view rust/hg-core/src/dirstate_tree/owning.rs @ 48061:060cd909439f

dirstate: drop all logic around the "non-normal" sets The dirstate has a lot of code to compute a set of all "non-normal" and "from_other_parent" entries. This is all used in one, unique, location, when `setparent` is called and moved from a merge to a non merge. At that time, any "merge related" information has to be dropped. This is mostly useful for command like `graft` or `shelve` that move to a single-parent state -before- the commit. Otherwise the commit will already have removed all traces of the merge information in the dirstate (e.g. for a regular merges). The bookkeeping for these sets is quite invasive. And it seems simpler to just drop it and do the full computation in the single location where we actually use it (since we have to do the computation at least once anyway). This simplify the code a lot, and clarify why this kind of computation is needed. The possible drawback compared to the previous code are: - if the operation happens in a loop, we will end up doing it multiple time, - the C code to detect entry of interest have been dropped, for now. It will be re-introduced later, with a processing code directly in C for even faster operation. Differential Revision: https://phab.mercurial-scm.org/D11507
author Pierre-Yves David <pierre-yves.david@octobus.net>
date Tue, 28 Sep 2021 20:05:37 +0200
parents 37a41267d000
children 3d0a9c6e614d
line wrap: on
line source

use super::dirstate_map::DirstateMap;
use stable_deref_trait::StableDeref;
use std::ops::Deref;

/// Keep a `DirstateMap<'on_disk>` next to the `on_disk` buffer that it
/// borrows.
///
/// This is similar to [`OwningRef`] which is more limited because it
/// represents exactly one `&T` reference next to the value it borrows, as
/// opposed to a struct that may contain an arbitrary number of references in
/// arbitrarily-nested data structures.
///
/// [`OwningRef`]: https://docs.rs/owning_ref/0.4.1/owning_ref/struct.OwningRef.html
pub struct OwningDirstateMap {
    /// Owned handle to a bytes buffer with a stable address.
    ///
    /// See <https://docs.rs/owning_ref/0.4.1/owning_ref/trait.StableAddress.html>.
    on_disk: Box<dyn Deref<Target = [u8]> + Send>,

    /// Pointer for `Box<DirstateMap<'on_disk>>`, typed-erased because the
    /// language cannot represent a lifetime referencing a sibling field.
    /// This is not quite a self-referencial struct (moving this struct is not
    /// a problem as it doesn’t change the address of the bytes buffer owned
    /// by `PyBytes`) but touches similar borrow-checker limitations.
    ptr: *mut (),
}

impl OwningDirstateMap {
    pub fn new_empty<OnDisk>(on_disk: OnDisk) -> Self
    where
        OnDisk: Deref<Target = [u8]> + StableDeref + Send + 'static,
    {
        let on_disk = Box::new(on_disk);
        let bytes: &'_ [u8] = &on_disk;
        let map = DirstateMap::empty(bytes);

        // Like in `bytes` above, this `'_` lifetime parameter borrows from
        // the bytes buffer owned by `on_disk`.
        let ptr: *mut DirstateMap<'_> = Box::into_raw(Box::new(map));

        // Erase the pointed type entirely in order to erase the lifetime.
        let ptr: *mut () = ptr.cast();

        Self { on_disk, ptr }
    }

    pub fn get_mut_pair<'a>(
        &'a mut self,
    ) -> (&'a [u8], &'a mut DirstateMap<'a>) {
        // SAFETY: We cast the type-erased pointer back to the same type it had
        // in `new`, except with a different lifetime parameter. This time we
        // connect the lifetime to that of `self`. This cast is valid because
        // `self` owns the same `PyBytes` whose buffer `DirstateMap`
        // references. That buffer has a stable memory address because the byte
        // string value of a `PyBytes` is immutable.
        let ptr: *mut DirstateMap<'a> = self.ptr.cast();
        // SAFETY: we dereference that pointer, connecting the lifetime of the
        // new   `&mut` to that of `self`. This is valid because the
        // raw pointer is   to a boxed value, and `self` owns that box.
        (&self.on_disk, unsafe { &mut *ptr })
    }

    pub fn get_mut<'a>(&'a mut self) -> &'a mut DirstateMap<'a> {
        self.get_mut_pair().1
    }

    pub fn get<'a>(&'a self) -> &'a DirstateMap<'a> {
        // SAFETY: same reasoning as in `get_mut` above.
        let ptr: *mut DirstateMap<'a> = self.ptr.cast();
        unsafe { &*ptr }
    }

    pub fn on_disk<'a>(&'a self) -> &'a [u8] {
        &self.on_disk
    }
}

impl Drop for OwningDirstateMap {
    fn drop(&mut self) {
        // Silence a "field is never read" warning, and demonstrate that this
        // value is still alive.
        let _ = &self.on_disk;
        // SAFETY: this cast is the same as in `get_mut`, and is valid for the
        // same reason. `self.on_disk` still exists at this point, drop glue
        // will drop it implicitly after this `drop` method returns.
        let ptr: *mut DirstateMap<'_> = self.ptr.cast();
        // SAFETY: `Box::from_raw` takes ownership of the box away from `self`.
        // This is fine because drop glue does nothig for `*mut ()` and we’re
        // in `drop`, so `get` and `get_mut` cannot be called again.
        unsafe { drop(Box::from_raw(ptr)) }
    }
}

fn _static_assert_is_send<T: Send>() {}

fn _static_assert_fields_are_send() {
    _static_assert_is_send::<Box<DirstateMap<'_>>>();
}

// SAFETY: we don’t get this impl implicitly because `*mut (): !Send` because
// thread-safety of raw pointers is unknown in the general case. However this
// particular raw pointer represents a `Box<DirstateMap<'on_disk>>` that we
// own. Since that `Box` is `Send` as shown in above, it is sound to mark
// this struct as `Send` too.
unsafe impl Send for OwningDirstateMap {}