In-tree CI builds report compatibility errors after whitespace cleanup:
Linux 7.3:
conflicting types for 'wr'; have 'struct ib_rdma_wr'
'sgv_shrinker' is a pointer; did you mean to use '->'?
Linux 6.18.57, dev_handlers/scst_vdisk_main.c:330:
too many arguments to function 'blkdev_issue_flush'
Two assumptions in specialize-patch change branch selection when only
whitespace changes in otherwise equivalent preprocessor expressions.
The dispatch expression accepts only spaces before '#'. A tab-indented
opening directive is copied as ordinary text, while its closing directive
still changes the conditional nesting state. Accept all horizontal
whitespace as the expression evaluator already does, including when an
indented #elif becomes #if after discarding its original opening branch.
The evaluator also substitutes an empty string for absent RHEL_MAJOR,
RHEL_RELEASE_CODE and MOFED version macros. Thus 'RHEL_MAJOR - 0' becomes
'- 0', which its signed-number patterns cannot reduce correctly. This
leaves the two-argument blkdev_issue_flush() call in a modern upstream
patch. Substitute zero for these absent numeric macros, matching C
preprocessor semantics and the existing RHEL_MINOR handling. Keep their
defined() results false and preserve supplied vendor version values.
Add isolated regressions for spaces and tabs around '#', nested unknown
guards, promoted #elif branches and absent vendor versions. Cover all
three flush signatures at the upstream 5.8/5.12 and RHEL 8.4/8.6 boundaries,
plus EL7, RHEL 9, upstream 6.18.57 and 7.3, with both '-0' and '- 0'. Apply
each resulting patch in a temporary directory in removal and blank-
preserving modes, checking selected branches, surviving guards and hunk
lengths without compiling or loading SCST.
Install patch in the CI build stage used by these generator tests. Without
this dependency, all 98 specialization subcases abort before applying the
patches, with the following error in the minimal Fedora build image:
FileNotFoundError: [Errno 2] No such file or directory: 'patch'
The basic-or-smoke PR filter selects 67 specs. Recent successful PR
runs spent more than 41 minutes executing their VM tests, in addition
to image preparation. Broad functional labels keep growing this set
whenever a new basic or smoke regression is added.
Select explicit pr-labelled specs instead: 11 local and 8 iSCSI cases.
Keep representative device, target, LUN and group management, repeated
configuration application, failed fileio creation, discovery and login,
MD5/SHA256 CHAP, portal validation, READ CAPACITY, read-only rejection,
and verified fileio/blockio I/O. Preserve each test's workload, assertions
and cleanup. Exclude manual, race and stress scenarios from the PR set.
The master and manual selections remain 103 and 106 specs respectively.
Limit the complete PR job to 30 minutes, its image build to 10 minutes
and VM execution to 18 minutes. Preserve the longer push/manual limits.
The intended end-to-end PR duration is 15-20 minutes; confirm it on CI
rather than treating dry-run counts as runtime evidence.
Validate selection in the Go runner and Python VM supervisor tests.
Framework checks, all 59 CI tests, actionlint and selection dry runs pass.
Add a parallel make checkpatch job for pushes to master and manual
runs. Fail analyzer jobs on remaining source findings with status 1;
keep tool, build and coverage failures separate with status 2.
Validate the current report, Make status and native coverage before
classifying results. Checkpatch needs its script and data, not a
prepared KUnit kernel. Update the isolated CI report fixtures.
Run SCST style fixtures and make scs_lint in a separate job without a
kernel build, alongside the native analyzers. Pin uv and its setup
action; the linter supplies its pinned parser dependencies.
Retain the pull-request Checkpatch workflow for patch and commit
checks. Remove the standalone push Checkpatch workflow, whose role
is now covered by the common static-analyzer workflow.
Rename Tooling to Analyzer script tests and clarify that analyzer
jobs include KUnit sources without executing KUnit. Update Checkout
and Upload Artifact to v7 to replace their deprecated Node.js 20
runtimes, and align the CI documentation with the actual jobs.
Reserve the legacy-kernel fork checks for manual runs instead of adding
two VM jobs and extra kernel builds to the regular CI selection. Remove
the Linux 5.8/5.9 Ginkgo jobs and regression matrix entries, retaining the
kernel pins and local-only VM runner for reproduction.
Label the fork spec manual and exclude that label from both ci-pr and
ci-push. Move smoke to the READ CAPACITY entries so their existing CI
coverage remains intact. Keep full manual runs inclusive, update the
selection checks, and document focused manual execution.
scst_user pins its data buffers from Linux 5.9, the first release in
which fork() gives the child a copy of a pinned page (commit
70e806e4e645, "mm: Do early cow for pinned pages during fork() for
ptes"). The Ginkgo "handler forks" spec checks this, but CI boots only
the current Fedora kernel, so the boundary itself is never tested.
Add a Ginkgo job matrix that runs the local "user" specs on Fedora 33's
Linux 5.8.18 and 5.9.16, on both sides of the boundary. On 5.9.16 the
fork spec relies on the pin alone; on 5.8.18 the handler protects its
buffers with MADV_DONTFORK, as required there. Both must keep correct
geometry and data. The jobs are named after their kernels, like the
regression jobs.
- kernel-5.8.env, kernel-5.9.env and their SHA256SUMS pin those kernels'
rpms from Fedora's koji, like kernel.env and kernel.SHA256SUMS.
Fedora's only 5.9.0 build is a rawhide kernel that defines RHEL_MAJOR 8
and RHEL_MINOR 99, so SCST's backports treat it as RHEL and the build
fails; 5.9.16 has the same fork() behavior.
- build-image takes an optional pin name, such as "kernel-5.9", for
Ginkgo images. The image is otherwise built the same way.
- The Dockerfile downloads exactly the rpms listed in the selected
checksum file, since 5.9 has no kernel-modules-core package.
- vm.py --local-only boots only the target guest, without the iSCSI
initiator or the private network, and expects only the local report.
The guest then writes a config without an iscsi section. The label
filter is restricted to local specs: "local && (<filter>)", or "local"
without one.
Run isolated tooling checks and separate Sparse and Smatch Make targets
on pushes to master and manual dispatch. Use the pinned kernel with
KUnit enabled and preserve full logs, policy and tool provenance.
During the source-fix rollout, completed source findings are advisory.
Tool, build, parser and coverage failures remain blocking, including an
empty scan or missing KUnit inputs. Local Make targets stay strict.
Validate the current run's report together with Make's exit status,
since Make maps both findings and incomplete analysis to status 2.
Prepare the runner and container for Coccinelle and Cppcheck as well,
but leave their matrix entries commented out. Cover all four report
formats and GNU Make exit mapping with isolated tooling fixtures.
Build standalone Sparse and Smatch from recorded revisions, applying the
reviewed SCST Smatch frontend correction. Install matching native Smatch
data without replacing standalone Sparse or adding project models.
Reuse the existing checksum-pinned kernel headers. Record source
commits, patch and binary hashes, and installed model contents for CI
evidence. Fail preparation instead of silently falling back to different
tools; no SCST modules are built or loaded while preparing this image.
Pull requests that only change development guides or independent tooling
still compile kernels and run the VM suites. Select Ginkgo, KUnit, Clang
and kernel regression through one shared path policy before expensive
work. Compare the whole PR against its merge base, retaining deleted
paths and both sides of renames. Run checks for unclassified inputs.
Keep path rules in .github/ci-checks.toml and validate their structure,
check names, literal paths and conflicting entries before selection.
Provide a validation-only command. Changes to the config or selector
exercise every check.
Select kernel regression once in the preparation job and gate the build
job before expanding its matrix. Unrelated PRs do not allocate a runner
for every kernel. Selection failures fail the workflow and successful
skips include a reason in the job summary.
Run the Packages workflow only after pushes to master, preserving its
DEB/RPM builds and publication. Keep the separate Clang DEB build on PRs.
Retain full master and manual checks and the existing Checkpatch runs.
Cover path ownership, config validation and Git histories with tests,
and shorten the CI guide around all workflows, selection and results.
Quote the CC environment key so Checkpatch does not interpret unchanged
YAML diff context as a Cc trailer.
Run usr/fileio/fileio_tgt over scst_local with an LD_PRELOAD hook that
replaces the block length of one READ CAPACITY(10) or (16) reply with
809052976, as a user space backend did in production. The invalid reply
must be rejected with INTERNAL TARGET FAILURE sense, and afterwards READ
CAPACITY(10) and (16), reads and writes must return the right geometry
and data.
The local suite now builds usr and loads scst_user, and the CI image
builds usr for the Ginkgo guest.
Keep CI selection policy in shared ci-pr and ci-push label sets. Resolve
them to ordinary Ginkgo expressions for local discovery and execution,
and before booting CI guests so report validation still checks the exact
filter. Preserve ordinary expressions and the full manual selection.
Mark both ACG teardown scenarios as race and exclude that category from
master push runs. Keep their strict overlap, recovery and cleanup checks
unchanged. PR runs retain basic/smoke coverage through ci-pr.
Document the named selections and check their filtering behavior and
propagation through the VM supervisor.
Build kernel-matched images on GitHub-hosted runners and execute tests
under an external QEMU supervisor. Run KUnit and Ginkgo in independent
jobs that each build and test SCST. Keep guest images local to each job
and upload build logs and test reports.
Capture serial logs outside the guests, require complete matching
reports, and bound guest execution independently of kernel health.
Exercise guest panic and timeout handling before KUnit. Give Ginkgo an
isolated iSCSI peer and disposable block storage.
Print native test output and summaries in the job log. On failure, include
supervisor errors, kernel diagnostics and bounded serial tails while
retaining complete artifacts and the original test command exit status.
Select basic management and smoke SCSI/I/O tests on both transports for
pull requests. Run every Ginkgo spec after pushes to master and on manual
runs. Pass the label filter into the guest and require both transport
reports to confirm the requested selection.
Gleb Chesnokov
wenlxieandGleb Chesnokov