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 passfdimpl
  13 import utmp
  14 import vserverimpl, vduimpl
  15 import cpulimit, bwlimit
  16
  17
  18
  19 #
  20 # these are the flags taken from the kernel linux/vserver/legacy.h
  21 #
  22 FLAGS_LOCK = 1
  23 FLAGS_SCHED = 2  # XXX - defined in util-vserver/src/chcontext.c
  24 FLAGS_NPROC = 4
  25 FLAGS_PRIVATE = 8
  26 FLAGS_INIT = 16
  27 FLAGS_HIDEINFO = 32
  28 FLAGS_ULIMIT = 64
  29 FLAGS_NAMESPACE = 128
  30
  31
  32
  33 class VServer:
  34
  35     INITSCRIPTS = [('/etc/rc.vinit', 'start'),
  36                    ('/etc/rc.d/rc', '%(runlevel)d')]
  37
  38     def __init__(self, name, vm_id, vm_running = False, resources = {}):
  39
  40         self.name = name
  41         self.config_file = "/etc/vservers/%s.conf" % name
  42         self.dir = "%s/%s" % (vserverimpl.VSERVER_BASEDIR, name)
  43         if not (os.path.isdir(self.dir) and
  44                 os.access(self.dir, os.R_OK | os.W_OK | os.X_OK)):
  45             raise Exception, "no such vserver: " + name
  46         self.config = {}
  47         for config_file in ["/etc/vservers.conf", self.config_file]:
  48             try:
  49                 self.config.update(self.__read_config_file(config_file))
  50             except IOError, ex:
  51                 if ex.errno != errno.ENOENT:
  52                     raise
  53         self.remove_caps = ~vserverimpl.CAP_SAFE;
  54         self.ctx = vm_id
  55         self.vm_running = vm_running
  56         self.resources = resources
  57
  58     config_var_re = re.compile(r"^ *([A-Z_]+)=(.*)\n?$", re.MULTILINE)
  59
  60     def __read_config_file(self, filename):
  61
  62         f = open(filename, "r")
  63         data = f.read()
  64         f.close()
  65         config = {}
  66         for m in self.config_var_re.finditer(data):
  67             (key, val) = m.groups()
  68             config[key] = val.strip('"')
  69         return config
  70
  71     def __update_config_file(self, filename, newvars):
  72
  73         # read old file, apply changes
  74         f = open(filename, "r")
  75         data = f.read()
  76         f.close()
  77         todo = newvars.copy()
  78         changed = False
  79         for m in self.config_var_re.finditer(data):
  80             (key, val) = m.groups()
  81             newval = todo.pop(key, None)
  82             if newval != None:
  83                 data = data[:m.start(2)] + str(newval) + data[m.end(2):]
  84                 changed = True
  85         for (newkey, newval) in todo.items():
  86             data += "%s=%s\n" % (newkey, newval)
  87             changed = True
  88
  89         if not changed:
  90             return
  91
  92         # write new file
  93         newfile = filename + ".new"
  94         f = open(newfile, "w")
  95         f.write(data)
  96         f.close()
  97
  98         # 'copy' original file, rename new to original
  99         backup = filename + ".old"
 100         try:
 101             os.unlink(backup)
 102         except OSError, ex:
 103             if ex.errno != errno.ENOENT:
 104                 raise
 105         os.link(filename, backup)
 106         os.rename(newfile, filename)
 107
 108     def __do_chroot(self):
 109
 110         os.chroot(self.dir)
 111         os.chdir("/")
 112
 113     def set_disklimit(self, block_limit):
 114
 115         # block_limit is in kB
 116         if block_limit == 0:
 117             vserverimpl.unsetdlimit(self.dir, self.ctx)
 118             return
 119
 120         if self.vm_running:
 121             block_usage = vserverimpl.DLIMIT_KEEP
 122             inode_usage = vserverimpl.DLIMIT_KEEP
 123         else:
 124             # init_disk_info() must have been called to get usage values
 125             block_usage = self.disk_blocks
 126             inode_usage = self.disk_inodes
 127
 128         vserverimpl.setdlimit(self.dir,
 129                               self.ctx,
 130                               block_usage,
 131                               block_limit,
 132                               inode_usage,
 133                               vserverimpl.DLIMIT_INF,  # inode limit
 134                               2)   # %age reserved for root
 135
 136     def get_disklimit(self):
 137
 138         try:
 139             blocksused, blocktotal, inodesused, inodestotal, reserved = \
 140                         vserverimpl.getdlimit(self.dir, self.ctx)
 141         except OSError, ex:
 142             if ex.errno != errno.ESRCH:
 143                 raise
 144             # get here if no vserver disk limit has been set for xid
 145             blocktotal = -1
 146
 147         return blocktotal
 148
 149     def set_sched(self, cpu_share):
 150
 151         if cpu_share == int(self.config.get("CPULIMIT", -1)):
 152             return
 153
 154         self.__update_config_file(self.config_file, { "CPULIMIT": cpu_share })
 155         if self.vm_running:
 156             vserverimpl.setsched(self.ctx, self.resources)
 157
 158     def get_sched(self):
 159         # have no way of querying scheduler right now on a per vserver basis
 160         return (-1, False)
 161
 162     def set_memlimit(self, limit):
 163         ret = vserverimpl.setrlimit(self.ctx,5,limit)
 164         return ret
 165
 166     def get_memlimit(self):
 167         ret = vserverimpl.getrlimit(self.ctx,5)
 168         return ret
 169
 170     def set_tasklimit(self, limit):
 171         ret = vserverimpl.setrlimit(self.ctx,6,limit)
 172         return ret
 173
 174     def get_tasklimit(self):
 175         ret = vserverimpl.getrlimit(self.ctx,6)
 176         return ret
 177
 178     def set_bwlimit(self, share, minrate = 1, maxrate = None, dev = "eth0"):
 179
 180         if share:
 181             bwlimit.on(self.ctx, dev, share, minrate, maxrate)
 182         else:
 183             bwlimit.off(self.ctx, dev)
 184
 185     def get_bwlimit(self, eth):
 186         # not implemented yet
 187         bwlimit = -1
 188         cap = "unknown"
 189         minrate = "unknown"
 190         maxrate = "unknown"
 191         return (bwlimit, cap, minrate, maxrate)
 192
 193     def open(self, filename, mode = "r", bufsize = -1):
 194
 195         (sendsock, recvsock) = passfdimpl.socketpair()
 196         child_pid = os.fork()
 197         if child_pid == 0:
 198             try:
 199                 # child process
 200                 self.__do_chroot()
 201                 f = open(filename, mode)
 202                 passfdimpl.sendmsg(f.fileno(), sendsock)
 203                 os._exit(0)
 204             except EnvironmentError, ex:
 205                 (result, errmsg) = (ex.errno, ex.strerror)
 206             except Exception, ex:
 207                 (result, errmsg) = (255, str(ex))
 208             os.write(sendsock, errmsg)
 209             os._exit(result)
 210
 211         # parent process
 212
 213         # XXX - need this since a lambda can't raise an exception
 214         def __throw(ex):
 215             raise ex
 216
 217         os.close(sendsock)
 218         throw = lambda : __throw(Exception(errmsg))
 219         while True:
 220             try:
 221                 (pid, status) = os.waitpid(child_pid, 0)
 222                 if os.WIFEXITED(status):
 223                     result = os.WEXITSTATUS(status)
 224                     if result != 255:
 225                         errmsg = os.strerror(result)
 226                         throw = lambda : __throw(IOError(result, errmsg))
 227                     else:
 228                         errmsg = "unexpected exception in child"
 229                 else:
 230                     result = -1
 231                     errmsg = "child killed"
 232                 break
 233             except OSError, ex:
 234                 if ex.errno != errno.EINTR:
 235                     os.close(recvsock)
 236                     raise ex
 237         fcntl.fcntl(recvsock, fcntl.F_SETFL, os.O_NONBLOCK)
 238         try:
 239             (fd, errmsg) = passfdimpl.recvmsg(recvsock)
 240         except OSError, ex:
 241             if ex.errno != errno.EAGAIN:
 242                 throw = lambda : __throw(ex)
 243             fd = 0
 244         os.close(recvsock)
 245         if not fd:
 246             throw()
 247
 248         return os.fdopen(fd, mode, bufsize)
 249
 250     def __do_chcontext(self, state_file):
 251
 252         vserverimpl.chcontext(self.ctx, self.resources)
 253
 254         if not state_file:
 255             return
 256         print >>state_file, "S_CONTEXT=%d" % self.ctx
 257         print >>state_file, "S_PROFILE=%s" % self.config.get("S_PROFILE", "")
 258         state_file.close()
 259
 260     def __prep(self, runlevel, log):
 261
 262         """ Perform all the crap that the vserver script does before
 263         actually executing the startup scripts. """
 264
 265         # remove /var/run and /var/lock/subsys files
 266         # but don't remove utmp from the top-level /var/run
 267         RUNDIR = "/var/run"
 268         LOCKDIR = "/var/lock/subsys"
 269         filter_fn = lambda fs: filter(lambda f: f != 'utmp', fs)
 270         garbage = reduce((lambda (out, ff), (dir, subdirs, files):
 271                           (out + map((dir + "/").__add__, ff(files)),
 272                            lambda fs: fs)),
 273                          list(os.walk(RUNDIR)),
 274                          ([], filter_fn))[0]
 275         garbage += filter(os.path.isfile, map((LOCKDIR + "/").__add__,
 276                                               os.listdir(LOCKDIR)))
 277         for f in garbage:
 278             os.unlink(f)
 279
 280         # set the initial runlevel
 281         f = open(RUNDIR + "/utmp", "w")
 282         utmp.set_runlevel(f, runlevel)
 283         f.close()
 284
 285         # mount /proc and /dev/pts
 286         self.__do_mount("none", "/proc", "proc")
 287         # XXX - magic mount options
 288         self.__do_mount("none", "/dev/pts", "devpts", 0, "gid=5,mode=0620")
 289
 290     def __do_mount(self, *mount_args):
 291
 292         try:
 293             mountimpl.mount(*mount_args)
 294         except OSError, ex:
 295             if ex.errno == errno.EBUSY:
 296                 # assume already mounted
 297                 return
 298             raise ex
 299
 300     def enter(self):
 301
 302         state_file = open("/var/run/vservers/%s.ctx" % self.name, "w")
 303         self.__do_chroot()
 304         self.__do_chcontext(state_file)
 305
 306     def start(self, wait, runlevel = 3):
 307
 308         self.vm_running = True
 309
 310         child_pid = os.fork()
 311         if child_pid == 0:
 312             # child process
 313             try:
 314                 # get a new session
 315                 os.setsid()
 316
 317                 # open state file to record vserver info
 318                 state_file = open("/var/run/vservers/%s.ctx" % self.name, "w")
 319
 320                 # use /dev/null for stdin, /var/log/boot.log for stdout/err
 321                 os.close(0)
 322                 os.close(1)
 323                 os.open("/dev/null", os.O_RDONLY)
 324                 self.__do_chroot()
 325                 log = open("/var/log/boot.log", "w", 0)
 326                 os.dup2(1, 2)
 327                 # XXX - close all other fds
 328
 329                 print >>log, ("%s: starting the virtual server %s" %
 330                               (time.asctime(time.gmtime()), self.name))
 331
 332                 # perform pre-init cleanup
 333                 self.__prep(runlevel, log)
 334
 335                 # execute each init script in turn
 336                 # XXX - we don't support all scripts that vserver script does
 337                 cmd_pid = 0
 338                 first_child = True
 339                 for cmd in self.INITSCRIPTS + [None]:
 340                     # wait for previous command to terminate, unless it
 341                     # is the last one and the caller has specified to wait
 342                     if cmd_pid and (cmd != None or wait):
 343                         try:
 344                             os.waitpid(cmd_pid, 0)
 345                         except:
 346                             print >>log, "error waiting for %s:" % cmd_pid
 347                             traceback.print_exc()
 348
 349                     # end of list
 350                     if cmd == None:
 351                         os._exit(0)
 352
 353                     # fork and exec next command
 354                     cmd_pid = os.fork()
 355                     if cmd_pid == 0:
 356                         try:
 357                             # enter vserver context
 358                             self.__do_chcontext(state_file)
 359                             arg_subst = { 'runlevel': runlevel }
 360                             cmd_args = [cmd[0]] + map(lambda x: x % arg_subst,
 361                                                       cmd[1:])
 362                             print >>log, "executing '%s'" % " ".join(cmd_args)
 363                             os.execl(cmd[0], *cmd_args)
 364                         except:
 365                             traceback.print_exc()
 366                             os._exit(1)
 367                     else:
 368                         # don't want to write state_file multiple times
 369                         state_file = None
 370
 371             # we get here due to an exception in the top-level child process
 372             except Exception, ex:
 373                 traceback.print_exc()
 374             os._exit(0)
 375
 376         # parent process
 377         return child_pid
 378
 379     def update_resources(self, resources):
 380
 381         self.config.update(resources)
 382
 383         # write new values to configuration file
 384         self.__update_config_file(self.config_file, resources)
 385
 386     def init_disk_info(self):
 387
 388         (self.disk_inodes, self.disk_blocks, size) = vduimpl.vdu(self.dir)
 389
 390         return size