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 import cpulimit, bwlimit
  17
  18 from util_vserver_vars import *
  19
  20 CAP_SAFE = (linuxcaps.CAP_CHOWN |
  21             linuxcaps.CAP_DAC_OVERRIDE |
  22             linuxcaps.CAP_DAC_READ_SEARCH |
  23             linuxcaps.CAP_FOWNER |
  24             linuxcaps.CAP_FSETID |
  25             linuxcaps.CAP_KILL |
  26             linuxcaps.CAP_SETGID |
  27             linuxcaps.CAP_SETUID |
  28             linuxcaps.CAP_SETPCAP |
  29             linuxcaps.CAP_SYS_TTY_CONFIG |
  30             linuxcaps.CAP_LEASE |
  31             linuxcaps.CAP_SYS_CHROOT |
  32             linuxcaps.CAP_SYS_PTRACE)
  33
  34 #
  35 # these are the flags taken from the kernel linux/vserver/legacy.h
  36 #
  37 FLAGS_LOCK = 1
  38 FLAGS_SCHED = 2  # XXX - defined in util-vserver/src/chcontext.c
  39 FLAGS_NPROC = 4
  40 FLAGS_PRIVATE = 8
  41 FLAGS_INIT = 16
  42 FLAGS_HIDEINFO = 32
  43 FLAGS_ULIMIT = 64
  44 FLAGS_NAMESPACE = 128
  45
  46 # default values for new vserver scheduler
  47 SCHED_TOKENS_MIN = 50
  48 SCHED_TOKENS_MAX = 100
  49 SCHED_TOKENS = 100
  50 SCHED_INTERVAL = 1000
  51
  52
  53 class VServer:
  54
  55     INITSCRIPTS = [('/etc/rc.vinit', 'start'),
  56                    ('/etc/rc.d/rc', '%(runlevel)d')]
  57
  58     def __init__(self, name):
  59
  60         self.name = name
  61         self.config = self.__read_config_file("/etc/vservers.conf")
  62         self.config.update(self.__read_config_file("/etc/vservers/%s.conf" %
  63                                                    self.name))
  64         self.flags = 0
  65         flags = self.config["S_FLAGS"].split(" ")
  66         if "lock" in flags:
  67             self.flags |= FLAGS_LOCK
  68         if "nproc" in flags:
  69             self.flags |= FLAGS_NPROC
  70         self.remove_caps = ~CAP_SAFE
  71         self.ctx = int(self.config["S_CONTEXT"])
  72
  73     config_var_re = re.compile(r"^ *([A-Z_]+)=(.*)\n?$", re.MULTILINE)
  74
  75     def __read_config_file(self, filename):
  76
  77         f = open(filename, "r")
  78         data = f.read()
  79         f.close()
  80         config = {}
  81         for m in self.config_var_re.finditer(data):
  82             (key, val) = m.groups()
  83             config[key] = val.strip('"')
  84         return config
  85
  86     def __do_chroot(self):
  87
  88         return os.chroot("%s/%s" % (DEFAULT_VSERVERDIR, self.name))
  89
  90     def set_disklimit(self, blocktotal):
  91         path = "%s/%s" % (DEFAULT_VSERVERDIR, self.name)
  92         inodes, blockcount, size = vduimpl.vdu(path)
  93         blockcount = blockcount >> 1
  94
  95         if blocktotal > blockcount:
  96             vserverimpl.setdlimit(path, self.ctx, blockcount>>1, \
  97                                   blocktotal, inodes, -1, 2)
  98         else:
  99             # should raise some error value
 100             print "block limit (%d) ignored for vserver %s" %(blocktotal,self.name)
 101
 102     def get_disklimit(self):
 103         path = "%s/%s" % (DEFAULT_VSERVERDIR, self.name)
 104         try:
 105             blocksused, blocktotal, inodesused, inodestotal, reserved = \
 106                         vserverimpl.getdlimit(path,self.ctx)
 107         except OSError, ex:
 108             if ex.errno == 3:
 109                 # get here if no vserver disk limit has been set for xid
 110                 # set blockused to -1 to indicate no limit
 111                 blocktotal = -1
 112
 113         return blocktotal
 114
 115     def set_sched(self, shares = 32, besteffort = True):
 116         # for the old CKRM scheduler
 117         if cpulimit.checkckrm() is True:
 118             cpulimit.cpuinit()
 119             cpulimit.vs2ckrm_on(self.name)
 120             try:
 121                 cpulimit.cpulimit(self.name,shares)
 122             except OSError, ex:
 123                 if ex.errno == 22:
 124                     print "invalid shares argument"
 125                     # should re-raise exception?!
 126
 127         # for the new vserver scheduler
 128         else:
 129             global SCHED_TOKENS_MIN, SCHED_TOKENS_MAX, SCHED_TOKENS, SCHED_INTERVAL
 130             tokensmin = SCHED_TOKENS_MIN
 131             tokensmax = SCHED_TOKENS_MAX
 132             tokens    = SCHED_TOKENS
 133             interval  = SCHED_INTERVAL
 134             fillrate = shares
 135
 136             if besteffort is True:
 137                 cpuguaranteed = 0
 138             else:
 139                 cpuguaranteed = 1
 140
 141             try:
 142                 vserverimpl.setsched(self.ctx,fillrate,interval,tokens,tokensmin,tokensmax,cpuguaranteed)
 143             except OSError, ex:
 144                 if ex.errno == 22:
 145                     print "kernel does not support vserver scheduler"
 146                 else:
 147                     raise ex
 148
 149     def get_sched(self):
 150         # have no way of querying scheduler right now on a per vserver basis
 151         return (-1, False)
 152
 153     def set_memlimit(self, limit):
 154         ret = vserverimpl.setrlimit(self.ctx,5,limit)
 155         return ret
 156
 157     def get_memlimit(self):
 158         ret = vserverimpl.getrlimit(self.ctx,5)
 159         return ret
 160
 161     def set_tasklimit(self, limit):
 162         ret = vserverimpl.setrlimit(self.ctx,6,limit)
 163         return ret
 164
 165     def get_tasklimit(self):
 166         ret = vserverimpl.getrlimit(self.ctx,6)
 167         return ret
 168
 169     def set_bwlimit(self, eth, limit, cap, minrate, maxrate):
 170         if cap == "-1":
 171             bwlimit.off(self.ctx,eth)
 172         else:
 173             bwlimit.on(self.ctx, eth, limit, cap, minrate, maxrate)
 174
 175     def get_bwlimit(self, eth):
 176         # not implemented yet
 177         bwlimit = -1
 178         cap = "unknown"
 179         minrate = "unknown"
 180         maxrate = "unknown"
 181         return (bwlimit, cap, minrate, maxrate)
 182
 183     def open(self, filename, mode = "r", bufsize = -1):
 184
 185         (sendsock, recvsock) = passfdimpl.socketpair()
 186         child_pid = os.fork()
 187         if child_pid == 0:
 188             try:
 189                 # child process
 190                 self.__do_chroot()
 191                 f = open(filename, mode)
 192                 passfdimpl.sendmsg(f.fileno(), sendsock)
 193                 os._exit(0)
 194             except EnvironmentError, ex:
 195                 (result, errmsg) = (ex.errno, ex.strerror)
 196             except Exception, ex:
 197                 (result, errmsg) = (255, str(ex))
 198             os.write(sendsock, errmsg)
 199             os._exit(result)
 200
 201         # parent process
 202
 203         # XXX - need this since a lambda can't raise an exception
 204         def __throw(ex):
 205             raise ex
 206
 207         os.close(sendsock)
 208         throw = lambda : __throw(Exception(errmsg))
 209         while True:
 210             try:
 211                 (pid, status) = os.waitpid(child_pid, 0)
 212                 if os.WIFEXITED(status):
 213                     result = os.WEXITSTATUS(status)
 214                     if result != 255:
 215                         errmsg = os.strerror(result)
 216                         throw = lambda : __throw(IOError(result, errmsg))
 217                     else:
 218                         errmsg = "unexpected exception in child"
 219                 else:
 220                     result = -1
 221                     errmsg = "child killed"
 222                 break
 223             except OSError, ex:
 224                 if ex.errno != errno.EINTR:
 225                     os.close(recvsock)
 226                     raise ex
 227         fcntl.fcntl(recvsock, fcntl.F_SETFL, os.O_NONBLOCK)
 228         try:
 229             (fd, errmsg) = passfdimpl.recvmsg(recvsock)
 230         except OSError, ex:
 231             if ex.errno != errno.EAGAIN:
 232                 throw = lambda : __throw(ex)
 233             fd = 0
 234         os.close(recvsock)
 235         if not fd:
 236             throw()
 237
 238         return os.fdopen(fd, mode, bufsize)
 239
 240     def __do_chcontext(self, state_file = None):
 241
 242         vserverimpl.chcontext(self.ctx)
 243
 244         if not state_file:
 245             return
 246         print >>state_file, "S_CONTEXT=%d" % self.ctx
 247         print >>state_file, "S_PROFILE=%s" % self.config.get("S_PROFILE", "")
 248         state_file.close()
 249
 250     def __prep(self, runlevel, log):
 251
 252         """ Perform all the crap that the vserver script does before
 253         actually executing the startup scripts. """
 254
 255         # remove /var/run and /var/lock/subsys files
 256         # but don't remove utmp from the top-level /var/run
 257         RUNDIR = "/var/run"
 258         LOCKDIR = "/var/lock/subsys"
 259         filter_fn = lambda fs: filter(lambda f: f != 'utmp', fs)
 260         garbage = reduce((lambda (out, ff), (dir, subdirs, files):
 261                           (out + map((dir + "/").__add__, ff(files)),
 262                            lambda fs: fs)),
 263                          list(os.walk(RUNDIR)),
 264                          ([], filter_fn))[0]
 265         garbage += filter(os.path.isfile, map((LOCKDIR + "/").__add__,
 266                                               os.listdir(LOCKDIR)))
 267         for f in garbage:
 268             os.unlink(f)
 269
 270         # set the initial runlevel
 271         f = open(RUNDIR + "/utmp", "w")
 272         utmp.set_runlevel(f, runlevel)
 273         f.close()
 274
 275         # mount /proc and /dev/pts
 276         self.__do_mount("none", "/proc", "proc")
 277         # XXX - magic mount options
 278         self.__do_mount("none", "/dev/pts", "devpts", 0, "gid=5,mode=0620")
 279
 280     def __do_mount(self, *mount_args):
 281
 282         try:
 283             mountimpl.mount(*mount_args)
 284         except OSError, ex:
 285             if ex.errno == errno.EBUSY:
 286                 # assume already mounted
 287                 return
 288             raise ex
 289
 290     def enter(self):
 291
 292         state_file = open("/var/run/vservers/%s.ctx" % self.name, "w")
 293         self.__do_chroot()
 294         self.__do_chcontext(state_file)
 295
 296     def start(self, wait, runlevel = 3):
 297
 298         child_pid = os.fork()
 299         if child_pid == 0:
 300             # child process
 301             try:
 302                 # get a new session
 303                 os.setsid()
 304
 305                 # open state file to record vserver info
 306                 state_file = open("/var/run/vservers/%s.ctx" % self.name, "w")
 307
 308                 # use /dev/null for stdin, /var/log/boot.log for stdout/err
 309                 os.close(0)
 310                 os.close(1)
 311                 os.open("/dev/null", os.O_RDONLY)
 312                 self.__do_chroot()
 313                 log = open("/var/log/boot.log", "w", 0)
 314                 os.dup2(1, 2)
 315
 316                 print >>log, ("%s: starting the virtual server %s" %
 317                               (time.asctime(time.gmtime()), self.name))
 318
 319                 # perform pre-init cleanup
 320                 self.__prep(runlevel, log)
 321
 322                 # execute each init script in turn
 323                 # XXX - we don't support all scripts that vserver script does
 324                 cmd_pid = 0
 325                 for cmd in self.INITSCRIPTS + [None]:
 326                     # wait for previous command to terminate, unless it
 327                     # is the last one and the caller has specified to wait
 328                     if cmd_pid and (cmd != None or wait):
 329                         try:
 330                             os.waitpid(cmd_pid, 0)
 331                         except:
 332                             print >>log, "error waiting for %s:" % cmd_pid
 333                             traceback.print_exc()
 334
 335                     # end of list
 336                     if cmd == None:
 337                         os._exit(0)
 338
 339                     # fork and exec next command
 340                     cmd_pid = os.fork()
 341                     if cmd_pid == 0:
 342                         try:
 343                             # enter vserver context
 344                             self.__do_chcontext(state_file)
 345                             arg_subst = { 'runlevel': runlevel }
 346                             cmd_args = [cmd[0]] + map(lambda x: x % arg_subst,
 347                                                       cmd[1:])
 348                             print >>log, "executing '%s'" % " ".join(cmd_args)
 349                             os.execl(cmd[0], *cmd_args)
 350                         except:
 351                             traceback.print_exc()
 352                             os._exit(1)
 353                     else:
 354                         # don't want to write state_file multiple times
 355                         state_file = None
 356
 357             # we get here due to an exception in the top-level child process
 358             except Exception, ex:
 359                 traceback.print_exc()
 360             os._exit(0)
 361
 362         # parent process
 363         return child_pid