python/vserver.py

   1 # Copyright 2005 Princeton University
   2
   3 import errno
   4 import fcntl
   5 import os
   6 import re
   7 import sys
   8 import time
   9 import traceback
  10
  11 import mountimpl
  12 import linuxcaps
  13 import passfdimpl
  14 import utmp
  15 import vserverimpl, vduimpl
  16
  17 from util_vserver_vars import *
  18
  19 CAP_SAFE = (linuxcaps.CAP_CHOWN |
  20             linuxcaps.CAP_DAC_OVERRIDE |
  21             linuxcaps.CAP_DAC_READ_SEARCH |
  22             linuxcaps.CAP_FOWNER |
  23             linuxcaps.CAP_FSETID |
  24             linuxcaps.CAP_KILL |
  25             linuxcaps.CAP_SETGID |
  26             linuxcaps.CAP_SETUID |
  27             linuxcaps.CAP_SETPCAP |
  28             linuxcaps.CAP_SYS_TTY_CONFIG |
  29             linuxcaps.CAP_LEASE |
  30             linuxcaps.CAP_SYS_CHROOT |
  31             linuxcaps.CAP_SYS_PTRACE)
  32
  33 #
  34 # these are the flags taken from the kernel linux/vserver/legacy.h
  35 #
  36 FLAGS_LOCK = 1
  37 FLAGS_SCHED = 2  # XXX - defined in util-vserver/src/chcontext.c
  38 FLAGS_NPROC = 4
  39 FLAGS_PRIVATE = 8
  40 FLAGS_INIT = 16
  41 FLAGS_HIDEINFO = 32
  42 FLAGS_ULIMIT = 64
  43 FLAGS_NAMESPACE = 128
  44
  45 # default values for new vserver scheduler
  46 SCHED_TOKENS_MIN = 50
  47 SCHED_TOKENS_MAX = 100
  48
  49
  50 class VServer:
  51
  52     INITSCRIPTS = [('/etc/rc.vinit', 'start'),
  53                    ('/etc/rc.d/rc', '%(runlevel)d')]
  54
  55     def __init__(self, name):
  56
  57         self.name = name
  58         self.config = self.__read_config_file("/etc/vservers.conf")
  59         self.config.update(self.__read_config_file("/etc/vservers/%s.conf" %
  60                                                    self.name))
  61         self.flags = 0
  62         flags = self.config["S_FLAGS"].split(" ")
  63         if "lock" in flags:
  64             self.flags |= FLAGS_LOCK
  65         if "nproc" in flags:
  66             self.flags |= FLAGS_NPROC
  67         self.remove_caps = ~CAP_SAFE
  68         self.ctx = int(self.config["S_CONTEXT"])
  69
  70     config_var_re = re.compile(r"^ *([A-Z_]+)=(.*)\n?$", re.MULTILINE)
  71
  72     def __read_config_file(self, filename):
  73
  74         f = open(filename, "r")
  75         data = f.read()
  76         f.close()
  77         config = {}
  78         for m in self.config_var_re.finditer(data):
  79             (key, val) = m.groups()
  80             config[key] = val.strip('"')
  81         return config
  82
  83     def __do_chroot(self):
  84
  85         return os.chroot("%s/%s" % (VROOTDIR, self.name))
  86
  87     def set_dlimit(self, blocktotal):
  88         path = "%s/%s" % (VROOTDIR, self.name)
  89         inodes, blockcount, size = vduimpl.vdu(path)
  90         blockcount = blockcount >> 1
  91
  92         if blocktotal > blockcount:
  93             vserverimpl.setdlimit(path, self.ctx, blockcount>>1, \
  94                                   blocktotal, inodes, -1, 2)
  95         else:
  96             # should raise some error value
  97             print "block limit (%d) ignored for vserver %s" %(blocktotal,self.name)
  98
  99     def get_dlimit(self):
 100         path = "%s/%s" % (VROOTDIR, self.name)
 101         try:
 102             blocksused, blocktotal, inodesused, inodestotal, reserved = \
 103                         vserverimpl.getdlimit(path,self.ctx)
 104         except OSError, ex:
 105             if ex.errno == 3:
 106                 # get here if no vserver disk limit has been set for xid
 107                 # set blockused to -1 to indicate no limit
 108                 blocktotal = -1
 109
 110         return blocktotal
 111
 112     def set_sched(self, shares, besteffort = True):
 113         global SCHED_TOKENS_MIN, SCHED_TOKENS_MAX
 114         tokensmin = SCHED_TOKENS_MIN
 115         tokensmax = SCHED_TOKENS_MAX
 116
 117         if besteffort is True:
 118             # magic "interval" value for Andy's scheduler to denote besteffort
 119             interval = 1000
 120             fillrate = shares
 121         else:
 122             interval = 1001
 123             fillrate = shares
 124
 125         try:
 126             vserverimpl.setsched(self.ctx,fillrate,interval,tokensmin,tokensmax)
 127         except OSError, ex:
 128             if ex.errno == 22:
 129                 print "kernel does not support vserver scheduler"
 130             else:
 131                 raise ex
 132
 133     def get_sched(self):
 134         # have no way of querying scheduler right now on a per vserver basis
 135         return -1, False
 136
 137     def set_memlimit(self, limit):
 138         ret = vserverimpl.setrlimit(self.ctx,5,limit)
 139         return ret
 140
 141     def get_memlimit(self):
 142         ret = vserverimpl.getrlimit(self.ctx,5)
 143         return ret
 144
 145     def set_tasklimit(self, limit):
 146         ret = vserverimpl.setrlimit(self.ctx,6,limit)
 147         return ret
 148
 149     def get_tasklimit(self):
 150         ret = vserverimpl.getrlimit(self.ctx,6)
 151         return ret
 152
 153     def open(self, filename, mode = "r", bufsize = -1):
 154
 155         (sendsock, recvsock) = passfdimpl.socketpair()
 156         child_pid = os.fork()
 157         if child_pid == 0:
 158             try:
 159                 # child process
 160                 self.__do_chroot()
 161                 f = open(filename, mode)
 162                 passfdimpl.sendmsg(f.fileno(), sendsock)
 163                 os._exit(0)
 164             except EnvironmentError, ex:
 165                 (result, errmsg) = (ex.errno, ex.strerror)
 166             except Exception, ex:
 167                 (result, errmsg) = (255, str(ex))
 168             os.write(sendsock, errmsg)
 169             os._exit(result)
 170
 171         # parent process
 172
 173         # XXX - need this since a lambda can't raise an exception
 174         def __throw(ex):
 175             raise ex
 176
 177         os.close(sendsock)
 178         throw = lambda : __throw(Exception(errmsg))
 179         while True:
 180             try:
 181                 (pid, status) = os.waitpid(child_pid, 0)
 182                 if os.WIFEXITED(status):
 183                     result = os.WEXITSTATUS(status)
 184                     if result != 255:
 185                         errmsg = os.strerror(result)
 186                         throw = lambda : __throw(IOError(result, errmsg))
 187                     else:
 188                         errmsg = "unexpected exception in child"
 189                 else:
 190                     result = -1
 191                     errmsg = "child killed"
 192                 break
 193             except OSError, ex:
 194                 if ex.errno != errno.EINTR:
 195                     os.close(recvsock)
 196                     raise ex
 197         fcntl.fcntl(recvsock, fcntl.F_SETFL, os.O_NONBLOCK)
 198         try:
 199             (fd, errmsg) = passfdimpl.recvmsg(recvsock)
 200         except OSError, ex:
 201             if ex.errno != errno.EAGAIN:
 202                 throw = lambda : __throw(ex)
 203             fd = 0
 204         os.close(recvsock)
 205         if not fd:
 206             throw()
 207
 208         return os.fdopen(fd, mode, bufsize)
 209
 210     def __do_chcontext(self, state_file = None):
 211
 212         vserverimpl.chcontext(self.ctx, self.remove_caps)
 213         if not state_file:
 214             return
 215         print >>state_file, "S_CONTEXT=%d" % self.ctx
 216         print >>state_file, "S_PROFILE=%s" % self.config.get("S_PROFILE", "")
 217         state_file.close()
 218
 219     def __prep(self, runlevel, log):
 220
 221         """ Perform all the crap that the vserver script does before
 222         actually executing the startup scripts. """
 223
 224         # remove /var/run and /var/lock/subsys files
 225         # but don't remove utmp from the top-level /var/run
 226         RUNDIR = "/var/run"
 227         LOCKDIR = "/var/lock/subsys"
 228         filter_fn = lambda fs: filter(lambda f: f != 'utmp', fs)
 229         garbage = reduce((lambda (out, ff), (dir, subdirs, files):
 230                           (out + map((dir + "/").__add__, ff(files)),
 231                            lambda fs: fs)),
 232                          list(os.walk(RUNDIR)),
 233                          ([], filter_fn))[0]
 234         garbage += filter(os.path.isfile, map((LOCKDIR + "/").__add__,
 235                                               os.listdir(LOCKDIR)))
 236         for f in garbage:
 237             os.unlink(f)
 238
 239         # set the initial runlevel
 240         f = open(RUNDIR + "/utmp", "w")
 241         utmp.set_runlevel(f, runlevel)
 242         f.close()
 243
 244         # mount /proc and /dev/pts
 245         self.__do_mount("none", "/proc", "proc")
 246         # XXX - magic mount options
 247         self.__do_mount("none", "/dev/pts", "devpts", 0, "gid=5,mode=0620")
 248
 249     def __do_mount(self, *mount_args):
 250
 251         try:
 252             mountimpl.mount(*mount_args)
 253         except OSError, ex:
 254             if ex.errno == errno.EBUSY:
 255                 # assume already mounted
 256                 return
 257             raise ex
 258
 259     def enter(self):
 260
 261         state_file = open("/var/run/vservers/%s.ctx" % self.name, "w")
 262         self.__do_chroot()
 263         self.__do_chcontext(state_file)
 264
 265     def start(self, wait, runlevel = 3):
 266
 267         child_pid = os.fork()
 268         if child_pid == 0:
 269             # child process
 270             try:
 271                 # get a new session
 272                 os.setsid()
 273
 274                 # open state file to record vserver info
 275                 state_file = open("/var/run/vservers/%s.ctx" % self.name, "w")
 276
 277                 # use /dev/null for stdin, /var/log/boot.log for stdout/err
 278                 os.close(0)
 279                 os.close(1)
 280                 os.open("/dev/null", os.O_RDONLY)
 281                 self.__do_chroot()
 282                 log = open("/var/log/boot.log", "w", 0)
 283                 os.dup2(1, 2)
 284
 285                 print >>log, ("%s: starting the virtual server %s" %
 286                               (time.asctime(time.gmtime()), self.name))
 287
 288                 # perform pre-init cleanup
 289                 self.__prep(runlevel, log)
 290
 291                 # execute each init script in turn
 292                 # XXX - we don't support all scripts that vserver script does
 293                 cmd_pid = 0
 294                 for cmd in self.INITSCRIPTS + [None]:
 295                     # wait for previous command to terminate, unless it
 296                     # is the last one and the caller has specified to wait
 297                     if cmd_pid and (cmd != None or wait):
 298                         try:
 299                             os.waitpid(cmd_pid, 0)
 300                         except:
 301                             print >>log, "error waiting for %s:" % cmd_pid
 302                             traceback.print_exc()
 303
 304                     # end of list
 305                     if cmd == None:
 306                         os._exit(0)
 307
 308                     # fork and exec next command
 309                     cmd_pid = os.fork()
 310                     if cmd_pid == 0:
 311                         try:
 312                             # enter vserver context
 313                             self.__do_chcontext(state_file)
 314                             arg_subst = { 'runlevel': runlevel }
 315                             cmd_args = [cmd[0]] + map(lambda x: x % arg_subst,
 316                                                       cmd[1:])
 317                             print >>log, "executing '%s'" % " ".join(cmd_args)
 318                             os.execl(cmd[0], *cmd_args)
 319                         except:
 320                             traceback.print_exc()
 321                             os._exit(1)
 322                     else:
 323                         # don't want to write state_file multiple times
 324                         state_file = None
 325
 326             # we get here due to an exception in the top-level child process
 327             except Exception, ex:
 328                 traceback.print_exc()
 329             os._exit(0)
 330
 331         # parent process
 332         return child_pid