bootstrapFiles should be updated periodically to avoid arcane bugs in stale binaries

[trofi]

Current bootstrapFiles in pkgs/stdenv/linux/bootstrap-files are not that current. Some of then come from 2019, some of them are more recent.

This causes various issues when newer packages (usually toolchains like gcc) fail to build against old tools.

A few recent examples:

• i686-linux bootstrap fails on filesystems with 64-bit inodes #253274: outdated patchelf does not understand 64-bit inodes (newer binary is fine)
• Bootstrap Reform and aarch64-linux's GCC Upgrade #208412: outdated libgcc does not provide basic atopic primitives needed by newer toolchain
• gcc: install info files serially #229898 (comment): outdated make enforces serial build when freshed one whould Just Work

Suggestion: update bootstrapFiles as part of the NixOS release process at least for the platforms supported by Hydra. That way we would get guaranteed 6 month cadence of updates.

[trofi]

Another example where update of bootstrapTools update is worth it is missing files to produce PIE and static-PIE files:

• stdenv: Enable PIE by default #252310

[trofi]

Yet another example is future transition to 64-bit time_t where would would need to rebuild bootstrapFiles a few more times as individual upstream packages are switching to 64-bit time_t internally:

• https://www.gnu.org/software/libc/manual/html_node/64_002dbit-time-symbol-handling.html

[Artturin]

@lovesegfault

I have created this script for you to hopefully help you update the bootstrap files.

#!/usr/bin/env bash

# before running git check out the latest eval revision here https://hydra.nixos.org/jobset/nixpkgs/cross-trunk
# hydra-check only checks the newest evaluation

crossSystems=()
mapfile -t crossSystems < <(nix eval --impure --expr 'let rel-cross = (import ./pkgs/top-level/release-cross.nix {}).bootstrapTools; in builtins.attrNames rel-cross' --json | jq '.[]' --raw-output)

whatToGet=(bootstrapFiles.bootstrapTools bootstrapFiles.busybox)

for system in "${crossSystems[@]}"; do
    echo ""
    echo "$system"
    for what in "${whatToGet[@]}"; do
        if nix run "github:nix-community/hydra-check" -- --arch "" --jobset "nixpkgs/cross-trunk" "bootstrapTools.${system}.${what}" 1>/dev/null; then
            echo -n "${what}:"
            output=$(nix build --quiet --no-link --json -f pkgs/top-level/release-cross.nix "bootstrapTools.${system}.${what}" | jq '.[] | .outputs | .out' --raw-output)
            sha256sum "$output"
        else
            echo "${what} for ${system} did not build in this evaluation and may have to be built manually"
            echo "\$ nix build --quiet --no-link --json -f pkgs/top-level/release-cross.nix \"bootstrapTools.${system}.${what}\" | jq '.[] | .outputs | .out' --raw-output"
            echo ""
        fi
    done
done

in #151399 (comment) you asked for the sha256sum of all the on-server components, however shouldn't it be okay to directly upload the packages from bootstrapFiles

nixpkgs/pkgs/stdenv/linux/make-bootstrap-tools.nix

Lines 230 to 233 in e009fad

	bootstrapFiles = {
	# Make them their own store paths to test that busybox still works when the binary is named /nix/store/HASH-busybox
	busybox = runCommand "busybox" {} "cp ${build}/on-server/busybox $out";
	bootstrapTools = runCommand "bootstrap-tools.tar.xz" {} "cp ${build}/on-server/bootstrap-tools.tar.xz $out";

Please join https://matrix.to/#/#stdenv:nixos.org so we can figure out a way to make the process more convenient for you.

[Ma27]

cc @NixOS/nixos-release-managers and @vcunat

[vcunat]

For the record, I believe it is practical for us to regenerate bootstrap tools once in a while, at least for the better supported platforms. Though twice a year might be an excessive rate IMO; every other year would probably still be OK.

[vcunat]

I know that some people in the community would prefer to go the other way and reduce the binaries as much as possible. (quick example)

However as it is now, making the bootstrap sequence much longer would make it very painful to change some parts like stdenv's setup.sh, even if the early bootstrap phases don't need them. If people are interested in that, I think we'd need to do those steps "separately"... basically like an independent method of producing good-enough tools that would then be plugged into our current bootstrapping instead of the binaries.

[Artturin]

I know that some people in the community would prefer to go the other way and reduce the binaries as much as possible. (quick example)

However as it is now, making the bootstrap sequence much longer would make it very painful to change some parts like stdenv's setup.sh, even if the early bootstrap phases don't need them. If people are interested in that, I think we'd need to do those steps "separately"... basically like an independent method of producing good-enough tools that would then be plugged into our current bootstrapping instead of the binaries.

It'll eventually be done once minimal-bootstrap is ready #227914

[vcunat]

Ah, thanks. I missed or forgot that one.

[trofi]

Another casualty of stale binaries:

gcc-8 (used in bootstrapTools) uses response files for a linker as soon as there are response files in it's inputs.
gcc-12 does not exhibit this behaviour. Normally it would not be an issue if expand-response-params would alsays be available. But it's not available in early bootstrap.

This makes it harder to use response files unconditionally for things like #255192

[lovesegfault]

Alright, taking this weekend to handle this, also I re-installed Element on my devices so that folks can reach me on Matrix.

Sorry for all the delays, work has been insane :)

[alyssais]

Oh, convenient that this is going on — I've just got Hydra to start building native bootstrap binaries for x86_64-unknown-linux-musl and aarch64-unknown-linux-musl (previously these were built, uh, not by Hydra), so if those could also be included in any bootstrap update, that'd be great! The process should be exactly the same as for all the other platforms with native Hydra builds.

[alyssais]

Do we still have time to get this done for 23.11?

[RaitoBezarius]

It's unclear to me what should be done to unblock this, are we waiting on OP or a merge?

[alyssais]

AIUI it's a matter of somebody with the permissions to upload to tarballs.nixos.org having the time and knowledge to do that. Usually that's @lovesegfault — I don't know if there's anybody else who meets those requirements.

[trofi]

I think we need a process to be:

• documented
• automated with a script
• executed periodically

[RaitoBezarius]

I think we need a process to be:

• documented
• automated with a script
• executed periodically

Agreed, I am trying to understand if we would like to get this cycle done even manually and who would be up to build this process and submit it to infrastructure people so we can implement it for 24.05 probably.

[Artturin]

One of the native bootstrap-files is failing to build
#258032
Would fix it but I don't know if it's the correct way
(A cross job should be added for it, but it'll still fail)

One of the cross bootstrap-files is failing to build

nix build --impure --expr 'with import ./pkgs/top-level/release-cross.nix {supportedSystems = [builtins.currentSystem];}; builtins.mapAttrs (k: v: v.bootstrapFiles.bootstrapTools) bootstrapTools'

Fails at the check phase of diffutils-x86_64-unknown-linux-musl

[ghost]

I see libgcc failures on ppc32 and alpha.

@trofi, it's not a regression if there is no test coverage for it. Please add a test if you care about these platforms. Nobody else has, so it stands to reason that nobody else does. You can be the first!

You've been dredging up a sequence of increasingly-obscure platforms in order to argue for reverting a project I invested an extraordinary amount of time in. Two of these platforms are no longer available for sale and the third of which is a multi-million-dollar mainframe. I find it very hard to believe that you own an S390 or use a DEC Alpha on a regular basis.

Taking into account all of the above, you seem to be going out of your way to make this personal. Please don't do that.

[trofi]

It has nothing personal about it. Debian occasionally asks me to debug the bugs on ports they care about. I think it's reasonable to try and help them. I only point out they used to work and proposed a possible solution to make them work again.

But I can stop responding to any threads you are in on github (or anywhere else) if you think it's personal and it will make you feel better.

[vcunat]

I agree we shouldn't block progress of anything on regressions for very "exotic" platforms (e.g. alpha, s390*). Some basic guidance was agreed on on https://github.com/nixos/rfcs/blob/master/rfcs/0046-platform-support-tiers.md

[trofi]

Absolutely. I would like to see at least a subset of bootstrapFiles to be updated on a regular basis. The to-be-written update script might need to be adapted to skip broken unsupported targets.

[lovesegfault]

FWIW, even without a script, I am willing to take a whole day to update all bootstrap files on each release of NixOS for the foreseeable future.

If that's what folks want, I can propose a process that would allow me to do it in a single day. I don't think what we're lacking is automation, primarily, but coordination.

[alyssais]

What coordination is lacking? Isn't it a case of downloading the latest builds of stdenvBootstrapTools (and maybe the cross versions for non-Hydra architectures), uploading them to tarballs.nixos.org, and updating the URLs and hashes? What is it that you need from other people?

[lovesegfault]

Well, that's roughly it, but there are two important bits missing:

1. I can't keep up with everything that's going on. Someone needs to prepare a "bootstrapFiles manifest" with everything that needs uploading and the associated info (see below for what that means)
2. I can upload bootstrap tarballs, but I cannot remove them, so I am cautious when performing the upload as I cannot undo mistakes.

Here's the process I follow to upload a new tarball, @amjoseph-nixpkgs is probably the person most familiar with this, since I've uploaded a number of tarballs for them:

1. Look at the Hydra build linked on GH, make sure nothing is off (built the
   right thing, on the right pkgs instance/channel/etc)
2. Check the build logs on Hydra, make sure there are no glaring issues that did
   not end up failing the build
3. Get the nix-store path, and sha256's from GH
4. nix build -L $storePath
5. Get the nixpkgs commit which generated the tarball
6. Check it out, build it, make sure it's reproducible
7. Figure out the correct subpath for this tarball
   1. x86_64-darwin → stdenv-darwin/x86_64
8. Check that the hashes within match the ones from GH
9. Construct the upload path
   1. s3://nixpkgs-tarballs/$os/$arch/$nixpkgs_commit
   2. e.g. s3://nixpkgs-tarballs/stdenv-darwin/x86_64/05ef940b94fe76e7ac06ea45a625adc8e4be96f9
10. Upload it
    1. e.g. aws s3 cp --recursive --acl public-read /nix/store/9h4d7s313wv3gkfwi493yr1wvdsz9lf2-stdenv-bootstrap-tools/on-server/ s3://nixpkgs-tarballs/stdenv-darwin/x86_64/05ef940b94fe76e7ac06ea45a625adc8e4be96f9
11. Download the uploaded tarballs, check hashes from Hydra match the hashes from the local build and match the hashes that were downloaded.

Here's an example of this process being followed: #188334

This is very manual, and is a lot of work for me.

I think the coordination I'd like to see is:

1. A boostrapTarballManifest file, which contains all the info above for all the tarballs we want to upload.
2. A review process where, after the manifest is created for a given nixpkgs revision, folks share the burden of validating the tarballs (i.e. the work I do of looking through build logs, re-building the tarballs and verifying that they are reproducible, etc.)
3. A tool I can use to ingest said manifest after it's been reviewed, and upload all the tarballs

I am happy to hack on (3), but I can't do the work of figuring out what (1) and (2) should look like, I think the stakeholders here should collaborate to define that.

[lovesegfault]

A follow-on thought, one thing that deserves consideration is the structure of the files in s3, it's kind of nonsensical right now.

If the plan for the future is "we always update all the tarballs periodically", perhaps it should be:

bootstrap/<nixpkgs-commit>/<arch>/<contents>

So, for example, we'd have:

bootstrap/05ef940b94fe76e7ac06ea45a625adc8e4be96f9/x86_64-apple-darwin/<stuff>
bootstrap/05ef940b94fe76e7ac06ea45a625adc8e4be96f9/aarch64-apple-darwin/<stuff>
bootstrap/05ef940b94fe76e7ac06ea45a625adc8e4be96f9/x86_64-unknown-linux-gnu/<stuff>
...

(cc. @Ericson2314, @amjoseph-nixpkgs b/c of platform naming)

[sdht0]

Any chance this is still happening for 23.11?

[trofi]

Filed #272750 to provide a mechanism to override bootstrapFiles on user's side.

[trofi]

I wrote a tiny PoC to extract all the info from hydra release for a single target (I'll try to document one-target-at-a-time).

I found a few inconsistencies across cross/native jobsets. #284090 should fix one of them.

The PoC itself

cat pkgs/stdenv/linux/bootstrap-files/refresh-tarballs.bash
#!/usr/bin/env bash

# How it works:
#
# For a given <target>:
# 1. fetch latest successful '.dist` job that contains reference tp busybox and bootstrapFiles
# 2. fetch oldest evaluation that contained that path, extract nixpkgs commit
# 3. fetch the `.dist` artifacts
# 4. canculate hashes and crash the commit

usage() {
    echo "Usage: $0 [ --commit ] <target>" >&2
    echo "Examples:" >&2
    echo "    $0 --commit i686-unknown-linux-gnu" >&2
    echo "    $0 aarch64-unknown-linux-gnu" >&2
    exit 1
}

die() {
    echo "ERROR: $*" >&2
    exit 1
}

info() {
    echo "INFO: $*" >&2
}

[[ ${#@} -eq 0 ]] && usage

target=$1

# Native and cross jobsets differ a bit. We'll have to pick the concrete
# one based on target.
jobset=unknown-jobset
job=unknown-job
s3_prefix=unknown-s3-prefix

case $target in
    # native targets:
    i686-unknown-linux-gnu)
        jobset="nixpkgs/trunk"
        job="stdenvBootstrapTools.${target}.dist"
        s3_prefix="stdenv-linux/i686"
        ;;
    riscv64-unknown-linux-gnu)
        jobset="nixpkgs/cross-trunk"
        job="bootstrapTools.${target}.dist"
        s3_prefix="stdenv-linux/riscv64"
        ;;

    # cross targets:
    *)
        die "Unsupported '$target': please extend '$0' for it's support"
        ;;
esac

latest_build_uri="https://hydra.nixos.org/job/$jobset/$job/latest-finished"
latest_build="$target.latest-build"
info "Fetching latest successful build from '${latest_build_uri}'"
curl -s -H "Content-Type: application/json" -L "$latest_build_uri" > "$latest_build"
[[ $? -ne 0 ]] && die "Failed to fetch latest successful build"

# We pick oldest instead of latest to make the result more stable across
# across unrelated updates. Ideally two subsequent runs should produce
# the same output (provided there are no bootstrapTools updates
# committed between the two).
latest_build_id=$(jq '.id' < "$latest_build")
oldest_eval_id=$(jq '.jobsetevals|min' < "$latest_build")
build_uri="https://hydra.nixos.org/build/${latest_build_uri}"

eval_uri="https://hydra.nixos.org/eval/${oldest_eval_id}"
eval_meta="$target.eval-meta"
info "Fetching oldest eval details from '${eval_uri}' (can take a minute)"
curl -s -H "Content-Type: application/json"  -L "${eval_uri}" > "$eval_meta"
[[ $? -ne 0 ]] && die "Failed to fetch eval metadata"

nixpkgs_revision=$(jq --raw-output ".jobsetevalinputs.nixpkgs.revision" < "$eval_meta")

drvpath=$(jq --raw-output '.drvpath' < "${latest_build}")
outpath=$(jq --raw-output '.buildoutputs.out.path' < "${latest_build}")
build_timestamp=$(jq --raw-output '.timestamp' < "${latest_build}")
build_time=$(TZ=UTC LANG=C date --date="@${build_timestamp}" --rfc-email)

info "Fetching bootstrap tools manifest to calculate hashes from '${outpath}'"
bootstrap_manifest=$(nix-store --realize "${outpath}")/nix-support/hydra-build-products

# The file format is the following:
#    file tarball /nix/store/...-stdenv-bootstrap-tools/on-server/bootstrap-tools.tar.xz
#    file busybox /nix/store/...-stdenv-bootstrap-tools/on-server/busybox

tools_dir=unknown-tools-dir
bootstrap_tools_hash=unknown-bootstrap-tools-hash
busybox_hash=unknown-busybox-hash
while read type flavour path; do
    case "$type:$flavour" in
        file:tarball)
            tools_dir=$(dirname "$path")
            bootstrap_tools_hash=$(nix-hash --to-sri sha256:$(nix-prefetch-url --name bootstrap-tools.tar.xz "file://$path"))
            ;;
        file:busybox)
            busybox_hash=$(nix-hash --to-sri sha256:$(nix-prefetch-url --executable --name busybox "file://$path"))
            ;;
        *)
            die "'$type:$flavour' is unhandled file type"
    esac
done < "$bootstrap_manifest"

# Calculate hashes:
#   $ nix-hash --to-sri sha256:...

target_file="${target}.nix"
info "Writing '${target_file}'"
cat > "${target_file}" <<EOF
# Autogenerated by pkgs/stdenv/linux/bootstrap-files/refresh-tarballs.bash as:
# $ ./refresh-tarballs.bash ${target}
#
# Metadata:
# - latest hydra build: ${latest_build_uri}
# - resolved hydra build: ${build_uri}
# - build time: ${build_time}
# - nixpkgs revision: ${nixpkgs_revision}
# - manifest derivation: ${drvpath}
# - manifest output: ${outpath}
# - tools directory: ${tools_dir}
{
  busybox = import <nix/fetchurl.nix> {
    url = "http://tarballs.nixos.org/${s3_prefix}/${nixpkgs_revision}/busybox";
    hash = "${busybox_hash}";
    executable = true;
  };

  bootstrapTools = import <nix/fetchurl.nix> {
    url = "http://tarballs.nixos.org/${s3_prefix}/${nixpkgs_revision}/bootstrap-tools.tar.xz";
    hash = "${bootstrap_tools_hash}";
  };
}
EOF

I'll send PoC itself for review if it's a reasonable direction and will update at least i686 and musl targets to check the procedure with @lovesegfault

[trofi]

Proposed a script that automatically generates bootstrap script update as:

• maintainers/scripts/bootstrap-files: documentation and a script to up… #284541

Example PR for i686-unknown-linux-gnu:

• pkgs/stdenv/linux: update i686-unknown-linux-gnu bootstrap-files #284542

[trofi]

Proposed a small unification on darwin side to get closer to linux WRT seed file location:

• pkgs/stdenv/darwin: move bootstrap files definitions to bootstrap-fil… #284628