount(fs_path)):

190 print "Error disabling cgroup access"

191 exit(1)

192

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

194 print "Error disabling cgroup access"

195 exit(1)

196

197 pid = os.fork()

198

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

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

201 #

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

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

204 # ngroups = MAX_GROUPS;

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

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

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

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

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

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

211

212 def getuid (slicename):

213 import pwd

214 try:

215 return pwd.getpwnam(slicename).pw_uid

216 except:

217 return

218

219 if (pid == 0):

220 cap_arg = '--drop='+drop_capabilities

221

222 if (not args.root):

223 uid = getuid (slice_name)

224 if not uid:

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

226 exit(1)

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

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

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

230 else:

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

232

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

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

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

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

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

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

239 else:

240 setns.proc_umount()

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

242 exit(os.WEXITSTATUS(status))

243

244 if __name__ == '__main__':

245 main()