revlogutils: for issue6528 fix, pre-cache nullrev as metadata-free

mercurial/revlogutils/rewrite.py

@@ -1,886 +1,886
 # censor code related to censoring revision
 # coding: utf8
 #
 # Copyright 2021 Pierre-Yves David <[e-mail]>
 # Copyright 2015 Google, Inc <[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 annotations
 
 import binascii
 import contextlib
 import os
 import struct
 
 from ..node import (
     nullrev,
 )
 from .constants import (
     COMP_MODE_PLAIN,
     ENTRY_DATA_COMPRESSED_LENGTH,
     ENTRY_DATA_COMPRESSION_MODE,
     ENTRY_DATA_OFFSET,
     ENTRY_DATA_UNCOMPRESSED_LENGTH,
     ENTRY_DELTA_BASE,
     ENTRY_LINK_REV,
     ENTRY_NODE_ID,
     ENTRY_PARENT_1,
     ENTRY_PARENT_2,
     ENTRY_SIDEDATA_COMPRESSED_LENGTH,
     ENTRY_SIDEDATA_COMPRESSION_MODE,
     ENTRY_SIDEDATA_OFFSET,
     REVIDX_ISCENSORED,
     REVLOGV0,
     REVLOGV1,
 )
 from ..i18n import _
 
 from .. import (
     error,
     mdiff,
     pycompat,
     revlogutils,
     util,
 )
 from ..utils import (
     storageutil,
 )
 from . import (
     constants,
     deltas,
 )
 
 
 def v1_censor(rl, tr, censor_nodes, tombstone=b''):
     """censors a revision in a "version 1" revlog"""
     assert rl._format_version == constants.REVLOGV1, rl._format_version
 
     # avoid cycle
     from .. import revlog
 
     censor_revs = set(rl.rev(node) for node in censor_nodes)
     tombstone = storageutil.packmeta({b'censored': tombstone}, b'')
 
     # Rewriting the revlog in place is hard. Our strategy for censoring is
     # to create a new revlog, copy all revisions to it, then replace the
     # revlogs on transaction close.
     #
     # This is a bit dangerous. We could easily have a mismatch of state.
     newrl = revlog.revlog(
         rl.opener,
         target=rl.target,
         radix=rl.radix,
         postfix=b'tmpcensored',
         censorable=True,
         data_config=rl.data_config,
         delta_config=rl.delta_config,
         feature_config=rl.feature_config,
         may_inline=rl._inline,
     )
     # inline splitting will prepare some transaction work that will get
     # confused by the final file move. So if there is a risk of not being
     # inline at the end, we prevent the new revlog to be inline in the first
     # place.
     assert not (newrl._inline and not rl._inline)
 
     for rev in rl.revs():
         node = rl.node(rev)
         p1, p2 = rl.parents(node)
 
         if rev in censor_revs:
             newrl.addrawrevision(
                 tombstone,
                 tr,
                 rl.linkrev(rev),
                 p1,
                 p2,
                 node,
                 constants.REVIDX_ISCENSORED,
             )
 
             if newrl.deltaparent(rev) != nullrev:
                 m = _(b'censored revision stored as delta; cannot censor')
                 h = _(
                     b'censoring of revlogs is not fully implemented;'
                     b' please report this bug'
                 )
                 raise error.Abort(m, hint=h)
             continue
 
         if rl.iscensored(rev):
             if rl.deltaparent(rev) != nullrev:
                 m = _(
                     b'cannot censor due to censored '
                     b'revision having delta stored'
                 )
                 raise error.Abort(m)
             rawtext = rl._inner._chunk(rev)
         else:
             rawtext = rl.rawdata(rev)
 
         newrl.addrawrevision(
             rawtext, tr, rl.linkrev(rev), p1, p2, node, rl.flags(rev)
         )
 
     tr.addbackup(rl._indexfile, location=b'store')
     if not rl._inline:
         tr.addbackup(rl._datafile, location=b'store')
 
     rl.opener.rename(newrl._indexfile, rl._indexfile)
     if newrl._inline:
         assert rl._inline
     else:
         assert not rl._inline
         rl.opener.rename(newrl._datafile, rl._datafile)
 
     rl.clearcaches()
     chunk_cache = rl._loadindex()
     rl._load_inner(chunk_cache)
 
 
 def v2_censor(revlog, tr, censor_nodes, tombstone=b''):
     """censors a revision in a "version 2" revlog"""
     assert revlog._format_version != REVLOGV0, revlog._format_version
     assert revlog._format_version != REVLOGV1, revlog._format_version
 
     censor_revs = {revlog.rev(node) for node in censor_nodes}
     _rewrite_v2(revlog, tr, censor_revs, tombstone)
 
 
 def _rewrite_v2(revlog, tr, censor_revs, tombstone=b''):
     """rewrite a revlog to censor some of its content
 
     General principle
 
     We create new revlog files (index/data/sidedata) to copy the content of
     the existing data without the censored data.
 
     We need to recompute new delta for any revision that used the censored
     revision as delta base. As the cumulative size of the new delta may be
     large, we store them in a temporary file until they are stored in their
     final destination.
 
     All data before the censored data can be blindly copied. The rest needs
     to be copied as we go and the associated index entry needs adjustement.
     """
     assert revlog._format_version != REVLOGV0, revlog._format_version
     assert revlog._format_version != REVLOGV1, revlog._format_version
 
     old_index = revlog.index
     docket = revlog._docket
 
     tombstone = storageutil.packmeta({b'censored': tombstone}, b'')
 
     first_excl_rev = min(censor_revs)
 
     first_excl_entry = revlog.index[first_excl_rev]
     index_cutoff = revlog.index.entry_size * first_excl_rev
     data_cutoff = first_excl_entry[ENTRY_DATA_OFFSET] >> 16
     sidedata_cutoff = revlog.sidedata_cut_off(first_excl_rev)
 
     with pycompat.unnamedtempfile(mode=b"w+b") as tmp_storage:
         # rev → (new_base, data_start, data_end, compression_mode)
         rewritten_entries = _precompute_rewritten_delta(
             revlog,
             old_index,
             censor_revs,
             tmp_storage,
         )
 
         all_files = _setup_new_files(
             revlog,
             index_cutoff,
             data_cutoff,
             sidedata_cutoff,
         )
 
         # we dont need to open the old index file since its content already
         # exist in a usable form in `old_index`.
         with all_files() as open_files:
             (
                 old_data_file,
                 old_sidedata_file,
                 new_index_file,
                 new_data_file,
                 new_sidedata_file,
             ) = open_files
 
             # writing the censored revision
 
             # Writing all subsequent revisions
             for rev in range(first_excl_rev, len(old_index)):
                 if rev in censor_revs:
                     _rewrite_censor(
                         revlog,
                         old_index,
                         open_files,
                         rev,
                         tombstone,
                     )
                 else:
                     _rewrite_simple(
                         revlog,
                         old_index,
                         open_files,
                         rev,
                         rewritten_entries,
                         tmp_storage,
                     )
     docket.write(transaction=None, stripping=True)
 
 
 def _precompute_rewritten_delta(
     revlog,
     old_index,
     excluded_revs,
     tmp_storage,
 ):
     """Compute new delta for revisions whose delta is based on revision that
     will not survive as is.
 
     Return a mapping: {rev → (new_base, data_start, data_end, compression_mode)}
     """
     dc = deltas.deltacomputer(revlog)
     rewritten_entries = {}
     first_excl_rev = min(excluded_revs)
     with revlog.reading():
         for rev in range(first_excl_rev, len(old_index)):
             if rev in excluded_revs:
                 # this revision will be preserved as is, so we don't need to
                 # consider recomputing a delta.
                 continue
             entry = old_index[rev]
             if entry[ENTRY_DELTA_BASE] not in excluded_revs:
                 continue
             # This is a revision that use the censored revision as the base
             # for its delta. We need a need new deltas
             if entry[ENTRY_DATA_UNCOMPRESSED_LENGTH] == 0:
                 # this revision is empty, we can delta against nullrev
                 rewritten_entries[rev] = (nullrev, 0, 0, COMP_MODE_PLAIN)
             else:
                 text = revlog.rawdata(rev)
                 info = revlogutils.revisioninfo(
                     node=entry[ENTRY_NODE_ID],
                     p1=revlog.node(entry[ENTRY_PARENT_1]),
                     p2=revlog.node(entry[ENTRY_PARENT_2]),
                     btext=[text],
                     textlen=len(text),
                     cachedelta=None,
                     flags=entry[ENTRY_DATA_OFFSET] & 0xFFFF,
                 )
                 d = dc.finddeltainfo(
                     info, excluded_bases=excluded_revs, target_rev=rev
                 )
                 default_comp = revlog._docket.default_compression_header
                 comp_mode, d = deltas.delta_compression(default_comp, d)
                 # using `tell` is a bit lazy, but we are not here for speed
                 start = tmp_storage.tell()
                 tmp_storage.write(d.data[1])
                 end = tmp_storage.tell()
                 rewritten_entries[rev] = (d.base, start, end, comp_mode)
     return rewritten_entries
 
 
 def _setup_new_files(
     revlog,
     index_cutoff,
     data_cutoff,
     sidedata_cutoff,
 ):
     """
 
     return a context manager to open all the relevant files:
     - old_data_file,
     - old_sidedata_file,
     - new_index_file,
     - new_data_file,
     - new_sidedata_file,
 
     The old_index_file is not here because it is accessed through the
     `old_index` object if the caller function.
     """
     docket = revlog._docket
     old_index_filepath = revlog.opener.join(docket.index_filepath())
     old_data_filepath = revlog.opener.join(docket.data_filepath())
     old_sidedata_filepath = revlog.opener.join(docket.sidedata_filepath())
 
     new_index_filepath = revlog.opener.join(docket.new_index_file())
     new_data_filepath = revlog.opener.join(docket.new_data_file())
     new_sidedata_filepath = revlog.opener.join(docket.new_sidedata_file())
 
     util.copyfile(old_index_filepath, new_index_filepath, nb_bytes=index_cutoff)
     util.copyfile(old_data_filepath, new_data_filepath, nb_bytes=data_cutoff)
     util.copyfile(
         old_sidedata_filepath,
         new_sidedata_filepath,
         nb_bytes=sidedata_cutoff,
     )
     revlog.opener.register_file(docket.index_filepath())
     revlog.opener.register_file(docket.data_filepath())
     revlog.opener.register_file(docket.sidedata_filepath())
 
     docket.index_end = index_cutoff
     docket.data_end = data_cutoff
     docket.sidedata_end = sidedata_cutoff
 
     # reload the revlog internal information
     revlog.clearcaches()
     revlog._loadindex(docket=docket)
 
     @contextlib.contextmanager
     def all_files_opener():
         # hide opening in an helper function to please check-code, black
         # and various python version at the same time
         with open(old_data_filepath, 'rb') as old_data_file:
             with open(old_sidedata_filepath, 'rb') as old_sidedata_file:
                 with open(new_index_filepath, 'r+b') as new_index_file:
                     with open(new_data_filepath, 'r+b') as new_data_file:
                         with open(
                             new_sidedata_filepath, 'r+b'
                         ) as new_sidedata_file:
                             new_index_file.seek(0, os.SEEK_END)
                             assert new_index_file.tell() == index_cutoff
                             new_data_file.seek(0, os.SEEK_END)
                             assert new_data_file.tell() == data_cutoff
                             new_sidedata_file.seek(0, os.SEEK_END)
                             assert new_sidedata_file.tell() == sidedata_cutoff
                             yield (
                                 old_data_file,
                                 old_sidedata_file,
                                 new_index_file,
                                 new_data_file,
                                 new_sidedata_file,
                             )
 
     return all_files_opener
 
 
 def _rewrite_simple(
     revlog,
     old_index,
     all_files,
     rev,
     rewritten_entries,
     tmp_storage,
 ):
     """append a normal revision to the index after the rewritten one(s)"""
     (
         old_data_file,
         old_sidedata_file,
         new_index_file,
         new_data_file,
         new_sidedata_file,
     ) = all_files
     entry = old_index[rev]
     flags = entry[ENTRY_DATA_OFFSET] & 0xFFFF
     old_data_offset = entry[ENTRY_DATA_OFFSET] >> 16
 
     if rev not in rewritten_entries:
         old_data_file.seek(old_data_offset)
         new_data_size = entry[ENTRY_DATA_COMPRESSED_LENGTH]
         new_data = old_data_file.read(new_data_size)
         data_delta_base = entry[ENTRY_DELTA_BASE]
         d_comp_mode = entry[ENTRY_DATA_COMPRESSION_MODE]
     else:
         (
             data_delta_base,
             start,
             end,
             d_comp_mode,
         ) = rewritten_entries[rev]
         new_data_size = end - start
         tmp_storage.seek(start)
         new_data = tmp_storage.read(new_data_size)
 
     # It might be faster to group continuous read/write operation,
     # however, this is censor, an operation that is not focussed
     # around stellar performance. So I have not written this
     # optimisation yet.
     new_data_offset = new_data_file.tell()
     new_data_file.write(new_data)
 
     sidedata_size = entry[ENTRY_SIDEDATA_COMPRESSED_LENGTH]
     new_sidedata_offset = new_sidedata_file.tell()
     if 0 < sidedata_size:
         old_sidedata_offset = entry[ENTRY_SIDEDATA_OFFSET]
         old_sidedata_file.seek(old_sidedata_offset)
         new_sidedata = old_sidedata_file.read(sidedata_size)
         new_sidedata_file.write(new_sidedata)
 
     data_uncompressed_length = entry[ENTRY_DATA_UNCOMPRESSED_LENGTH]
     sd_com_mode = entry[ENTRY_SIDEDATA_COMPRESSION_MODE]
     assert data_delta_base <= rev, (data_delta_base, rev)
 
     new_entry = revlogutils.entry(
         flags=flags,
         data_offset=new_data_offset,
         data_compressed_length=new_data_size,
         data_uncompressed_length=data_uncompressed_length,
         data_delta_base=data_delta_base,
         link_rev=entry[ENTRY_LINK_REV],
         parent_rev_1=entry[ENTRY_PARENT_1],
         parent_rev_2=entry[ENTRY_PARENT_2],
         node_id=entry[ENTRY_NODE_ID],
         sidedata_offset=new_sidedata_offset,
         sidedata_compressed_length=sidedata_size,
         data_compression_mode=d_comp_mode,
         sidedata_compression_mode=sd_com_mode,
     )
     revlog.index.append(new_entry)
     entry_bin = revlog.index.entry_binary(rev)
     new_index_file.write(entry_bin)
 
     revlog._docket.index_end = new_index_file.tell()
     revlog._docket.data_end = new_data_file.tell()
     revlog._docket.sidedata_end = new_sidedata_file.tell()
 
 
 def _rewrite_censor(
     revlog,
     old_index,
     all_files,
     rev,
     tombstone,
 ):
     """rewrite and append a censored revision"""
     (
         old_data_file,
         old_sidedata_file,
         new_index_file,
         new_data_file,
         new_sidedata_file,
     ) = all_files
     entry = old_index[rev]
 
     # XXX consider trying the default compression too
     new_data_size = len(tombstone)
     new_data_offset = new_data_file.tell()
     new_data_file.write(tombstone)
 
     # we are not adding any sidedata as they might leak info about the censored version
 
     link_rev = entry[ENTRY_LINK_REV]
 
     p1 = entry[ENTRY_PARENT_1]
     p2 = entry[ENTRY_PARENT_2]
 
     new_entry = revlogutils.entry(
         flags=constants.REVIDX_ISCENSORED,
         data_offset=new_data_offset,
         data_compressed_length=new_data_size,
         data_uncompressed_length=new_data_size,
         data_delta_base=rev,
         link_rev=link_rev,
         parent_rev_1=p1,
         parent_rev_2=p2,
         node_id=entry[ENTRY_NODE_ID],
         sidedata_offset=0,
         sidedata_compressed_length=0,
         data_compression_mode=COMP_MODE_PLAIN,
         sidedata_compression_mode=COMP_MODE_PLAIN,
     )
     revlog.index.append(new_entry)
     entry_bin = revlog.index.entry_binary(rev)
     new_index_file.write(entry_bin)
     revlog._docket.index_end = new_index_file.tell()
     revlog._docket.data_end = new_data_file.tell()
 
 
 def _get_filename_from_filelog_index(path):
     # Drop the extension and the `data/` prefix
     path_part = path.rsplit(b'.', 1)[0].split(b'/', 1)
     if len(path_part) < 2:
         msg = _(b"cannot recognize filelog from filename: '%s'")
         msg %= path
         raise error.Abort(msg)
 
     return path_part[1]
 
 
 def _filelog_from_filename(repo, path):
     """Returns the filelog for the given `path`. Stolen from `engine.py`"""
 
     from .. import filelog  # avoid cycle
 
     fl = filelog.filelog(repo.svfs, path)
     return fl
 
 
 def _write_swapped_parents(repo, rl, rev, offset, fp):
     """Swaps p1 and p2 and overwrites the revlog entry for `rev` in `fp`"""
     from ..pure import parsers  # avoid cycle
 
     if repo._currentlock(repo._lockref) is None:
         # Let's be paranoid about it
         msg = "repo needs to be locked to rewrite parents"
         raise error.ProgrammingError(msg)
 
     index_format = parsers.IndexObject.index_format
     entry = rl.index[rev]
     new_entry = list(entry)
     new_entry[5], new_entry[6] = entry[6], entry[5]
     packed = index_format.pack(*new_entry[:8])
     fp.seek(offset)
     fp.write(packed)
 
 
 def _reorder_filelog_parents(repo, fl, to_fix):
     """
     Swaps p1 and p2 for all `to_fix` revisions of filelog `fl` and writes the
     new version to disk, overwriting the old one with a rename.
     """
     from ..pure import parsers  # avoid cycle
 
     ui = repo.ui
     assert len(to_fix) > 0
     rl = fl._revlog
     if rl._format_version != constants.REVLOGV1:
         msg = "expected version 1 revlog, got version '%d'" % rl._format_version
         raise error.ProgrammingError(msg)
 
     index_file = rl._indexfile
     new_file_path = index_file + b'.tmp-parents-fix'
     repaired_msg = _(b"repaired revision %d of 'filelog %s'\n")
 
     with ui.uninterruptible():
         try:
             util.copyfile(
                 rl.opener.join(index_file),
                 rl.opener.join(new_file_path),
                 checkambig=rl.data_config.check_ambig,
             )
 
             with rl.opener(new_file_path, mode=b"r+") as fp:
                 if rl._inline:
                     index = parsers.InlinedIndexObject(fp.read())
                     for rev in fl.revs():
                         if rev in to_fix:
                             offset = index._calculate_index(rev)
                             _write_swapped_parents(repo, rl, rev, offset, fp)
                             ui.write(repaired_msg % (rev, index_file))
                 else:
                     index_format = parsers.IndexObject.index_format
                     for rev in to_fix:
                         offset = rev * index_format.size
                         _write_swapped_parents(repo, rl, rev, offset, fp)
                         ui.write(repaired_msg % (rev, index_file))
 
             rl.opener.rename(new_file_path, index_file)
             rl.clearcaches()
             rl._loadindex()
         finally:
             util.tryunlink(new_file_path)
 
 
 def _is_revision_affected(fl, filerev, metadata_cache=None):
     full_text = lambda: fl._revlog.rawdata(filerev)
     parent_revs = lambda: fl._revlog.parentrevs(filerev)
     return _is_revision_affected_inner(
         full_text, parent_revs, filerev, metadata_cache
     )
 
 
 def _is_revision_affected_inner(
     full_text,
     parents_revs,
     filerev,
     metadata_cache=None,
 ):
     """Mercurial currently (5.9rc0) uses `p1 == nullrev and p2 != nullrev` as a
     special meaning compared to the reverse in the context of filelog-based
     copytracing. issue6528 exists because new code assumed that parent ordering
     didn't matter, so this detects if the revision contains metadata (since
     it's only used for filelog-based copytracing) and its parents are in the
     "wrong" order."""
     try:
         raw_text = full_text()
     except error.CensoredNodeError:
         # We don't care about censored nodes as they never carry metadata
         return False
 
     # raw text can be a `memoryview`, which doesn't implement `startswith`
     has_meta = bytes(raw_text[:2]) == b'\x01\n'
     if metadata_cache is not None:
         metadata_cache[filerev] = has_meta
     if has_meta:
         (p1, p2) = parents_revs()
         if p1 != nullrev and p2 == nullrev:
             return True
     return False
 
 
 def _is_revision_affected_fast(repo, fl, filerev, metadata_cache):
     rl = fl._revlog
     is_censored = lambda: rl.iscensored(filerev)
     delta_base = lambda: rl.deltaparent(filerev)
     delta = lambda: rl._inner._chunk(filerev)
     full_text = lambda: rl.rawdata(filerev)
     parent_revs = lambda: rl.parentrevs(filerev)
     return _is_revision_affected_fast_inner(
         is_censored,
         delta_base,
         delta,
         full_text,
         parent_revs,
         filerev,
         metadata_cache,
     )
 
 
 def _is_revision_affected_fast_inner(
     is_censored,
     delta_base,
     delta,
     full_text,
     parent_revs,
     filerev,
     metadata_cache,
 ):
     """Optimization fast-path for `_is_revision_affected`.
 
     `metadata_cache` is a dict of `{rev: has_metadata}` which allows any
     revision to check if its base has metadata, saving computation of the full
     text, instead looking at the current delta.
 
     This optimization only works if the revisions are looked at in order."""
 
     if is_censored():
         # Censored revisions don't contain metadata, so they cannot be affected
         metadata_cache[filerev] = False
         return False
 
     p1, p2 = parent_revs()
     if p1 == nullrev or p2 != nullrev:
         return False
 
     delta_parent = delta_base()
     parent_has_metadata = metadata_cache.get(delta_parent)
     if parent_has_metadata is None:
         return _is_revision_affected_inner(
             full_text,
             parent_revs,
             filerev,
             metadata_cache,
         )
 
     chunk = delta()
     if not len(chunk):
         # No diff for this revision
         metadata_cache[filerev] = parent_has_metadata
         return parent_has_metadata
 
     header_length = 12
     if len(chunk) < header_length:
         raise error.Abort(_(b"patch cannot be decoded"))
 
     start, _end, _length = struct.unpack(b">lll", chunk[:header_length])
 
     if start < 2:  # len(b'\x01\n') == 2
         # This delta does *something* to the metadata marker (if any).
         # Check it the slow way
         is_affected = _is_revision_affected_inner(
             full_text,
             parent_revs,
             filerev,
             metadata_cache,
         )
         return is_affected
 
     # The diff did not remove or add the metadata header, it's then in the same
     # situation as its parent
     metadata_cache[filerev] = parent_has_metadata
     return parent_has_metadata
 
 
 def _from_report(ui, repo, context, from_report, dry_run):
     """
     Fix the revisions given in the `from_report` file, but still checks if the
     revisions are indeed affected to prevent an unfortunate cyclic situation
     where we'd swap well-ordered parents again.
 
     See the doc for `debug_fix_issue6528` for the format documentation.
     """
     ui.write(_(b"loading report file '%s'\n") % from_report)
 
     with context(), open(from_report, mode='rb') as f:
         for line in f.read().split(b'\n'):
             if not line:
                 continue
             filenodes, filename = line.split(b' ', 1)
             fl = _filelog_from_filename(repo, filename)
             to_fix = set(
                 fl.rev(binascii.unhexlify(n)) for n in filenodes.split(b',')
             )
             excluded = set()
 
             for filerev in to_fix:
                 if _is_revision_affected(fl, filerev):
                     msg = b"found affected revision %d for filelog '%s'\n"
                     ui.warn(msg % (filerev, filename))
                 else:
                     msg = _(b"revision %s of file '%s' is not affected\n")
                     msg %= (binascii.hexlify(fl.node(filerev)), filename)
                     ui.warn(msg)
                     excluded.add(filerev)
 
             to_fix = to_fix - excluded
             if not to_fix:
                 msg = _(b"no affected revisions were found for '%s'\n")
                 ui.write(msg % filename)
                 continue
             if not dry_run:
                 _reorder_filelog_parents(repo, fl, sorted(to_fix))
 
 
 def filter_delta_issue6528(revlog, deltas_iter):
     """filter incomind deltas to repaire issue 6528 on the fly"""
-    metadata_cache = {}
+    metadata_cache = {nullrev: False}
 
     deltacomputer = deltas.deltacomputer(revlog)
 
     for rev, d in enumerate(deltas_iter, len(revlog)):
         (
             node,
             p1_node,
             p2_node,
             linknode,
             deltabase,
             delta,
             flags,
             sidedata,
         ) = d
 
         if not revlog.index.has_node(deltabase):
             raise error.LookupError(
                 deltabase, revlog.radix, _(b'unknown parent')
             )
         base_rev = revlog.rev(deltabase)
         if not revlog.index.has_node(p1_node):
             raise error.LookupError(p1_node, revlog.radix, _(b'unknown parent'))
         p1_rev = revlog.rev(p1_node)
         if not revlog.index.has_node(p2_node):
             raise error.LookupError(p2_node, revlog.radix, _(b'unknown parent'))
         p2_rev = revlog.rev(p2_node)
 
         is_censored = lambda: bool(flags & REVIDX_ISCENSORED)
         delta_base = lambda: revlog.rev(delta_base)
         delta_base = lambda: base_rev
         parent_revs = lambda: (p1_rev, p2_rev)
 
         def full_text():
             # note: being able to reuse the full text computation in the
             # underlying addrevision would be useful however this is a bit too
             # intrusive the for the "quick" issue6528 we are writing before the
             # 5.8 release
             textlen = mdiff.patchedsize(revlog.size(base_rev), delta)
 
             revinfo = revlogutils.revisioninfo(
                 node,
                 p1_node,
                 p2_node,
                 [None],
                 textlen,
                 (base_rev, delta),
                 flags,
             )
             return deltacomputer.buildtext(revinfo)
 
         is_affected = _is_revision_affected_fast_inner(
             is_censored,
             delta_base,
             lambda: delta,
             full_text,
             parent_revs,
             rev,
             metadata_cache,
         )
         if is_affected:
             d = (
                 node,
                 p2_node,
                 p1_node,
                 linknode,
                 deltabase,
                 delta,
                 flags,
                 sidedata,
             )
         yield d
 
 
 def repair_issue6528(
     ui, repo, dry_run=False, to_report=None, from_report=None, paranoid=False
 ):
     @contextlib.contextmanager
     def context():
         if dry_run or to_report:  # No need for locking
             yield
         else:
             with repo.wlock(), repo.lock():
                 yield
 
     if from_report:
         return _from_report(ui, repo, context, from_report, dry_run)
 
     report_entries = []
 
     with context():
         files = list(
             entry
             for entry in repo.store.data_entries()
             if entry.is_revlog and entry.is_filelog
         )
 
         progress = ui.makeprogress(
             _(b"looking for affected revisions"),
             unit=_(b"filelogs"),
             total=len(files),
         )
         found_nothing = True
 
         for entry in files:
             progress.increment()
             filename = entry.target_id
             fl = _filelog_from_filename(repo, entry.target_id)
 
             # Set of filerevs (or hex filenodes if `to_report`) that need fixing
             to_fix = set()
-            metadata_cache = {}
+            metadata_cache = {nullrev: False}
             for filerev in fl.revs():
                 affected = _is_revision_affected_fast(
                     repo, fl, filerev, metadata_cache
                 )
                 if paranoid:
                     slow = _is_revision_affected(fl, filerev)
                     if slow != affected:
                         msg = _(b"paranoid check failed for '%s' at node %s")
                         node = binascii.hexlify(fl.node(filerev))
                         raise error.Abort(msg % (filename, node))
                 if affected:
                     msg = b"found affected revision %d for file '%s'\n"
                     ui.warn(msg % (filerev, filename))
                     found_nothing = False
                     if not dry_run:
                         if to_report:
                             to_fix.add(binascii.hexlify(fl.node(filerev)))
                         else:
                             to_fix.add(filerev)
 
             if to_fix:
                 to_fix = sorted(to_fix)
                 if to_report:
                     report_entries.append((filename, to_fix))
                 else:
                     _reorder_filelog_parents(repo, fl, to_fix)
 
         if found_nothing:
             ui.write(_(b"no affected revisions were found\n"))
 
         if to_report and report_entries:
             with open(to_report, mode="wb") as f:
                 for path, to_fix in report_entries:
                     f.write(b"%s %s\n" % (b",".join(to_fix), path))
 
         progress.complete()