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

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 drop_capabilities='cap_sys_admin,cap_sys_boot,cap_sys_module'

11

12 # can set to True here, but also use the -d option

13 debug = False

14

15 def getarch(f):

16 output = os.popen('readelf -h %s 2>&1'%f).readlines()

17 classlines = [x for x in output if ('Class' in x.split(':')[0])]

18 line = classlines[0]

19 c = line.split(':')[1]

20 if ('ELF64' in c):

21 return 'x86_64'

22 elif ('ELF32' in c):

23 return 'i686'

24 else:

25 raise Exception('Could not determine architecture')

26

27 def get_cgroup_subdirs_for_pid(pid):

28 cgroup_info_file = '/proc/%s/cgroup'%pid

29 cgroup_lines = open(cgroup_info_file).read().splitlines()

30

31 subdirs = {}

32 for line in cgroup_lines:

33 try:

34 _, cgroup_name, subdir = line.split(':')

35 subdirs[cgroup_name] = subdir

36 except Exception, e:

37 print "Error reading cgroup info: %s"%str(e)

38 pass

39

40 return subdirs

41

42

43 def umount(fs_dir, opts=''):

44 output = os.popen('/bin/umount %s %s 2>&1'%(opts, fs_dir)).read()

45 return ('device is busy' not in output)

46

47 def main ():

48 parser = ArgumentParser()

49 parser.add_argument("-n", "--nonet",

50 action="store_true", dest="no_netns", default=False,

51 help="Don't enter network namespace")

52 parser.add_argument("-m", "--nomnt",

53 action="store_true", dest="no_mntns", default=False,

54 help="Don't enter mount namespace")

55 parser.add_argument("-p", "--nopid",

56 action="store_true", dest="no_pidns", default=False,

57 help="Don't enter pid namespace")

58 parser.add_argument("-r", "--root",

59 action="store_true", dest="root", default=False,

60 help="Enter as root: be careful")

61 parser.add_argument("-i","--internal",

62 action="store_true", dest="internal", default=False,

63 help="does *not* prepend '-- -c' to arguments - or invoke lxcsu-internal")

64 parser.add_argument("-d","--debug",

65 action='store_true', dest='debug', default=False,

66 help="debug option")

67 parser.add_argument("-s","--nosliceuid",

68 action='store_true', dest="nosliceuid", default=False,

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

70 parser.add_argument("-o","--noslicehome",

71 action='store_true', dest="noslicehome", default=False,

72 help="do not change to slice home directory inside of slice")

73

74 if os.path.exists("/etc/lxcsu_default"):

75 defaults = parser.parse_args(file("/etc/lxcsu_default","r").read().split())

76 parser.set_defaults(**defaults.__dict__)

77

78 parser.add_argument ("slice_name")

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

80

81 args = parser.parse_args()

82 slice_name=args.slice_name

83

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

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

86

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

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

89 # networking in a slice.

90 args.nosliceuid = True

91

92 # plain lxcsu

93 if not args.internal:

94 # no command given: enter interactive shell

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

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

97

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

99 if args.debug:

100 global debug

101 debug=True

102

103 try:

104 cmd = '/usr/bin/virsh --connect lxc:/// domid %s'%slice_name

105 # convert to int as a minimal raincheck

106 driver_pid = int(os.popen(cmd).read().strip())

107 # locate the pid for the - expected - single child, that would be the init for that VM

108 init_pid = int(open("/proc/%s/task/%s/children"%(driver_pid,driver_pid)).read().strip())

109 # Thierry: I am changing the code below to use child_pid instead of driver_pid

110 # for the namespace handling features, that I was able to check

111 # I've left the other ones as they were, i.e. using driver_pid, but I suspect

112 # they chould be changed as well

113

114 except:

115 print "Domain %s not found"%slice_name

116 exit(1)

117

118 if not driver_pid or not init_pid:

119 print "Domain %s not started"%slice_name

120 exit(1)

121

122 if debug: print "Found driver_pid",driver_pid,'and init_pid=',init_pid

123 # xxx probably init_pid here too

124 arch = getarch('/proc/%s/exe'%driver_pid)

125

126 # Set sysctls specific to slice

127 sysctls = []

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

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

130 entries = os.listdir(sysctl_dir)

131 for e in entries:

132 prefix = 'vsys_sysctl.'

133 if (e.startswith(prefix)):

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

135 sysctl_name = e[len(prefix):]

136 sysctl_val = open(sysctl_file).read()

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

138

139 # xxx probably init_pid here too

140 subdirs = get_cgroup_subdirs_for_pid(driver_pid)

141 sysfs_root = '/sys/fs/cgroup'

142

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

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

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

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

147 # Enter cgroups

148 current_cgroup = ''

149 for subsystem in ['cpuset','memory','blkio','cpuacct','cpuacct,cpu','freezer']:

150 try:

151 current_cgroup = subsystem

152

153 # There seems to be a bug in the cgroup schema: cpuacct,cpu can become cpu,cpuacct

154 # We need to handle both

155 task_path_alt = None

156 try:

157 subsystem_comps = subsystem.split(',')

158 subsystem_comps.reverse()

159 subsystem_alt = ','.join(subsystem_comps)

160 tasks_path_alt = [sysfs_root, subsystem_alt, subdirs[subsystem], 'tasks']

161 except Exception,e:

162 pass

163

164 tasks_path = [sysfs_root,subsystem,subdirs[subsystem],'tasks']

165 tasks_path_str = '/'.join(tasks_path)

166

167 try:

168 f = open(tasks_path_str, 'w')

169 except:

170 tasks_path_alt_str = '/'.join(tasks_path_alt)

171 f = open(tasks_path_alt_str, 'w')

172

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

174 if (subsystem=='freezer'):

175 f.close()

176

177 except Exception,e:

178 if (not subdirs.has_key(subsystem)):

179 pass

180 else:

181 if debug: print e

182 print "Error assigning cgroup %s (pid=%s) for slice %s"%(current_cgroup,driver_pid, slice_name)

183 exit(1)

184

185

186 def chcontext (path):

187 retcod = setns.chcontext (path)

188 if retcod != 0:

189 print 'WARNING - setns(%s)=>%s (ignored)'%(path,retcod)

190 return retcod

191

192 # Use init_pid and not driver_pid to locate reference namespaces

193 ref_ns = "/proc/%s/ns/"%init_pid

194

195 if True: chcontext(ref_ns+'uts')

196 if True: chcontext(ref_ns+'ipc')

197

198 if (not args.no_pidns): chcontext(ref_ns+'pid')

199 if (not args.no_netns): chcontext(ref_ns+'net')

200 if (not args.no_mntns): chcontext(ref_ns+'mnt')

201

202 proc_mounted = False

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

204 proc_mounted = True

205 setns.proc_mount()

206

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

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

209 if os.path.exists(fn):

210 os.system('%s -w %s=%s >/dev/null 2>&1'%(fn, sysctl_name,sysctl_val))

211 break

212 else:

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

214

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

216 # strategy of unmounting the filesystem

217

218 umount_result = True

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

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

221 if (not umount(fs_path,'-l')):

222 print 'WARNING - umount failed (ignored) with path=',fs_path

223 pass

224 # Leaving these comments for historical reference

225 #print "Error disabling cgroup access"

226 #exit(1) - Don't need this because failure here implies failure in the call to umount /sys/fs/cgroup

227

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

229 print "Error disabling cgroup access"

230 exit(1)

231

232 fork_pid = os.fork()

233

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

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

236 #

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

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

239 # ngroups = MAX_GROUPS;

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

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

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

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

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

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

246

247 def getuid (slicename):

248 import pwd

249 try:

250 return pwd.getpwnam(slicename).pw_uid

251 except:

252 return

253

254 if (fork_pid == 0):

255 cap_arg = '--drop='+drop_capabilities

256

257 if (not args.root):

258 if (args.nosliceuid):

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

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

261 else:

262 uid = getuid (slice_name)

263 if not uid:

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

265 exit(1)

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

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

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

269 else:

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

271

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

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

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

275 if not args.noslicehome:

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

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

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

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

280 else:

281 setns.proc_umount()

282 _,status = os.waitpid(fork_pid,0)

283 exit(os.WEXITSTATUS(status))

284

285 if __name__ == '__main__':

286 main()