import fcntl
import os
import re
+import pwd
+import signal
import sys
import time
import traceback
import mountimpl
-import passfdimpl
+import runcmd
import utmp
import vserverimpl, vduimpl
import cpulimit, bwlimit
+from vserverimpl import VS_SCHED_CPU_GUARANTEED as SCHED_CPU_GUARANTEED
+from vserverimpl import DLIMIT_INF
+
#
FLAGS_ULIMIT = 64
FLAGS_NAMESPACE = 128
-# default values for new vserver scheduler
-SCHED_TOKENS_MIN = 50
-SCHED_TOKENS_MAX = 100
-SCHED_TOKENS = 100
-SCHED_INTERVAL = 1000
-
+
+class NoSuchVServer(Exception): pass
+
+
+
class VServer:
INITSCRIPTS = [('/etc/rc.vinit', 'start'),
('/etc/rc.d/rc', '%(runlevel)d')]
- def __init__(self, name):
+ def __init__(self, name, vm_id = None, vm_running = False):
self.name = name
self.config_file = "/etc/vservers/%s.conf" % name
self.dir = "%s/%s" % (vserverimpl.VSERVER_BASEDIR, name)
if not (os.path.isdir(self.dir) and
os.access(self.dir, os.R_OK | os.W_OK | os.X_OK)):
- raise Exception, "no such vserver: " + name
- self.config = self.__read_config_file("/etc/vservers.conf")
- self.config.update(self.__read_config_file(self.config_file))
- self.flags = 0
- flags = self.config["S_FLAGS"].split(" ")
- if "lock" in flags:
- self.flags |= FLAGS_LOCK
- if "nproc" in flags:
- self.flags |= FLAGS_NPROC
+ raise NoSuchVServer, "no such vserver: " + name
+ self.config = {}
+ for config_file in ["/etc/vservers.conf", self.config_file]:
+ try:
+ self.config.update(self.__read_config_file(config_file))
+ except IOError, ex:
+ if ex.errno != errno.ENOENT:
+ raise
self.remove_caps = ~vserverimpl.CAP_SAFE;
- self.ctx = int(self.config["S_CONTEXT"])
+ if vm_id == None:
+ vm_id = int(self.config['S_CONTEXT'])
+ self.ctx = vm_id
+ self.vm_running = vm_running
config_var_re = re.compile(r"^ *([A-Z_]+)=(.*)\n?$", re.MULTILINE)
f.close()
todo = newvars.copy()
changed = False
+ offset = 0
for m in self.config_var_re.finditer(data):
(key, val) = m.groups()
newval = todo.pop(key, None)
if newval != None:
- data = data[:m.start(2)] + newval + data[m.end(2):]
+ data = data[:offset+m.start(2)] + str(newval) + data[offset+m.end(2):]
+ offset += len(str(newval)) - (m.end(2)-m.start(2))
changed = True
for (newkey, newval) in todo.items():
data += "%s=%s\n" % (newkey, newval)
f.write(data)
f.close()
- # 'copy' original file, rename new to original
- backup = filename + ".old"
- try:
- os.unlink(backup)
- except OSError, ex:
- if ex.errno != errno.ENOENT:
- raise
- os.link(filename, backup)
+ # replace old file with new
os.rename(newfile, filename)
def __do_chroot(self):
- return os.chroot(self.dir)
+ os.chroot(self.dir)
+ os.chdir("/")
+
+ def chroot_call(self, fn, *args):
+
+ cwd_fd = os.open(".", os.O_RDONLY)
+ try:
+ root_fd = os.open("/", os.O_RDONLY)
+ try:
+ self.__do_chroot()
+ result = fn(*args)
+ finally:
+ os.fchdir(root_fd)
+ os.chroot(".")
+ os.fchdir(cwd_fd)
+ os.close(root_fd)
+ finally:
+ os.close(cwd_fd)
+ return result
def set_disklimit(self, block_limit):
- # block_limit is in kB, get_disk_usage() must have been called
- if self.disk_usage_set:
+ # block_limit is in kB
+ if block_limit == 0:
+ vserverimpl.unsetdlimit(self.dir, self.ctx)
+ return
+
+ if self.vm_running:
block_usage = vserverimpl.DLIMIT_KEEP
inode_usage = vserverimpl.DLIMIT_KEEP
else:
+ # init_disk_info() must have been called to get usage values
block_usage = self.disk_blocks
inode_usage = self.disk_inodes
- if block_limit < block_usage:
- raise Exception, ("%s disk usage (%u blocks) > limit (%u)" %
- (self.name, block_usage, block_limit))
- self.disk_usage_set = True
vserverimpl.setdlimit(self.dir,
self.ctx,
block_usage,
block_limit,
inode_usage,
- -1, # inode limit
+ vserverimpl.DLIMIT_INF, # inode limit
2) # %age reserved for root
def get_disklimit(self):
try:
- blocksused, blocktotal, inodesused, inodestotal, reserved = \
- vserverimpl.getdlimit(self.dir, self.ctx)
+ (self.disk_blocks, block_limit, self.disk_inodes, inode_limit,
+ reserved) = vserverimpl.getdlimit(self.dir, self.ctx)
except OSError, ex:
- if ex.errno == errno.ESRCH:
- # get here if no vserver disk limit has been set for xid
- # set blockused to -1 to indicate no limit
- blocktotal = -1
-
- return blocktotal
-
- def set_sched(self, shares = 32, besteffort = True):
- # for the old CKRM scheduler
- if cpulimit.checkckrm() is True:
- cpulimit.cpuinit()
- cpulimit.vs2ckrm_on(self.name)
- try:
- cpulimit.cpulimit(self.name,shares)
- except OSError, ex:
- if ex.errno == 22:
- print "invalid shares argument"
- # should re-raise exception?!
+ if ex.errno != errno.ESRCH:
+ raise
+ # get here if no vserver disk limit has been set for xid
+ block_limit = -1
- # for the new vserver scheduler
- else:
- global SCHED_TOKENS_MIN, SCHED_TOKENS_MAX, SCHED_TOKENS, SCHED_INTERVAL
- tokensmin = SCHED_TOKENS_MIN
- tokensmax = SCHED_TOKENS_MAX
- tokens = SCHED_TOKENS
- interval = SCHED_INTERVAL
- fillrate = shares
-
- if besteffort is True:
- cpuguaranteed = 0
- else:
- cpuguaranteed = 1
+ return block_limit
- try:
- vserverimpl.setsched(self.ctx,fillrate,interval,tokens,tokensmin,tokensmax,cpuguaranteed)
- except OSError, ex:
- if ex.errno == 22:
- print "kernel does not support vserver scheduler"
- else:
- raise ex
+ def set_sched_config(self, cpu_share, sched_flags):
+
+ """ Write current CPU scheduler parameters to the vserver
+ configuration file. This method does not modify the kernel CPU
+ scheduling parameters for this context. """
+
+ if cpu_share == int(self.config.get("CPULIMIT", -1)):
+ return
+ cpu_guaranteed = sched_flags & SCHED_CPU_GUARANTEED
+ cpu_config = { "CPULIMIT": cpu_share, "CPUGUARANTEED": cpu_guaranteed }
+ self.update_resources(cpu_config)
+ if self.vm_running:
+ self.set_sched(cpu_share, sched_flags)
+
+ def set_sched(self, cpu_share, sched_flags = 0):
+
+ """ Update kernel CPU scheduling parameters for this context. """
+
+ vserverimpl.setsched(self.ctx, cpu_share, sched_flags)
def get_sched(self):
# have no way of querying scheduler right now on a per vserver basis
ret = vserverimpl.getrlimit(self.ctx,6)
return ret
- def set_bwlimit(self, eth, limit, cap, minrate, maxrate):
- if cap == "-1":
- bwlimit.off(self.ctx,eth)
- else:
- bwlimit.on(self.ctx, eth, limit, cap, minrate, maxrate)
-
- def get_bwlimit(self, eth):
- # not implemented yet
- bwlimit = -1
- cap = "unknown"
- minrate = "unknown"
- maxrate = "unknown"
- return (bwlimit, cap, minrate, maxrate)
-
- def open(self, filename, mode = "r", bufsize = -1):
+ def set_bwlimit(self, minrate = bwlimit.bwmin, maxrate = None,
+ exempt_min = None, exempt_max = None,
+ share = None, dev = "eth0"):
- (sendsock, recvsock) = passfdimpl.socketpair()
- child_pid = os.fork()
- if child_pid == 0:
- try:
- # child process
- self.__do_chroot()
- f = open(filename, mode)
- passfdimpl.sendmsg(f.fileno(), sendsock)
- os._exit(0)
- except EnvironmentError, ex:
- (result, errmsg) = (ex.errno, ex.strerror)
- except Exception, ex:
- (result, errmsg) = (255, str(ex))
- os.write(sendsock, errmsg)
- os._exit(result)
+ if minrate is None:
+ bwlimit.off(self.ctx, dev)
+ else:
+ bwlimit.on(self.ctx, dev, share,
+ minrate, maxrate, exempt_min, exempt_max)
- # parent process
+ def get_bwlimit(self, dev = "eth0"):
- # XXX - need this since a lambda can't raise an exception
- def __throw(ex):
- raise ex
+ result = bwlimit.get(self.ctx)
+ # result of bwlimit.get is (ctx, share, minrate, maxrate)
+ if result:
+ result = result[1:]
+ return result
- os.close(sendsock)
- throw = lambda : __throw(Exception(errmsg))
- while True:
- try:
- (pid, status) = os.waitpid(child_pid, 0)
- if os.WIFEXITED(status):
- result = os.WEXITSTATUS(status)
- if result != 255:
- errmsg = os.strerror(result)
- throw = lambda : __throw(IOError(result, errmsg))
- else:
- errmsg = "unexpected exception in child"
- else:
- result = -1
- errmsg = "child killed"
- break
- except OSError, ex:
- if ex.errno != errno.EINTR:
- os.close(recvsock)
- raise ex
- fcntl.fcntl(recvsock, fcntl.F_SETFL, os.O_NONBLOCK)
- try:
- (fd, errmsg) = passfdimpl.recvmsg(recvsock)
- except OSError, ex:
- if ex.errno != errno.EAGAIN:
- throw = lambda : __throw(ex)
- fd = 0
- os.close(recvsock)
- if not fd:
- throw()
+ def open(self, filename, mode = "r", bufsize = -1):
- return os.fdopen(fd, mode, bufsize)
+ return self.chroot_call(open, filename, mode, bufsize)
- def __do_chcontext(self, state_file = None):
+ def __do_chcontext(self, state_file):
- vserverimpl.chcontext(self.ctx)
+ if state_file:
+ print >>state_file, "S_CONTEXT=%u" % self.ctx
+ print >>state_file, "S_PROFILE="
+ state_file.close()
- if not state_file:
- return
- print >>state_file, "S_CONTEXT=%d" % self.ctx
- print >>state_file, "S_PROFILE=%s" % self.config.get("S_PROFILE", "")
- state_file.close()
+ if vserverimpl.chcontext(self.ctx):
+ self.set_resources()
+ vserverimpl.setup_done(self.ctx)
def __prep(self, runlevel, log):
def start(self, wait, runlevel = 3):
- # XXX - temporary hack
- self.set_disklimit(int(self.config.get("DISKLIMIT", 5000000)))
+ self.vm_running = True
child_pid = os.fork()
if child_pid == 0:
self.__do_chroot()
log = open("/var/log/boot.log", "w", 0)
os.dup2(1, 2)
- # XXX - close all other fds
print >>log, ("%s: starting the virtual server %s" %
(time.asctime(time.gmtime()), self.name))
# parent process
return child_pid
+ def set_resources(self):
+
+ """ Called when vserver context is entered for first time,
+ should be overridden by subclass. """
+
+ pass
+
def update_resources(self, resources):
self.config.update(resources)
# write new values to configuration file
self.__update_config_file(self.config_file, resources)
- # disklimit can be applied without a process in context
- disklimit = resources.get("DISKLIMIT", 0)
- if disklimit:
- self.set_disklimit(disklimit)
-
- #
- # Figure out if any processes are active in context, apply new
- # values if there are.
- #
-
def init_disk_info(self):
(self.disk_inodes, self.disk_blocks, size) = vduimpl.vdu(self.dir)
- self.disk_usage_set = False
return size
+
+ def stop(self, signal = signal.SIGKILL):
+
+ vserverimpl.killall(self.ctx, signal)
+ self.vm_running = False
+
+
+
+def create(vm_name, static = False, ctor = VServer):
+
+ options = []
+ if static:
+ options += ['--static']
+ runcmd.run('vuseradd', options + [vm_name])
+ vm_id = pwd.getpwnam(vm_name)[2]
+
+ return ctor(vm_name, vm_id)