1230         # Since the multiple files involved in the transaction cannot be

  1231         # written atomically (with current repository format), there is a race

  1232         # condition here.

  1233         #

  1234         # 1) changelog content A is read

  1235         # 2) outside transaction update changelog to content B

  1236         # 3) outside transaction update bookmark file referring to content B

  1237         # 4) bookmarks file content is read and filtered against changelog-A

  1238         #

  1239         # When this happens, bookmarks against nodes missing from A are dropped.

  1240         #

  1241         # Having this happening during read is not great, but it become worse

  1242         # when this happen during write because the bookmarks to the "unknown"

  1243         # nodes will be dropped for good. However, writes happen within locks.

  1244         # This locking makes it possible to have a race free consistent read.

  1245         # For this purpose data read from disc before locking  are

  1246         # "invalidated" right after the locks are taken. This invalidations are

  1247         # "light", the `filecache` mechanism keep the data in memory and will

  1248         # reuse them if the underlying files did not changed. Not parsing the

  1249         # same data multiple times helps performances.

  1250         #

  1251         # Unfortunately in the case describe above, the files tracked by the

  1252         # bookmarks file cache might not have changed, but the in-memory

  1253         # content is still "wrong" because we used an older changelog content

  1254         # to process the on-disk data. So after locking, the changelog would be

  1255         # refreshed but `_bookmarks` would be preserved.

  1256         # Adding `00changelog.i` to the list of tracked file is not

  1257         # enough, because at the time we build the content for `_bookmarks` in

  1258         # (4), the changelog file has already diverged from the content used

  1259         # for loading `changelog` in (1)

  1260         #

  1261         # To prevent the issue, we force the changelog to be explicitly

  1262         # reloaded while computing `_bookmarks`. The data race can still happen

  1263         # without the lock (with a narrower window), but it would no longer go

  1264         # undetected during the lock time refresh.

  1265         #

  1266         # The new schedule is as follow

  1267         #

  1268         # 1) filecache logic detect that `_bookmarks` needs to be computed

  1269         # 2) cachestat for `bookmarks` and `changelog` are captured (for book)

  1270         # 3) We force `changelog` filecache to be tested

  1271         # 4) cachestat for `changelog` are captured (for changelog)

  1272         # 5) `_bookmarks` is computed and cached

  1273         #

  1274         # The step in (3) ensure we have a changelog at least as recent as the

  1275         # cache stat computed in (1). As a result at locking time:

  1276         #  * if the changelog did not changed since (1) -> we can reuse the data

  1277         #  * otherwise -> the bookmarks get refreshed.

  1278         self._refreshchangelog()