import os import sys import time import stat import random import signal import logging import errno from errno import ENOENT, ENODATA from threading import currentThread, Condition, Lock from gconf import gconf from syncdutils import FreeObject, Thread, GsyncdError, boolify URXTIME = (-1, 0) class GMaster(object): """class impementling master role""" KFGN = 0 KNAT = 1 def get_sys_volinfo(self): """query volume marks on fs root err out on multiple foreign masters """ fgn_vis, nat_vi = self.master.server.foreign_volume_infos(), \ self.master.server.native_volume_info() fgn_vi = None if fgn_vis: if len(fgn_vis) > 1: raise GsyncdError("cannot work with multiple foreign masters") fgn_vi = fgn_vis[0] return fgn_vi, nat_vi @property def uuid(self): if self.volinfo: return self.volinfo['uuid'] @property def volmark(self): if self.volinfo: return self.volinfo['volume_mark'] @property def inter_master(self): """decide if we are an intermediate master in a cascading setup """ return self.volinfo_state[self.KFGN] and True or False def xtime(self, path, *a, **opts): """get amended xtime as of amending, we can create missing xtime, or determine a valid value if what we get is expired (as of the volume mark expiry); way of amendig depends on @opts and on subject of query (master or slave). """ if a: rsc = a[0] else: rsc = self.master if not 'create' in opts: opts['create'] = (rsc == self.master and not self.inter_master) if not 'default_xtime' in opts: if rsc == self.master and self.inter_master: opts['default_xtime'] = ENODATA else: opts['default_xtime'] = URXTIME xt = rsc.server.xtime(path, self.uuid) if isinstance(xt, int) and xt != ENODATA: return xt invalid_xtime = (xt == ENODATA or xt < self.volmark) if invalid_xtime: if opts['create']: t = time.time() sec = int(t) nsec = int((t - sec) * 1000000) xt = (sec, nsec) rsc.server.set_xtime(path, self.uuid, xt) else: xt = opts['default_xtime'] return xt def __init__(self, master, slave): self.master = master self.slave = slave self.jobtab = {} self.syncer = Syncer(slave) # crawls vs. turns: # - self.crawls is simply the number of crawl() invocations on root # - one turn is a maximal consecutive sequence of crawls so that each # crawl in it detects a change to be synced # - self.turns is the number of turns since start # - self.total_turns is a limit so that if self.turns reaches it, then # we exit (for diagnostic purposes) # so, eg., if the master fs changes unceasingly, self.turns will remain 0. self.crawls = 0 self.turns = 0 self.total_turns = int(gconf.turns) self.lastreport = {'crawls': 0, 'turns': 0} self.start = None self.change_seen = None # the authoritative (foreign, native) volinfo pair # which lets us deduce what to do when we refetch # the volinfos from system uuid_preset = getattr(gconf, 'volume_id', None) self.volinfo_state = (uuid_preset and {'uuid': uuid_preset}, None) # the actual volinfo we make use of self.volinfo = None self.terminate = False def crawl_loop(self): """start the keep-alive thread and iterate .crawl""" timo = int(gconf.timeout or 0) if timo > 0: def keep_alive(): while True: gap = timo * 0.5 # first grab a reference as self.volinfo # can be changed in main thread vi = self.volinfo if vi: # then have a private copy which we can mod vi = vi.copy() vi['timeout'] = int(time.time()) + timo else: # send keep-alives more frequently to # avoid a delay in announcing our volume info # to slave if it becomes established in the # meantime gap = min(10, gap) self.slave.server.keep_alive(vi) time.sleep(gap) t = Thread(target=keep_alive) t.start() self.lastreport['time'] = time.time() while not self.terminate: self.crawl() def add_job(self, path, label, job, *a, **kw): """insert @job function to job table at @path with @label""" if self.jobtab.get(path) == None: self.jobtab[path] = [] self.jobtab[path].append((label, a, lambda : job(*a, **kw))) def add_failjob(self, path, label): """invoke .add_job with a job that does nothing just fails""" logging.debug('salvaged: ' + label) self.add_job(path, label, lambda: False) def wait(self, path, *args): """perform jobs registered for @path Reset jobtab entry for @path, determine success as the conjuction of success of all the jobs. In case of success, call .sendmark on @path """ jobs = self.jobtab.pop(path, []) succeed = True for j in jobs: ret = j[-1]() if not ret: succeed = False if succeed: self.sendmark(path, *args) return succeed def sendmark(self, path, mark, adct=None): """update slave side xtime for @path to master side xtime also can send a setattr payload (see Server.setattr). """ if adct: self.slave.server.setattr(path, adct) self.slave.server.set_xtime(path, self.uuid, mark) @staticmethod def volinfo_state_machine(volinfo_state, volinfo_sys): """compute new volinfo_state from old one and incoming as of current system state, also indicating if there was a change regarding which volume mark is the authoritative one @volinfo_state, @volinfo_sys are pairs of volume mark dicts (foreign, native). Note this method is marked as static, ie. the computation is pure, without reliance on any excess implicit state. State transitions which are deemed as ambiguous or banned will raise an exception. """ # store the value below "boxed" to emulate proper closures # (variables of the enclosing scope are available inner functions # provided they are no reassigned; mutation is OK). param = FreeObject(relax_mismatch = False, state_change = None, index=-1) def select_vi(vi0, vi): param.index += 1 if vi and (not vi0 or vi0['uuid'] == vi['uuid']): if not vi0 and not param.relax_mismatch: param.state_change = param.index # valid new value found; for the rest, we are graceful about # uuid mismatch param.relax_mismatch = True return vi if vi0 and vi and vi0['uuid'] != vi['uuid'] and not param.relax_mismatch: # uuid mismatch for master candidate, bail out raise GsyncdError("aborting on uuid change from %s to %s" % \ (vi0['uuid'], vi['uuid'])) # fall back to old return vi0 newstate = tuple(select_vi(*vip) for vip in zip(volinfo_state, volinfo_sys)) srep = lambda vi: vi and vi['uuid'][0:8] logging.debug('(%s, %s) << (%s, %s) -> (%s, %s)' % \ tuple(srep(vi) for vi in volinfo_state + volinfo_sys + newstate)) return newstate, param.state_change def crawl(self, path='.', xtl=None): """crawling... Standing around All the right people Crawling Tennis on Tuesday The ladder is long It is your nature You've gotta suntan Football on Sunday Society boy Recursively walk the master side tree and check if updates are needed due to xtime differences. One invocation of crawl checks children of @path and do a recursive enter only on those directory children where there is an update needed. Way of updates depend on file type: - for symlinks, sync them directy and synchronously - for regular children, register jobs for @path (cf. .add_job) to start and wait on their rsync - for directory children, register a job for @path which waits (.wait) on jobs for the given child (other kind of filesystem nodes are not considered) Those slave side children which do not exist on master are simply purged (see Server.purge). Behavior is fault tolerant, synchronization is adaptive: if some action fails, just go on relentlessly, adding a fail job (see .add_failjob) which will prevent the .sendmark on @path, so when the next crawl will arrive to @path it will not see it as up-to-date and will try to sync it again. While this semantics can be supported by funky design principles (http://c2.com/cgi/wiki?LazinessImpatienceHubris), the ultimate reason which excludes other possibilities is simply transience: we cannot assert that the file systems (master / slave) underneath do not change and actions taken upon some condition will not lose their context by the time they are performed. """ if path == '.': if self.start: self.crawls += 1 logging.debug("... crawl #%d done, took %.6f seconds" % \ (self.crawls, time.time() - self.start)) time.sleep(1) self.start = time.time() should_display_info = self.start - self.lastreport['time'] >= 60 if should_display_info: logging.info("completed %d crawls, %d turns", self.crawls - self.lastreport['crawls'], self.turns - self.lastreport['turns']) self.lastreport.update(crawls = self.crawls, turns = self.turns, time = self.start) volinfo_sys = self.get_sys_volinfo() self.volinfo_state, state_change = self.volinfo_state_machine(self.volinfo_state, volinfo_sys) if self.inter_master: self.volinfo = volinfo_sys[self.KFGN] else: self.volinfo = volinfo_sys[self.KNAT] if state_change == self.KFGN or (state_change == self.KNAT and not self.inter_master): logging.info('new master is %s', self.uuid) if self.volinfo: logging.info("%s master with volume id %s ..." % \ (self.inter_master and "intermediate" or "primary", self.uuid)) if state_change == self.KFGN: gconf.configinterface.set('volume_id', self.uuid) if self.volinfo: if self.volinfo['retval']: raise GsyncdError ("master is corrupt") else: if should_display_info or self.crawls == 0: if self.inter_master: logging.info("waiting for being synced from %s ..." % \ self.volinfo_state[self.KFGN]['uuid']) else: logging.info("waiting for volume info ...") return logging.debug("entering " + path) if not xtl: xtl = self.xtime(path) if isinstance(xtl, int): self.add_failjob(path, 'no-local-node') return xtr0 = self.xtime(path, self.slave) if isinstance(xtr0, int): if xtr0 != ENOENT: self.slave.server.purge(path) try: self.slave.server.mkdir(path) except OSError: self.add_failjob(path, 'no-remote-node') return xtr = URXTIME else: xtr = xtr0 if xtr > xtl: raise GsyncdError("timestamp corruption for " + path) if xtl == xtr: if path == '.' and self.change_seen: self.turns += 1 self.change_seen = False if self.total_turns: logging.info("finished turn #%s/%s" % \ (self.turns, self.total_turns)) if self.turns == self.total_turns: logging.info("reached turn limit") self.terminate = True return if path == '.': self.change_seen = True try: dem = self.master.server.entries(path) except OSError: self.add_failjob(path, 'local-entries-fail') return random.shuffle(dem) try: des = self.slave.server.entries(path) except OSError: self.slave.server.purge(path) try: self.slave.server.mkdir(path) des = self.slave.server.entries(path) except OSError: self.add_failjob(path, 'remote-entries-fail') return dd = set(des) - set(dem) if dd and not boolify(gconf.ignore_deletes): self.slave.server.purge(path, dd) chld = [] for e in dem: e = os.path.join(path, e) xte = self.xtime(e) if isinstance(xte, int): logging.warn("irregular xtime for %s: %s" % (e, errno.errorcode[xte])) elif xte > xtr: chld.append((e, xte)) def indulgently(e, fnc, blame=None): if not blame: blame = path try: return fnc(e) except (IOError, OSError): ex = sys.exc_info()[1] if ex.errno == ENOENT: logging.warn("salvaged ENOENT for " + e) self.add_failjob(blame, 'by-indulgently') return False else: raise for e, xte in chld: st = indulgently(e, lambda e: os.lstat(e)) if st == False: continue mo = st.st_mode adct = {'own': (st.st_uid, st.st_gid)} if stat.S_ISLNK(mo): if indulgently(e, lambda e: self.slave.server.symlink(os.readlink(e), e)) == False: continue self.sendmark(e, xte, adct) elif stat.S_ISREG(mo): logging.debug("syncing %s ..." % e) pb = self.syncer.add(e) def regjob(e, xte, pb): if pb.wait(): logging.debug("synced " + e) self.sendmark(e, xte) return True else: logging.warn("failed to sync " + e) self.add_job(path, 'reg', regjob, e, xte, pb) elif stat.S_ISDIR(mo): adct['mode'] = mo if indulgently(e, lambda e: (self.add_job(path, 'cwait', self.wait, e, xte, adct), self.crawl(e, xte), True)[-1], blame=e) == False: continue else: # ignore fifos, sockets and special files pass if path == '.': self.wait(path, xtl) class BoxClosedErr(Exception): pass class PostBox(list): """synchronized collection for storing things thought of as "requests" """ def __init__(self, *a): list.__init__(self, *a) # too bad Python stdlib does not have read/write locks... # it would suffivce to grab the lock in .append as reader, in .close as writer self.lever = Condition() self.open = True self.done = False def wait(self): """wait on requests to be processed""" self.lever.acquire() if not self.done: self.lever.wait() self.lever.release() return self.result def wakeup(self, data): """wake up requestors with the result""" self.result = data self.lever.acquire() self.done = True self.lever.notifyAll() self.lever.release() def append(self, e): """post a request""" self.lever.acquire() if not self.open: raise BoxClosedErr list.append(self, e) self.lever.release() def close(self): """prohibit the posting of further requests""" self.lever.acquire() self.open = False self.lever.release() class Syncer(object): """a staged queue to relay rsync requests to rsync workers By "staged queue" its meant that when a consumer comes to the queue, it takes _all_ entries, leaving the queue empty. (I don't know if there is an official term for this pattern.) The queue uses a PostBox to accumulate incoming items. When a consumer (rsync worker) comes, a new PostBox is set up and the old one is passed on to the consumer. Instead of the simplistic scheme of having one big lock which synchronizes both the addition of new items and PostBox exchanges, use a separate lock to arbitrate consumers, and rely on PostBox's synchronization mechanisms take care about additions. There is a corner case racy situation, producers vs. consumers, which is not handled by this scheme: namely, when the PostBox exchange occurs in between being passed to the producer for posting and the post placement. But that's what Postbox.close is for: such a posting will find the PostBox closed, in which case the producer can re-try posting against the actual PostBox of the queue. To aid accumlation of items in the PostBoxen before grabbed by an rsync worker, the worker goes to sleep a bit after each completed syncjob. """ def __init__(self, slave): """spawn worker threads""" self.slave = slave self.lock = Lock() self.pb = PostBox() for i in range(int(gconf.sync_jobs)): t = Thread(target=self.syncjob) t.start() def syncjob(self): """the life of a worker""" while True: pb = None while True: self.lock.acquire() if self.pb: pb, self.pb = self.pb, PostBox() self.lock.release() if pb: break time.sleep(0.5) pb.close() po = self.slave.rsync(pb) if po.returncode == 0: ret = True elif po.returncode in (23, 24): # partial transfer (cf. rsync(1)), that's normal ret = False else: po.errfail() pb.wakeup(ret) def add(self, e): while True: pb = self.pb try: pb.append(e) return pb except BoxClosedErr: pass