mercurial/changegroup.py

bundle2: increase payload part chunk size to 32kb

Bundle2 payload parts are framed chunks. Esentially, we obtain
data in equal size chunks of size `preferedchunksize` and emit those
to a generator. That generator is fed into a compressor (which can
be the no-op compressor, which just re-emits the generator). And
the output from the compressor likely goes to a file descriptor
or socket.

What this means is that small chunk sizes create more Python objects
and Python function calls than larger chunk sizes. And as we know,
Python object and function call overhead in performance sensitive
code matters (at least with CPython).

This commit increases the bundle2 part payload chunk size from 4k
to 32k. Practically speaking, this means that the chunks we feed
into a compressor (implemented in C code) or feed directly into a
file handle or socket write() are larger. It's possible the chunks
might be larger than what the receiver can handle in one logical
operation. But at that point, we're in C code, which is much more
efficient at dealing with splitting up the chunk and making multiple
function calls than Python is.

A downside to larger chunks is that the receiver has to wait for that
much data to arrive (either raw or from a decompressor) before it
can process the chunk. But 32kb still feels like a small buffer to
have to wait for. And in many cases, the client will convert from
8 read(4096) to 1 read(32768). That's happening in Python land. So
we cut down on the number of Python objects and function calls,
making the client faster as well. I don't think there are any
significant concerns to increasing the payload chunk size to 32kb.

The impact of this change on performance significant. Using `curl`
to obtain a stream clone bundle2 payload from a server on localhost
serving the mozilla-unified repository:

before: 20.78 user; 7.71 system;  80.5 MB/s
after:  13.90 user; 3.51 system; 132 MB/s
legacy:  9.72 user; 8.16 system; 132 MB/s

bundle2 stream clone generation is still more resource intensive than
legacy stream clone (that's likely because of the use of a
util.chunkbuffer). But the throughput is the same. We might
be in territory we're this is effectively a benchmark of the
networking stack or Python's syscall throughput.

From the client perspective, `hg clone -U --stream`:

before: 33.50 user; 7.95 system; 53.3 MB/s
after:  22.82 user; 7.33 system; 72.7 MB/s
legacy: 29.96 user; 7.94 system; 58.0 MB/s

And for `hg clone --stream` with a working directory update of
~230k files:

after:  119.55 user; 26.47 system; 0:57.08 wall
legacy: 126.98 user; 26.94 system; 1:05.56 wall

So, it appears that bundle2's stream clone is now definitively faster
than legacy stream clone!

Differential Revision: https://phab.mercurial-scm.org/D1932

# changegroup.py - Mercurial changegroup manipulation functions
#
#  Copyright 2006 Matt Mackall <[e-mail]>
#
# This software may be used and distributed according to the terms of the
# GNU General Public License version 2 or any later version.

from __future__ import absolute_import

import os
import struct
import tempfile
import weakref

from .i18n import _
from .node import (
    hex,
    nullrev,
    short,
)

from . import (
    dagutil,
    error,
    mdiff,
    phases,
    pycompat,
    util,
)

_CHANGEGROUPV1_DELTA_HEADER = "20s20s20s20s"
_CHANGEGROUPV2_DELTA_HEADER = "20s20s20s20s20s"
_CHANGEGROUPV3_DELTA_HEADER = ">20s20s20s20s20sH"

readexactly = util.readexactly

def getchunk(stream):
    """return the next chunk from stream as a string"""
    d = readexactly(stream, 4)
    l = struct.unpack(">l", d)[0]
    if l <= 4:
        if l:
            raise error.Abort(_("invalid chunk length %d") % l)
        return ""
    return readexactly(stream, l - 4)

def chunkheader(length):
    """return a changegroup chunk header (string)"""
    return struct.pack(">l", length + 4)

def closechunk():
    """return a changegroup chunk header (string) for a zero-length chunk"""
    return struct.pack(">l", 0)

def writechunks(ui, chunks, filename, vfs=None):
    """Write chunks to a file and return its filename.

    The stream is assumed to be a bundle file.
    Existing files will not be overwritten.
    If no filename is specified, a temporary file is created.
    """
    fh = None
    cleanup = None
    try:
        if filename:
            if vfs:
                fh = vfs.open(filename, "wb")
            else:
                # Increase default buffer size because default is usually
                # small (4k is common on Linux).
                fh = open(filename, "wb", 131072)
        else:
            fd, filename = tempfile.mkstemp(prefix="hg-bundle-", suffix=".hg")
            fh = os.fdopen(fd, pycompat.sysstr("wb"))
        cleanup = filename
        for c in chunks:
            fh.write(c)
        cleanup = None
        return filename
    finally:
        if fh is not None:
            fh.close()
        if cleanup is not None:
            if filename and vfs:
                vfs.unlink(cleanup)
            else:
                os.unlink(cleanup)

class cg1unpacker(object):
    """Unpacker for cg1 changegroup streams.

    A changegroup unpacker handles the framing of the revision data in
    the wire format. Most consumers will want to use the apply()
    method to add the changes from the changegroup to a repository.

    If you're forwarding a changegroup unmodified to another consumer,
    use getchunks(), which returns an iterator of changegroup
    chunks. This is mostly useful for cases where you need to know the
    data stream has ended by observing the end of the changegroup.

    deltachunk() is useful only if you're applying delta data. Most
    consumers should prefer apply() instead.

    A few other public methods exist. Those are used only for
    bundlerepo and some debug commands - their use is discouraged.
    """
    deltaheader = _CHANGEGROUPV1_DELTA_HEADER
    deltaheadersize = struct.calcsize(deltaheader)
    version = '01'
    _grouplistcount = 1 # One list of files after the manifests

    def __init__(self, fh, alg, extras=None):
        if alg is None:
            alg = 'UN'
        if alg not in util.compengines.supportedbundletypes:
            raise error.Abort(_('unknown stream compression type: %s')
                             % alg)
        if alg == 'BZ':
            alg = '_truncatedBZ'

        compengine = util.compengines.forbundletype(alg)
        self._stream = compengine.decompressorreader(fh)
        self._type = alg
        self.extras = extras or {}
        self.callback = None

    # These methods (compressed, read, seek, tell) all appear to only
    # be used by bundlerepo, but it's a little hard to tell.
    def compressed(self):
        return self._type is not None and self._type != 'UN'
    def read(self, l):
        return self._stream.read(l)
    def seek(self, pos):
        return self._stream.seek(pos)
    def tell(self):
        return self._stream.tell()
    def close(self):
        return self._stream.close()

    def _chunklength(self):
        d = readexactly(self._stream, 4)
        l = struct.unpack(">l", d)[0]
        if l <= 4:
            if l:
                raise error.Abort(_("invalid chunk length %d") % l)
            return 0
        if self.callback:
            self.callback()
        return l - 4

    def changelogheader(self):
        """v10 does not have a changelog header chunk"""
        return {}

    def manifestheader(self):
        """v10 does not have a manifest header chunk"""
        return {}

    def filelogheader(self):
        """return the header of the filelogs chunk, v10 only has the filename"""
        l = self._chunklength()
        if not l:
            return {}
        fname = readexactly(self._stream, l)
        return {'filename': fname}

    def _deltaheader(self, headertuple, prevnode):
        node, p1, p2, cs = headertuple
        if prevnode is None:
            deltabase = p1
        else:
            deltabase = prevnode
        flags = 0
        return node, p1, p2, deltabase, cs, flags

    def deltachunk(self, prevnode):
        l = self._chunklength()
        if not l:
            return {}
        headerdata = readexactly(self._stream, self.deltaheadersize)
        header = struct.unpack(self.deltaheader, headerdata)
        delta = readexactly(self._stream, l - self.deltaheadersize)
        node, p1, p2, deltabase, cs, flags = self._deltaheader(header, prevnode)
        return (node, p1, p2, cs, deltabase, delta, flags)

    def getchunks(self):
        """returns all the chunks contains in the bundle

        Used when you need to forward the binary stream to a file or another
        network API. To do so, it parse the changegroup data, otherwise it will
        block in case of sshrepo because it don't know the end of the stream.
        """
        # For changegroup 1 and 2, we expect 3 parts: changelog, manifestlog,
        # and a list of filelogs. For changegroup 3, we expect 4 parts:
        # changelog, manifestlog, a list of tree manifestlogs, and a list of
        # filelogs.
        #
        # Changelog and manifestlog parts are terminated with empty chunks. The
        # tree and file parts are a list of entry sections. Each entry section
        # is a series of chunks terminating in an empty chunk. The list of these
        # entry sections is terminated in yet another empty chunk, so we know
        # we've reached the end of the tree/file list when we reach an empty
        # chunk that was proceeded by no non-empty chunks.

        parts = 0
        while parts < 2 + self._grouplistcount:
            noentries = True
            while True:
                chunk = getchunk(self)
                if not chunk:
                    # The first two empty chunks represent the end of the
                    # changelog and the manifestlog portions. The remaining
                    # empty chunks represent either A) the end of individual
                    # tree or file entries in the file list, or B) the end of
                    # the entire list. It's the end of the entire list if there
                    # were no entries (i.e. noentries is True).
                    if parts < 2:
                        parts += 1
                    elif noentries:
                        parts += 1
                    break
                noentries = False
                yield chunkheader(len(chunk))
                pos = 0
                while pos < len(chunk):
                    next = pos + 2**20
                    yield chunk[pos:next]
                    pos = next
            yield closechunk()

    def _unpackmanifests(self, repo, revmap, trp, prog, numchanges):
        # We know that we'll never have more manifests than we had
        # changesets.
        self.callback = prog(_('manifests'), numchanges)
        # no need to check for empty manifest group here:
        # if the result of the merge of 1 and 2 is the same in 3 and 4,
        # no new manifest will be created and the manifest group will
        # be empty during the pull
        self.manifestheader()
        deltas = self.deltaiter()
        repo.manifestlog._revlog.addgroup(deltas, revmap, trp)
        repo.ui.progress(_('manifests'), None)
        self.callback = None

    def apply(self, repo, tr, srctype, url, targetphase=phases.draft,
              expectedtotal=None):
        """Add the changegroup returned by source.read() to this repo.
        srctype is a string like 'push', 'pull', or 'unbundle'.  url is
        the URL of the repo where this changegroup is coming from.

        Return an integer summarizing the change to this repo:
        - nothing changed or no source: 0
        - more heads than before: 1+added heads (2..n)
        - fewer heads than before: -1-removed heads (-2..-n)
        - number of heads stays the same: 1
        """
        repo = repo.unfiltered()
        def csmap(x):
            repo.ui.debug("add changeset %s\n" % short(x))
            return len(cl)

        def revmap(x):
            return cl.rev(x)

        changesets = files = revisions = 0

        try:
            # The transaction may already carry source information. In this
            # case we use the top level data. We overwrite the argument
            # because we need to use the top level value (if they exist)
            # in this function.
            srctype = tr.hookargs.setdefault('source', srctype)
            url = tr.hookargs.setdefault('url', url)
            repo.hook('prechangegroup',
                      throw=True, **pycompat.strkwargs(tr.hookargs))

            # write changelog data to temp files so concurrent readers
            # will not see an inconsistent view
            cl = repo.changelog
            cl.delayupdate(tr)
            oldheads = set(cl.heads())

            trp = weakref.proxy(tr)
            # pull off the changeset group
            repo.ui.status(_("adding changesets\n"))
            clstart = len(cl)
            class prog(object):
                def __init__(self, step, total):
                    self._step = step
                    self._total = total
                    self._count = 1
                def __call__(self):
                    repo.ui.progress(self._step, self._count, unit=_('chunks'),
                                     total=self._total)
                    self._count += 1
            self.callback = prog(_('changesets'), expectedtotal)

            efiles = set()
            def onchangelog(cl, node):
                efiles.update(cl.readfiles(node))

            self.changelogheader()
            deltas = self.deltaiter()
            cgnodes = cl.addgroup(deltas, csmap, trp, addrevisioncb=onchangelog)
            efiles = len(efiles)

            if not cgnodes:
                repo.ui.develwarn('applied empty changegroup',
                                  config='warn-empty-changegroup')
            clend = len(cl)
            changesets = clend - clstart
            repo.ui.progress(_('changesets'), None)
            self.callback = None

            # pull off the manifest group
            repo.ui.status(_("adding manifests\n"))
            self._unpackmanifests(repo, revmap, trp, prog, changesets)

            needfiles = {}
            if repo.ui.configbool('server', 'validate'):
                cl = repo.changelog
                ml = repo.manifestlog
                # validate incoming csets have their manifests
                for cset in xrange(clstart, clend):
                    mfnode = cl.changelogrevision(cset).manifest
                    mfest = ml[mfnode].readdelta()
                    # store file cgnodes we must see
                    for f, n in mfest.iteritems():
                        needfiles.setdefault(f, set()).add(n)

            # process the files
            repo.ui.status(_("adding file changes\n"))
            newrevs, newfiles = _addchangegroupfiles(
                repo, self, revmap, trp, efiles, needfiles)
            revisions += newrevs
            files += newfiles

            deltaheads = 0
            if oldheads:
                heads = cl.heads()
                deltaheads = len(heads) - len(oldheads)
                for h in heads:
                    if h not in oldheads and repo[h].closesbranch():
                        deltaheads -= 1
            htext = ""
            if deltaheads:
                htext = _(" (%+d heads)") % deltaheads

            repo.ui.status(_("added %d changesets"
                             " with %d changes to %d files%s\n")
                             % (changesets, revisions, files, htext))
            repo.invalidatevolatilesets()

            if changesets > 0:
                if 'node' not in tr.hookargs:
                    tr.hookargs['node'] = hex(cl.node(clstart))
                    tr.hookargs['node_last'] = hex(cl.node(clend - 1))
                    hookargs = dict(tr.hookargs)
                else:
                    hookargs = dict(tr.hookargs)
                    hookargs['node'] = hex(cl.node(clstart))
                    hookargs['node_last'] = hex(cl.node(clend - 1))
                repo.hook('pretxnchangegroup',
                          throw=True, **pycompat.strkwargs(hookargs))

            added = [cl.node(r) for r in xrange(clstart, clend)]
            phaseall = None
            if srctype in ('push', 'serve'):
                # Old servers can not push the boundary themselves.
                # New servers won't push the boundary if changeset already
                # exists locally as secret
                #
                # We should not use added here but the list of all change in
                # the bundle
                if repo.publishing():
                    targetphase = phaseall = phases.public
                else:
                    # closer target phase computation

                    # Those changesets have been pushed from the
                    # outside, their phases are going to be pushed
                    # alongside. Therefor `targetphase` is
                    # ignored.
                    targetphase = phaseall = phases.draft
            if added:
                phases.registernew(repo, tr, targetphase, added)
            if phaseall is not None:
                phases.advanceboundary(repo, tr, phaseall, cgnodes)

            if changesets > 0:

                def runhooks():
                    # These hooks run when the lock releases, not when the
                    # transaction closes. So it's possible for the changelog
                    # to have changed since we last saw it.
                    if clstart >= len(repo):
                        return

                    repo.hook("changegroup", **pycompat.strkwargs(hookargs))

                    for n in added:
                        args = hookargs.copy()
                        args['node'] = hex(n)
                        del args['node_last']
                        repo.hook("incoming", **pycompat.strkwargs(args))

                    newheads = [h for h in repo.heads()
                                if h not in oldheads]
                    repo.ui.log("incoming",
                                "%s incoming changes - new heads: %s\n",
                                len(added),
                                ', '.join([hex(c[:6]) for c in newheads]))

                tr.addpostclose('changegroup-runhooks-%020i' % clstart,
                                lambda tr: repo._afterlock(runhooks))
        finally:
            repo.ui.flush()
        # never return 0 here:
        if deltaheads < 0:
            ret = deltaheads - 1
        else:
            ret = deltaheads + 1
        return ret

    def deltaiter(self):
        """
        returns an iterator of the deltas in this changegroup

        Useful for passing to the underlying storage system to be stored.
        """
        chain = None
        for chunkdata in iter(lambda: self.deltachunk(chain), {}):
            # Chunkdata: (node, p1, p2, cs, deltabase, delta, flags)
            yield chunkdata
            chain = chunkdata[0]

class cg2unpacker(cg1unpacker):
    """Unpacker for cg2 streams.

    cg2 streams add support for generaldelta, so the delta header
    format is slightly different. All other features about the data
    remain the same.
    """
    deltaheader = _CHANGEGROUPV2_DELTA_HEADER
    deltaheadersize = struct.calcsize(deltaheader)
    version = '02'

    def _deltaheader(self, headertuple, prevnode):
        node, p1, p2, deltabase, cs = headertuple
        flags = 0
        return node, p1, p2, deltabase, cs, flags

class cg3unpacker(cg2unpacker):
    """Unpacker for cg3 streams.

    cg3 streams add support for exchanging treemanifests and revlog
    flags. It adds the revlog flags to the delta header and an empty chunk
    separating manifests and files.
    """
    deltaheader = _CHANGEGROUPV3_DELTA_HEADER
    deltaheadersize = struct.calcsize(deltaheader)
    version = '03'
    _grouplistcount = 2 # One list of manifests and one list of files

    def _deltaheader(self, headertuple, prevnode):
        node, p1, p2, deltabase, cs, flags = headertuple
        return node, p1, p2, deltabase, cs, flags

    def _unpackmanifests(self, repo, revmap, trp, prog, numchanges):
        super(cg3unpacker, self)._unpackmanifests(repo, revmap, trp, prog,
                                                  numchanges)
        for chunkdata in iter(self.filelogheader, {}):
            # If we get here, there are directory manifests in the changegroup
            d = chunkdata["filename"]
            repo.ui.debug("adding %s revisions\n" % d)
            dirlog = repo.manifestlog._revlog.dirlog(d)
            deltas = self.deltaiter()
            if not dirlog.addgroup(deltas, revmap, trp):
                raise error.Abort(_("received dir revlog group is empty"))

class headerlessfixup(object):
    def __init__(self, fh, h):
        self._h = h
        self._fh = fh
    def read(self, n):
        if self._h:
            d, self._h = self._h[:n], self._h[n:]
            if len(d) < n:
                d += readexactly(self._fh, n - len(d))
            return d
        return readexactly(self._fh, n)

class cg1packer(object):
    deltaheader = _CHANGEGROUPV1_DELTA_HEADER
    version = '01'
    def __init__(self, repo, bundlecaps=None):
        """Given a source repo, construct a bundler.

        bundlecaps is optional and can be used to specify the set of
        capabilities which can be used to build the bundle. While bundlecaps is
        unused in core Mercurial, extensions rely on this feature to communicate
        capabilities to customize the changegroup packer.
        """
        # Set of capabilities we can use to build the bundle.
        if bundlecaps is None:
            bundlecaps = set()
        self._bundlecaps = bundlecaps
        # experimental config: bundle.reorder
        reorder = repo.ui.config('bundle', 'reorder')
        if reorder == 'auto':
            reorder = None
        else:
            reorder = util.parsebool(reorder)
        self._repo = repo
        self._reorder = reorder
        self._progress = repo.ui.progress
        if self._repo.ui.verbose and not self._repo.ui.debugflag:
            self._verbosenote = self._repo.ui.note
        else:
            self._verbosenote = lambda s: None

    def close(self):
        return closechunk()

    def fileheader(self, fname):
        return chunkheader(len(fname)) + fname

    # Extracted both for clarity and for overriding in extensions.
    def _sortgroup(self, revlog, nodelist, lookup):
        """Sort nodes for change group and turn them into revnums."""
        # for generaldelta revlogs, we linearize the revs; this will both be
        # much quicker and generate a much smaller bundle
        if (revlog._generaldelta and self._reorder is None) or self._reorder:
            dag = dagutil.revlogdag(revlog)
            return dag.linearize(set(revlog.rev(n) for n in nodelist))
        else:
            return sorted([revlog.rev(n) for n in nodelist])

    def group(self, nodelist, revlog, lookup, units=None):
        """Calculate a delta group, yielding a sequence of changegroup chunks
        (strings).

        Given a list of changeset revs, return a set of deltas and
        metadata corresponding to nodes. The first delta is
        first parent(nodelist[0]) -> nodelist[0], the receiver is
        guaranteed to have this parent as it has all history before
        these changesets. In the case firstparent is nullrev the
        changegroup starts with a full revision.

        If units is not None, progress detail will be generated, units specifies
        the type of revlog that is touched (changelog, manifest, etc.).
        """
        # if we don't have any revisions touched by these changesets, bail
        if len(nodelist) == 0:
            yield self.close()
            return

        revs = self._sortgroup(revlog, nodelist, lookup)

        # add the parent of the first rev
        p = revlog.parentrevs(revs[0])[0]
        revs.insert(0, p)

        # build deltas
        total = len(revs) - 1
        msgbundling = _('bundling')
        for r in xrange(len(revs) - 1):
            if units is not None:
                self._progress(msgbundling, r + 1, unit=units, total=total)
            prev, curr = revs[r], revs[r + 1]
            linknode = lookup(revlog.node(curr))
            for c in self.revchunk(revlog, curr, prev, linknode):
                yield c

        if units is not None:
            self._progress(msgbundling, None)
        yield self.close()

    # filter any nodes that claim to be part of the known set
    def prune(self, revlog, missing, commonrevs):
        rr, rl = revlog.rev, revlog.linkrev
        return [n for n in missing if rl(rr(n)) not in commonrevs]

    def _packmanifests(self, dir, mfnodes, lookuplinknode):
        """Pack flat manifests into a changegroup stream."""
        assert not dir
        for chunk in self.group(mfnodes, self._repo.manifestlog._revlog,
                                lookuplinknode, units=_('manifests')):
            yield chunk

    def _manifestsdone(self):
        return ''

    def generate(self, commonrevs, clnodes, fastpathlinkrev, source):
        '''yield a sequence of changegroup chunks (strings)'''
        repo = self._repo
        cl = repo.changelog

        clrevorder = {}
        mfs = {} # needed manifests
        fnodes = {} # needed file nodes
        changedfiles = set()

        # Callback for the changelog, used to collect changed files and manifest
        # nodes.
        # Returns the linkrev node (identity in the changelog case).
        def lookupcl(x):
            c = cl.read(x)
            clrevorder[x] = len(clrevorder)
            n = c[0]
            # record the first changeset introducing this manifest version
            mfs.setdefault(n, x)
            # Record a complete list of potentially-changed files in
            # this manifest.
            changedfiles.update(c[3])
            return x

        self._verbosenote(_('uncompressed size of bundle content:\n'))
        size = 0
        for chunk in self.group(clnodes, cl, lookupcl, units=_('changesets')):
            size += len(chunk)
            yield chunk
        self._verbosenote(_('%8.i (changelog)\n') % size)

        # We need to make sure that the linkrev in the changegroup refers to
        # the first changeset that introduced the manifest or file revision.
        # The fastpath is usually safer than the slowpath, because the filelogs
        # are walked in revlog order.
        #
        # When taking the slowpath with reorder=None and the manifest revlog
        # uses generaldelta, the manifest may be walked in the "wrong" order.
        # Without 'clrevorder', we would get an incorrect linkrev (see fix in
        # cc0ff93d0c0c).
        #
        # When taking the fastpath, we are only vulnerable to reordering
        # of the changelog itself. The changelog never uses generaldelta, so
        # it is only reordered when reorder=True. To handle this case, we
        # simply take the slowpath, which already has the 'clrevorder' logic.
        # This was also fixed in cc0ff93d0c0c.
        fastpathlinkrev = fastpathlinkrev and not self._reorder
        # Treemanifests don't work correctly with fastpathlinkrev
        # either, because we don't discover which directory nodes to
        # send along with files. This could probably be fixed.
        fastpathlinkrev = fastpathlinkrev and (
            'treemanifest' not in repo.requirements)

        for chunk in self.generatemanifests(commonrevs, clrevorder,
                fastpathlinkrev, mfs, fnodes, source):
            yield chunk
        mfs.clear()
        clrevs = set(cl.rev(x) for x in clnodes)

        if not fastpathlinkrev:
            def linknodes(unused, fname):
                return fnodes.get(fname, {})
        else:
            cln = cl.node
            def linknodes(filerevlog, fname):
                llr = filerevlog.linkrev
                fln = filerevlog.node
                revs = ((r, llr(r)) for r in filerevlog)
                return dict((fln(r), cln(lr)) for r, lr in revs if lr in clrevs)

        for chunk in self.generatefiles(changedfiles, linknodes, commonrevs,
                                        source):
            yield chunk

        yield self.close()

        if clnodes:
            repo.hook('outgoing', node=hex(clnodes[0]), source=source)

    def generatemanifests(self, commonrevs, clrevorder, fastpathlinkrev, mfs,
                          fnodes, source):
        """Returns an iterator of changegroup chunks containing manifests.

        `source` is unused here, but is used by extensions like remotefilelog to
        change what is sent based in pulls vs pushes, etc.
        """
        repo = self._repo
        mfl = repo.manifestlog
        dirlog = mfl._revlog.dirlog
        tmfnodes = {'': mfs}

        # Callback for the manifest, used to collect linkrevs for filelog
        # revisions.
        # Returns the linkrev node (collected in lookupcl).
        def makelookupmflinknode(dir, nodes):
            if fastpathlinkrev:
                assert not dir
                return mfs.__getitem__

            def lookupmflinknode(x):
                """Callback for looking up the linknode for manifests.

                Returns the linkrev node for the specified manifest.

                SIDE EFFECT:

                1) fclnodes gets populated with the list of relevant
                   file nodes if we're not using fastpathlinkrev
                2) When treemanifests are in use, collects treemanifest nodes
                   to send

                Note that this means manifests must be completely sent to
                the client before you can trust the list of files and
                treemanifests to send.
                """
                clnode = nodes[x]
                mdata = mfl.get(dir, x).readfast(shallow=True)
                for p, n, fl in mdata.iterentries():
                    if fl == 't': # subdirectory manifest
                        subdir = dir + p + '/'
                        tmfclnodes = tmfnodes.setdefault(subdir, {})
                        tmfclnode = tmfclnodes.setdefault(n, clnode)
                        if clrevorder[clnode] < clrevorder[tmfclnode]:
                            tmfclnodes[n] = clnode
                    else:
                        f = dir + p
                        fclnodes = fnodes.setdefault(f, {})
                        fclnode = fclnodes.setdefault(n, clnode)
                        if clrevorder[clnode] < clrevorder[fclnode]:
                            fclnodes[n] = clnode
                return clnode
            return lookupmflinknode

        size = 0
        while tmfnodes:
            dir, nodes = tmfnodes.popitem()
            prunednodes = self.prune(dirlog(dir), nodes, commonrevs)
            if not dir or prunednodes:
                for x in self._packmanifests(dir, prunednodes,
                                             makelookupmflinknode(dir, nodes)):
                    size += len(x)
                    yield x
        self._verbosenote(_('%8.i (manifests)\n') % size)
        yield self._manifestsdone()

    # The 'source' parameter is useful for extensions
    def generatefiles(self, changedfiles, linknodes, commonrevs, source):
        repo = self._repo
        progress = self._progress
        msgbundling = _('bundling')

        total = len(changedfiles)
        # for progress output
        msgfiles = _('files')
        for i, fname in enumerate(sorted(changedfiles)):
            filerevlog = repo.file(fname)
            if not filerevlog:
                raise error.Abort(_("empty or missing revlog for %s") % fname)

            linkrevnodes = linknodes(filerevlog, fname)
            # Lookup for filenodes, we collected the linkrev nodes above in the
            # fastpath case and with lookupmf in the slowpath case.
            def lookupfilelog(x):
                return linkrevnodes[x]

            filenodes = self.prune(filerevlog, linkrevnodes, commonrevs)
            if filenodes:
                progress(msgbundling, i + 1, item=fname, unit=msgfiles,
                         total=total)
                h = self.fileheader(fname)
                size = len(h)
                yield h
                for chunk in self.group(filenodes, filerevlog, lookupfilelog):
                    size += len(chunk)
                    yield chunk
                self._verbosenote(_('%8.i  %s\n') % (size, fname))
        progress(msgbundling, None)

    def deltaparent(self, revlog, rev, p1, p2, prev):
        return prev

    def revchunk(self, revlog, rev, prev, linknode):
        node = revlog.node(rev)
        p1, p2 = revlog.parentrevs(rev)
        base = self.deltaparent(revlog, rev, p1, p2, prev)

        prefix = ''
        if revlog.iscensored(base) or revlog.iscensored(rev):
            try:
                delta = revlog.revision(node, raw=True)
            except error.CensoredNodeError as e:
                delta = e.tombstone
            if base == nullrev:
                prefix = mdiff.trivialdiffheader(len(delta))
            else:
                baselen = revlog.rawsize(base)
                prefix = mdiff.replacediffheader(baselen, len(delta))
        elif base == nullrev:
            delta = revlog.revision(node, raw=True)
            prefix = mdiff.trivialdiffheader(len(delta))
        else:
            delta = revlog.revdiff(base, rev)
        p1n, p2n = revlog.parents(node)
        basenode = revlog.node(base)
        flags = revlog.flags(rev)
        meta = self.builddeltaheader(node, p1n, p2n, basenode, linknode, flags)
        meta += prefix
        l = len(meta) + len(delta)
        yield chunkheader(l)
        yield meta
        yield delta
    def builddeltaheader(self, node, p1n, p2n, basenode, linknode, flags):
        # do nothing with basenode, it is implicitly the previous one in HG10
        # do nothing with flags, it is implicitly 0 for cg1 and cg2
        return struct.pack(self.deltaheader, node, p1n, p2n, linknode)

class cg2packer(cg1packer):
    version = '02'
    deltaheader = _CHANGEGROUPV2_DELTA_HEADER

    def __init__(self, repo, bundlecaps=None):
        super(cg2packer, self).__init__(repo, bundlecaps)
        if self._reorder is None:
            # Since generaldelta is directly supported by cg2, reordering
            # generally doesn't help, so we disable it by default (treating
            # bundle.reorder=auto just like bundle.reorder=False).
            self._reorder = False

    def deltaparent(self, revlog, rev, p1, p2, prev):
        dp = revlog.deltaparent(rev)
        if dp == nullrev and revlog.storedeltachains:
            # Avoid sending full revisions when delta parent is null. Pick prev
            # in that case. It's tempting to pick p1 in this case, as p1 will
            # be smaller in the common case. However, computing a delta against
            # p1 may require resolving the raw text of p1, which could be
            # expensive. The revlog caches should have prev cached, meaning
            # less CPU for changegroup generation. There is likely room to add
            # a flag and/or config option to control this behavior.
            return prev
        elif dp == nullrev:
            # revlog is configured to use full snapshot for a reason,
            # stick to full snapshot.
            return nullrev
        elif dp not in (p1, p2, prev):
            # Pick prev when we can't be sure remote has the base revision.
            return prev
        else:
            return dp

    def builddeltaheader(self, node, p1n, p2n, basenode, linknode, flags):
        # Do nothing with flags, it is implicitly 0 in cg1 and cg2
        return struct.pack(self.deltaheader, node, p1n, p2n, basenode, linknode)

class cg3packer(cg2packer):
    version = '03'
    deltaheader = _CHANGEGROUPV3_DELTA_HEADER

    def _packmanifests(self, dir, mfnodes, lookuplinknode):
        if dir:
            yield self.fileheader(dir)

        dirlog = self._repo.manifestlog._revlog.dirlog(dir)
        for chunk in self.group(mfnodes, dirlog, lookuplinknode,
                                units=_('manifests')):
            yield chunk

    def _manifestsdone(self):
        return self.close()

    def builddeltaheader(self, node, p1n, p2n, basenode, linknode, flags):
        return struct.pack(
            self.deltaheader, node, p1n, p2n, basenode, linknode, flags)

_packermap = {'01': (cg1packer, cg1unpacker),
             # cg2 adds support for exchanging generaldelta
             '02': (cg2packer, cg2unpacker),
             # cg3 adds support for exchanging revlog flags and treemanifests
             '03': (cg3packer, cg3unpacker),
}

def allsupportedversions(repo):
    versions = set(_packermap.keys())
    if not (repo.ui.configbool('experimental', 'changegroup3') or
            repo.ui.configbool('experimental', 'treemanifest') or
            'treemanifest' in repo.requirements):
        versions.discard('03')
    return versions

# Changegroup versions that can be applied to the repo
def supportedincomingversions(repo):
    return allsupportedversions(repo)

# Changegroup versions that can be created from the repo
def supportedoutgoingversions(repo):
    versions = allsupportedversions(repo)
    if 'treemanifest' in repo.requirements:
        # Versions 01 and 02 support only flat manifests and it's just too
        # expensive to convert between the flat manifest and tree manifest on
        # the fly. Since tree manifests are hashed differently, all of history
        # would have to be converted. Instead, we simply don't even pretend to
        # support versions 01 and 02.
        versions.discard('01')
        versions.discard('02')
    return versions

def localversion(repo):
    # Finds the best version to use for bundles that are meant to be used
    # locally, such as those from strip and shelve, and temporary bundles.
    return max(supportedoutgoingversions(repo))

def safeversion(repo):
    # Finds the smallest version that it's safe to assume clients of the repo
    # will support. For example, all hg versions that support generaldelta also
    # support changegroup 02.
    versions = supportedoutgoingversions(repo)
    if 'generaldelta' in repo.requirements:
        versions.discard('01')
    assert versions
    return min(versions)

def getbundler(version, repo, bundlecaps=None):
    assert version in supportedoutgoingversions(repo)
    return _packermap[version][0](repo, bundlecaps)

def getunbundler(version, fh, alg, extras=None):
    return _packermap[version][1](fh, alg, extras=extras)

def _changegroupinfo(repo, nodes, source):
    if repo.ui.verbose or source == 'bundle':
        repo.ui.status(_("%d changesets found\n") % len(nodes))
    if repo.ui.debugflag:
        repo.ui.debug("list of changesets:\n")
        for node in nodes:
            repo.ui.debug("%s\n" % hex(node))

def makechangegroup(repo, outgoing, version, source, fastpath=False,
                    bundlecaps=None):
    cgstream = makestream(repo, outgoing, version, source,
                          fastpath=fastpath, bundlecaps=bundlecaps)
    return getunbundler(version, util.chunkbuffer(cgstream), None,
                        {'clcount': len(outgoing.missing) })

def makestream(repo, outgoing, version, source, fastpath=False,
               bundlecaps=None):
    bundler = getbundler(version, repo, bundlecaps=bundlecaps)

    repo = repo.unfiltered()
    commonrevs = outgoing.common
    csets = outgoing.missing
    heads = outgoing.missingheads
    # We go through the fast path if we get told to, or if all (unfiltered
    # heads have been requested (since we then know there all linkrevs will
    # be pulled by the client).
    heads.sort()
    fastpathlinkrev = fastpath or (
            repo.filtername is None and heads == sorted(repo.heads()))

    repo.hook('preoutgoing', throw=True, source=source)
    _changegroupinfo(repo, csets, source)
    return bundler.generate(commonrevs, csets, fastpathlinkrev, source)

def _addchangegroupfiles(repo, source, revmap, trp, expectedfiles, needfiles):
    revisions = 0
    files = 0
    for chunkdata in iter(source.filelogheader, {}):
        files += 1
        f = chunkdata["filename"]
        repo.ui.debug("adding %s revisions\n" % f)
        repo.ui.progress(_('files'), files, unit=_('files'),
                         total=expectedfiles)
        fl = repo.file(f)
        o = len(fl)
        try:
            deltas = source.deltaiter()
            if not fl.addgroup(deltas, revmap, trp):
                raise error.Abort(_("received file revlog group is empty"))
        except error.CensoredBaseError as e:
            raise error.Abort(_("received delta base is censored: %s") % e)
        revisions += len(fl) - o
        if f in needfiles:
            needs = needfiles[f]
            for new in xrange(o, len(fl)):
                n = fl.node(new)
                if n in needs:
                    needs.remove(n)
                else:
                    raise error.Abort(
                        _("received spurious file revlog entry"))
            if not needs:
                del needfiles[f]
    repo.ui.progress(_('files'), None)

    for f, needs in needfiles.iteritems():
        fl = repo.file(f)
        for n in needs:
            try:
                fl.rev(n)
            except error.LookupError:
                raise error.Abort(
                    _('missing file data for %s:%s - run hg verify') %
                    (f, hex(n)))

    return revisions, files