view hgext/remotefilelog/datapack.py @ 51420:ac1c75188440

phases: invalidate the phases set less often on retract boundary

We already have the information to update the phase set, so we do so directly
instead of invalidating the cache.

This show a sizeable speedup in our `perf::unbundle` benchmark on the
many-draft mozilla-try repository.

### data-env-vars.name            = mozilla-try-2023-03-22-zstd-sparse-revlog
  # benchmark.name                = hg.perf.perf-unbundle
  # bin-env-vars.hg.flavor        = no-rust
  # bin-env-vars.hg.py-re2-module = default
  # benchmark.variants.issue6528  = disabled

  # benchmark.variants.revs = last-10
before: 2.055259 seconds
after:  1.887064 seconds (-8.18%)
  # benchmark.variants.revs = last-100
before: 2.409239 seconds
after:  2.222429 seconds (-7.75%)
  # benchmark.variants.revs = last-1000
before: 3.945648 seconds
after:  3.762480 seconds (-4.64%)

import struct
import zlib

from mercurial.node import (
    hex,
    sha1nodeconstants,
)
from mercurial.i18n import _
from mercurial import (
    util,
)
from . import (
    basepack,
    constants,
    shallowutil,
)

NODELENGTH = 20

# The indicator value in the index for a fulltext entry.
FULLTEXTINDEXMARK = -1
NOBASEINDEXMARK = -2

INDEXSUFFIX = b'.dataidx'
PACKSUFFIX = b'.datapack'


class datapackstore(basepack.basepackstore):
    INDEXSUFFIX = INDEXSUFFIX
    PACKSUFFIX = PACKSUFFIX

    def __init__(self, ui, path):
        super(datapackstore, self).__init__(ui, path)

    def getpack(self, path):
        return datapack(path)

    def get(self, name, node):
        raise RuntimeError(b"must use getdeltachain with datapackstore")

    def getmeta(self, name, node):
        for pack in self.packs:
            try:
                return pack.getmeta(name, node)
            except KeyError:
                pass

        for pack in self.refresh():
            try:
                return pack.getmeta(name, node)
            except KeyError:
                pass

        raise KeyError((name, hex(node)))

    def getdelta(self, name, node):
        for pack in self.packs:
            try:
                return pack.getdelta(name, node)
            except KeyError:
                pass

        for pack in self.refresh():
            try:
                return pack.getdelta(name, node)
            except KeyError:
                pass

        raise KeyError((name, hex(node)))

    def getdeltachain(self, name, node):
        for pack in self.packs:
            try:
                return pack.getdeltachain(name, node)
            except KeyError:
                pass

        for pack in self.refresh():
            try:
                return pack.getdeltachain(name, node)
            except KeyError:
                pass

        raise KeyError((name, hex(node)))

    def add(self, name, node, data):
        raise RuntimeError(b"cannot add to datapackstore")


class datapack(basepack.basepack):
    INDEXSUFFIX = INDEXSUFFIX
    PACKSUFFIX = PACKSUFFIX

    # Format is <node><delta offset><pack data offset><pack data size>
    # See the mutabledatapack doccomment for more details.
    INDEXFORMAT = b'!20siQQ'
    INDEXENTRYLENGTH = 40

    SUPPORTED_VERSIONS = [2]

    def getmissing(self, keys):
        missing = []
        for name, node in keys:
            value = self._find(node)
            if not value:
                missing.append((name, node))

        return missing

    def get(self, name, node):
        raise RuntimeError(
            b"must use getdeltachain with datapack (%s:%s)" % (name, hex(node))
        )

    def getmeta(self, name, node):
        value = self._find(node)
        if value is None:
            raise KeyError((name, hex(node)))

        node, deltabaseoffset, offset, size = value
        rawentry = self._data[offset : offset + size]

        # see docstring of mutabledatapack for the format
        offset = 0
        offset += struct.unpack_from(b'!H', rawentry, offset)[0] + 2  # filename
        offset += 40  # node, deltabase node
        offset += struct.unpack_from(b'!Q', rawentry, offset)[0] + 8  # delta

        metalen = struct.unpack_from(b'!I', rawentry, offset)[0]
        offset += 4

        meta = shallowutil.parsepackmeta(rawentry[offset : offset + metalen])

        return meta

    def getdelta(self, name, node):
        value = self._find(node)
        if value is None:
            raise KeyError((name, hex(node)))

        node, deltabaseoffset, offset, size = value
        entry = self._readentry(offset, size, getmeta=True)
        filename, node, deltabasenode, delta, meta = entry

        # If we've read a lot of data from the mmap, free some memory.
        self.freememory()

        return delta, filename, deltabasenode, meta

    def getdeltachain(self, name, node):
        value = self._find(node)
        if value is None:
            raise KeyError((name, hex(node)))

        params = self.params

        # Precompute chains
        chain = [value]
        deltabaseoffset = value[1]
        entrylen = self.INDEXENTRYLENGTH
        while (
            deltabaseoffset != FULLTEXTINDEXMARK
            and deltabaseoffset != NOBASEINDEXMARK
        ):
            loc = params.indexstart + deltabaseoffset
            value = struct.unpack(
                self.INDEXFORMAT, self._index[loc : loc + entrylen]
            )
            deltabaseoffset = value[1]
            chain.append(value)

        # Read chain data
        deltachain = []
        for node, deltabaseoffset, offset, size in chain:
            filename, node, deltabasenode, delta = self._readentry(offset, size)
            deltachain.append((filename, node, filename, deltabasenode, delta))

        # If we've read a lot of data from the mmap, free some memory.
        self.freememory()

        return deltachain

    def _readentry(self, offset, size, getmeta=False):
        rawentry = self._data[offset : offset + size]
        self._pagedin += len(rawentry)

        # <2 byte len> + <filename>
        lengthsize = 2
        filenamelen = struct.unpack(b'!H', rawentry[:2])[0]
        filename = rawentry[lengthsize : lengthsize + filenamelen]

        # <20 byte node> + <20 byte deltabase>
        nodestart = lengthsize + filenamelen
        deltabasestart = nodestart + NODELENGTH
        node = rawentry[nodestart:deltabasestart]
        deltabasenode = rawentry[deltabasestart : deltabasestart + NODELENGTH]

        # <8 byte len> + <delta>
        deltastart = deltabasestart + NODELENGTH
        rawdeltalen = rawentry[deltastart : deltastart + 8]
        deltalen = struct.unpack(b'!Q', rawdeltalen)[0]

        delta = rawentry[deltastart + 8 : deltastart + 8 + deltalen]
        delta = self._decompress(delta)

        if getmeta:
            metastart = deltastart + 8 + deltalen
            metalen = struct.unpack_from(b'!I', rawentry, metastart)[0]

            rawmeta = rawentry[metastart + 4 : metastart + 4 + metalen]
            meta = shallowutil.parsepackmeta(rawmeta)
            return filename, node, deltabasenode, delta, meta
        else:
            return filename, node, deltabasenode, delta

    def _decompress(self, data):
        return zlib.decompress(data)

    def add(self, name, node, data):
        raise RuntimeError(b"cannot add to datapack (%s:%s)" % (name, node))

    def _find(self, node):
        params = self.params
        fanoutkey = struct.unpack(
            params.fanoutstruct, node[: params.fanoutprefix]
        )[0]
        fanout = self._fanouttable

        start = fanout[fanoutkey] + params.indexstart
        indexend = self._indexend

        # Scan forward to find the first non-same entry, which is the upper
        # bound.
        for i in range(fanoutkey + 1, params.fanoutcount):
            end = fanout[i] + params.indexstart
            if end != start:
                break
        else:
            end = indexend

        # Bisect between start and end to find node
        index = self._index
        startnode = index[start : start + NODELENGTH]
        endnode = index[end : end + NODELENGTH]
        entrylen = self.INDEXENTRYLENGTH
        if startnode == node:
            entry = index[start : start + entrylen]
        elif endnode == node:
            entry = index[end : end + entrylen]
        else:
            while start < end - entrylen:
                mid = start + (end - start) // 2
                mid = mid - ((mid - params.indexstart) % entrylen)
                midnode = index[mid : mid + NODELENGTH]
                if midnode == node:
                    entry = index[mid : mid + entrylen]
                    break
                if node > midnode:
                    start = mid
                elif node < midnode:
                    end = mid
            else:
                return None

        return struct.unpack(self.INDEXFORMAT, entry)

    def markledger(self, ledger, options=None):
        for filename, node in self:
            ledger.markdataentry(self, filename, node)

    def cleanup(self, ledger):
        entries = ledger.sources.get(self, [])
        allkeys = set(self)
        repackedkeys = {
            (e.filename, e.node) for e in entries if e.datarepacked or e.gced
        }

        if len(allkeys - repackedkeys) == 0:
            if self.path not in ledger.created:
                util.unlinkpath(self.indexpath, ignoremissing=True)
                util.unlinkpath(self.packpath, ignoremissing=True)

    def __iter__(self):
        for f, n, deltabase, deltalen in self.iterentries():
            yield f, n

    def iterentries(self):
        # Start at 1 to skip the header
        offset = 1
        data = self._data
        while offset < self.datasize:
            oldoffset = offset

            # <2 byte len> + <filename>
            filenamelen = struct.unpack(b'!H', data[offset : offset + 2])[0]
            offset += 2
            filename = data[offset : offset + filenamelen]
            offset += filenamelen

            # <20 byte node>
            node = data[offset : offset + constants.NODESIZE]
            offset += constants.NODESIZE
            # <20 byte deltabase>
            deltabase = data[offset : offset + constants.NODESIZE]
            offset += constants.NODESIZE

            # <8 byte len> + <delta>
            rawdeltalen = data[offset : offset + 8]
            deltalen = struct.unpack(b'!Q', rawdeltalen)[0]
            offset += 8

            # TODO(augie): we should store a header that is the
            # uncompressed size.
            uncompressedlen = len(
                self._decompress(data[offset : offset + deltalen])
            )
            offset += deltalen

            # <4 byte len> + <metadata-list>
            metalen = struct.unpack_from(b'!I', data, offset)[0]
            offset += 4 + metalen

            yield (filename, node, deltabase, uncompressedlen)

            # If we've read a lot of data from the mmap, free some memory.
            self._pagedin += offset - oldoffset
            if self.freememory():
                data = self._data


class mutabledatapack(basepack.mutablebasepack):
    """A class for constructing and serializing a datapack file and index.

    A datapack is a pair of files that contain the revision contents for various
    file revisions in Mercurial. It contains only revision contents (like file
    contents), not any history information.

    It consists of two files, with the following format. All bytes are in
    network byte order (big endian).

    .datapack
        The pack itself is a series of revision deltas with some basic header
        information on each. A revision delta may be a fulltext, represented by
        a deltabasenode equal to the nullid.

        datapack = <version: 1 byte>
                   [<revision>,...]
        revision = <filename len: 2 byte unsigned int>
                   <filename>
                   <node: 20 byte>
                   <deltabasenode: 20 byte>
                   <delta len: 8 byte unsigned int>
                   <delta>
                   <metadata-list len: 4 byte unsigned int> [1]
                   <metadata-list>                          [1]
        metadata-list = [<metadata-item>, ...]
        metadata-item = <metadata-key: 1 byte>
                        <metadata-value len: 2 byte unsigned>
                        <metadata-value>

        metadata-key could be METAKEYFLAG or METAKEYSIZE or other single byte
        value in the future.

    .dataidx
        The index file consists of two parts, the fanout and the index.

        The index is a list of index entries, sorted by node (one per revision
        in the pack). Each entry has:

        - node (The 20 byte node of the entry; i.e. the commit hash, file node
                hash, etc)
        - deltabase index offset (The location in the index of the deltabase for
                                  this entry. The deltabase is the next delta in
                                  the chain, with the chain eventually
                                  terminating in a full-text, represented by a
                                  deltabase offset of -1. This lets us compute
                                  delta chains from the index, then do
                                  sequential reads from the pack if the revision
                                  are nearby on disk.)
        - pack entry offset (The location of this entry in the datapack)
        - pack content size (The on-disk length of this entry's pack data)

        The fanout is a quick lookup table to reduce the number of steps for
        bisecting the index. It is a series of 4 byte pointers to positions
        within the index. It has 2^16 entries, which corresponds to hash
        prefixes [0000, 0001,..., FFFE, FFFF]. Example: the pointer in slot
        4F0A points to the index position of the first revision whose node
        starts with 4F0A. This saves log(2^16)=16 bisect steps.

        dataidx = <fanouttable>
                  <index>
        fanouttable = [<index offset: 4 byte unsigned int>,...] (2^16 entries)
        index = [<index entry>,...]
        indexentry = <node: 20 byte>
                     <deltabase location: 4 byte signed int>
                     <pack entry offset: 8 byte unsigned int>
                     <pack entry size: 8 byte unsigned int>

    [1]: new in version 1.
    """

    INDEXSUFFIX = INDEXSUFFIX
    PACKSUFFIX = PACKSUFFIX

    # v[01] index format: <node><delta offset><pack data offset><pack data size>
    INDEXFORMAT = datapack.INDEXFORMAT
    INDEXENTRYLENGTH = datapack.INDEXENTRYLENGTH

    # v1 has metadata support
    SUPPORTED_VERSIONS = [2]

    def _compress(self, data):
        return zlib.compress(data)

    def add(self, name, node, deltabasenode, delta, metadata=None):
        # metadata is a dict, ex. {METAKEYFLAG: flag}
        if len(name) > 2 ** 16:
            raise RuntimeError(_(b"name too long %s") % name)
        if len(node) != 20:
            raise RuntimeError(_(b"node should be 20 bytes %s") % node)

        if node in self.entries:
            # The revision has already been added
            return

        # TODO: allow configurable compression
        delta = self._compress(delta)

        rawdata = b''.join(
            (
                struct.pack(b'!H', len(name)),  # unsigned 2 byte int
                name,
                node,
                deltabasenode,
                struct.pack(b'!Q', len(delta)),  # unsigned 8 byte int
                delta,
            )
        )

        # v1 support metadata
        rawmeta = shallowutil.buildpackmeta(metadata)
        rawdata += struct.pack(b'!I', len(rawmeta))  # unsigned 4 byte
        rawdata += rawmeta

        offset = self.packfp.tell()

        size = len(rawdata)

        self.entries[node] = (deltabasenode, offset, size)

        self.writeraw(rawdata)

    def createindex(self, nodelocations, indexoffset):
        entries = sorted(
            (n, db, o, s) for n, (db, o, s) in self.entries.items()
        )

        rawindex = b''
        fmt = self.INDEXFORMAT
        for node, deltabase, offset, size in entries:
            if deltabase == sha1nodeconstants.nullid:
                deltabaselocation = FULLTEXTINDEXMARK
            else:
                # Instead of storing the deltabase node in the index, let's
                # store a pointer directly to the index entry for the deltabase.
                deltabaselocation = nodelocations.get(
                    deltabase, NOBASEINDEXMARK
                )

            entry = struct.pack(fmt, node, deltabaselocation, offset, size)
            rawindex += entry

        return rawindex