view relnotes/5.8 @ 49000:dd6b67d5c256 stable

rust: fix unsound `OwningDirstateMap`

As per the previous patch, `OwningDirstateMap` is unsound. Self-referential
structs are difficult to implement correctly in Rust since the compiler is
free to move structs around as much as it wants to. They are also very rarely
needed in practice, so the state-of-the-art on how they should be done within
the Rust rules is still a bit new.

The crate `ouroboros` is an attempt at providing a safe way (in the Rust sense)
of declaring self-referential structs. It is getting a lot attention and was
improved very quickly when soundness issues were found in the past: rather than
relying on our own (limited) review circle, we might as well use the de-facto
common crate to fix this problem. This will give us a much better chance of
finding issues should any new ones be discovered as well as the benefit of
fewer `unsafe` APIs of our own.

I was starting to think about how I would present a safe API to the old struct
but soon realized that the callback-based approach was already done in
`ouroboros`, along with a lot more care towards refusing incorrect structs.

In short: we don't return a mutable reference to the `DirstateMap` anymore, we
expect users of its API to pass a `FnOnce` that takes the map as an argument.
This allows our `OwningDirstateMap` to control the input and output lifetimes
of the code that modifies it to prevent such issues.

Changing to `ouroboros` meant changing every API with it, but it is relatively
low churn in the end. It correctly identified the example buggy modification of
`copy_map_insert` outlined in the previous patch as violating the borrow rules.

Differential Revision: https://phab.mercurial-scm.org/D12429

== New Features ==
 
 * `hg purge` is now a core command using `--confirm` by default.
 
 * The `rev-branch-cache` is now updated incrementally whenever changesets
   are added.

 * The new options `experimental.bundlecompthreads` and
   `experimental.bundlecompthreads.<engine>` can be used to instruct
   the compression engines for bundle operations to use multiple threads
   for compression. The default is single threaded operation. Currently
   only supported for zstd.

== Default Format Change ==

These changes affects newly created repositories (or new clone) done with
Mercurial 5.8.

  * The `ZSTD` compression will now be used by default for new repositories
    when available. This compression format was introduced in Mercurial 5.0,
    released in May 2019. See `hg help config.format.revlog-compression` for
    details.

  * Mercurial installation built with the Rust parts will now use the
    "persistent nodemap" feature by default. This feature was introduced in
    Mercurial 5.4 (May 2020). However Mercurial instalation built without the
    fast Rust implementation will refuse to interract with them by default.
    This restriction can be lifted through configuration.

    See `hg help config.format.use-persistent-nodemap` for details

== New Experimental Features ==

 * There's a new `diff.merge` config option to show the changes
    relative to an automerge for merge changesets. This makes it
    easier to detect and review manual changes performed in merge
    changesets. It is supported by `hg diff --change`, `hg log -p`
    `hg incoming -p`, and `hg outgoing -p` so far.


== Bug Fixes ==

 * gracefully recover from inconsistent persistent-nodemap data from disk.


== Backwards Compatibility Changes ==

 * In normal repositories, the first parent of a changeset is not null,
   unless both parents are null (like the first changeset). Some legacy
   repositories violate this condition. The revlog code will now
   silentely swap the parents if this condition is tested. This can
   change the output of `hg log` when explicitly asking for first or
   second parent. The changesets "nodeid" are not affected.


== Internal API Changes ==

 * `changelog.branchinfo` is deprecated and will be removed after 5.8.
   It is superseded by `changelogrevision.branchinfo`.

 * Callbacks for revlog.addgroup and the changelog._nodeduplicatecallback hook
   now get a revision number as argument instead of a node.

 * revlog.addrevision returns the revision number instead of the node.

 * `nodes.nullid` and related constants are being phased out as part of
   the deprecation of SHA1. Repository instances and related classes
   provide access via `nodeconstants` and in some cases `nullid` attributes.