revlog: improve documentation of the entry tuple The code in revlog, and outside revlog directly use the index's entry tuple, with direct integer indexing. This is a voluntary trade off to obtains better performance from the Python code at the expense of the developers sanity. Let's at least have a clear and central documentation about what this tuple is about. Differential Revision: https://phab.mercurial-scm.org/D10643

revlog: simplify entry update logic in `rewrite_sidedata` We don't need to rebuild the full entry. So lets just gather the data we need to pass to `replace_sidedata_info`. Differential Revision: https://phab.mercurial-scm.org/D10642

revlog: simplify the replace_sidedata_info code We already know how to serialize and de-serialize and enty. So lets just do that and modify the entry tuple directly. This avoid having to duplicated binary operation in complicated code. Differential Revision: https://phab.mercurial-scm.org/D10641

revlog: use `rev` instead of `i` in replace_sidedata_info This is clearer. Differential Revision: https://phab.mercurial-scm.org/D10640

revlogv2: also keep track for the size of the "data" file This is useful to make sure we always start writing at the right location, without effort. Differential Revision: https://phab.mercurial-scm.org/D10632

revlogv2: track pending write in the docket and expose it to hooks The docket is now able to write pending data. We could have used a distinct intermediate files, however keeping everything in the same file will make it simpler to keep track of the various involved files if necessary. However it might prove more complicated for streaming clone. This will be dealt with later. Note that we lifted the stderr redirection in the test since we no longer suffer from "unkown working directory parent" message. Differential Revision: https://phab.mercurial-scm.org/D10631

revlog: move the `trypending` logic from the `changelog` to the `revlog` We move the -reading- logic for the pending's '.a' suffixed index within the revlog class. This is motivated by the fact the logic could be simpler and cleaner if directly handled by the revlog docket. Before we can do so, we need to teach the revlog code about reading "pending" changes. To be honest, we already needed some special casing of the `.a` postfix, so this does not adds much complexity. The logic around -writing- the special '00changelog.i.a' remains in the `changelog` class. Note that the revlog-v2 logic no longer use this logic. The only remaining user of the `postfix` argument is the `censor` logic. We could probably also make the revlog full aware of it (most of the code is already implemented in revlog anyway) and get rid of the `postfix` argument and logic. However this is an adventure for another time. Since we have more information, we add more, paranoid, Programming error in case we detect such "pending reader" trying to do a read (which does not happens anyways). Differential Revision: https://phab.mercurial-scm.org/D10630

revlogv2: delay the update of the changelog docket to transaction end This prevent external reader to see the transaction content before it is commited. However this also prevent the hooks to see the transaction content. We will fix this in later changesets. We have to temporarily suppress the error output of the command ran during the transaction as they sometimes get confused about unknown working directory and sometimes issue message on std-err in unspecified order. Differential Revision: https://phab.mercurial-scm.org/D10629

revlogv2: track current index size in the docket This help use to fix transaction safety on repos. See next changesets for details. Differential Revision: https://phab.mercurial-scm.org/D10628

revlogv2: also test transactionality of revlog v2 Without surprise, this is broken :-) Now that it is tested, we can fix it. Differential Revision: https://phab.mercurial-scm.org/D10627