Change default cargo artifact installation to use symlinks where poss…

…ible (#150) * Define a new installation mode which symlinks identical artifacts against a directory of previously generated ones * This allows for linear space usage in the Nix store across many chained derivations (as opposed to using a zstd compressed tarball which uses quadratic space across many chained derivations) * This new installation mode is the new default for all cargo based builds. The previous behavior is still available by setting `installCargoArtifactsMode = "use-zstd";` on a derivation * `buildPackage` will continue to use zstd compressed tarballs while building dependencies (unless either of `cargoArtifacts` or `installCargoArtifactsMode` is defined, in which case they will be honored)
ipetkov · Oct 29, 2022 · 975bda7 · 975bda7
1 parent f48777a
commit 975bda7
Show file tree

Hide file tree

Showing 6 changed files with 175 additions and 26 deletions.
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -6,6 +6,22 @@ and this project adheres to [Semantic Versioning](http://semver.org/spec/v2.0.0.
 
 ## Unreleased
 
+### Added
+* A new installation mode has been defined which symlinks identical cargo
+  artifacts against previously generated ones. This allows for linear space
+  usage in the Nix store across many chained derivations (as opposed to using a
+  zstd compressed tarball which uses quadratic space across many chained
+  derivations).
+
+### Changed
+* **Breaking**: all cargo-based derivations will now default to using symlinking
+  their installed artifacts together instead of using zstd compressed tarballs.
+  To get the old behavior back, set `installCargoArtifactsMode = "use-zstd";` in
+  the derivation.
+  - Note that `buildPackage` will continue to use zstd compressed tarballs while
+    building dependencies (unless either of `cargoArtifacts` or
+    `installCargoArtifactsMode` is defined, in which case they will be honored)
+
 ## [0.9.0] - 2022-10-29
 
 ### Changed

diff --git a/ci/run-tests.sh b/ci/run-tests.sh
@@ -11,7 +11,10 @@ main() {
   runLocked="1"
 
   while [ $# -gt 0 ]; do
-    case "$1" in
+    local arg="$1";
+    shift
+
+    case "${arg}" in
       "--locked")
         runLocked="1"
         runStable=""
@@ -21,7 +24,7 @@ main() {
         runStable="1"
         ;;
       *)
-        echo "unrecognized option $1"
+        echo "unrecognized option ${arg}"
         exit 1
         ;;
     esac
@@ -58,4 +61,4 @@ runtests() {
   done
 }
 
-main
+main "$@"
diff --git a/docs/API.md b/docs/API.md
@@ -168,7 +168,8 @@ subsequently install from that log.
   `target` directory, which will be reused at the start of the derivation.
   Useful for caching incremental cargo builds.
   - Default value: the result of `buildDepsOnly` after applying the arguments
-    set (with the respective default values)
+    set (with the respective default values). `installCargoArtifactsMode` will
+    be set to `"use-zstd"` if not specified.
 * `cargoBuildCommand`: A cargo invocation to run during the derivation's build
   phase
   - Default value: `"cargo build --profile release"`
@@ -1095,7 +1096,14 @@ directory using a previous derivation. It takes two positional arguments:
    * If not specified, the value of `$cargoArtifacts` will be used
    * If `cargoArtifacts` is not specified, an error will be raised
    * If the specified path is a directory which contains a file called
-     `target.tar.zst`, then that file will be used during unpacking
+     `target.tar.zst`, then that file will be used as specified below
+   * If the specified path is a file (and not a directory) it is assumed that it
+     contains a zstd compressed tarball and will be decompressed and unpacked
+     into the specified cargo artifacts directory
+   * If the specified path is a directory which contains another directory
+     called `target`, then that directory will be used as specified below
+   * If the specified path is a directory, its contents will be copied into the
+     specified cargo artifacts directory
    * The previously prepared artifacts are expected to be a zstd compressed
      tarball
 1. the path to cargo's artifact directory, where the previously prepared
@@ -1112,18 +1120,56 @@ post patch hook.
 
 ### `lib.installCargoArtifactsHook`
 
-Defines `prepareAndInstallCargoArtifactsDir()` which handles installing cargo's
-artifact directory to the derivation's output. It takes two positional
-arguments:
+Defines `compressAndInstallCargoArtifactsDir()` which handles installing
+cargo's artifact directory to the derivation's output as a zstd compressed
+tarball. It takes two positional arguments:
 1. the installation directory for the output.
-   * If not specified, the value of `$out` will be used
+   * An error will be raised if not specified
    * Cargo's artifact directory will be compressed as a reproducible tarball
      with zstd compression. It will be written to this directory and named
      `target.tar.zstd`
+1. the path to cargo's artifact directory
+   * An error will be raised if not specified
+
+Defines `dedupAndInstallCargoArtifactsDir()` which handles installing
+cargo's artifact directory to the derivation's output after deduplicating
+identical files against a directory of previously prepared cargo artifacts.
+It takes three positional arguments:
+1. the installation directory for the output.
+   * An error will be raised if not specified
+   * If the specified path is a directory which exists then the current cargo
+     artifacts will be compared with the contents of said directory. Any files
+     whose contents and paths match will be symbolically linked together to
+     reduce the size of the data stored in the Nix store.
+1. the path to cargo's artifact directory
+   * An error will be raised if not specified
+1. a path to the previously prepared cargo artifacts
+   * An error will be raised if not specified
+   * `/dev/null` can be specified here if there is no previous directory to
+     deduplicate against
+
+Defines `prepareAndInstallCargoArtifactsDir()` which handles installing cargo's
+artifact directory to the derivation's output. It takes three positional
+arguments:
+1. the installation directory for the output.
+   * If not specified, the value of `$out` will be used
+   * Cargo's artifact directory will be installed based on the installation mode
+     selected below
 1. the path to cargo's artifact directory
    * If not specified, the value of `$CARGO_TARGET_DIR` will be used
    * If `CARGO_TARGET_DIR` is not set, cargo's default target location  (i.e.
      `./target`) will be used.
+1. the installation mode to apply
+   * If not specified, the value of `$installCargoArtifactsMode` will be used.
+     If `$installCargoArtifactsMode` is not specified, a default value of
+     `"use-symlink" will be used
+   * If set to "use-symlink" then `dedupAndInstallCargoArtifactsDir()` will be
+     used.
+     - If `$cargoArtifacts` is defined and `$cargoArtifacts/target` is a valid
+       directory, it will be used during file deduplication
+   * If set to "use-zstd" then `compressAndInstallCargoArtifactsDir()` will be
+     used.
+   * Otherwise an error will be raised if the mode is not recognized
 
 **Automatic behavior:** if `doInstallCargoArtifacts` is set to `1`, then
 `prepareAndInstallCargoArtifactsDir "$out" "$CARGO_TARGET_DIR"` will be run as a

diff --git a/lib/buildPackage.nix b/lib/buildPackage.nix
@@ -35,7 +35,9 @@ mkCargoDerivation (cleanedArgs // memoizedArgs // {
   doCheck = args.doCheck or true;
   doInstallCargoArtifacts = args.doInstallCargoArtifacts or false;
 
-  cargoArtifacts = args.cargoArtifacts or (buildDepsOnly args // memoizedArgs);
+  cargoArtifacts = args.cargoArtifacts or (buildDepsOnly (args // memoizedArgs // {
+    installCargoArtifactsMode = args.installCargoArtifactsMode or "use-zstd";
+  }));
 
   buildPhaseCargoCommand = args.buildPhaseCargoCommand or ''
     cargoBuildLog=$(mktemp cargoBuildLogXXXX.json)

diff --git a/lib/setupHooks/inheritCargoArtifactsHook.sh b/lib/setupHooks/inheritCargoArtifactsHook.sh
@@ -5,15 +5,41 @@ inheritCargoArtifacts() {
   local cargoTargetDir="${2:-${CARGO_TARGET_DIR:-target}}"
 
   if [ -d "${preparedArtifacts}" ]; then
-    local preparedArtifacts="${preparedArtifacts}/target.tar.zst"
+    local candidateTarZst="${preparedArtifacts}/target.tar.zst"
+    local candidateTargetDir="${preparedArtifacts}/target"
+
+    if [ -f "${candidateTarZst}" ]; then
+      local preparedArtifacts="${candidateTarZst}"
+    elif [ -d "${candidateTargetDir}" ]; then
+      local preparedArtifacts="${candidateTargetDir}"
+    fi
   fi
 
+  mkdir -p "${cargoTargetDir}"
   if [ -f "${preparedArtifacts}" ]; then
-    mkdir -p "${cargoTargetDir}"
-    echo "copying cargo artifacts from ${preparedArtifacts} to ${cargoTargetDir}"
-
+    echo "decompressing cargo artifacts from ${preparedArtifacts} to ${cargoTargetDir}"
+
     zstd -d "${preparedArtifacts}" --stdout | \
       tar -x -C "${cargoTargetDir}" --strip-components=1
+  elif [ -d "${preparedArtifacts}" ]; then
+    echo "copying cargo artifacts from ${preparedArtifacts} to ${cargoTargetDir}"
+
+    # NB: rustc doesn't like it when artifacts are either symlinks or hardlinks to the store
+    # (it tries to truncate files instead of unlinking and recreating them)
+    # so we're forced to do a full copy here :(
+    #
+    # Notes:
+    # - --no-target-directory to avoid nesting (i.e. `./target/target`)
+    # - preserve timestamps to avoid rebuilding
+    # - no-preserve mode to ensure copies are writable
+    cp -r "${preparedArtifacts}" \
+      --no-target-directory "${cargoTargetDir}" \
+      --preserve=timestamps \
+      --no-preserve=mode
+
+    # NB: cargo also doesn't like it if `.cargo-lock` files remain with a
+    # timestamp in the distant past so we need to delete them here
+    find "${cargoTargetDir}" -name '.cargo-lock' -delete
   else
     echo unable to copy cargo artifacts, \"${preparedArtifacts}\" looks invalid
     false

diff --git a/lib/setupHooks/installCargoArtifactsHook.sh b/lib/setupHooks/installCargoArtifactsHook.sh
@@ -1,23 +1,79 @@
+compressAndInstallCargoArtifactsDir() {
+  local dir="${1:?destination directory not defined}"
+  local cargoTargetDir="${2:?cargoTargetDir not defined}"
+
+  mkdir -p "${dir}"
+
+  local dest="${dir}/target.tar.zst"
+  echo "compressing ${cargoTargetDir} to ${dest}"
+  (
+    export SOURCE_DATE_EPOCH=1
+    tar --sort=name \
+      --mtime="@${SOURCE_DATE_EPOCH}" \
+      --owner=0 \
+      --group=0 \
+      --numeric-owner \
+      --pax-option=exthdr.name=%d/PaxHeaders/%f,delete=atime,delete=ctime \
+      -c "${cargoTargetDir}" | zstd -o "${dest}"
+  )
+}
+
+dedupAndInstallCargoArtifactsDir() {
+  local dest="${1:?destination directory not defined}"
+  local cargoTargetDir="${2:?cargoTargetDir not defined}"
+  local prevCargoTargetDir="${3:?prevCargoTargetDir not defined}"
+
+  mkdir -p "${dest}"
+
+  if [ -d "${prevCargoTargetDir}" ]; then
+    echo "symlinking duplicates in ${cargoTargetDir} to ${prevCargoTargetDir}"
+
+    while read -r fullTargetFile; do
+      # Strip the common prefix of the current target directory
+      local targetFile="${fullTargetFile#"${cargoTargetDir}"}"
+      # Join the path and ensure we don't have a duplicate `/` separator
+      local candidateOrigFile="${prevCargoTargetDir}/${targetFile#/}"
+
+      if cmp --silent "${candidateOrigFile}" "${fullTargetFile}"; then
+        ln --symbolic --force --logical "${candidateOrigFile}" "${fullTargetFile}"
+      fi
+    done < <(find "${cargoTargetDir}" -type f)
+  fi
+
+  echo installing "${cargoTargetDir}" to "${dest}"
+  mv "${cargoTargetDir}" --target-directory="${dest}"
+}
+
 prepareAndInstallCargoArtifactsDir() {
   # Allow for calling with customized parameters
   # or fall back to defaults if none are provided
   local dir="${1:-${out}}"
   local cargoTargetDir="${2:-${CARGO_TARGET_DIR:-target}}"
-  local dest="${dir}/target.tar.zst"
-
-  echo "copying ${cargoTargetDir} to ${dest}"
+  local mode="${3:-${installCargoArtifactsMode:-use-symlink}}"
 
-  export SOURCE_DATE_EPOCH=1
   mkdir -p "${dir}"
 
-  # See: https://reproducible-builds.org/docs/archives/
-  tar --sort=name \
-    --mtime="@${SOURCE_DATE_EPOCH}" \
-    --owner=0 \
-    --group=0 \
-    --numeric-owner \
-    --pax-option=exthdr.name=%d/PaxHeaders/%f,delete=atime,delete=ctime \
-    -c "${cargoTargetDir}" | zstd -o "${dest}"
+  case "${mode}" in
+    "use-zstd")
+      compressAndInstallCargoArtifactsDir "${dir}" "${cargoTargetDir}"
+      ;;
+
+    "use-symlink")
+      # Placeholder if previous artifacts aren't present
+      local prevCargoTargetDir="/dev/null"
+
+      if [ -n "${cargoArtifacts}" ] && [ -d "${cargoArtifacts}/target" ]; then
+        local prevCargoTargetDir="${cargoArtifacts}/target"
+      fi
+
+      dedupAndInstallCargoArtifactsDir "${dir}" "${cargoTargetDir}" "${prevCargoTargetDir}"
+      ;;
+
+    *)
+      echo "unknown mode: \"${mode}\""
+      false
+      ;;
+  esac
 }
 
 if [ "1" = "${doInstallCargoArtifacts-}" ]; then