cargo/workspace/

registry.rs

//! Types that hold source information for a group of packages.
//!
//! The primary type you're looking for is [`PackageRegistry`]. It is an
//! abstraction over multiple [`Source`]s. [`PackageRegistry`] also implements
//! the [`Registry`] trait, allowing a dependency resolver to query necessary
//! package metadata (i.e., [Summary]) from it.
//!
//! Not to be confused with [`crate::sources::registry`] and [`crate::ops::registry`].
//! The former is just one kind of source,
//! while the latter involves operations on the registry Web API.

use cargo_util::collections::{HashMap, HashSet};
use std::cell::RefCell;

use crate::sources::IndexSummary;
use crate::sources::config::SourceConfigMap;
use crate::sources::source::QueryKind;
use crate::sources::source::Source;
use crate::sources::source::SourceMap;
use crate::util::VersionReqMatchMode;
use crate::util::errors::CargoResult;
use crate::util::interning::InternedString;
use crate::util::{CanonicalUrl, GlobalContext};
use crate::workspace::{Dependency, PackageId, PackageSet, Patch, SourceId, Summary};
use anyhow::Context as _;
use cargo_util_terminal::report::Level;
use futures::stream::FuturesUnordered;
use itertools::Itertools;
use tracing::{debug, trace};
use url::Url;

/// An abstraction provides a set of methods for querying source information
/// about a group of packages, without leaking too much implementation details
/// of the actual registry.
///
/// As of 2024-04, only [`PackageRegistry`] and `MyRegistry` in resolver-tests
/// are found implementing this.
///
/// See also the [`Source`] trait, as many of the methods here mirror and
/// abstract over its functionalities.
#[allow(async_fn_in_trait)]
pub trait Registry {
    /// Attempt to find the packages that match a dependency request.
    async fn query(
        &self,
        dep: &Dependency,
        kind: QueryKind,
        f: &mut dyn FnMut(IndexSummary),
    ) -> CargoResult<()>;

    /// Gathers the result from [`Registry::query`] as a list of [`IndexSummary`] items
    /// when they become available.
    async fn query_vec(&self, dep: &Dependency, kind: QueryKind) -> CargoResult<Vec<IndexSummary>> {
        let mut ret = Vec::new();
        self.query(dep, kind, &mut |s| ret.push(s))
            .await
            .map(|()| ret)
    }

    /// Gets the description of a source, to provide useful messages.
    fn describe_source(&self, source: SourceId) -> String;

    /// Checks if a source is replaced with some other source.
    fn is_replaced(&self, source: SourceId) -> bool;
}

/// This structure represents a registry of known packages. It internally
/// contains a number of [`Source`] instances which are used to load a
/// [`Package`] from.
///
/// The resolution phase of Cargo uses this to drive knowledge about new
/// packages as well as querying for lists of new packages.
/// It is here that sources are updated (e.g., network operations) and
/// overrides/patches are handled.
///
/// The general idea behind this registry is that it is centered around the
/// [`SourceMap`] structure, contained within which is a mapping of a [`SourceId`]
/// to a [`Source`]. Each [`Source`] in the map has been updated (using network
/// operations if necessary) and is ready to be queried for packages.
///
/// [`Package`]: crate::workspace::Package
pub struct PackageRegistry<'gctx> {
    gctx: &'gctx GlobalContext,
    sources: RefCell<SourceMap<'gctx>>,

    /// A list of sources which are considered "path-overrides" which take
    /// precedent when querying for packages.
    overrides: RefCell<Vec<SourceId>>,

    /// Use for tracking sources that are already loaded into the registry.
    // Note that each SourceId does not take into account its `precise` field
    // when hashing or testing for equality. When adding a new `SourceId`, we
    // want to avoid duplicates in the `SourceMap` (to prevent re-updating the
    // same git repo twice for example), but we also want to ensure that the
    // loaded source is always updated.
    //
    // Sources with a `precise` field normally don't need to be updated because
    // their contents are already on disk, but sources without a `precise` field
    // almost always need to be updated. If we have a cached `Source` for a
    // precise `SourceId`, then when we add a new `SourceId` that is not precise
    // we want to ensure that the underlying source is updated.
    //
    // This is basically a long-winded way of saying that we want to know
    // precisely what the keys of `sources` are, so this is a mapping of key to
    // what exactly the key is.
    source_ids: RefCell<HashMap<SourceId, (SourceId, Kind)>>,

    /// This is constructed via [`PackageRegistry::register_lock`].
    /// See also [`LockedMap`].
    locked: LockedMap,
    source_config: SourceConfigMap<'gctx>,

    /// Patches registered during calls to [`PackageRegistry::patch`].
    ///
    /// These are available for `query` after calling [`PackageRegistry::lock_patches`],
    /// which `lock`s them all to specific versions.
    patches: HashMap<CanonicalUrl, Vec<Summary>>,
    /// Whether patches are locked. That is, they are available to resolution.
    ///
    /// See [`PackageRegistry::lock_patches`] and [`PackageRegistry::patch`] for more.
    patches_locked: bool,
    /// Patches available for each source.
    ///
    /// This is for determining whether a dependency entry from a lockfile
    /// happened through `[patch]`, during calls to [`lock`] to rewrite
    /// summaries to point directly at these patched entries.
    ///
    /// This is constructed during calls to [`PackageRegistry::patch`],
    /// along with the `patches` field, thoough these entries never get locked.
    patches_available: HashMap<CanonicalUrl, Vec<PackageId>>,
}

/// A map of all "locked packages" which is filled in when parsing a lock file
/// and is used to guide dependency resolution by altering summaries as they're
/// queried from this source.
///
/// This map can be thought of as a glorified `Vec<MySummary>` where `MySummary`
/// has a `PackageId` for which package it represents as well as a list of
/// `PackageId` for the resolved dependencies. The hash map is otherwise
/// structured though for easy access throughout this registry.
type LockedMap = HashMap<
    // The first level of key-ing done in this hash map is the source that
    // dependencies come from, identified by a `SourceId`.
    // The next level is keyed by the name of the package...
    (SourceId, InternedString),
    // ... and the value here is a list of tuples. The first element of each
    // tuple is a package which has the source/name used to get to this
    // point. The second element of each tuple is the list of locked
    // dependencies that the first element has.
    Vec<(PackageId, Vec<PackageId>)>,
>;

/// Kinds of sources a [`PackageRegistry`] has loaded.
#[derive(PartialEq, Eq, Clone, Copy)]
enum Kind {
    /// A source from a [path override].
    ///
    /// [path overrides]: https://doc.rust-lang.org/nightly/cargo/reference/overriding-dependencies.html#paths-overrides
    Override,
    /// A source that is locked and not going to change.
    ///
    /// For example, sources of workspace members are loaded during the
    /// workspace initialization, so not allowed to change.
    Locked,
    /// A source that is not locked nor a path-override.
    Normal,
}

/// This tuple is an argument to [`PackageRegistry::patch`].
///
/// * The first element is the patch definition straight from the manifest.
/// * The second element is an optional variant where the patch has been locked.
///   It is the patch locked to a specific version found in Cargo.lock.
///   This will be `None` if `Cargo.lock` doesn't exist,
///   or the patch did not match any existing entries in `Cargo.lock`.
pub type PatchDependency<'a> = (&'a Patch, Option<LockedPatchDependency>);

/// Argument to [`PackageRegistry::patch`] which is information about a `[patch]`
/// directive that we found in a lockfile, if present.
pub struct LockedPatchDependency {
    /// The original `Dependency` directive, except "locked" so it's version
    /// requirement is Locked to `foo` and its `SourceId` has a "precise" listed.
    pub dependency: Dependency,
    /// The `PackageId` that was previously found in a lock file which
    /// `dependency` matches.
    pub package_id: PackageId,
    /// Something only used for backwards compatibility with the v2 lock file
    /// format where `branch=master` is considered the same as `DefaultBranch`.
    /// For more comments on this see the code in `ops/resolve.rs`.
    pub alt_package_id: Option<PackageId>,
}

impl<'gctx> PackageRegistry<'gctx> {
    pub fn new_with_source_config(
        gctx: &'gctx GlobalContext,
        source_config: SourceConfigMap<'gctx>,
    ) -> CargoResult<PackageRegistry<'gctx>> {
        Ok(PackageRegistry {
            gctx,
            sources: RefCell::new(SourceMap::new()),
            source_ids: RefCell::new(HashMap::default()),
            overrides: RefCell::new(Vec::new()),
            source_config,
            locked: HashMap::default(),
            patches: HashMap::default(),
            patches_locked: false,
            patches_available: HashMap::default(),
        })
    }

    pub fn get(self, package_ids: &[PackageId]) -> CargoResult<PackageSet<'gctx>> {
        trace!("getting packages; sources={}", self.sources.borrow().len());
        PackageSet::new(package_ids, self.sources.into_inner(), self.gctx)
    }

    /// Ensures the [`Source`] of the given [`SourceId`] is loaded.
    fn ensure_loaded(&self, namespace: SourceId, kind: Kind) -> CargoResult<()> {
        match self.source_ids.borrow().get(&namespace) {
            // We've previously loaded this source, and we've already locked it,
            // so we're not allowed to change it even if `namespace` has a
            // slightly different precise version listed.
            Some((_, Kind::Locked)) => {
                debug!("load/locked   {}", namespace);
                return Ok(());
            }

            // If the previous source was not a precise source, then we can be
            // sure that it's already been updated if we've already loaded it.
            Some((previous, _)) if !previous.has_precise() => {
                debug!("load/precise  {}", namespace);
                return Ok(());
            }

            // If the previous source has the same precise version as we do,
            // then we're done, otherwise we need to move forward
            // updating this source.
            Some((previous, _)) => {
                if previous.has_same_precise_as(namespace) {
                    debug!("load/match    {}", namespace);
                    return Ok(());
                }
                debug!("load/mismatch {}", namespace);
            }
            None => {
                debug!("load/missing  {}", namespace);
            }
        }

        self.load(namespace, kind)?;

        // Ensure `shell` is not already in use,
        // regardless of which source is used and how it happens to behave this time
        self.gctx.debug_assert_shell_not_borrowed();
        Ok(())
    }

    pub fn add_sources(&mut self, ids: impl IntoIterator<Item = SourceId>) -> CargoResult<()> {
        for id in ids {
            self.ensure_loaded(id, Kind::Locked)?;
        }
        Ok(())
    }

    /// Adds a source which will be locked.
    /// Useful for path sources such as the source of a workspace member.
    pub fn add_preloaded(&mut self, source: Box<dyn Source + 'gctx>) {
        self.add_source(source, Kind::Locked);
    }

    /// Adds a source to the registry.
    fn add_source(&self, source: Box<dyn Source + 'gctx>, kind: Kind) {
        let id = source.source_id();
        self.sources.borrow_mut().insert(source);
        self.source_ids.borrow_mut().insert(id, (id, kind));
    }

    /// Adds a source from a [path override].
    ///
    /// [path override]: https://doc.rust-lang.org/nightly/cargo/reference/overriding-dependencies.html#paths-overrides
    pub fn add_override(&mut self, source: Box<dyn Source + 'gctx>) {
        self.overrides.borrow_mut().push(source.source_id());
        self.add_source(source, Kind::Override);
    }

    /// remove all residual state from previous lock files.
    pub fn clear_lock(&mut self) {
        trace!("clear_lock");
        self.locked = HashMap::default();
    }

    /// Registers one "locked package" to the registry, for guiding the
    /// dependency resolution. See [`LockedMap`] for more.
    pub fn register_lock(&mut self, id: PackageId, deps: Vec<PackageId>) {
        trace!("register_lock: {}", id);
        for dep in deps.iter() {
            trace!("\t-> {}", dep);
        }
        let sub_vec = self
            .locked
            .entry((id.source_id(), id.name()))
            .or_insert_with(Vec::new);
        sub_vec.push((id, deps));
    }

    /// Insert a `[patch]` section into this registry.
    ///
    /// This method will insert a `[patch]` section for the `url` specified,
    /// with the given list of dependencies. The `url` specified is the URL of
    /// the source to patch (for example this is `crates-io` in the manifest).
    /// The `deps` is an array of all the entries in the `[patch]` section of
    /// the manifest.
    ///
    /// Here the `patch_deps` will be resolved to a precise version and stored
    /// internally for future calls to `query` below.
    ///
    /// Note that the patch list specified here *will not* be available to
    /// [`Registry::query`] until [`PackageRegistry::lock_patches`] is called
    /// below, which should be called once all patches have been added.
    ///
    /// The return value is a `Vec` of patches that should *not* be locked.
    /// This happens when the patch is locked, but the patch has been updated
    /// so the locked value is no longer correct.
    #[tracing::instrument(skip(self, patch_deps))]
    pub fn patch(
        &mut self,
        url: &Url,
        patch_deps: &[PatchDependency<'_>],
    ) -> CargoResult<Vec<(Patch, PackageId)>> {
        // NOTE: None of this code is aware of required features. If a patch
        // is missing a required feature, you end up with an "unused patch"
        // warning, which is very hard to understand. Ideally the warning
        // would be tailored to indicate *why* it is unused.
        let canonical = CanonicalUrl::new(url)?;

        // Return value of patches that shouldn't be locked.
        let mut unlock_patches = Vec::new();

        // First up we need to actually resolve each `patch_deps` specification
        // to precisely one summary. We're not using the `query` method below
        // as it internally uses maps we're building up as part of this method
        // (`patches_available` and `patches`). Instead we're going straight to
        // the source to load information from it.
        //
        // Remember that each dependency listed in `[patch]` has to resolve to
        // precisely one package, so that's why we're just creating a flat list
        // of summaries which should be the same length as `deps` above.

        let pending = FuturesUnordered::new();

        for (orig_patch, locked) in patch_deps {
            // Use the locked patch if it exists, otherwise use the original.
            let dep = match locked {
                Some(lock) => &lock.dependency,
                None => &orig_patch.dep,
            };
            debug!(
                "registering a patch for `{}` with `{}`",
                url,
                dep.package_name()
            );

            let mut unused_fields = Vec::new();
            if dep.features().len() != 0 {
                unused_fields.push("`features`");
            }
            if !dep.uses_default_features() {
                unused_fields.push("`default-features`")
            }
            if !unused_fields.is_empty() {
                self.source_config.gctx().shell().print_report(
                    &[Level::WARNING
                        .secondary_title(format!(
                            "unused field in patch for `{}`: {}",
                            dep.package_name(),
                            unused_fields.join(", ")
                        ))
                        .element(Level::HELP.message(format!(
                            "configure {} in the `dependencies` entry",
                            unused_fields.join(", ")
                        )))],
                    false,
                )?;
            }

            // Go straight to the source for resolving `dep`. Load it as we
            // normally would and then ask it directly for the list of summaries
            // corresponding to this `dep`.
            self.ensure_loaded(dep.source_id(), Kind::Normal)
                .with_context(|| {
                    format!(
                        "failed to load source for dependency `{}`",
                        dep.package_name()
                    )
                })?;

            let source = self
                .sources
                .borrow()
                .get(dep.source_id())
                .expect("loaded source not present")
                .clone();
            pending.push(async move {
                let mut summaries = Vec::new();
                source
                    .query(&dep, QueryKind::Exact, &mut |s| {
                        if let IndexSummary::Candidate(summary) = s {
                            summaries.push(summary)
                        }
                    })
                    .await
                    .with_context(|| format!("unable to update {}", source.source_id()))
                    .with_context(|| {
                        format!(
                            "failed to load source for dependency `{}`",
                            dep.package_name()
                        )
                    })?;

                let (summary, should_unlock) =
                    summary_for_patch(&orig_patch, url, &locked, summaries, source.as_ref())
                        .await?;
                Ok::<_, anyhow::Error>((orig_patch, dep, summary, should_unlock))
            });
        }

        let unlocked_summaries = crate::util::block_on_stream(pending).map(|next| {
            let (orig_patch, dep, summary, should_unlock) = next?;
            debug!(
                "patch summary is {:?} should_unlock={:?}",
                summary, should_unlock
            );
            if let Some(unlock_id) = should_unlock {
                unlock_patches.push(((*orig_patch).clone(), unlock_id));
            }

            if *summary.package_id().source_id().canonical_url() == canonical {
                return Err(anyhow::anyhow!(
                    "patch for `{}` points to the same source, but patches must point to different sources\n\
                    help: check `{}` patch definition for `{}` in `{}`",
                    dep.package_name(),
                    dep.package_name(),
                    url,
                    orig_patch.loc
                ));
            }
            Ok(summary)
        }).collect::<CargoResult<Vec<_>>>()?;

        let mut name_and_version = HashSet::default();
        for summary in unlocked_summaries.iter() {
            let name = summary.package_id().name();
            let version = summary.package_id().version();
            if !name_and_version.insert((name, version)) {
                let duplicate_locations = patch_deps
                    .iter()
                    .filter(|&p| p.0.dep.package_name() == name)
                    .map(|p| format!("`{}`", p.0.loc))
                    .unique()
                    .join(", ");
                return Err(anyhow::anyhow!(
                    "several `[patch]` entries resolving to same version `{} v{}`\n\
                    help: check `{}` patch definitions for `{}` in {}",
                    name,
                    version,
                    name,
                    url,
                    duplicate_locations
                ));
            }
        }

        // Calculate a list of all patches available for this source.
        let mut ids = Vec::new();
        for (summary, (_, lock)) in unlocked_summaries.iter().zip(patch_deps) {
            ids.push(summary.package_id());
            // This is subtle where the list of `ids` for a canonical URL is
            // extend with possibly two ids per summary. This is done to handle
            // the transition from the v2->v3 lock file format where in v2
            // DefaultBranch was either DefaultBranch or Branch("master") for
            // git dependencies. In this case if `summary.package_id()` is
            // Branch("master") then alt_package_id will be DefaultBranch. This
            // signifies that there's a patch available for either of those
            // dependency directives if we see them in the dependency graph.
            if let Some(lock) = lock {
                ids.extend(lock.alt_package_id);
            }
        }
        self.patches_available.insert(canonical.clone(), ids);

        // Note that we do not use `lock` here to lock summaries! That step
        // happens later once `lock_patches` is invoked. In the meantime though
        // we want to fill in the `patches_available` map (later used in the
        // `lock` method) and otherwise store the unlocked summaries in
        // `patches` to get locked in a future call to `lock_patches`.
        self.patches.insert(canonical, unlocked_summaries);

        Ok(unlock_patches)
    }

    /// Lock all patch summaries added via [`patch`](Self::patch),
    /// making them available to resolution via [`Registry::query`].
    pub fn lock_patches(&mut self) {
        assert!(!self.patches_locked);

        let mode = self.gctx.version_req_match_mode();
        for summaries in self.patches.values_mut() {
            for summary in summaries {
                debug!("locking patch {:?}", summary);
                *summary = lock(&self.locked, &self.patches_available, summary.clone(), mode);
            }
        }
        self.patches_locked = true;
    }

    /// Gets all patches grouped by the source URLs they are going to patch.
    ///
    /// These patches are mainly collected from [`patch`](Self::patch).
    /// They might not be the same as patches actually used during dependency resolving.
    pub fn patches(&self) -> &HashMap<CanonicalUrl, Vec<Summary>> {
        &self.patches
    }

    /// Loads the [`Source`] for a given [`SourceId`] to this registry, making
    /// them available to resolution.
    fn load(&self, source_id: SourceId, kind: Kind) -> CargoResult<()> {
        debug!("loading source {}", source_id);
        let source = self
            .source_config
            .load(source_id)
            .with_context(|| format!("unable to update {}", source_id))?;
        assert_eq!(source.source_id(), source_id);

        if kind == Kind::Override {
            self.overrides.borrow_mut().push(source_id);
        }
        self.add_source(source, kind);

        // If we have an imprecise version then we don't know what we're going
        // to look for, so we always attempt to perform an update here.
        //
        // If we have a precise version, then we'll update lazily during the
        // querying phase. Note that precise in this case is only
        // `"locked"` as other values indicate a `cargo update
        // --precise` request
        if !source_id.has_locked_precise() {
            self.sources
                .borrow()
                .get(source_id)
                .unwrap()
                .invalidate_cache();
        } else {
            debug!("skipping update due to locked registry");
        }
        Ok(())
    }

    /// Queries path overrides from this registry.
    async fn query_overrides(&self, dep: &Dependency) -> CargoResult<Option<Summary>> {
        let overrides = self.overrides.borrow();
        for &s in overrides.iter() {
            let dep = Dependency::new_override(dep.package_name(), s);
            let mut results = None;
            self.sources
                .borrow()
                .get(s)
                .unwrap()
                .query(&dep, QueryKind::Exact, &mut |s| {
                    if let IndexSummary::Candidate(s) = s {
                        results = Some(s);
                    }
                })
                .await?;
            if results.is_some() {
                return Ok(results);
            }
        }
        Ok(None)
    }

    /// This function is used to transform a summary to another locked summary
    /// if possible. This is where the concept of a lock file comes into play.
    ///
    /// If a summary points at a package ID which was previously locked, then we
    /// override the summary's ID itself, as well as all dependencies, to be
    /// rewritten to the locked versions. This will transform the summary's
    /// source to a precise source (listed in the locked version) as well as
    /// transforming all of the dependencies from range requirements on
    /// imprecise sources to exact requirements on precise sources.
    ///
    /// If a summary does not point at a package ID which was previously locked,
    /// or if any dependencies were added and don't have a previously listed
    /// version, we still want to avoid updating as many dependencies as
    /// possible to keep the graph stable. In this case we map all of the
    /// summary's dependencies to be rewritten to a locked version wherever
    /// possible. If we're unable to map a dependency though, we just pass it on
    /// through.
    pub fn lock(&self, summary: Summary) -> Summary {
        assert!(self.patches_locked);
        let mode = self.gctx.version_req_match_mode();
        lock(&self.locked, &self.patches_available, summary, mode)
    }

    fn warn_bad_override(
        &self,
        override_summary: &Summary,
        real_summary: &Summary,
    ) -> CargoResult<()> {
        let mut real_deps = real_summary.dependencies().iter().collect::<Vec<_>>();

        let boilerplate = "\
This is currently allowed but is known to produce buggy behavior with spurious
recompiles and changes to the crate graph. Path overrides unfortunately were
never intended to support this feature, so for now this message is just a
warning. In the future, however, this message will become a hard error.

To change the dependency graph via an override it's recommended to use the
`[patch]` feature of Cargo instead of the path override feature. This is
documented online at the url below for more information.

https://doc.rust-lang.org/cargo/reference/overriding-dependencies.html
";

        for dep in override_summary.dependencies() {
            if let Some(i) = real_deps.iter().position(|d| dep == *d) {
                real_deps.remove(i);
                continue;
            }
            let msg = format!(
                "path override for crate `{}` has altered the original list of\n\
                 dependencies; the dependency on `{}` was either added or\n\
                 modified to not match the previously resolved version\n\n\
                 {}",
                override_summary.package_id().name(),
                dep.package_name(),
                boilerplate
            );
            self.source_config.gctx().shell().warn(&msg)?;
            return Ok(());
        }

        if let Some(dep) = real_deps.get(0) {
            let msg = format!(
                "path override for crate `{}` has altered the original list of\n\
                 dependencies; the dependency on `{}` was removed\n\n\
                 {}",
                override_summary.package_id().name(),
                dep.package_name(),
                boilerplate
            );
            self.source_config.gctx().shell().warn(&msg)?;
            return Ok(());
        }

        Ok(())
    }
}

impl<'gctx> Registry for PackageRegistry<'gctx> {
    async fn query(
        &self,
        dep: &Dependency,
        kind: QueryKind,
        f: &mut dyn FnMut(IndexSummary),
    ) -> CargoResult<()> {
        assert!(self.patches_locked);
        // Look for an override and get ready to query the real source.
        let override_summary = self.query_overrides(dep).await?;

        // Next up on our list of candidates is to check the `[patch]` section
        // of the manifest. Here we look through all patches relevant to the
        // source that `dep` points to, and then we match name/version. Note
        // that we don't use `dep.matches(..)` because the patches, by definition,
        // come from a different source. This means that `dep.matches(..)` will
        // always return false, when what we really care about is the name/version match.
        let mut patches = Vec::<Summary>::new();
        if let Some(extra) = self.patches.get(dep.source_id().canonical_url()) {
            let mode = self.gctx.version_req_match_mode();
            patches.extend(
                extra
                    .iter()
                    .filter(|s| dep.matches_ignoring_source(s.package_id(), mode))
                    .cloned(),
            );
        }

        // A crucial feature of the `[patch]` feature is that we don't query the
        // actual registry if we have a "locked" dependency. A locked dep basically
        // just means a version constraint of `=a.b.c`, and because patches take
        // priority over the actual source then if we have a candidate we're done.
        if patches.len() == 1 && dep.is_locked() {
            let patch = patches.remove(0);
            match override_summary {
                Some(override_summary) => {
                    self.warn_bad_override(&override_summary, &patch)?;
                    let override_summary = self.lock(override_summary);
                    f(IndexSummary::Candidate(override_summary));
                }
                None => f(IndexSummary::Candidate(patch)),
            }

            return Ok(());
        }

        if !patches.is_empty() {
            debug!(
                "found {} patches with an unlocked dep on `{}` at {} \
                     with `{}`, \
                     looking at sources",
                patches.len(),
                dep.package_name(),
                dep.source_id(),
                dep.version_req()
            );
        }

        // Ensure the requested source_id is loaded
        self.ensure_loaded(dep.source_id(), Kind::Normal)
            .with_context(|| {
                format!(
                    "failed to load source for dependency `{}`",
                    dep.package_name()
                )
            })?;

        // Helper function to add context for query errors.
        async fn query_with_context(
            source: &dyn Source,
            dep: &Dependency,
            kind: QueryKind,
            f: &mut dyn FnMut(IndexSummary),
        ) -> CargoResult<()> {
            source
                .query(dep, kind, f)
                .await
                .with_context(|| format!("unable to update {}", source.source_id()))
                .with_context(|| {
                    format!(
                        "failed to load source for dependency `{}`",
                        dep.package_name()
                    )
                })
        }

        let source = self.sources.borrow().get(dep.source_id()).cloned();
        match (override_summary, source) {
            (Some(_), None) => {
                return Err(anyhow::anyhow!("override found but no real ones"));
            }
            (None, None) => return Ok(()),

            // If we don't have an override then we just ship everything upstairs after locking the summary
            (None, Some(source)) => {
                for patch in patches.iter() {
                    f(IndexSummary::Candidate(patch.clone()));
                }

                // Our sources shouldn't ever come back to us with two summaries
                // that have the same version. We could, however, have an `[patch]`
                // section which is in use to override a version in the registry.
                // This means that if our `summary` in this loop has the same
                // version as something in `patches` that we've already selected,
                // then we skip this `summary`.
                let locked = &self.locked;
                let all_patches = &self.patches_available;
                let mode = self.gctx.version_req_match_mode();
                let callback = &mut |summary: IndexSummary| {
                    for patch in patches.iter() {
                        let patch = patch.package_id().version();
                        if summary.package_id().version() == patch {
                            return;
                        }
                    }
                    let summary =
                        summary.map_summary(|summary| lock(locked, all_patches, summary, mode));
                    f(summary)
                };
                return query_with_context(&*source, dep, kind, callback).await;
            }

            // If we have an override summary then we query the source to sanity check its results.
            // We don't actually use any of the summaries it gives us though.
            (Some(override_summary), Some(source)) => {
                if !patches.is_empty() {
                    return Err(anyhow::anyhow!("found patches and a path override"));
                }
                let mut n = 0;
                let mut to_warn = None;
                let callback = &mut |summary| {
                    n += 1;
                    match summary {
                        IndexSummary::Candidate(summary)
                        | IndexSummary::Yanked(summary)
                        | IndexSummary::Offline(summary)
                        | IndexSummary::Unsupported(summary, _)
                        | IndexSummary::Invalid(summary) => {
                            to_warn = Some(summary);
                        }
                    }
                };
                query_with_context(&*source, dep, kind, callback).await?;
                if n > 1 {
                    return Err(anyhow::anyhow!("found an override with a non-locked list"));
                }
                if let Some(to_warn) = to_warn {
                    self.warn_bad_override(&override_summary, &to_warn)?;
                }
                let override_summary = self.lock(override_summary);
                f(IndexSummary::Candidate(override_summary));
            }
        }

        Ok(())
    }

    fn describe_source(&self, id: SourceId) -> String {
        match self.sources.borrow().get(id) {
            Some(src) => src.describe(),
            None => id.to_string(),
        }
    }

    fn is_replaced(&self, id: SourceId) -> bool {
        match self.sources.borrow().get(id) {
            Some(src) => src.is_replaced(),
            None => false,
        }
    }
}

/// See [`PackageRegistry::lock`].
fn lock(
    locked: &LockedMap,
    patches: &HashMap<CanonicalUrl, Vec<PackageId>>,
    summary: Summary,
    mode: VersionReqMatchMode,
) -> Summary {
    let pair = locked
        .get(&(summary.source_id(), summary.name()))
        .and_then(|vec| vec.iter().find(|&&(id, _)| id == summary.package_id()));

    trace!("locking summary of {}", summary.package_id());

    // Lock the summary's ID if possible
    let summary = match pair {
        Some((precise, _)) => summary.override_id(*precise),
        None => summary,
    };
    summary.map_dependencies(|dep| {
        trace!(
            "\t{}/{}/{}",
            dep.package_name(),
            dep.version_req(),
            dep.source_id()
        );

        // If we've got a known set of overrides for this summary, then
        // one of a few cases can arise:
        //
        // 1. We have a lock entry for this dependency from the same
        //    source as it's listed as coming from. In this case we make
        //    sure to lock to precisely the given package ID.
        //
        // 2. We have a lock entry for this dependency, but it's from a
        //    different source than what's listed, or the version
        //    requirement has changed. In this case we must discard the
        //    locked version because the dependency needs to be
        //    re-resolved.
        //
        // 3. We have a lock entry for this dependency, but it's from a
        //    different source than what's listed. This lock though happens
        //    through `[patch]`, so we want to preserve it.
        //
        // 4. We don't have a lock entry for this dependency, in which
        //    case it was likely an optional dependency which wasn't
        //    included previously so we just pass it through anyway.
        //
        // Cases 1/2 are handled by `matches_id`, case 3 is handled specially,
        // and case 4 is handled by falling through to the logic below.
        if let Some((_, locked_deps)) = pair {
            let locked = locked_deps.iter().find(|&&id| {
                // If the dependency matches the package id exactly then we've
                // found a match, this is the id the dependency was previously
                // locked to.
                if dep.matches_id(id, mode) {
                    return true;
                }

                // If the name/version doesn't match, then we definitely don't
                // have a match whatsoever. Otherwise we need to check
                // `[patch]`...
                if !dep.matches_ignoring_source(id, mode) {
                    return false;
                }

                // ... so here we look up the dependency url in the patches
                // map, and we see if `id` is contained in the list of patches
                // for that url. If it is then this lock is still valid,
                // otherwise the lock is no longer valid.
                match patches.get(dep.source_id().canonical_url()) {
                    Some(list) => list.contains(&id),
                    None => false,
                }
            });

            if let Some(&locked) = locked {
                trace!("\tfirst hit on {}", locked);
                let mut dep = dep;

                // If we found a locked version where the sources match, then
                // we can `lock_to` to get an exact lock on this dependency.
                // Otherwise we got a lock via `[patch]` so we only lock the
                // version requirement, not the source.
                if locked.source_id() == dep.source_id() {
                    dep.lock_to(locked, mode);
                } else {
                    dep.lock_version(locked.version(), mode);
                }
                return dep;
            }
        }

        // If this dependency did not have a locked version, then we query
        // all known locked packages to see if they match this dependency.
        // If anything does then we lock it to that and move on.
        let v = locked
            .get(&(dep.source_id(), dep.package_name()))
            .and_then(|vec| vec.iter().find(|&&(id, _)| dep.matches_id(id, mode)));
        if let Some(&(id, _)) = v {
            trace!("\tsecond hit on {}", id);
            let mut dep = dep;
            dep.lock_to(id, mode);
            return dep;
        }

        trace!("\tnope, unlocked");
        dep
    })
}

/// A helper for selecting the summary, or generating a helpful error message.
///
/// Returns a tuple that the first element is the summary selected. The second
/// is a package ID indicating that the patch entry should be unlocked. This
/// happens when a match cannot be found with the `locked` one, but found one
/// via the original patch, so we need to inform the resolver to "unlock" it.
async fn summary_for_patch(
    original_patch: &Patch,
    orig_patch_url: &Url,
    locked: &Option<LockedPatchDependency>,
    mut summaries: Vec<Summary>,
    source: &dyn Source,
) -> CargoResult<(Summary, Option<PackageId>)> {
    if summaries.len() == 1 {
        return Ok((summaries.pop().unwrap(), None));
    }
    if summaries.len() > 1 {
        // TODO: In the future, it might be nice to add all of these
        // candidates so that version selection would just pick the
        // appropriate one. However, as this is currently structured, if we
        // added these all as patches, the unselected versions would end up in
        // the "unused patch" listing, and trigger a warning. It might take a
        // fair bit of restructuring to make that work cleanly, and there
        // isn't any demand at this time to support that.
        let mut vers: Vec<_> = summaries.iter().map(|summary| summary.version()).collect();
        vers.sort();
        let versions: Vec<_> = vers.into_iter().map(|v| v.to_string()).collect();
        return Err(anyhow::anyhow!(
            "patch for `{}` in `{}` resolved to more than one candidate\n\
            note: found versions: {}\n\
            help: check `{}` patch definition for `{}` in `{}`\n\
            help: select only one package using `version = \"={}\"`",
            &original_patch.dep.package_name(),
            &original_patch.dep.source_id(),
            versions.join(", "),
            &original_patch.dep.package_name(),
            orig_patch_url,
            original_patch.loc,
            versions.last().unwrap()
        ));
    }
    assert!(summaries.is_empty());
    // No summaries found, try to help the user figure out what is wrong.
    if let Some(locked) = locked {
        // Since the locked patch did not match anything, try the unlocked one.
        let orig_matches = source
            .query_vec(&original_patch.dep, QueryKind::Exact)
            .await
            .unwrap_or_else(|e| {
                tracing::warn!(
                    "could not determine unlocked summaries for dep {:?}: {:?}",
                    &original_patch.dep,
                    e
                );
                Vec::new()
            });

        let orig_matches = orig_matches
            .into_iter()
            .filter_map(|s| match s {
                IndexSummary::Candidate(s) => Some(s),
                _ => None,
            })
            .collect();

        let summary = Box::pin(summary_for_patch(
            original_patch,
            orig_patch_url,
            &None,
            orig_matches,
            source,
        ))
        .await?;
        return Ok((summary.0, Some(locked.package_id)));
    }
    // Try checking if there are *any* packages that match this by name.
    let name_only_dep = Dependency::new_override(
        original_patch.dep.package_name(),
        original_patch.dep.source_id(),
    );

    let name_summaries = source
        .query_vec(&name_only_dep, QueryKind::Exact)
        .await
        .unwrap_or_else(|e| {
            tracing::warn!(
                "failed to do name-only summary query for {:?}: {:?}",
                name_only_dep,
                e
            );
            Vec::new()
        });
    let mut vers = name_summaries
        .iter()
        .filter_map(|s| match s {
            IndexSummary::Candidate(s) => Some(s.version()),
            _ => None,
        })
        .collect::<Vec<_>>();
    let found = match vers.len() {
        0 => "".to_string(),
        1 => format!("version `{}`", vers[0]),
        _ => {
            vers.sort();
            let strs: Vec<_> = vers.into_iter().map(|v| v.to_string()).collect();
            format!("versions `{}`", strs.join(", "))
        }
    };
    Err(if found.is_empty() {
        anyhow::anyhow!(
            "patch location `{}` does not contain packages matching `{}`\n\
            help: check `{}` patch definition for `{}` in `{}`",
            &original_patch.dep.source_id(),
            &original_patch.dep.package_name(),
            &original_patch.dep.package_name(),
            orig_patch_url,
            original_patch.loc
        )
    } else {
        anyhow::anyhow!(
            "patch `{}` version mismatch\n\
            note: patch location contains {}, but patch definition requires `{}`\n\
            help: check patch location `{}`\n\
            help: check `{}` patch definition for `{}` in `{}`",
            &original_patch.dep.package_name(),
            found,
            &original_patch.dep.version_req(),
            &original_patch.dep.source_id(),
            &original_patch.dep.package_name(),
            orig_patch_url,
            original_patch.loc
        )
    })
}