X-Git-Url: http://git.onelab.eu/?p=util-vserver.git;a=blobdiff_plain;f=python%2Fvserver.py;h=ff8752200c503a8f3262d9ef849738da5ecfeaba;hp=d77abbab5d60130617bdb33f1ad2e604ca13bd1c;hb=ec4370f7ebd7fb0ce7f002f5bf2c74f03acd3ec1;hpb=c67a2a042e09fd8799563d9b877efce6629a86eb diff --git a/python/vserver.py b/python/vserver.py index d77abba..ff87522 100644 --- a/python/vserver.py +++ b/python/vserver.py @@ -4,17 +4,35 @@ import errno 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 vserverimpl import cpulimit, bwlimit - +from vserverimpl import VS_SCHED_CPU_GUARANTEED as SCHED_CPU_GUARANTEED +from vserverimpl import DLIMIT_INF +from vserverimpl import VC_LIM_KEEP + +from vserverimpl import RLIMIT_CPU +from vserverimpl import RLIMIT_RSS +from vserverimpl import RLIMIT_NPROC +from vserverimpl import RLIMIT_NOFILE +from vserverimpl import RLIMIT_MEMLOCK +from vserverimpl import RLIMIT_AS +from vserverimpl import RLIMIT_LOCKS +from vserverimpl import RLIMIT_SIGPENDING +from vserverimpl import RLIMIT_MSGQUEUE +from vserverimpl import VLIMIT_NSOCK +from vserverimpl import VLIMIT_OPENFD +from vserverimpl import VLIMIT_ANON +from vserverimpl import VLIMIT_SHMEM # # these are the flags taken from the kernel linux/vserver/legacy.h @@ -28,256 +46,274 @@ FLAGS_HIDEINFO = 32 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 +RLIMITS = {"CPU": RLIMIT_CPU, + "RSS": RLIMIT_RSS, + "NPROC": RLIMIT_NPROC, + "NOFILE": RLIMIT_NOFILE, + "MEMLOCK": RLIMIT_MEMLOCK, + "AS": RLIMIT_AS, + "LOCKS": RLIMIT_LOCKS, + "SIGPENDING": RLIMIT_SIGPENDING, + "MSGQUEUE": RLIMIT_MSGQUEUE, + "NSOCK": VLIMIT_NSOCK, + "OPENFD": VLIMIT_OPENFD, + "ANON": VLIMIT_ANON, + "SHMEM": VLIMIT_SHMEM} + +class NoSuchVServer(Exception): pass + + +class VServerConfig: + mapping = { + 'S_CONTEXT': 'context', + 'VS_WHITELISTED': 'whitelisted', + 'CPULIMIT': 'sched/fill-rate2', + 'CPUGUARANTEED': 'sched/fill-rate', + 'VS_DISK_MAX': 'dlimits/0/space_total', + } + + def __init__(self, name, directory): + global RLIMITS + for i in RLIMITS.keys(): + for j in ('HARD', 'SOFT', 'MINIMUM'): + self.mapping['VS_%s_%s' % (i, j)] = 'rlimits/%s.%s' % (i.lower(), j.replace('MINIMUM', 'min').lower()) + self.name = name + self.dir = directory + + def get(self, option, default = None): + try: + filename = self.mapping[option] + f = open(os.path.join(self.dir, filename), "r") + buf = f.readline().rstrip() + f.close() + return buf + except KeyError, e: + # No mapping exists for this option + raise e + except IOError, e: + if default is not None: + return default + else: + raise KeyError, "Key %s is not set for %s" % (option, self.name) + + def update(self, vars): + for (option, value) in vars.iteritems(): + try: + old_umask = os.umask(0022) + filename = os.path.join(self.dir, self.mapping[option]) + try: + os.makedirs(os.path.dirname(filename), 0755) + except: + pass + f = open(filename, 'w') + f.write("%s\n" % value) + f.close() + os.umask(old_umask) + except KeyError, e: + raise KeyError, "Don't know how to handle %s, sorry" % option + - 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.rlimits_changed = False 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 = VServerConfig(name, "/etc/vservers/%s" % name) self.remove_caps = ~vserverimpl.CAP_SAFE; - self.ctx = int(self.config["S_CONTEXT"]) - - config_var_re = re.compile(r"^ *([A-Z_]+)=(.*)\n?$", re.MULTILINE) + if vm_id == None: + vm_id = int(self.config.get('S_CONTEXT')) + self.ctx = vm_id + self.vm_running = vm_running - def __read_config_file(self, filename): + def have_limits_changed(self): + return self.rlimits_changed - f = open(filename, "r") - data = f.read() - f.close() - config = {} - for m in self.config_var_re.finditer(data): - (key, val) = m.groups() - config[key] = val.strip('"') - return config - - def __update_config_file(self, filename, newvars): - - # read old file, apply changes - f = open(filename, "r") - data = f.read() - f.close() - todo = newvars.copy() + def set_rlimit_limit(self,type,hard,soft,minimum): + """Generic set resource limit function for vserver""" + global RLIMITS changed = False - 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):] - changed = True - for (newkey, newval) in todo.items(): - data += "%s=%s\n" % (newkey, newval) - changed = True - - if not changed: - return + try: + old_hard, old_soft, old_minimum = self.get_rlimit_limit(type) + if old_hard != VC_LIM_KEEP and old_hard <> hard: changed = True + if old_soft != VC_LIM_KEEP and old_soft <> soft: changed = True + if old_minimum != VC_LIM_KEEP and old_minimum <> minimum: changed = True + self.rlimits_changed = self.rlimits_changed or changed + except OSError, e: + if self.is_running(): print "Unexpected error with getrlimit for running context %d" % self.ctx + + resource_type = RLIMITS[type] + try: + ret = vserverimpl.setrlimit(self.ctx,resource_type,hard,soft,minimum) + except OSError, e: + if self.is_running(): print "Unexpected error with setrlimit for running context %d" % self.ctx + + def set_rlimit_config(self,type,hard,soft,minimum): + """Generic set resource limit function for vserver""" + resources = {} + if hard <> VC_LIM_KEEP: + resources["VS_%s_HARD"%type] = hard + if soft <> VC_LIM_KEEP: + resources["VS_%s_SOFT"%type] = soft + if minimum <> VC_LIM_KEEP: + resources["VS_%s_MINIMUM"%type] = minimum + if len(resources)>0: + self.update_resources(resources) + self.set_rlimit_limit(type,hard,soft,minimum) + + def get_rlimit_limit(self,type): + """Generic get resource configuration function for vserver""" + global RLIMITS + resource_type = RLIMITS[type] + try: + ret = vserverimpl.getrlimit(self.ctx,resource_type) + except OSError, e: + print "Unexpected error with getrlimit for context %d" % self.ctx + ret = self.get_rlimit_config(type) + return ret - # write new file - newfile = filename + ".new" - f = open(newfile, "w") - f.write(data) - f.close() + def get_rlimit_config(self,type): + """Generic get resource configuration function for vserver""" + hard = int(self.config.get("VS_%s_HARD"%type,VC_LIM_KEEP)) + soft = int(self.config.get("VS_%s_SOFT"%type,VC_LIM_KEEP)) + minimum = int(self.config.get("VS_%s_MINIMUM"%type,VC_LIM_KEEP)) + return (hard,soft,minimum) - # 'copy' original file, rename new to original - backup = filename + ".old" - os.unlink(backup) - os.link(filename, backup) - os.rename(newfile, filename) + def set_WHITELISTED_config(self,whitelisted): + resources = {'VS_WHITELISTED': whitelisted} + self.update_resources(resources) 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 + if block_limit == 0: + try: + vserverimpl.unsetdlimit(self.dir, self.ctx) + except OSError, e: + print "Unexpected error with unsetdlimit for context %d" % self.ctx + return - # block_limit is in kB, get_disk_usage() must have been called - if self.disk_usage_set: + 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 - 2) # %age reserved for root + + try: + vserverimpl.setdlimit(self.dir, + self.ctx, + block_usage, + block_limit, + inode_usage, + vserverimpl.DLIMIT_INF, # inode limit + 2) # %age reserved for root + except OSError, e: + print "Unexpected error with setdlimit for context %d" % self.ctx + + + resources = {'VS_DISK_MAX': block_limit} + self.update_resources(resources) + + def is_running(self): + return vserverimpl.isrunning(self.ctx) + 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 return (-1, False) - def set_memlimit(self, limit): - ret = vserverimpl.setrlimit(self.ctx,5,limit) - return ret + def set_bwlimit(self, minrate = bwlimit.bwmin, maxrate = None, + exempt_min = None, exempt_max = None, + share = None, dev = "eth0"): - def get_memlimit(self): - ret = vserverimpl.getrlimit(self.ctx,5) - return ret - - def set_tasklimit(self, limit): - ret = vserverimpl.setrlimit(self.ctx,6,limit) - return ret - - def get_tasklimit(self): - 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) + if minrate is None: + bwlimit.off(self.ctx, dev) 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): + bwlimit.on(self.ctx, dev, share, + minrate, maxrate, exempt_min, exempt_max) - (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) + def get_bwlimit(self, dev = "eth0"): - # parent process + result = bwlimit.get(self.ctx) + # result of bwlimit.get is (ctx, share, minrate, maxrate) + if result: + result = result[1:] + return result - # XXX - need this since a lambda can't raise an exception - def __throw(ex): - raise ex - - 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, "%u" % self.ctx + 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): @@ -320,15 +356,12 @@ class VServer: raise ex def enter(self): - - state_file = open("/var/run/vservers/%s.ctx" % self.name, "w") self.__do_chroot() - self.__do_chcontext(state_file) + self.__do_chcontext(None) def start(self, wait, runlevel = 3): - - # XXX - temporary hack - self.set_disklimit(self.config.get("DISKLIMIT", 5000000)) + self.vm_running = True + self.rlimits_changed = False child_pid = os.fork() if child_pid == 0: @@ -338,7 +371,7 @@ class VServer: os.setsid() # open state file to record vserver info - state_file = open("/var/run/vservers/%s.ctx" % self.name, "w") + state_file = open("/var/run/vservers/%s" % self.name, "w") # use /dev/null for stdin, /var/log/boot.log for stdout/err os.close(0) @@ -347,7 +380,6 @@ class VServer: 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)) @@ -357,39 +389,18 @@ class VServer: # execute each init script in turn # XXX - we don't support all scripts that vserver script does - cmd_pid = 0 - first_child = True - for cmd in self.INITSCRIPTS + [None]: - # wait for previous command to terminate, unless it - # is the last one and the caller has specified to wait - if cmd_pid and (cmd != None or wait): - try: - os.waitpid(cmd_pid, 0) - except: - print >>log, "error waiting for %s:" % cmd_pid - traceback.print_exc() - - # end of list - if cmd == None: - os._exit(0) - - # fork and exec next command - cmd_pid = os.fork() - if cmd_pid == 0: - try: - # enter vserver context - self.__do_chcontext(state_file) - arg_subst = { 'runlevel': runlevel } - cmd_args = [cmd[0]] + map(lambda x: x % arg_subst, - cmd[1:]) - print >>log, "executing '%s'" % " ".join(cmd_args) - os.execl(cmd[0], *cmd_args) - except: - traceback.print_exc() - os._exit(1) - else: - # don't want to write state_file multiple times - state_file = None + self.__do_chcontext(state_file) + for cmd in self.INITSCRIPTS + [None]: + try: + # enter vserver context + arg_subst = { 'runlevel': runlevel } + cmd_args = [cmd[0]] + map(lambda x: x % arg_subst, + cmd[1:]) + print >>log, "executing '%s'" % " ".join(cmd_args) + os.spawnvp(os.P_WAIT,cmd[0],*cmd_args) + except: + traceback.print_exc() + os._exit(1) # we get here due to an exception in the top-level child process except Exception, ex: @@ -399,26 +410,44 @@ class VServer: # 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) + def init_disk_info(self): + cmd = "/usr/sbin/vdu --script --space --inodes --blocksize 1024 --xid %d %s" % (self.ctx, self.dir) + (child_stdin, child_stdout, child_stderr) = os.popen3(cmd) + child_stdin.close() + line = child_stdout.readline() + if not line: + sys.stderr.write(child_stderr.readline()) + (space, inodes) = line.split() + self.disk_inodes = int(inodes) + self.disk_blocks = int(space) + #(self.disk_inodes, self.disk_blocks) = vduimpl.vdu(self.dir) - # disklimit can be applied without a process in context - disklimit = resources.get("DISKLIMIT", 0) - if disklimit: - self.set_disklimit(disklimit) + return self.disk_blocks * 1024 - # - # Figure out if any processes are active in context, apply new - # values if there are. - # + def stop(self, signal = signal.SIGKILL): + vserverimpl.killall(self.ctx, signal) + self.vm_running = False + self.rlimits_changed = False - def init_disk_info(self): - (self.disk_inodes, self.disk_blocks, size) = vduimpl.vdu(self.dir) - self.disk_usage_set = False - return size +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)