Support cargo vendor

[ehuss]

The cargo vendor command currently does not know about std dependencies. Will need to figure out how to handle the root dependencies, since they are special.

Source replacement may also be broken, I haven't tried it.

[Gankra]

Firefox is running into this for tsan CI

I've currently hacked up a script that can weave in the std tomls to a project's vendor, but yeah it'd be nice if this was handled automatically: https://gist.github.com/Gankra/6a025b5cce9d9b047e46a5caeded3050

[alexcrichton]

One thing to consider for this is that vendoring the standard library's dependencies effectively locks you to a compiler version since the dependencies are likely to drift over time. This may or may not be something that Cargo wants to provide a first-class warning/error about rather than a bland "couldn't find crate version in registry" error.

In theory this integration wouldn't be too too hard though, mostly just loading up the Resolve for the standard library, around here-ish

[ehuss]

Copying my comment from Zulip: We have recently discussed adding all the dependencies to the rust-src component for other reasons (matklad was interested in rust-analyzer support). That would be a potential solution of just not changing cargo vendor, and keeping the standard library in sync with the compiler.

I'm uncertain how we can wire that in to the resolver, though. It may be tricky to use source replacement or paths overrides (they cannot currently be customized independently). An option would be to add patches, but it doesn't know what to patch until after it loads the resolver. Another option would be to surgically modify the Dependency information to be path dependencies when the standard library workspace is loaded. None of those sound terribly appealing to me, though.

[alexcrichton]

Oh actually I think that's a great idea to inclue the dependencies in rust-src! I think that adding patch entries for all members in the vendor directory of rust-src should work perhaps? I don't think there should be much downside to just iterating through the vendor directory and adding a patch-per-crate. We would want to make sure that any "patch not used" warnings were suppressed as well.

[Gankra]

One thing to consider for this is that vendoring the standard library's dependencies effectively locks you to a compiler version since the dependencies are likely to drift over time. This may or may not be something that Cargo wants to provide a first-class warning/error about rather than a bland "couldn't find crate version in registry" error.

This isn't an issue for firefox's usecase in particular. We only need these vendors for our tsan CI, and that uses pinned versions of all the toolchains. Although it is a problem if we want these vendoring checked in and we want people to be able to bump non-std dependencies without having the nightly we use for CI. So it's possible we want to vendor live in CI as part of the build. Or we need to carefully identify deps which are only used by std so we can preserve them across non-std-vendors.

Certainly if the dependencies were just part of rust-src and automagically weaved into the build, that would solve the issue. No need to worry about weaving together different "kinds" of vendor or anything else.

[alexcrichton]

Yeah pinning versions is too wonky and restrictive, so I think including the dependencies in rust-src is a great solution.

[Gankra]

Anything I can do to help the plan along?

[ehuss]

I think the first step is getting the dependencies included in the rust-src component (the relevant code is here). Unfortunately, cargo vendor doesn't have a way to only vendor a subset of packages. I'm not sure how to address that. Some ideas:

• It could call cargo vendor and then just delete everything it doesn't need.
• cargo vendor could be enhanced to only vendor a subset.
• The standard library could be split into a separate workspace (matklad is promoting this, but I am concerned about the downsides).
• Some hacky solution of manually extracting the minimum set of dependencies in rustbuild.

None of these options seem easy.

[Gankra]

I believe the approach I'm using to test out tsan builds right now would work for building "subset vendors": https://gist.github.com/Gankra/6a025b5cce9d9b047e46a5caeded3050

Essentially just temporarily copying the root lockfile to each individual library and then cargo vendor -s'ing them in. Vendor doesn't pull things in just because they're in the lockfile, but it will respect the constraints it applies to the dependency graph.

This would be a fairly maintainable hack as you would only need to directly -s in std and test (and maybe panic_abort for good measure?). Everything else is transitively pulled in by those ones. (hmm, maybe only test is actually needed since it explicitly depends on std and both the panic crates?)

What do you think of shipping that?

[alexcrichton]

I think that's probably bordering on the 4th bullet of @ehuss's comment above, but I think something along those lines may be workable. As we do for build-std in general we're just trying to load one "root" package and vendor that, so the filtering in cargo-vendor probably isn't the hardest thing in the world. Once you're inside cargo-vendor you've got access to lots of fun internals of Cargo as well which the CLI may not expose.

[Gankra]

Would -Zbuild-std also need to be reworked, or would the existence of a vendoring .cargo/config at the root of the rust-src directory automagically do the right thing?

[alexcrichton]

We would need to update Cargo's implementation of -Zbuild-std as well to automatically inject [patch] entries for crates found in rust-src

[Gankra]

Ok I'll be working on this for the next week or two, hopefully won't be too messy.

[ehuss]

Ah OK so this is what is tripping us up. I saw https://doc.rust-lang.org/cargo/reference/source-replacement.html#local-registry-sources which says it's OK to have this subset mirror.

By "subset" it is saying that you don't have to mirror all of crates.io locally. It doesn't mean that it will allow something to be missing from the subset. Source replacement works as a "full replacement", not as an overlay.

Also, a local-registry is not what we are talking about here. The vendoring mechanism we are talking about is a directory-source (ala cargo vendor).

[Ericson2314]

@ehuss OK thanks for the clarifications. Two more questions:

1. @Gankra's PR both changed the rust-src component format along with trying not make cargo vendor, work right?
2. The regression just when using cargo vendor, or even without using it? I thought it was the latter, and due to the way the new rust-src was patched in.

My previous comments frankly forgot about the original issue --- cargo vendor --- and were all about what went wrong in the regression case, which as mentioned I thought happened regardless of whether cargo vendor was actually attempted.

[ehuss]

The original changes made Cargo treat the standard library dependencies as always vendored (with -Z build-std), whether or not there is a local source replacement. So it is independent if you run cargo vendor locally.

There were multiple regressions as detailed in #23 (comment). These happened with all -Z build-std builds, and also affected distribution builds of rustc itself.

[Gankra]

Hmm ok so that was all a bunch of confusion and didn't actually help us figure out the "right" solution, eh?

@ehuss this stuff is really out of my league -- have you had a chance to mull over the possible solutions?

As far as the partial vendoring thing goes, I'm starting to lean towards just giving up on partial vendoring and just grabbing the whole rustc graph. AIUI that should be fairly unproblematic and also not need to pull in the really heavy stuff like llvm?

Making patching in the vendor totally seamless is the nasty bit I really can't wrap my head around. Is it potentially as simple as combining the WIP patches you and Alex posted?

If we don't have any clear solutions, firefox does have the nuclear option of maintaining patches for cargo for our builds, since my solution did work fine for us. But obviously that's a gross mess we'd rather not maintain.

[cuviper]

As far as the partial vendoring thing goes, I'm starting to lean towards just giving up on partial vendoring and just grabbing the whole rustc graph. AIUI that should be fairly unproblematic and also not need to pull in the really heavy stuff like llvm?

There's still some heavy stuff in there -- you can see that in the full rustc-1.50.0-src.tar.xz in vendor/:

$ dua | tail
   7.69 MB jemalloc-sys
   7.88 MB winapi
   9.75 MB vte
  11.63 MB libgit2-sys
  12.16 MB libnghttp2-sys
  25.01 MB curl-sys
  27.65 MB openssl-src
  53.83 MB winapi-i686-pc-windows-gnu
  56.12 MB winapi-x86_64-pc-windows-gnu
 382.91 MB total

[raphaelcohn]

This is clearly hard to fix - and after my error report in #65 I discovered how hard. @ehuss is completely right about the collision of dependencies in a workspace when vendored - my workspace was using hashbrown 0.9.1, yet the std dependencies needed 0.9.0 and bang! conflict. Which surprised me, until I recalled the point of a workspace is to provide consistent versions.

That means in practice, either I have to track the exact same matching package versions as std (not a particularly easy thing to simply know without a bit of digging) or give up building std. Or provide a sensible mechanism for 'overriding' package version choices in std. Also not nice. Or providing the ideal of cascading workspaces, which works, but is complex. Since std does not expose its private dependencies (eg hashbrown), this is viable and defensible - but tough.

Along with several common packages not playing nicely with cross compilation (libc and cc, I'm looking at you) without wrapper scripts setting up env vars et al, this is just a bit too hard. Giving up for now, personally. All I wanted to do was find a better way to build linux x86_64 musl targets with a few more target features (eg avx512 stuff)...

[Ericson2314]

@Gankra

Hmm ok so that was all a bunch of confusion and didn't actually help us figure out the "right" solution, eh?

Yes, haha.

But, still believe #64 is a solution here

Which surprised me, until I recalled the point of a workspace is to provide consistent versions.

Heh, well !64 would at least make us have to face down this problem regardless of vendoring. Without knowing all the details, this seems wrong. Per https://github.com/rust-lang/rfcs/blob/master/text/1977-public-private-dependencies.md std should have private dependencies and that is the basis by which this is allowed. Any vendoring, what's in/not-in the workspace, etc. should be completely transparent and independent of the "platonic" build plan.

[baloo]

For nix users where cargo-vendor is the only option available, I found a "workaround" in downloading rust's compiler dependencies manually and merging them.
The result can be found here:
baloo/nixos-firmware@e108c50

[kgrech]

The easiest workaround is probably something like this:

RUST_DIR="$(rustc --print=sysroot)"

# Copy the Cargo.lock for Rust to place the `cargo vendor` will see
cp "$RUST_DIR"/lib/rustlib/src/rust/Cargo.lock "$RUST_DIR"/lib/rustlib/src/rust/library/test/

cargo vendor \
    --manifest-path Cargo.toml \
    --sync "$RUST_DIR"/lib/rustlib/src/rust/library/test/Cargo.toml \
    cargo_vendor

# Remove copied file
rm "$RUST_DIR"/lib/rustlib/src/rust/library/test/Cargo.lock

[Dev380]

@kgrech sorry if this is a dumb question, but how would I make cargo use the vendored stdlib when building? cargo vendor says

[source.crates-io]
replace-with = "vendored-sources"

[source.vendored-sources]
directory = "vendor"

but obviously it's not from crates.io and I don't know the registry to tell it to replace

[kgrech]

@Dev380 as far as I remember, the std library sources need to be installed using rustup. I think they never fetched from crates.io. it is only for dependencies of std lib, but not stdlib itself.

Not sure I fully understand you question, but you just put what cargo vendor prints to your configuration. vendor is just a folder created by cargo vendor

[kleisauke]

• The standard library could be split into a separate workspace (matklad is promoting this, but I am concerned about the downsides).

This has been done via PR rust-lang/rust#128534 (nightly-2024-08-05).

[bjorn3]

I think it would make sense to provide pre-vendored versions of all standard library dependencies in the rust-src component and have -Zbuild-std unconditionally use those instead of downloading them from crates.io. Unlike adding support for vendoring the standard library to cargo vendor, this will ensure that the right versions for the toolchain that will build the project are available even if a different toolchain version vendored all dependencies.

[weihanglo]

have -Zbuild-std unconditionally use those instead of downloading them from crates.io.

Haven't followed the entire thread, but I think we may want vendoring as an opt-in feature. Tarball size may not matter (285M vs 297M gz tarball size). However, people may want to patch dependencies when building std. We should at this ensure that case to work.

[hoodmane]

I'm running into this now and it's quite a knot to try to work around.