python/vserver.py

   1 # Copyright 2005 Princeton University
   2
   3 #$Id: vserver.py,v 1.72 2007/08/02 16:01:59 dhozac Exp $
   4
   5 import errno
   6 import fcntl
   7 import os
   8 import re
   9 import pwd
  10 import signal
  11 import sys
  12 import time
  13 import traceback
  14 import subprocess
  15 import resource
  16
  17 import vserverimpl
  18 import cpulimit, bwlimit
  19
  20 from vserverimpl import DLIMIT_INF
  21 from vserverimpl import VC_LIM_KEEP
  22 from vserverimpl import VLIMIT_NSOCK
  23 from vserverimpl import VLIMIT_OPENFD
  24 from vserverimpl import VLIMIT_ANON
  25 from vserverimpl import VLIMIT_SHMEM
  26
  27 #
  28 # these are the flags taken from the kernel linux/vserver/legacy.h
  29 #
  30 FLAGS_LOCK = 1
  31 FLAGS_SCHED = 2  # XXX - defined in util-vserver/src/chcontext.c
  32 FLAGS_NPROC = 4
  33 FLAGS_PRIVATE = 8
  34 FLAGS_INIT = 16
  35 FLAGS_HIDEINFO = 32
  36 FLAGS_ULIMIT = 64
  37 FLAGS_NAMESPACE = 128
  38
  39 RLIMITS = { "NSOCK": VLIMIT_NSOCK,
  40             "OPENFD": VLIMIT_OPENFD,
  41             "ANON": VLIMIT_ANON,
  42             "SHMEM": VLIMIT_SHMEM}
  43
  44 # add in the platform supported rlimits
  45 for entry in resource.__dict__.keys():
  46     if entry.find("RLIMIT_")==0:
  47         k = entry[len("RLIMIT_"):]
  48         if not RLIMITS.has_key(k):
  49             RLIMITS[k]=resource.__dict__[entry]
  50         else:
  51             print "WARNING: duplicate RLIMITS key %s" % k
  52
  53 class NoSuchVServer(Exception): pass
  54
  55
  56 class VServerConfig:
  57     def __init__(self, name, directory):
  58         self.name = name
  59         self.dir = directory
  60         self.cache = None
  61         if not (os.path.isdir(self.dir) and
  62                 os.access(self.dir, os.R_OK | os.W_OK | os.X_OK)):
  63             raise NoSuchVServer, "%s does not exist" % self.dir
  64
  65     def get(self, option, default = None):
  66         try:
  67             if self.cache:
  68                 return self.cache[option]
  69             else:
  70                 f = open(os.path.join(self.dir, option), "r")
  71                 buf = f.read().rstrip()
  72                 f.close()
  73                 return buf
  74         except:
  75             if default is not None:
  76                 return default
  77             else:
  78                 raise KeyError, "Key %s is not set for %s" % (option, self.name)
  79
  80     def update(self, option, value):
  81         if self.cache:
  82             return
  83
  84         try:
  85             old_umask = os.umask(0022)
  86             filename = os.path.join(self.dir, option)
  87             try:
  88                 os.makedirs(os.path.dirname(filename), 0755)
  89             except:
  90                 pass
  91             f = open(filename, 'w')
  92             if isinstance(value, list):
  93                 f.write("%s\n" % "\n".join(value))
  94             else:
  95                 f.write("%s\n" % value)
  96             f.close()
  97             os.umask(old_umask)
  98         except:
  99             raise
 100
 101     def unset(self, option):
 102         if self.cache:
 103             return
 104
 105         try:
 106             filename = os.path.join(self.dir, option)
 107             os.unlink(filename)
 108             try:
 109                 os.removedirs(os.path.dirname(filename))
 110             except:
 111                 pass
 112             return True
 113         except:
 114             return False
 115
 116     def cache_it(self):
 117         self.cache = {}
 118         def add_to_cache(cache, dirname, fnames):
 119             for file in fnames:
 120                 full_name = os.path.join(dirname, file)
 121                 if os.path.islink(full_name):
 122                     fnames.remove(file)
 123                 elif (os.path.isfile(full_name) and
 124                       os.access(full_name, os.R_OK)):
 125                     f = open(full_name, "r")
 126                     cache[full_name.replace(os.path.join(self.dir, ''),
 127                                             '')] = f.read().rstrip()
 128                     f.close()
 129         os.path.walk(self.dir, add_to_cache, self.cache)
 130
 131
 132 class VServer:
 133
 134     INITSCRIPTS = [('/etc/rc.vinit', 'start'),
 135                    ('/etc/rc.d/rc', '%(runlevel)d')]
 136
 137     def __init__(self, name, vm_id = None, vm_running = None, logfile=None):
 138
 139         self.name = name
 140         self.dir = "%s/%s" % (vserverimpl.VSERVER_BASEDIR, name)
 141         if not (os.path.isdir(self.dir) and
 142                 os.access(self.dir, os.R_OK | os.W_OK | os.X_OK)):
 143             raise NoSuchVServer, "no such vserver: " + name
 144         self.config = VServerConfig(name, "/etc/vservers/%s" % name)
 145         self.remove_caps = ~vserverimpl.CAP_SAFE;
 146         if vm_id == None:
 147             vm_id = int(self.config.get('context'))
 148         self.ctx = vm_id
 149         if vm_running == None:
 150             vm_running = self.is_running()
 151         self.vm_running = vm_running
 152         self.logfile = logfile
 153
 154     # inspired from nodemanager's logger
 155     def log(self,msg):
 156         if self.logfile:
 157             try:
 158                 fd = os.open(self.logfile,os.O_WRONLY | os.O_CREAT | os.O_APPEND, 0600)
 159                 if not msg.endswith('\n'): msg += '\n'
 160                 os.write(fd, '%s: %s' % (time.asctime(time.gmtime()), msg))
 161                 os.close(fd)
 162             except:
 163                 print '%s: (%s failed to open) %s'%(time.asctime(time.gmtime()),self.logfile,msg)
 164
 165     def set_rlimit(self, type, hard, soft, min):
 166         """Generic set resource limit function for vserver"""
 167         global RLIMITS
 168         update = False
 169
 170         if hard <> VC_LIM_KEEP:
 171             self.config.update('rlimits/%s.hard' % type.lower(), hard)
 172             update = True
 173         if soft <> VC_LIM_KEEP:
 174             self.config.update('rlimits/%s.soft' % type.lower(), soft)
 175             update = True
 176         if min <> VC_LIM_KEEP:
 177             self.config.update('rlimits/%s.min' % type.lower(), min)
 178             update = True
 179
 180         if self.is_running() and update:
 181             resource_type = RLIMITS[type]
 182             try:
 183                 vserverimpl.setrlimit(self.ctx, resource_type, hard, soft, min)
 184             except OSError, e:
 185                 self.log("Error: setrlimit(%d, %s, %d, %d, %d): %s"
 186                          % (self.ctx, type.lower(), hard, soft, min))
 187
 188         return update
 189
 190     def get_prefix_from_capabilities(self, capabilities, prefix):
 191         split_caps = capabilities.split(',')
 192         return ",".join(["%s" % (c) for c in split_caps if c.startswith(prefix.upper()) or c.startswith(prefix.lower())])
 193
 194     def get_bcaps_from_capabilities(self, capabilities):
 195         return self.get_prefix_from_capabilities(capabilities, "cap_")
 196
 197     def get_ccaps_from_capabilities(self, capabilities):
 198         return self.get_prefix_from_capabilities(capabilities, "vxc_")
 199
 200     def set_capabilities_config(self, capabilities):
 201         bcaps = self.get_bcaps_from_capabilities(capabilities)
 202         ccaps = self.get_ccaps_from_capabilities(capabilities)
 203         self.config.update('bcapabilities', bcaps)
 204         self.config.update('ccapabilities', ccaps)
 205         ret = vserverimpl.setbcaps(self.ctx, vserverimpl.text2bcaps(bcaps))
 206         if ret > 0:
 207             return ret
 208         return vserverimpl.setccaps(self.ctx, vserverimpl.text2ccaps(ccaps))
 209
 210     def get_capabilities(self):
 211         bcaps = vserverimpl.bcaps2text(vserverimpl.getbcaps(self.ctx))
 212         ccaps = vserverimpl.ccaps2text(vserverimpl.getccaps(self.ctx))
 213         if bcaps and ccaps:
 214             ccaps = "," + ccaps
 215         return (bcaps + ccaps)
 216
 217     def get_capabilities_config(self):
 218         bcaps = self.config.get('bcapabilities', '')
 219         ccaps = self.config.get('ccapabilities', '')
 220         if bcaps and ccaps:
 221             ccaps = "," + ccaps
 222         return (bcaps + ccaps)
 223
 224     def set_ipaddresses(self, addresses):
 225         vserverimpl.netremove(self.ctx, "all")
 226         for a in addresses.split(","):
 227             vserverimpl.netadd(self.ctx, a)
 228
 229     def set_ipaddresses_config(self, addresses):
 230         i = 0
 231         for a in addresses.split(","):
 232             self.config.update("interfaces/%d/ip" % i, a)
 233             i += 1
 234         while self.config.unset("interfaces/%d/ip" % i):
 235             i += 1
 236         self.set_ipaddresses(addresses)
 237
 238     def get_ipaddresses_config(self):
 239         i = 0
 240         ret = []
 241         while True:
 242             r = self.config.get("interfaces/%d/ip" % i, '')
 243             if r == '':
 244                 break
 245             ret += [r]
 246             i += 1
 247         return ",".join(ret)
 248
 249     def get_ipaddresses(self):
 250         # No clean way to do this right now.
 251         return None
 252
 253     def __do_chroot(self):
 254         os.chroot(self.dir)
 255         os.chdir("/")
 256
 257     def chroot_call(self, fn, *args):
 258
 259         cwd_fd = os.open(".", os.O_RDONLY)
 260         try:
 261             root_fd = os.open("/", os.O_RDONLY)
 262             try:
 263                 self.__do_chroot()
 264                 result = fn(*args)
 265             finally:
 266                 os.fchdir(root_fd)
 267                 os.chroot(".")
 268                 os.fchdir(cwd_fd)
 269                 os.close(root_fd)
 270         finally:
 271             os.close(cwd_fd)
 272         return result
 273
 274     def set_disklimit(self, block_limit):
 275         # block_limit is in kB
 276         if block_limit == 0:
 277             try:
 278                 vserverimpl.unsetdlimit(self.dir, self.ctx)
 279             except OSError, e:
 280                 self.log("Unexpected error with unsetdlimit for context %d" % self.ctx)
 281             return
 282
 283         if self.vm_running:
 284             block_usage = vserverimpl.DLIMIT_KEEP
 285             inode_usage = vserverimpl.DLIMIT_KEEP
 286         else:
 287             # init_disk_info() must have been called to get usage values
 288             block_usage = self.disk_blocks
 289             inode_usage = self.disk_inodes
 290
 291         try:
 292             vserverimpl.setdlimit(self.dir,
 293                                   self.ctx,
 294                                   block_usage,
 295                                   block_limit,
 296                                   inode_usage,
 297                                   vserverimpl.DLIMIT_INF,  # inode limit
 298                                   2)   # %age reserved for root
 299         except OSError, e:
 300             self.log("Unexpected error with setdlimit for context %d" % self.ctx)
 301
 302
 303         self.config.update('dlimits/0/space_total', block_limit)
 304
 305     def is_running(self):
 306         return vserverimpl.isrunning(self.ctx)
 307
 308     def get_disklimit(self):
 309
 310         try:
 311             (self.disk_blocks, block_limit, self.disk_inodes, inode_limit,
 312              reserved) = vserverimpl.getdlimit(self.dir, self.ctx)
 313         except OSError, ex:
 314             if ex.errno != errno.ESRCH:
 315                 raise
 316             # get here if no vserver disk limit has been set for xid
 317             block_limit = -1
 318
 319         return block_limit
 320
 321     def set_sched_config(self, cpu_min, cpu_share):
 322
 323         """ Write current CPU scheduler parameters to the vserver
 324         configuration file. This method does not modify the kernel CPU
 325         scheduling parameters for this context. """
 326
 327         self.config.update('sched/fill-rate', cpu_min)
 328         self.config.update('sched/fill-rate2', cpu_share)
 329         if cpu_share == 0:
 330             self.config.unset('sched/idle-time')
 331
 332         if self.is_running():
 333             self.set_sched(cpu_min, cpu_share)
 334
 335     def set_sched(self, cpu_min, cpu_share):
 336         """ Update kernel CPU scheduling parameters for this context. """
 337         vserverimpl.setsched(self.ctx, cpu_min, cpu_share)
 338
 339     def get_sched(self):
 340         # have no way of querying scheduler right now on a per vserver basis
 341         return (-1, False)
 342
 343     def set_bwlimit(self, minrate = bwlimit.bwmin, maxrate = None,
 344                     exempt_min = None, exempt_max = None,
 345                     share = None, dev = "eth0"):
 346
 347         if minrate is None:
 348             bwlimit.off(self.ctx, dev)
 349         else:
 350             bwlimit.on(self.ctx, dev, share,
 351                        minrate, maxrate, exempt_min, exempt_max)
 352
 353     def get_bwlimit(self, dev = "eth0"):
 354
 355         result = bwlimit.get(self.ctx)
 356         # result of bwlimit.get is (ctx, share, minrate, maxrate)
 357         if result:
 358             result = result[1:]
 359         return result
 360
 361     def open(self, filename, mode = "r", bufsize = -1):
 362
 363         return self.chroot_call(open, filename, mode, bufsize)
 364
 365     def __do_chcontext(self, state_file):
 366
 367         if state_file:
 368             print >>state_file, "%u" % self.ctx
 369             state_file.close()
 370
 371         if vserverimpl.chcontext(self.ctx, vserverimpl.text2bcaps(self.get_capabilities_config())):
 372             self.set_resources()
 373             vserverimpl.setup_done(self.ctx)
 374
 375     def __prep(self, runlevel):
 376
 377         """ Perform all the crap that the vserver script does before
 378         actually executing the startup scripts. """
 379
 380         # remove /var/run and /var/lock/subsys files
 381         # but don't remove utmp from the top-level /var/run
 382         RUNDIR = "/var/run"
 383         LOCKDIR = "/var/lock/subsys"
 384         filter_fn = lambda fs: filter(lambda f: f != 'utmp', fs)
 385         garbage = reduce((lambda (out, ff), (dir, subdirs, files):
 386                           (out + map((dir + "/").__add__, ff(files)),
 387                            lambda fs: fs)),
 388                          list(os.walk(RUNDIR)),
 389                          ([], filter_fn))[0]
 390         garbage += filter(os.path.isfile, map((LOCKDIR + "/").__add__,
 391                                               os.listdir(LOCKDIR)))
 392         if False:
 393             for f in garbage:
 394                 os.unlink(f)
 395
 396         # set the initial runlevel
 397         vserverimpl.setrunlevel(RUNDIR + "/utmp", runlevel)
 398
 399         # mount /proc and /dev/pts
 400         self.__do_mount("none", self.dir, "/proc", "proc")
 401         # XXX - magic mount options
 402         self.__do_mount("none", self.dir, "/dev/pts", "devpts", 0, "gid=5,mode=0620")
 403
 404     def __do_mount(self, *mount_args):
 405
 406         try:
 407             vserverimpl.mount(*mount_args)
 408         except OSError, ex:
 409             if ex.errno == errno.EBUSY:
 410                 # assume already mounted
 411                 return
 412             raise ex
 413
 414     def enter(self):
 415         self.config.cache_it()
 416         self.__do_chroot()
 417         self.__do_chcontext(None)
 418
 419     def start(self, runlevel = 3):
 420
 421         if (os.fork() != 0):
 422             # Parent should just return.
 423             self.vm_running = True
 424             return
 425         else:
 426             # child process
 427             try:
 428                 # so we don't chcontext with priv'ed fds
 429                 close_nonstandard_fds()
 430
 431                 # get a new session
 432                 os.setsid()
 433
 434                 # open state file to record vserver info
 435                 state_file = open("/var/run/vservers/%s" % self.name, "w")
 436
 437                 # use /dev/null for stdin, /var/log/boot.log for stdout/err
 438                 fd = os.open("/dev/null", os.O_RDONLY)
 439                 if fd != 0:
 440                     os.dup2(fd, 0)
 441                     os.close(fd)
 442
 443                 # perform pre-init cleanup
 444                 self.__prep(runlevel)
 445
 446                 self.config.cache_it()
 447                 self.__do_chroot()
 448                 log = open("/var/log/boot.log", "a", 0)
 449                 if log.fileno() != 1:
 450                     os.dup2(log.fileno(), 1)
 451                 os.dup2(1, 2)
 452
 453                 print >>log, ("%s: starting the virtual server %s" %
 454                               (time.asctime(time.gmtime()), self.name))
 455                 # execute each init script in turn
 456                 # XXX - we don't support all scripts that vserver script does
 457                 self.__do_chcontext(state_file)
 458                 for cmd in self.INITSCRIPTS:
 459                     try:
 460                         # enter vserver context
 461                         arg_subst = { 'runlevel': runlevel }
 462                         cmd_args = [cmd[0]] + map(lambda x: x % arg_subst,
 463                                                    cmd[1:])
 464                         if os.path.isfile(cmd[0]):
 465                             print >>log, "executing '%s'" % " ".join(cmd_args)
 466                             os.spawnvp(os.P_NOWAIT,cmd[0],cmd_args)
 467                     except:
 468                         print >>log, traceback.format_exc()
 469
 470             # we get here due to an exception in the top-level child process
 471             except Exception, ex:
 472                 self.log(traceback.format_exc())
 473             os._exit(0)
 474
 475     def set_resources(self):
 476
 477         """ Called when vserver context is entered for first time,
 478         should be overridden by subclass. """
 479
 480         pass
 481
 482     def init_disk_info(self):
 483         try:
 484             dlimit = vserverimpl.getdlimit(self.dir, self.ctx)
 485             self.disk_blocks = dlimit[0]
 486             self.disk_inodes = dlimit[2]
 487             return self.disk_blocks * 1024
 488         except Exception, e:
 489             pass
 490         cmd = "/usr/sbin/vdu --script --space --inodes --blocksize 1024 --xid %d %s" % (self.ctx, self.dir)
 491         p = subprocess.Popen(cmd, shell=True, stdin=subprocess.PIPE,
 492                              stdout=subprocess.PIPE, stderr=subprocess.PIPE,
 493                              close_fds=True)
 494         p.stdin.close()
 495         line = p.stdout.readline()
 496         if not line:
 497             sys.stderr.write(p.stderr.read())
 498         p.stdout.close()
 499         p.stderr.close()
 500         ret = p.wait()
 501
 502         (space, inodes) = line.split()
 503         self.disk_inodes = int(inodes)
 504         self.disk_blocks = int(space)
 505         #(self.disk_inodes, self.disk_blocks) = vduimpl.vdu(self.dir)
 506
 507         return self.disk_blocks * 1024
 508
 509     def stop(self, signal = signal.SIGKILL):
 510         vserverimpl.killall(self.ctx, signal)
 511         self.vm_running = False
 512
 513     def setname(self, slice_id):
 514         '''Set vcVHI_CONTEXT field in kernel to slice_id'''
 515         vserverimpl.setname(self.ctx, slice_id)
 516
 517     def getname(self):
 518         '''Get vcVHI_CONTEXT field in kernel'''
 519         return vserverimpl.getname(self.ctx)
 520
 521
 522 def create(vm_name, static = False, ctor = VServer):
 523
 524     options = ['vuseradd']
 525     if static:
 526         options += ['--static']
 527     ret = os.spawnvp(os.P_WAIT, 'vuseradd', options + [vm_name])
 528     if not os.WIFEXITED(ret) or os.WEXITSTATUS(ret) != 0:
 529         out = "system command ('%s') " % options
 530         if os.WIFEXITED(ret):
 531             out += "failed, rc = %d" % os.WEXITSTATUS(ret)
 532         else:
 533             out += "killed by signal %d" % os.WTERMSIG(ret)
 534         raise SystemError, out
 535     vm_id = pwd.getpwnam(vm_name)[2]
 536
 537     return ctor(vm_name, vm_id)
 538
 539
 540 def close_nonstandard_fds():
 541     """Close all open file descriptors other than 0, 1, and 2."""
 542     _SC_OPEN_MAX = 4
 543     for fd in range(3, os.sysconf(_SC_OPEN_MAX)):
 544         try: os.close(fd)
 545         except OSError: pass  # most likely an fd that isn't open
 546