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