diff options
Diffstat (limited to '.github/workflows/regression_template.yml')
| -rw-r--r-- | .github/workflows/regression_template.yml | 106 |
1 files changed, 104 insertions, 2 deletions
diff --git a/.github/workflows/regression_template.yml b/.github/workflows/regression_template.yml index 392f6a96..decfadd4 100644 --- a/.github/workflows/regression_template.yml +++ b/.github/workflows/regression_template.yml @@ -41,6 +41,21 @@ on: default: 15 required: false type: number + # Minutes for the test step. Sixty covers every ThreadX suite, whose + # longest observed run is 37 minutes. NetX Duo runs all 43 configurations + # of its full suite in this one step, which took 201 minutes on 25 Oct + # 2026, so it passes a larger value. + test_timeout_minutes: + default: 60 + required: false + type: number + # Keep the .deb files apt downloads in the install step in the Actions + # cache, so a later run installs them without fetching them again. See the + # apt cache steps below. + apt_cache: + default: true + required: false + type: boolean skip_test: default: false required: false @@ -123,6 +138,59 @@ jobs: with: submodules: true + # The apt cache. The install scripts retry and bound every fetch, but a + # mirror can be slow without being down: on 23 Oct 2026 it served packages + # at 50 to 250 KB/s while indexes still came at 12 MB/s, and the ThreadX and + # RISC-V installs timed out on all three attempts inside an 11 MB and a 58 MB + # file. With the .deb files restored from the Actions cache, apt fetches + # only the indexes and whatever changed since the cache was saved. + # + # It is done here rather than in the install scripts so that every caller's + # script gets it unchanged: an apt.conf.d entry points apt-get's archive + # directory at the restored one. A restored file cannot alter what is + # installed, because apt checks every .deb against the hash in the signed + # index before using it, and fetches it again if the two differ. + # + # The key carries the runner image, whose package versions the cache + # matches, and the install script, whose package list it holds. A miss on + # either still restores the closest older cache, so a new image or one + # added package costs only the packages that changed. + - name: Prepare the apt cache key + id: apt_key + if: ${{ inputs.apt_cache }} + env: + INSTALL_SCRIPT: ${{ inputs.install_script }} + run: | + echo "dir=${RUNNER_TEMP}/apt-archives" >> "$GITHUB_OUTPUT" + echo "image=${ImageOS}-${ImageVersion}" >> "$GITHUB_OUTPUT" + echo "script=$( (sha256sum "$INSTALL_SCRIPT" 2> /dev/null || echo none) | cut -c1-16)" >> "$GITHUB_OUTPUT" + + - name: Restore the apt cache + id: apt_restore + if: ${{ inputs.apt_cache }} + uses: actions/cache/restore@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0 + with: + path: ${{ steps.apt_key.outputs.dir }} + key: apt-${{ steps.apt_key.outputs.image }}-${{ steps.apt_key.outputs.script }} + restore-keys: | + apt-${{ steps.apt_key.outputs.image }}- + apt- + + # apt downloads as the unprivileged _apt user, which must own partial/. + # Keep-Downloaded-Packages is set because the apt front end deletes what it + # installed by default; apt-get keeps it, but a caller's script may use + # either. + - name: Point apt at the cache + if: ${{ inputs.apt_cache }} + env: + APT_DIR: ${{ steps.apt_key.outputs.dir }} + run: | + sudo mkdir -p "$APT_DIR/partial" + sudo chown _apt:root "$APT_DIR/partial" + printf 'Dir::Cache::Archives "%s/";\nAPT::Keep-Downloaded-Packages "true";\nBinary::apt::APT::Keep-Downloaded-Packages "true";\n' "$APT_DIR" \ + | sudo tee /etc/apt/apt.conf.d/99-ci-archive-cache > /dev/null + echo "apt cache: $(find "$APT_DIR" -maxdepth 1 -name '*.deb' | wc -l) packages restored" + # apt and PyPI have stalled this step twice, for 55 minutes and for over two # hours, against a normal 27 to 152 seconds. Nothing here had a timeout, so a # stall ran until the six hour job limit and cost a whole run rather than @@ -131,6 +199,29 @@ jobs: timeout-minutes: ${{ inputs.install_timeout_minutes }} run: ${{ inputs.install_script }} + # Saved straight after the install rather than at the end of the job, so a + # run whose tests fail still leaves the packages for the next one. autoclean + # drops .deb files the indexes no longer list, so the cache does not grow + # with every superseded version. apt's lock file and partial/ are root-only + # and are removed so the cache action, which runs unprivileged, can read the + # rest. + - name: Save the apt cache + if: ${{ inputs.apt_cache && steps.apt_restore.outputs.cache-hit != 'true' }} + env: + APT_DIR: ${{ steps.apt_key.outputs.dir }} + run: | + sudo apt-get autoclean -qq + sudo rm -rf "$APT_DIR/partial" "$APT_DIR/lock" + sudo chmod -R a+rX "$APT_DIR" + echo "apt cache: $(find "$APT_DIR" -maxdepth 1 -name '*.deb' | wc -l) packages, $(du -sh "$APT_DIR" | cut -f1)" + + - name: Upload the apt cache + if: ${{ inputs.apt_cache && steps.apt_restore.outputs.cache-hit != 'true' }} + uses: actions/cache/save@55cc8345863c7cc4c66a329aec7e433d2d1c52a9 # v6.1.0 + with: + path: ${{ steps.apt_key.outputs.dir }} + key: apt-${{ steps.apt_key.outputs.image }}-${{ steps.apt_key.outputs.script }} + # TX_COVERAGE instruments every build configuration rather than only the one # whose name ends in _coverage. It has to be set for the build as well as the # test: the build is where -fprofile-arcs is decided, and the test run is @@ -142,7 +233,7 @@ jobs: run: ${{ inputs.build_script }} - name: Test - timeout-minutes: 60 + timeout-minutes: ${{ inputs.test_timeout_minutes }} env: TX_COVERAGE: ${{ inputs.skip_coverage && 'OFF' || 'ON' }} run: ${{ inputs.test_script }} @@ -276,9 +367,20 @@ jobs: # run_tests that is cancelled -- which is what a job that never gets a runner # ends up as -- leaves both false and lets the deploy proceed to download an # artifact the cancelled job never uploaded. + # + # The one exception is a caller passing skip_test, such as the deploy jobs + # of ThreadX and NetX Duo: run_tests is skipped by design, and the steps below + # download the coverage the caller's other jobs uploaded. The caller's needs + # gates those jobs. !cancelled() is required because an if without a status + # function implies success(), which is false when run_tests is skipped. deploy_code_coverage: runs-on: ubuntu-24.04 - if: ((github.event_name == 'push') || (github.event_name == 'workflow_dispatch')) && !inputs.skip_coverage && !inputs.skip_deploy && needs.run_tests.result == 'success' + if: >- + !cancelled() && + ((github.event_name == 'push') || (github.event_name == 'workflow_dispatch')) && + !inputs.skip_coverage && !inputs.skip_deploy && + (needs.run_tests.result == 'success' || + (inputs.skip_test && needs.run_tests.result == 'skipped')) needs: run_tests environment: name: github-pages |