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 # default values for new vserver scheduler
  32 SCHED_TOKENS_MIN = 50
  33 SCHED_TOKENS_MAX = 100
  34 SCHED_TOKENS = 100
  35 SCHED_INTERVAL = 1000
  36
  37
  38 class VServer:
  39
  40     INITSCRIPTS = [('/etc/rc.vinit', 'start'),
  41                    ('/etc/rc.d/rc', '%(runlevel)d')]
  42
  43     def __init__(self, name):
  44
  45         self.name = name
  46         self.config_file = "/etc/vservers/%s.conf" % name
  47         self.dir = "%s/%s" % (vserverimpl.VSERVER_BASEDIR, name)
  48         if not (os.path.isdir(self.dir) and
  49                 os.access(self.dir, os.R_OK | os.W_OK | os.X_OK)):
  50             raise Exception, "no such vserver: " + name
  51         self.config = self.__read_config_file("/etc/vservers.conf")
  52         self.config.update(self.__read_config_file(self.config_file))
  53         self.flags = 0
  54         flags = self.config["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 = int(self.config["S_CONTEXT"])
  61
  62     def __str__(self):
  63
  64         return self.name
  65
  66     config_var_re = re.compile(r"^ *([A-Z_]+)=(.*)\n?$", re.MULTILINE)
  67
  68     def __read_config_file(self, filename):
  69
  70         f = open(filename, "r")
  71         data = f.read()
  72         f.close()
  73         config = {}
  74         for m in self.config_var_re.finditer(data):
  75             (key, val) = m.groups()
  76             config[key] = val.strip('"')
  77         return config
  78
  79     def __update_config_file(self, filename, newvars):
  80
  81         # read old file, apply changes
  82         f = open(filename, "r")
  83         data = f.read()
  84         f.close()
  85         todo = newvars.copy()
  86         changed = False
  87         for m in self.config_var_re.finditer(data):
  88             (key, val) = m.groups()
  89             newval = todo.pop(key, None)
  90             if newval != None:
  91                 data = data[:m.start(2)] + newval + data[m.end(2):]
  92                 changed = True
  93         for (newkey, newval) in todo.items():
  94             data += "%s=%s\n" % (newkey, newval)
  95             changed = True
  96
  97         if not changed:
  98             return
  99
 100         # write new file
 101         newfile = filename + ".new"
 102         f = open(newfile, "w")
 103         f.write(data)
 104         f.close()
 105
 106         # 'copy' original file, rename new to original
 107         os.link(filename, filename + ".old")
 108         os.rename(newfile, filename)
 109
 110     def __do_chroot(self):
 111
 112         return os.chroot(self.dir)
 113
 114     def set_disklimit(self, block_limit):
 115
 116         # block_limit is in kB, get_disk_usage() must have been called
 117         if self.disk_usage_set:
 118             block_usage = vserverimpl.DLIMIT_KEEP
 119             inode_usage = vserverimpl.DLIMIT_KEEP
 120         else:
 121             block_usage = self.disk_blocks
 122             inode_usage = self.disk_inodes
 123             if block_limit < block_usage:
 124                 raise Exception, ("%s disk usage (%u blocks) > limit (%u)" %
 125                                   (self.name, block_usage, block_limit))
 126             self.disk_usage_set = True
 127
 128         vserverimpl.setdlimit(self.dir,
 129                               self.ctx,
 130                               block_usage,
 131                               block_limit,
 132                               inode_usage,
 133                               -1,  # 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                 # get here if no vserver disk limit has been set for xid
 144                 # set blockused to -1 to indicate no limit
 145                 blocktotal = -1
 146
 147         return blocktotal
 148
 149     def set_sched(self, shares = 32, besteffort = True):
 150         # for the old CKRM scheduler
 151         if cpulimit.checkckrm() is True:
 152             cpulimit.cpuinit()
 153             cpulimit.vs2ckrm_on(self.name)
 154             try:
 155                 cpulimit.cpulimit(self.name,shares)
 156             except OSError, ex:
 157                 if ex.errno == 22:
 158                     print "invalid shares argument"
 159                     # should re-raise exception?!
 160
 161         # for the new vserver scheduler
 162         else:
 163             global SCHED_TOKENS_MIN, SCHED_TOKENS_MAX, SCHED_TOKENS, SCHED_INTERVAL
 164             tokensmin = SCHED_TOKENS_MIN
 165             tokensmax = SCHED_TOKENS_MAX
 166             tokens    = SCHED_TOKENS
 167             interval  = SCHED_INTERVAL
 168             fillrate = shares
 169
 170             if besteffort is True:
 171                 cpuguaranteed = 0
 172             else:
 173                 cpuguaranteed = 1
 174
 175             try:
 176                 vserverimpl.setsched(self.ctx,fillrate,interval,tokens,tokensmin,tokensmax,cpuguaranteed)
 177             except OSError, ex:
 178                 if ex.errno == 22:
 179                     print "kernel does not support vserver scheduler"
 180                 else:
 181                     raise ex
 182
 183     def get_sched(self):
 184         # have no way of querying scheduler right now on a per vserver basis
 185         return (-1, False)
 186
 187     def set_memlimit(self, limit):
 188         ret = vserverimpl.setrlimit(self.ctx,5,limit)
 189         return ret
 190
 191     def get_memlimit(self):
 192         ret = vserverimpl.getrlimit(self.ctx,5)
 193         return ret
 194
 195     def set_tasklimit(self, limit):
 196         ret = vserverimpl.setrlimit(self.ctx,6,limit)
 197         return ret
 198
 199     def get_tasklimit(self):
 200         ret = vserverimpl.getrlimit(self.ctx,6)
 201         return ret
 202
 203     def set_bwlimit(self, eth, limit, cap, minrate, maxrate):
 204         if cap == "-1":
 205             bwlimit.off(self.ctx,eth)
 206         else:
 207             bwlimit.on(self.ctx, eth, limit, cap, minrate, maxrate)
 208
 209     def get_bwlimit(self, eth):
 210         # not implemented yet
 211         bwlimit = -1
 212         cap = "unknown"
 213         minrate = "unknown"
 214         maxrate = "unknown"
 215         return (bwlimit, cap, minrate, maxrate)
 216
 217     def open(self, filename, mode = "r", bufsize = -1):
 218
 219         (sendsock, recvsock) = passfdimpl.socketpair()
 220         child_pid = os.fork()
 221         if child_pid == 0:
 222             try:
 223                 # child process
 224                 self.__do_chroot()
 225                 f = open(filename, mode)
 226                 passfdimpl.sendmsg(f.fileno(), sendsock)
 227                 os._exit(0)
 228             except EnvironmentError, ex:
 229                 (result, errmsg) = (ex.errno, ex.strerror)
 230             except Exception, ex:
 231                 (result, errmsg) = (255, str(ex))
 232             os.write(sendsock, errmsg)
 233             os._exit(result)
 234
 235         # parent process
 236
 237         # XXX - need this since a lambda can't raise an exception
 238         def __throw(ex):
 239             raise ex
 240
 241         os.close(sendsock)
 242         throw = lambda : __throw(Exception(errmsg))
 243         while True:
 244             try:
 245                 (pid, status) = os.waitpid(child_pid, 0)
 246                 if os.WIFEXITED(status):
 247                     result = os.WEXITSTATUS(status)
 248                     if result != 255:
 249                         errmsg = os.strerror(result)
 250                         throw = lambda : __throw(IOError(result, errmsg))
 251                     else:
 252                         errmsg = "unexpected exception in child"
 253                 else:
 254                     result = -1
 255                     errmsg = "child killed"
 256                 break
 257             except OSError, ex:
 258                 if ex.errno != errno.EINTR:
 259                     os.close(recvsock)
 260                     raise ex
 261         fcntl.fcntl(recvsock, fcntl.F_SETFL, os.O_NONBLOCK)
 262         try:
 263             (fd, errmsg) = passfdimpl.recvmsg(recvsock)
 264         except OSError, ex:
 265             if ex.errno != errno.EAGAIN:
 266                 throw = lambda : __throw(ex)
 267             fd = 0
 268         os.close(recvsock)
 269         if not fd:
 270             throw()
 271
 272         return os.fdopen(fd, mode, bufsize)
 273
 274     def __do_chcontext(self, state_file = None):
 275
 276         vserverimpl.chcontext(self.ctx)
 277
 278         if not state_file:
 279             return
 280         print >>state_file, "S_CONTEXT=%d" % self.ctx
 281         print >>state_file, "S_PROFILE=%s" % self.config.get("S_PROFILE", "")
 282         state_file.close()
 283
 284     def __prep(self, runlevel, log):
 285
 286         """ Perform all the crap that the vserver script does before
 287         actually executing the startup scripts. """
 288
 289         # remove /var/run and /var/lock/subsys files
 290         # but don't remove utmp from the top-level /var/run
 291         RUNDIR = "/var/run"
 292         LOCKDIR = "/var/lock/subsys"
 293         filter_fn = lambda fs: filter(lambda f: f != 'utmp', fs)
 294         garbage = reduce((lambda (out, ff), (dir, subdirs, files):
 295                           (out + map((dir + "/").__add__, ff(files)),
 296                            lambda fs: fs)),
 297                          list(os.walk(RUNDIR)),
 298                          ([], filter_fn))[0]
 299         garbage += filter(os.path.isfile, map((LOCKDIR + "/").__add__,
 300                                               os.listdir(LOCKDIR)))
 301         for f in garbage:
 302             os.unlink(f)
 303
 304         # set the initial runlevel
 305         f = open(RUNDIR + "/utmp", "w")
 306         utmp.set_runlevel(f, runlevel)
 307         f.close()
 308
 309         # mount /proc and /dev/pts
 310         self.__do_mount("none", "/proc", "proc")
 311         # XXX - magic mount options
 312         self.__do_mount("none", "/dev/pts", "devpts", 0, "gid=5,mode=0620")
 313
 314     def __do_mount(self, *mount_args):
 315
 316         try:
 317             mountimpl.mount(*mount_args)
 318         except OSError, ex:
 319             if ex.errno == errno.EBUSY:
 320                 # assume already mounted
 321                 return
 322             raise ex
 323
 324     def enter(self):
 325
 326         state_file = open("/var/run/vservers/%s.ctx" % self.name, "w")
 327         self.__do_chroot()
 328         self.__do_chcontext(state_file)
 329
 330     def start(self, wait, runlevel = 3):
 331
 332         child_pid = os.fork()
 333         if child_pid == 0:
 334             # child process
 335             try:
 336                 # get a new session
 337                 os.setsid()
 338
 339                 # open state file to record vserver info
 340                 state_file = open("/var/run/vservers/%s.ctx" % self.name, "w")
 341
 342                 # use /dev/null for stdin, /var/log/boot.log for stdout/err
 343                 os.close(0)
 344                 os.close(1)
 345                 os.open("/dev/null", os.O_RDONLY)
 346                 self.__do_chroot()
 347                 log = open("/var/log/boot.log", "w", 0)
 348                 os.dup2(1, 2)
 349
 350                 print >>log, ("%s: starting the virtual server %s" %
 351                               (time.asctime(time.gmtime()), self.name))
 352
 353                 # perform pre-init cleanup
 354                 self.__prep(runlevel, log)
 355
 356                 # execute each init script in turn
 357                 # XXX - we don't support all scripts that vserver script does
 358                 cmd_pid = 0
 359                 for cmd in self.INITSCRIPTS + [None]:
 360                     # wait for previous command to terminate, unless it
 361                     # is the last one and the caller has specified to wait
 362                     if cmd_pid and (cmd != None or wait):
 363                         try:
 364                             os.waitpid(cmd_pid, 0)
 365                         except:
 366                             print >>log, "error waiting for %s:" % cmd_pid
 367                             traceback.print_exc()
 368
 369                     # end of list
 370                     if cmd == None:
 371                         os._exit(0)
 372
 373                     # fork and exec next command
 374                     cmd_pid = os.fork()
 375                     if cmd_pid == 0:
 376                         try:
 377                             # enter vserver context
 378                             self.__do_chcontext(state_file)
 379                             arg_subst = { 'runlevel': runlevel }
 380                             cmd_args = [cmd[0]] + map(lambda x: x % arg_subst,
 381                                                       cmd[1:])
 382                             print >>log, "executing '%s'" % " ".join(cmd_args)
 383                             os.execl(cmd[0], *cmd_args)
 384                         except:
 385                             traceback.print_exc()
 386                             os._exit(1)
 387                     else:
 388                         # don't want to write state_file multiple times
 389                         state_file = None
 390
 391             # we get here due to an exception in the top-level child process
 392             except Exception, ex:
 393                 traceback.print_exc()
 394             os._exit(0)
 395
 396         # parent process
 397         return child_pid
 398
 399     def update_resources(self, resources):
 400
 401         # write new values to configuration file
 402         self.__update_config_file(self.config_file, resources)
 403
 404         #
 405         # Figure out if any processes are active in context, apply new
 406         # values if there are.
 407         #
 408
 409     def init_disk_info(self):
 410
 411         (self.disk_inodes, self.disk_blocks, size) = vduimpl.vdu(self.dir)
 412         self.disk_usage_set = False
 413
 414         return size