Mercurial > hg
view hgext/fsmonitor/__init__.py @ 29039:e3055b46ed1b stable
tags: silence hgtagsfnodes reading failures
tryread() doesn't handle "is a directory" errors and presumably
others. We might not want to globally swallow such tryread errors, so
we replace with our own try/except handling.
An upcoming test will use directories as a portable stand-in for
various bizarre circumstances that cache read/write code should be
robust to.
author | Matt Mackall <mpm@selenic.com> |
---|---|
date | Thu, 28 Apr 2016 15:40:43 -0500 |
parents | 49d65663d7e4 |
children | a0939666b836 |
line wrap: on
line source
# __init__.py - fsmonitor initialization and overrides # # Copyright 2013-2016 Facebook, Inc. # # This software may be used and distributed according to the terms of the # GNU General Public License version 2 or any later version. '''Faster status operations with the Watchman file monitor (EXPERIMENTAL) Integrates the file-watching program Watchman with Mercurial to produce faster status results. On a particular Linux system, for a real-world repository with over 400,000 files hosted on ext4, vanilla `hg status` takes 1.3 seconds. On the same system, with fsmonitor it takes about 0.3 seconds. fsmonitor requires no configuration -- it will tell Watchman about your repository as necessary. You'll need to install Watchman from https://facebook.github.io/watchman/ and make sure it is in your PATH. The following configuration options exist: :: [fsmonitor] mode = {off, on, paranoid} When `mode = off`, fsmonitor will disable itself (similar to not loading the extension at all). When `mode = on`, fsmonitor will be enabled (the default). When `mode = paranoid`, fsmonitor will query both Watchman and the filesystem, and ensure that the results are consistent. :: [fsmonitor] timeout = (float) A value, in seconds, that determines how long fsmonitor will wait for Watchman to return results. Defaults to `2.0`. :: [fsmonitor] blacklistusers = (list of userids) A list of usernames for which fsmonitor will disable itself altogether. :: [fsmonitor] walk_on_invalidate = (boolean) Whether or not to walk the whole repo ourselves when our cached state has been invalidated, for example when Watchman has been restarted or .hgignore rules have been changed. Walking the repo in that case can result in competing for I/O with Watchman. For large repos it is recommended to set this value to false. You may wish to set this to true if you have a very fast filesystem that can outpace the IPC overhead of getting the result data for the full repo from Watchman. Defaults to false. fsmonitor is incompatible with the largefiles and eol extensions, and will disable itself if any of those are active. ''' # Platforms Supported # =================== # # **Linux:** *Stable*. Watchman and fsmonitor are both known to work reliably, # even under severe loads. # # **Mac OS X:** *Stable*. The Mercurial test suite passes with fsmonitor # turned on, on case-insensitive HFS+. There has been a reasonable amount of # user testing under normal loads. # # **Solaris, BSD:** *Alpha*. watchman and fsmonitor are believed to work, but # very little testing has been done. # # **Windows:** *Alpha*. Not in a release version of watchman or fsmonitor yet. # # Known Issues # ============ # # * fsmonitor will disable itself if any of the following extensions are # enabled: largefiles, inotify, eol; or if the repository has subrepos. # * fsmonitor will produce incorrect results if nested repos that are not # subrepos exist. *Workaround*: add nested repo paths to your `.hgignore`. # # The issues related to nested repos and subrepos are probably not fundamental # ones. Patches to fix them are welcome. from __future__ import absolute_import import os import stat import sys from mercurial import ( context, extensions, localrepo, merge, pathutil, scmutil, util, ) from mercurial import match as matchmod from mercurial.i18n import _ from . import ( state, watchmanclient, ) # Note for extension authors: ONLY specify testedwith = 'internal' for # extensions which SHIP WITH MERCURIAL. Non-mainline extensions should # be specifying the version(s) of Mercurial they are tested with, or # leave the attribute unspecified. testedwith = 'internal' # This extension is incompatible with the following blacklisted extensions # and will disable itself when encountering one of these: _blacklist = ['largefiles', 'eol'] def _handleunavailable(ui, state, ex): """Exception handler for Watchman interaction exceptions""" if isinstance(ex, watchmanclient.Unavailable): if ex.warn: ui.warn(str(ex) + '\n') if ex.invalidate: state.invalidate() ui.log('fsmonitor', 'Watchman unavailable: %s\n', ex.msg) else: ui.log('fsmonitor', 'Watchman exception: %s\n', ex) def _hashignore(ignore): """Calculate hash for ignore patterns and filenames If this information changes between Mercurial invocations, we can't rely on Watchman information anymore and have to re-scan the working copy. """ sha1 = util.sha1() if util.safehasattr(ignore, 'includepat'): sha1.update(ignore.includepat) sha1.update('\0\0') if util.safehasattr(ignore, 'excludepat'): sha1.update(ignore.excludepat) sha1.update('\0\0') if util.safehasattr(ignore, 'patternspat'): sha1.update(ignore.patternspat) sha1.update('\0\0') if util.safehasattr(ignore, '_files'): for f in ignore._files: sha1.update(f) sha1.update('\0') return sha1.hexdigest() def overridewalk(orig, self, match, subrepos, unknown, ignored, full=True): '''Replacement for dirstate.walk, hooking into Watchman. Whenever full is False, ignored is False, and the Watchman client is available, use Watchman combined with saved state to possibly return only a subset of files.''' def bail(): return orig(match, subrepos, unknown, ignored, full=True) if full or ignored or not self._watchmanclient.available(): return bail() state = self._fsmonitorstate clock, ignorehash, notefiles = state.get() if not clock: if state.walk_on_invalidate: return bail() # Initial NULL clock value, see # https://facebook.github.io/watchman/docs/clockspec.html clock = 'c:0:0' notefiles = [] def fwarn(f, msg): self._ui.warn('%s: %s\n' % (self.pathto(f), msg)) return False def badtype(mode): kind = _('unknown') if stat.S_ISCHR(mode): kind = _('character device') elif stat.S_ISBLK(mode): kind = _('block device') elif stat.S_ISFIFO(mode): kind = _('fifo') elif stat.S_ISSOCK(mode): kind = _('socket') elif stat.S_ISDIR(mode): kind = _('directory') return _('unsupported file type (type is %s)') % kind ignore = self._ignore dirignore = self._dirignore if unknown: if _hashignore(ignore) != ignorehash and clock != 'c:0:0': # ignore list changed -- can't rely on Watchman state any more if state.walk_on_invalidate: return bail() notefiles = [] clock = 'c:0:0' else: # always ignore ignore = util.always dirignore = util.always matchfn = match.matchfn matchalways = match.always() dmap = self._map nonnormalset = getattr(self, '_nonnormalset', None) copymap = self._copymap getkind = stat.S_IFMT dirkind = stat.S_IFDIR regkind = stat.S_IFREG lnkkind = stat.S_IFLNK join = self._join normcase = util.normcase fresh_instance = False exact = skipstep3 = False if matchfn == match.exact: # match.exact exact = True dirignore = util.always # skip step 2 elif match.files() and not match.anypats(): # match.match, no patterns skipstep3 = True if not exact and self._checkcase: # note that even though we could receive directory entries, we're only # interested in checking if a file with the same name exists. So only # normalize files if possible. normalize = self._normalizefile skipstep3 = False else: normalize = None # step 1: find all explicit files results, work, dirsnotfound = self._walkexplicit(match, subrepos) skipstep3 = skipstep3 and not (work or dirsnotfound) work = [d for d in work if not dirignore(d[0])] if not work and (exact or skipstep3): for s in subrepos: del results[s] del results['.hg'] return results # step 2: query Watchman try: # Use the user-configured timeout for the query. # Add a little slack over the top of the user query to allow for # overheads while transferring the data self._watchmanclient.settimeout(state.timeout + 0.1) result = self._watchmanclient.command('query', { 'fields': ['mode', 'mtime', 'size', 'exists', 'name'], 'since': clock, 'expression': [ 'not', [ 'anyof', ['dirname', '.hg'], ['name', '.hg', 'wholename'] ] ], 'sync_timeout': int(state.timeout * 1000), 'empty_on_fresh_instance': state.walk_on_invalidate, }) except Exception as ex: _handleunavailable(self._ui, state, ex) self._watchmanclient.clearconnection() return bail() else: # We need to propagate the last observed clock up so that we # can use it for our next query state.setlastclock(result['clock']) if result['is_fresh_instance']: if state.walk_on_invalidate: state.invalidate() return bail() fresh_instance = True # Ignore any prior noteable files from the state info notefiles = [] # for file paths which require normalization and we encounter a case # collision, we store our own foldmap if normalize: foldmap = dict((normcase(k), k) for k in results) switch_slashes = os.sep == '\\' # The order of the results is, strictly speaking, undefined. # For case changes on a case insensitive filesystem we may receive # two entries, one with exists=True and another with exists=False. # The exists=True entries in the same response should be interpreted # as being happens-after the exists=False entries due to the way that # Watchman tracks files. We use this property to reconcile deletes # for name case changes. for entry in result['files']: fname = entry['name'] if switch_slashes: fname = fname.replace('\\', '/') if normalize: normed = normcase(fname) fname = normalize(fname, True, True) foldmap[normed] = fname fmode = entry['mode'] fexists = entry['exists'] kind = getkind(fmode) if not fexists: # if marked as deleted and we don't already have a change # record, mark it as deleted. If we already have an entry # for fname then it was either part of walkexplicit or was # an earlier result that was a case change if fname not in results and fname in dmap and ( matchalways or matchfn(fname)): results[fname] = None elif kind == dirkind: if fname in dmap and (matchalways or matchfn(fname)): results[fname] = None elif kind == regkind or kind == lnkkind: if fname in dmap: if matchalways or matchfn(fname): results[fname] = entry elif (matchalways or matchfn(fname)) and not ignore(fname): results[fname] = entry elif fname in dmap and (matchalways or matchfn(fname)): results[fname] = None # step 3: query notable files we don't already know about # XXX try not to iterate over the entire dmap if normalize: # any notable files that have changed case will already be handled # above, so just check membership in the foldmap notefiles = set((normalize(f, True, True) for f in notefiles if normcase(f) not in foldmap)) visit = set((f for f in notefiles if (f not in results and matchfn(f) and (f in dmap or not ignore(f))))) if nonnormalset is not None and not fresh_instance: if matchalways: visit.update(f for f in nonnormalset if f not in results) visit.update(f for f in copymap if f not in results) else: visit.update(f for f in nonnormalset if f not in results and matchfn(f)) visit.update(f for f in copymap if f not in results and matchfn(f)) else: if matchalways: visit.update(f for f, st in dmap.iteritems() if (f not in results and (st[2] < 0 or st[0] != 'n' or fresh_instance))) visit.update(f for f in copymap if f not in results) else: visit.update(f for f, st in dmap.iteritems() if (f not in results and (st[2] < 0 or st[0] != 'n' or fresh_instance) and matchfn(f))) visit.update(f for f in copymap if f not in results and matchfn(f)) audit = pathutil.pathauditor(self._root).check auditpass = [f for f in visit if audit(f)] auditpass.sort() auditfail = visit.difference(auditpass) for f in auditfail: results[f] = None nf = iter(auditpass).next for st in util.statfiles([join(f) for f in auditpass]): f = nf() if st or f in dmap: results[f] = st for s in subrepos: del results[s] del results['.hg'] return results def overridestatus( orig, self, node1='.', node2=None, match=None, ignored=False, clean=False, unknown=False, listsubrepos=False): listignored = ignored listclean = clean listunknown = unknown def _cmpsets(l1, l2): try: if 'FSMONITOR_LOG_FILE' in os.environ: fn = os.environ['FSMONITOR_LOG_FILE'] f = open(fn, 'wb') else: fn = 'fsmonitorfail.log' f = self.opener(fn, 'wb') except (IOError, OSError): self.ui.warn(_('warning: unable to write to %s\n') % fn) return try: for i, (s1, s2) in enumerate(zip(l1, l2)): if set(s1) != set(s2): f.write('sets at position %d are unequal\n' % i) f.write('watchman returned: %s\n' % s1) f.write('stat returned: %s\n' % s2) finally: f.close() if isinstance(node1, context.changectx): ctx1 = node1 else: ctx1 = self[node1] if isinstance(node2, context.changectx): ctx2 = node2 else: ctx2 = self[node2] working = ctx2.rev() is None parentworking = working and ctx1 == self['.'] match = match or matchmod.always(self.root, self.getcwd()) # Maybe we can use this opportunity to update Watchman's state. # Mercurial uses workingcommitctx and/or memctx to represent the part of # the workingctx that is to be committed. So don't update the state in # that case. # HG_PENDING is set in the environment when the dirstate is being updated # in the middle of a transaction; we must not update our state in that # case, or we risk forgetting about changes in the working copy. updatestate = (parentworking and match.always() and not isinstance(ctx2, (context.workingcommitctx, context.memctx)) and 'HG_PENDING' not in os.environ) try: if self._fsmonitorstate.walk_on_invalidate: # Use a short timeout to query the current clock. If that # takes too long then we assume that the service will be slow # to answer our query. # walk_on_invalidate indicates that we prefer to walk the # tree ourselves because we can ignore portions that Watchman # cannot and we tend to be faster in the warmer buffer cache # cases. self._watchmanclient.settimeout(0.1) else: # Give Watchman more time to potentially complete its walk # and return the initial clock. In this mode we assume that # the filesystem will be slower than parsing a potentially # very large Watchman result set. self._watchmanclient.settimeout( self._fsmonitorstate.timeout + 0.1) startclock = self._watchmanclient.getcurrentclock() except Exception as ex: self._watchmanclient.clearconnection() _handleunavailable(self.ui, self._fsmonitorstate, ex) # boo, Watchman failed. bail return orig(node1, node2, match, listignored, listclean, listunknown, listsubrepos) if updatestate: # We need info about unknown files. This may make things slower the # first time, but whatever. stateunknown = True else: stateunknown = listunknown r = orig(node1, node2, match, listignored, listclean, stateunknown, listsubrepos) modified, added, removed, deleted, unknown, ignored, clean = r if updatestate: notefiles = modified + added + removed + deleted + unknown self._fsmonitorstate.set( self._fsmonitorstate.getlastclock() or startclock, _hashignore(self.dirstate._ignore), notefiles) if not listunknown: unknown = [] # don't do paranoid checks if we're not going to query Watchman anyway full = listclean or match.traversedir is not None if self._fsmonitorstate.mode == 'paranoid' and not full: # run status again and fall back to the old walk this time self.dirstate._fsmonitordisable = True # shut the UI up quiet = self.ui.quiet self.ui.quiet = True fout, ferr = self.ui.fout, self.ui.ferr self.ui.fout = self.ui.ferr = open(os.devnull, 'wb') try: rv2 = orig( node1, node2, match, listignored, listclean, listunknown, listsubrepos) finally: self.dirstate._fsmonitordisable = False self.ui.quiet = quiet self.ui.fout, self.ui.ferr = fout, ferr # clean isn't tested since it's set to True above _cmpsets([modified, added, removed, deleted, unknown, ignored, clean], rv2) modified, added, removed, deleted, unknown, ignored, clean = rv2 return scmutil.status( modified, added, removed, deleted, unknown, ignored, clean) def makedirstate(cls): class fsmonitordirstate(cls): def _fsmonitorinit(self, fsmonitorstate, watchmanclient): # _fsmonitordisable is used in paranoid mode self._fsmonitordisable = False self._fsmonitorstate = fsmonitorstate self._watchmanclient = watchmanclient def walk(self, *args, **kwargs): orig = super(fsmonitordirstate, self).walk if self._fsmonitordisable: return orig(*args, **kwargs) return overridewalk(orig, self, *args, **kwargs) def rebuild(self, *args, **kwargs): self._fsmonitorstate.invalidate() return super(fsmonitordirstate, self).rebuild(*args, **kwargs) def invalidate(self, *args, **kwargs): self._fsmonitorstate.invalidate() return super(fsmonitordirstate, self).invalidate(*args, **kwargs) return fsmonitordirstate def wrapdirstate(orig, self): ds = orig(self) # only override the dirstate when Watchman is available for the repo if util.safehasattr(self, '_fsmonitorstate'): ds.__class__ = makedirstate(ds.__class__) ds._fsmonitorinit(self._fsmonitorstate, self._watchmanclient) return ds def extsetup(ui): wrapfilecache(localrepo.localrepository, 'dirstate', wrapdirstate) if sys.platform == 'darwin': # An assist for avoiding the dangling-symlink fsevents bug extensions.wrapfunction(os, 'symlink', wrapsymlink) extensions.wrapfunction(merge, 'update', wrapupdate) def wrapsymlink(orig, source, link_name): ''' if we create a dangling symlink, also touch the parent dir to encourage fsevents notifications to work more correctly ''' try: return orig(source, link_name) finally: try: os.utime(os.path.dirname(link_name), None) except OSError: pass class state_update(object): ''' This context mananger is responsible for dispatching the state-enter and state-leave signals to the watchman service ''' def __init__(self, repo, node, distance, partial): self.repo = repo self.node = node self.distance = distance self.partial = partial def __enter__(self): self._state('state-enter') return self def __exit__(self, type_, value, tb): status = 'ok' if type_ is None else 'failed' self._state('state-leave', status=status) def _state(self, cmd, status='ok'): if not util.safehasattr(self.repo, '_watchmanclient'): return try: commithash = self.repo[self.node].hex() self.repo._watchmanclient.command(cmd, { 'name': 'hg.update', 'metadata': { # the target revision 'rev': commithash, # approximate number of commits between current and target 'distance': self.distance, # success/failure (only really meaningful for state-leave) 'status': status, # whether the working copy parent is changing 'partial': self.partial, }}) except Exception as e: # Swallow any errors; fire and forget self.repo.ui.log( 'watchman', 'Exception %s while running %s\n', e, cmd) # Bracket working copy updates with calls to the watchman state-enter # and state-leave commands. This allows clients to perform more intelligent # settling during bulk file change scenarios # https://facebook.github.io/watchman/docs/cmd/subscribe.html#advanced-settling def wrapupdate(orig, repo, node, branchmerge, force, ancestor=None, mergeancestor=False, labels=None, matcher=None, **kwargs): distance = 0 partial = True if matcher is None or matcher.always(): partial = False wc = repo[None] parents = wc.parents() if len(parents) == 2: anc = repo.changelog.ancestor(parents[0].node(), parents[1].node()) ancrev = repo[anc].rev() distance = abs(repo[node].rev() - ancrev) elif len(parents) == 1: distance = abs(repo[node].rev() - parents[0].rev()) with state_update(repo, node, distance, partial): return orig( repo, node, branchmerge, force, ancestor, mergeancestor, labels, matcher, *kwargs) def reposetup(ui, repo): # We don't work with largefiles or inotify exts = extensions.enabled() for ext in _blacklist: if ext in exts: ui.warn(_('The fsmonitor extension is incompatible with the %s ' 'extension and has been disabled.\n') % ext) return if util.safehasattr(repo, 'dirstate'): # We don't work with subrepos either. Note that we can get passed in # e.g. a statichttprepo, which throws on trying to access the substate. # XXX This sucks. try: # if repo[None].substate can cause a dirstate parse, which is too # slow. Instead, look for a file called hgsubstate, if repo.wvfs.exists('.hgsubstate') or repo.wvfs.exists('.hgsub'): return except AttributeError: return fsmonitorstate = state.state(repo) if fsmonitorstate.mode == 'off': return try: client = watchmanclient.client(repo) except Exception as ex: _handleunavailable(ui, fsmonitorstate, ex) return repo._fsmonitorstate = fsmonitorstate repo._watchmanclient = client # at this point since fsmonitorstate wasn't present, repo.dirstate is # not a fsmonitordirstate repo.dirstate.__class__ = makedirstate(repo.dirstate.__class__) # nuke the dirstate so that _fsmonitorinit and subsequent configuration # changes take effect on it del repo._filecache['dirstate'] delattr(repo.unfiltered(), 'dirstate') class fsmonitorrepo(repo.__class__): def status(self, *args, **kwargs): orig = super(fsmonitorrepo, self).status return overridestatus(orig, self, *args, **kwargs) repo.__class__ = fsmonitorrepo def wrapfilecache(cls, propname, wrapper): """Wraps a filecache property. These can't be wrapped using the normal wrapfunction. This should eventually go into upstream Mercurial. """ assert callable(wrapper) for currcls in cls.__mro__: if propname in currcls.__dict__: origfn = currcls.__dict__[propname].func assert callable(origfn) def wrap(*args, **kwargs): return wrapper(origfn, *args, **kwargs) currcls.__dict__[propname].func = wrap break if currcls is object: raise AttributeError( _("type '%s' has no property '%s'") % (cls, propname))