BTreeMap: prevent tree from ever being owned by non-root node

[ssomers]

This introduces a new marker type, Dying, which is used to note trees which are in the process of deallocation. On such trees, some fields may be in an inconsistent state as we are deallocating the tree. Unfortunately, there's not a great way to express conditional unsafety, so the methods for traversal can cause UB if not invoked correctly, but not marked as such. This is not a regression from the previous state, but rather isolates the destructive methods to solely being called on the dying state.

[ssomers]

Apart from the comment, a major tweak is to have only one trait and a kind of static assert instead of a trait blocking access to descend for Owned nodes. That is actually one small step away from realizing the old "FIXME(@gereeter) consider adding yet another type parameter to NodeRef that restricts the use of navigation methods on reborrowed pointers". But I'm not taking that step, because I'm old enough to know that a small step for man usually turns out to be a giant leap.

[bors]

☔ The latest upstream changes (presumably #81159) made this pull request unmergeable. Please resolve the merge conflicts.

[ssomers]

@rustbot modify labels: -S-waiting-on-author +S-waiting-on-review

[Mark-Simulacrum]

Can we explain this more? I'm not sure what you mean by pinned. It would probably also be good to find a different name, as Pin carries a pretty specific meaning in Rust. (Of course, if we match that meaning, then the name is fine; I am not sure we do).

[ssomers]

The link with Pin is rather loose. For the particular safety problem at hand, "descendible" would be enough, but I think it's more clear and useful to say "neither ascendable nor descendable", which implies "unmovable", "locked", "fixed", "frozen", "stuck", "constant". It seems somewhat fuzzy because it's not the borrow that is "unmovable", but the NodeRef that the borrow type qualifies. I think the language's equivalent difference is between a mut and an immutable variable of a reference type, but "immutable" is definitely not better than "pinned" because the borrow type Immut would not have the "immutable" property.

[Mark-Simulacrum]

I admit that I don't really follow what you wrote. What precisely do we need? It seems like this is really just saying that you cannot descend from owned nodes? Can we call this "permits traversal" or something like that?

[ssomers]

Oops, I mixed up value semantics with variable semantics. All NodeRef are immutable already, luckily. I like "PERMITS_TRAVERSAL", that really… pins it down.

[Mark-Simulacrum]

Overall I think this seems fine.

[Mark-Simulacrum]

It would be good to update the description of the PR to summarize the current diff -- I think there was a pretty major reworking since you wrote it.

[ssomers]

It would be good to update the description of the PR to summarize the current diff

It's still the same goal in my book. The fact it requires a distinction between dying and living ownership is., I think, useful on its own and could be another PR to jump in front.

[bors]

☔ The latest upstream changes (presumably #81217) made this pull request unmergeable. Please resolve the merge conflicts.

[ssomers]

@rustbot label: +S-waiting-on-review -S-waiting-on-author

[Mark-Simulacrum]

@bors r+ rollup=never

[bors]

📌 Commit 417eefe has been approved by Mark-Simulacrum

[bors]

⌛ Testing commit 417eefe with merge c6bc462...

[bors]

☀️ Test successful - checks-actions
Approved by: Mark-Simulacrum
Pushing c6bc462 to master...