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
  50
  51 class VServer:
  52
  53     INITSCRIPTS = [('/etc/rc.vinit', 'start'),
  54                    ('/etc/rc.d/rc', '%(runlevel)d')]
  55
  56     def __init__(self, name):
  57
  58         self.name = name
  59         self.config = self.__read_config_file("/etc/vservers.conf")
  60         self.config.update(self.__read_config_file("/etc/vservers/%s.conf" %
  61                                                    self.name))
  62         self.flags = 0
  63         flags = self.config["S_FLAGS"].split(" ")
  64         if "lock" in flags:
  65             self.flags |= FLAGS_LOCK
  66         if "nproc" in flags:
  67             self.flags |= FLAGS_NPROC
  68         self.remove_caps = ~CAP_SAFE
  69         self.ctx = int(self.config["S_CONTEXT"])
  70
  71     config_var_re = re.compile(r"^ *([A-Z_]+)=(.*)\n?$", re.MULTILINE)
  72
  73     def __read_config_file(self, filename):
  74
  75         f = open(filename, "r")
  76         data = f.read()
  77         f.close()
  78         config = {}
  79         for m in self.config_var_re.finditer(data):
  80             (key, val) = m.groups()
  81             config[key] = val.strip('"')
  82         return config
  83
  84     def __do_chroot(self):
  85
  86         return os.chroot("%s/%s" % (DEFAULT_VSERVERDIR, self.name))
  87
  88     def set_disklimit(self, blocktotal):
  89         path = "%s/%s" % (DEFAULT_VSERVERDIR, self.name)
  90         inodes, blockcount, size = vduimpl.vdu(path)
  91         blockcount = blockcount >> 1
  92
  93         if blocktotal > blockcount:
  94             vserverimpl.setdlimit(path, self.ctx, blockcount>>1, \
  95                                   blocktotal, inodes, -1, 2)
  96         else:
  97             # should raise some error value
  98             print "block limit (%d) ignored for vserver %s" %(blocktotal,self.name)
  99
 100     def get_disklimit(self):
 101         path = "%s/%s" % (DEFAULT_VSERVERDIR, self.name)
 102         try:
 103             blocksused, blocktotal, inodesused, inodestotal, reserved = \
 104                         vserverimpl.getdlimit(path,self.ctx)
 105         except OSError, ex:
 106             if ex.errno == 3:
 107                 # get here if no vserver disk limit has been set for xid
 108                 # set blockused to -1 to indicate no limit
 109                 blocktotal = -1
 110
 111         return blocktotal
 112
 113     def set_sched(self, shares, besteffort = True):
 114         # for the old CKRM scheduler
 115         if cpulimit.checkckrm() is True:
 116             cpulimit.cpuinit()
 117             cpulimit.vs2ckrm_on(self.name)
 118             try:
 119                 cpulimit.cpulimit(self.name,shares)
 120             except OSError, ex:
 121                 if ex.errno == 22:
 122                     print "invalid shares argument"
 123                     # should re-raise exception?!
 124
 125         # for the new vserver scheduler
 126         else:
 127             global SCHED_TOKENS_MIN, SCHED_TOKENS_MAX
 128             tokensmin = SCHED_TOKENS_MIN
 129             tokensmax = SCHED_TOKENS_MAX
 130
 131             if besteffort is True:
 132                 # magic "interval" value for Andy's scheduler to denote besteffort
 133                 interval = 1000
 134                 fillrate = shares
 135             else:
 136                 interval = 1001
 137                 fillrate = shares
 138
 139             try:
 140                 cpuguaranteed = 0 # need to set this from the conf file
 141                 vserverimpl.setsched(self.ctx,fillrate,interval,tokensmin,tokensmax,cpuguaranteed)
 142             except OSError, ex:
 143                 if ex.errno == 22:
 144                     print "kernel does not support vserver scheduler"
 145                 else:
 146                     raise ex
 147
 148     def get_sched(self):
 149         # have no way of querying scheduler right now on a per vserver basis
 150         return (-1, False)
 151
 152     def set_memlimit(self, limit):
 153         ret = vserverimpl.setrlimit(self.ctx,5,limit)
 154         return ret
 155
 156     def get_memlimit(self):
 157         ret = vserverimpl.getrlimit(self.ctx,5)
 158         return ret
 159
 160     def set_tasklimit(self, limit):
 161         ret = vserverimpl.setrlimit(self.ctx,6,limit)
 162         return ret
 163
 164     def get_tasklimit(self):
 165         ret = vserverimpl.getrlimit(self.ctx,6)
 166         return ret
 167
 168     def set_bwlimit(self, eth, limit, cap, minrate, maxrate):
 169         if cap == "-1":
 170             bwlimit.off(self.ctx,eth)
 171         else:
 172             bwlimit.on(self.ctx, eth, limit, cap, minrate, maxrate)
 173
 174     def get_bwlimit(self, eth):
 175         # not implemented yet
 176         bwlimit = -1
 177         cap = "unknown"
 178         minrate = "unknown"
 179         maxrate = "unknown"
 180         return (bwlimit, cap, minrate, maxrate)
 181
 182     def open(self, filename, mode = "r", bufsize = -1):
 183
 184         (sendsock, recvsock) = passfdimpl.socketpair()
 185         child_pid = os.fork()
 186         if child_pid == 0:
 187             try:
 188                 # child process
 189                 self.__do_chroot()
 190                 f = open(filename, mode)
 191                 passfdimpl.sendmsg(f.fileno(), sendsock)
 192                 os._exit(0)
 193             except EnvironmentError, ex:
 194                 (result, errmsg) = (ex.errno, ex.strerror)
 195             except Exception, ex:
 196                 (result, errmsg) = (255, str(ex))
 197             os.write(sendsock, errmsg)
 198             os._exit(result)
 199
 200         # parent process
 201
 202         # XXX - need this since a lambda can't raise an exception
 203         def __throw(ex):
 204             raise ex
 205
 206         os.close(sendsock)
 207         throw = lambda : __throw(Exception(errmsg))
 208         while True:
 209             try:
 210                 (pid, status) = os.waitpid(child_pid, 0)
 211                 if os.WIFEXITED(status):
 212                     result = os.WEXITSTATUS(status)
 213                     if result != 255:
 214                         errmsg = os.strerror(result)
 215                         throw = lambda : __throw(IOError(result, errmsg))
 216                     else:
 217                         errmsg = "unexpected exception in child"
 218                 else:
 219                     result = -1
 220                     errmsg = "child killed"
 221                 break
 222             except OSError, ex:
 223                 if ex.errno != errno.EINTR:
 224                     os.close(recvsock)
 225                     raise ex
 226         fcntl.fcntl(recvsock, fcntl.F_SETFL, os.O_NONBLOCK)
 227         try:
 228             (fd, errmsg) = passfdimpl.recvmsg(recvsock)
 229         except OSError, ex:
 230             if ex.errno != errno.EAGAIN:
 231                 throw = lambda : __throw(ex)
 232             fd = 0
 233         os.close(recvsock)
 234         if not fd:
 235             throw()
 236
 237         return os.fdopen(fd, mode, bufsize)
 238
 239     def __do_chcontext(self, state_file = None):
 240
 241         vserverimpl.chcontext(self.ctx)
 242         if not state_file:
 243             return
 244         print >>state_file, "S_CONTEXT=%d" % self.ctx
 245         print >>state_file, "S_PROFILE=%s" % self.config.get("S_PROFILE", "")
 246         state_file.close()
 247
 248     def __prep(self, runlevel, log):
 249
 250         """ Perform all the crap that the vserver script does before
 251         actually executing the startup scripts. """
 252
 253         # remove /var/run and /var/lock/subsys files
 254         # but don't remove utmp from the top-level /var/run
 255         RUNDIR = "/var/run"
 256         LOCKDIR = "/var/lock/subsys"
 257         filter_fn = lambda fs: filter(lambda f: f != 'utmp', fs)
 258         garbage = reduce((lambda (out, ff), (dir, subdirs, files):
 259                           (out + map((dir + "/").__add__, ff(files)),
 260                            lambda fs: fs)),
 261                          list(os.walk(RUNDIR)),
 262                          ([], filter_fn))[0]
 263         garbage += filter(os.path.isfile, map((LOCKDIR + "/").__add__,
 264                                               os.listdir(LOCKDIR)))
 265         for f in garbage:
 266             os.unlink(f)
 267
 268         # set the initial runlevel
 269         f = open(RUNDIR + "/utmp", "w")
 270         utmp.set_runlevel(f, runlevel)
 271         f.close()
 272
 273         # mount /proc and /dev/pts
 274         self.__do_mount("none", "/proc", "proc")
 275         # XXX - magic mount options
 276         self.__do_mount("none", "/dev/pts", "devpts", 0, "gid=5,mode=0620")
 277
 278     def __do_mount(self, *mount_args):
 279
 280         try:
 281             mountimpl.mount(*mount_args)
 282         except OSError, ex:
 283             if ex.errno == errno.EBUSY:
 284                 # assume already mounted
 285                 return
 286             raise ex
 287
 288     def enter(self):
 289
 290         state_file = open("/var/run/vservers/%s.ctx" % self.name, "w")
 291         self.__do_chroot()
 292         self.__do_chcontext(state_file)
 293
 294     def start(self, wait, runlevel = 3):
 295
 296         child_pid = os.fork()
 297         if child_pid == 0:
 298             # child process
 299             try:
 300                 # get a new session
 301                 os.setsid()
 302
 303                 # open state file to record vserver info
 304                 state_file = open("/var/run/vservers/%s.ctx" % self.name, "w")
 305
 306                 # use /dev/null for stdin, /var/log/boot.log for stdout/err
 307                 os.close(0)
 308                 os.close(1)
 309                 os.open("/dev/null", os.O_RDONLY)
 310                 self.__do_chroot()
 311                 log = open("/var/log/boot.log", "w", 0)
 312                 os.dup2(1, 2)
 313
 314                 print >>log, ("%s: starting the virtual server %s" %
 315                               (time.asctime(time.gmtime()), self.name))
 316
 317                 # perform pre-init cleanup
 318                 self.__prep(runlevel, log)
 319
 320                 # execute each init script in turn
 321                 # XXX - we don't support all scripts that vserver script does
 322                 cmd_pid = 0
 323                 for cmd in self.INITSCRIPTS + [None]:
 324                     # wait for previous command to terminate, unless it
 325                     # is the last one and the caller has specified to wait
 326                     if cmd_pid and (cmd != None or wait):
 327                         try:
 328                             os.waitpid(cmd_pid, 0)
 329                         except:
 330                             print >>log, "error waiting for %s:" % cmd_pid
 331                             traceback.print_exc()
 332
 333                     # end of list
 334                     if cmd == None:
 335                         os._exit(0)
 336
 337                     # fork and exec next command
 338                     cmd_pid = os.fork()
 339                     if cmd_pid == 0:
 340                         try:
 341                             # enter vserver context
 342                             self.__do_chcontext(state_file)
 343                             arg_subst = { 'runlevel': runlevel }
 344                             cmd_args = [cmd[0]] + map(lambda x: x % arg_subst,
 345                                                       cmd[1:])
 346                             print >>log, "executing '%s'" % " ".join(cmd_args)
 347                             os.execl(cmd[0], *cmd_args)
 348                         except:
 349                             traceback.print_exc()
 350                             os._exit(1)
 351                     else:
 352                         # don't want to write state_file multiple times
 353                         state_file = None
 354
 355             # we get here due to an exception in the top-level child process
 356             except Exception, ex:
 357                 traceback.print_exc()
 358             os._exit(0)
 359
 360         # parent process
 361         return child_pid