Repository navigation
arm64: enable the existing managed sandbox - #1510
Gautam Ahuja (ahuja-gautam) wants to merge 4 commits into
Conversation
This pull request changes existing testsIt modifies or removes 60 existing test asset(s), listed below. Review each with care: a change to a test can hide a regression that the test caught. See doc/contribute.md#changes-to-existing-tests. Test code
End-to-end tests
Relocated unchanged: 0; added: 21. Compared the merge base |
There was a problem hiding this comment.
🟡 Changes recommended
ARM artifact validation checks default directory paths rather than the paths actually launched, allowing a replacement x86 kernel to pass admission.
2 open issues
What changed in this PR
Phase 2 builds on the ARM64 cold-boot foundation in #1509 to enable the existing managed sandbox with native local artifacts. It preserves legacy x86 artifact defaults and existing sandbox interfaces; ARM release downloads, snapshots, and final combined qualification remain outside this layer.
Changes:
- Select and validate native ARM artifacts in the sandbox CLI and runtime adapter.
- Add ARM Ubuntu EROFS production and Docker exports while preserving the canonical scratch format.
- Adapt existing fixtures and extend architecture, lifecycle, persistence, and isolation coverage.
| File | Description |
|---|---|
tests/runtime-adapter/tests/runtime.rs |
Tests ARM paths separately from legacy x86 defaults. |
tests/runtime-adapter/tests/layered_launch.rs |
Uses native artifact fixtures for launch refusals. |
tests/runtime-adapter/tests/guest_session.rs |
Gates x86-specific announcer helpers. |
tests/runtime-adapter/tests/common/mod.rs |
Selects native runtime test artifacts. |
tests/runtime-adapter/src/runtime.rs |
Adds architecture-qualified runtime construction. |
tests/runtime-adapter/src/bin/runtime-adapter-host.rs |
Selects the native runtime architecture. |
scripts/test_nvx_tools.py |
Tests scratch normalization, exports, and ARM diagnostics. |
scripts/test_microvm_tests.py |
Tests layered lifecycle inputs and architecture dispatch. |
scripts/test_guest_binaries.py |
Adds ARM producer and artifact-selection regressions. |
scripts/nvx.py |
Adds distro-layer architecture selection. |
scripts/nvx_tools/sandbox_ram_upper_tests.py |
Passes architecture into RAM-upper acceptance. |
scripts/nvx_tools/sandbox_proxy_tests.py |
Passes architecture into proxy acceptance. |
scripts/nvx_tools/sandbox_lifecycle_tests.py |
Selects matching artifacts and provenance. |
scripts/nvx_tools/microvm_tests.py |
Adds architecture selection and managed fixture controls. |
scripts/nvx_tools/managed_exec_tests.py |
Loads architecture-qualified workload artifacts. |
scripts/nvx_tools/common.py |
Supports architecture-qualified artifact paths. |
scripts/nvx_tools/build.py |
Produces ARM layers and normalizes scratch hash flags. |
scripts/nvx_tools/build_config.py |
Adds architecture-aware distro build configuration. |
nvx-rs/tests/managed_lifecycle.rs |
Extends topology, persistence, cancellation, and reconnect checks. |
nvx-rs/tests/cli/validation.rs |
Uses native sandbox validation fixtures. |
nvx-rs/tests/cli/openvmm.rs |
Updates native command and block-record expectations. |
nvx-rs/tests/cli/one_shot.rs |
Translates native one-shot fixture paths. |
nvx-rs/tests/cli/lifecycle.rs |
Reads native lifecycle records and logs. |
nvx-rs/tests/cli/harness.rs |
Adds ARM fixtures and sandbox path translation. |
nvx-rs/tests/cli/formats.rs |
Uses native paths in saved-format tests. |
nvx-rs/tests/cli/dry_run.rs |
Uses native sandbox dry-run fixtures. |
nvx-rs/tests/cli/download.rs |
Tests ARM download refusal and dual-architecture cleanup. |
nvx-rs/tests/cli.rs |
Updates native lifecycle paths and PE-refusal expectations. |
nvx-rs/src/host/cli/sandbox/mod.rs |
Selects and validates native sandbox artifacts. |
nvx-rs/src/host/artifacts/mod.rs |
Adds ARM probing and exposes directory-based construction. |
nvx-rs/src/host/artifacts/identity.rs |
Makes PE validation available without the CLI feature. |
nvx-rs/src/host/artifacts/dir.rs |
Enables ARM sandbox artifacts and distro paths. |
nvx-rs/src/host/artifacts/architecture/tests.rs |
Tests ARM managed-artifact admission. |
nvx-rs/public-api/windows.txt |
Records the exposed artifact constructor. |
nvx-rs/public-api/linux.txt |
Records the exposed artifact constructor. |
guest/common/nvx-sandbox-smoke |
Adds ARM diagnostic and raw-device checks. |
docker/Dockerfile |
Adds ARM distro-layer and scratch exports. |
doc/usage.md |
Documents architecture options and scratch normalization. |
doc/run.md |
Explains architecture-specific sandbox smoke checks. |
doc/build.md |
Documents ARM layer production and sandbox scope. |
Review effort: Balanced
cb017cf to
f865578
Compare
f865578 to
7f1d398
Compare
There was a problem hiding this comment.
🟡 Changes recommended
The ARM debug-kernel path bypasses required watchdog validation, allowing a production kernel to yield misleading debug-kernel evidence; config validation and rejection tests are needed.
2 open issues
1 resolved since last review
Review effort: Balanced
7f1d398 to
4bef5ff
Compare
There was a problem hiding this comment.
🔵 Needs a closer look
Native ARM managed-sandbox behavior and canonical ext4 byte reproducibility still require hardware and producer/backend qualification beyond the supplied slice checks, with combined-stack runtime validation deferred.
2 open issues
Review effort: Balanced
There was a problem hiding this comment.
🔵 Needs a closer look
Native ARM managed execution and ext4 reproducibility require producer/backend qualification beyond static review, with fresh Linux guest evidence absent and combined OpenVMM-pin validation deferred to Phase 3.
2 open issues
Review effort: Balanced
Select native local artifacts in the shared sandbox CLI and runtime adapter. Produce ARM Ubuntu EROFS layers through the maintained pipeline and preserve the canonical ISA-neutral scratch format. Adapt existing managed, policy and CLI fixtures without changing public sandbox options, saved formats or control protocols. Retain x86 behavior and explicit ARM snapshot and release-catalog exclusions. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Reuse the remembered-directory path check for ARM probe admission without requiring runtime identity metadata. Cover every public path override, missing inputs, and explicit/x86 compatibility. Pad the inert Windows batch fixture so short TEMP paths reach the intended PE magic refusal while retaining the no-execution check. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Reuse the shared optional counter-architecture API before starting a process and carry the selected ISA through debug guest-boot checks. Recognize the maintained ARM Image-debug product for legacy callers while retaining explicit ISA precedence and lower-layer restore detection. Recommend matching ARM distro and namespaced scratch producers without changing x86 guidance or qualified scenario scopes. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Require the architecture-selected debug kernel config to enable the existing watchdog and hung-task detectors before running scenarios. Keep x86 admission and ISA monitoring unchanged, and cover a valid ARM config plus refusal for each missing detector without running a VM. Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
4135868 to
aacccd9
Compare
There was a problem hiding this comment.
🔵 Needs a closer look
The host/guest sandbox and ext4 changes require maintainer review because exact-head native ARM VM/debug-kernel and combined VMM/backend qualification remain unclaimed and cannot be established by mocked tests.
No open issues 🎉
2 resolved since last review
Review effort: Balanced
doc/contribute.md asks every pull request to reference an issue on the first line of its description, but nothing checks it: pull requests such as #1675, whose only references sit in the template's HTML comment, and #1516, whose Refs #1509 names a pull request, reference no issue, and a reviewer has to notice. Add the Issue link workflow, which pull_request_target runs, from the default branch, when a pull request is opened, reopened, pushed to, edited, labelled, or unlabelled. Its script, check_issue_link.py, reads the pull request through gh api, in UTF-8 whatever the locale, and passes it when GitHub links it to an issue of the repository (closingIssuesReferences) or when its description holds close, fix, resolve (in any form), Part of, or Refs, with an optional colon, then #N, microsoft/nvx#N, or the issue's URL, where N is an issue. A pull request's number, a missing number, a bare #N, and another repository's issue do not count. The script reads the description as GitHub renders it, through the Markdown API in the repository's context, and only the text that GitHub shows counts: not an HTML comment, an attribute such as a link's title, or code, inline, fenced, indented, or in <pre>, nor a keyword that code or the end of a block, such as a list item or a table cell, separates from its number. GitHub's own renderer keeps the check in step with what the pull request's page shows, which a parser of CommonMark, GFM tables, and raw HTML written here kept missing. It looks up at most the first 10 numbers, so a description cannot make it call the API without bound. Pull requests by dependabot[bot] or github-actions[bot], or labelled no-issue, are exempt, and closed ones are left alone. An unlinked pull request gets the needs-issue label and one marker comment that says how to reference an issue, and the check fails; editing the description reruns it, which then removes the label and rewrites the comment. Only a comment of github-actions[bot] that starts with the marker is the script's. The script does every read before its first write, so an API error fails the check and changes nothing, and its summary names the verdict and the numbers, never the description. The job has no condition, since a skipped run would replace the previous verdict, runs one check a pull request at a time, keeps no checkout credentials, and gets only the pull request's number. Its token reads contents and writes pull requests; reading an issue of a public repository needs no issues permission, so the nightly report and issue labelling stay the only workflows that write issues. Evidence: dry runs against the live repository pass #1687, #1686, and #1682, which GitHub links to issues, and #1679, whose description references #1676, and flag #1675, #1516, #1510, and #1509, as reading their descriptions confirms. A 64 KiB description renders in about a second, and its text reads in at most 60 ms. test_issue_link.py covers the reading of GitHub's HTML, as gh api markdown renders it: keywords and targets, other repositories, code and <pre>, attributes, the ends of blocks and table cells, and inline elements; the verdicts and the writes through a fake gh api, the render request, gh's input and output in UTF-8, the bound on lookups, an API error at each read, a comment that quotes the marker or has no author, and the workflow's triggers, permissions, checkout, concurrency, and inputs. validate-nvx runs it on Linux and Windows. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
doc/contribute.md asks every pull request to reference an issue on the first line of its description, but nothing checks it: pull requests such as #1675, whose only references sit in the template's HTML comment, and #1516, whose Refs #1509 names a pull request, reference no issue, and a reviewer has to notice. Add the Issue link workflow, which pull_request_target runs, from the default branch, when a pull request is opened, reopened, pushed to, edited, labelled, or unlabelled. Its script, check_issue_link.py, reads the pull request through gh api, in UTF-8 whatever the locale, and passes it when GitHub links it to an issue of the repository (closingIssuesReferences) or when its description holds close, fix, resolve (in any form), Part of, or Refs, with an optional colon, then #N, microsoft/nvx#N, or the issue's URL, where N is an issue. A pull request's number, a missing number, a bare #N, and another repository's issue do not count. The script reads the description as GitHub renders it, through the Markdown API in the repository's context, and only the text that GitHub shows counts: not an HTML comment, an attribute such as a link's title, or code, inline, fenced, indented, or in <pre>, nor a keyword that code or the end of a block, such as a list item or a table cell, separates from its number. A link counts by where it goes, whatever its text: a link to an issue or a pull request as OWNER/REPO#N, and any other link as code does. GitHub's own renderer keeps the check in step with what the pull request's page shows, which a parser of CommonMark, GFM tables, and raw HTML written here kept missing. It looks up at most the first 10 numbers, so a description cannot make it call the API without bound. Pull requests by dependabot[bot] or github-actions[bot], or labelled no-issue, are exempt, and closed ones are left alone. An unlinked pull request gets the needs-issue label and one marker comment that says how to reference an issue, and the check fails; editing the description reruns it, which then removes the label and rewrites the comment. Only a comment of github-actions[bot] that starts with the marker is the script's. The script does every read before its first write, so an API error fails the check and changes nothing, and its summary names the verdict and the numbers, never the description. The job has no condition, since a skipped run would replace the previous verdict, runs one check a pull request at a time, keeps no checkout credentials, and gets only the pull request's number. Its token reads contents and writes pull requests; reading an issue of a public repository needs no issues permission, so the nightly report and issue labelling stay the only workflows that write issues. Evidence: dry runs against the live repository pass #1687, #1686, and #1682, which GitHub links to issues, and #1679, whose description references #1676, and flag #1675, #1516, #1510, and #1509, as reading their descriptions confirms. A 64 KiB description renders in about a second, and its text reads in at most 60 ms. test_issue_link.py covers the reading of GitHub's HTML, as gh api markdown renders it: keywords and targets, other repositories, code and <pre>, attributes, links, the ends of blocks and table cells, and inline elements; the verdicts and the writes through a fake gh api, the render request, gh's input and output in UTF-8, the bound on lookups, an API error at each read, a comment that quotes the marker or has no author, and the workflow's triggers, permissions, checkout, concurrency, and inputs. validate-nvx runs it on Linux and Windows. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
doc/contribute.md asks every pull request to reference an issue on the first line of its description, but nothing checks it: pull requests such as #1675, whose only references sit in the template's HTML comment, and #1516, whose Refs #1509 names a pull request, reference no issue, and a reviewer has to notice. Add the Issue link workflow, which pull_request_target runs, from the default branch, when a pull request is opened, reopened, pushed to, edited, labelled, or unlabelled. Its script, check_issue_link.py, reads the pull request through gh api, in UTF-8 whatever the locale, and passes it when GitHub links it to an issue of the repository (closingIssuesReferences) or when its description holds close, fix, resolve (in any form), Part of, or Refs, with an optional colon, then #N, microsoft/nvx#N, or the issue's URL, where N is an issue. A pull request's number, a missing number, a bare #N, and another repository's issue do not count. The script reads the description as GitHub renders it, through the Markdown API in the repository's context, and only the text that GitHub shows counts: not an HTML comment, an attribute such as a link's title, or code, inline, fenced, indented, or in <pre>, nor a keyword that code or the end of a block, such as a list item or a table cell, separates from its number. A link counts by where it goes, and of its text only a keyword that starts it: a link to an issue or a pull request, absolute or root-relative, with nothing after the number other than a slash, a query, or a fragment, as that keyword and OWNER/REPO#N, and any other link as code does. GitHub's own renderer keeps the check in step with what the pull request's page shows, which a parser of CommonMark, GFM tables, and raw HTML written here kept missing. It looks up at most the first 10 numbers, so a description cannot make it call the API without bound. Pull requests by dependabot[bot] or github-actions[bot], or labelled no-issue, are exempt, and closed ones are left alone. An unlinked pull request gets the needs-issue label and one marker comment that says how to reference an issue, and the check fails; editing the description reruns it, which then removes the label and rewrites the comment. Only a comment of github-actions[bot] that starts with the marker is the script's. The script does every read before its first write, so an API error fails the check and changes nothing, and its summary names the verdict and the numbers, never the description. The job has no condition, since a skipped run would replace the previous verdict, runs one check a pull request at a time, keeps no checkout credentials, and gets only the pull request's number. Its token reads contents and writes pull requests; reading an issue of a public repository needs no issues permission, so the nightly report and issue labelling stay the only workflows that write issues. Evidence: dry runs against the live repository pass #1687, #1686, and #1682, which GitHub links to issues, and #1679, whose description references #1676, and flag #1675, #1516, #1510, and #1509, as reading their descriptions confirms. A 64 KiB description renders in about a second, and its text reads in at most 60 ms. test_issue_link.py covers the reading of GitHub's HTML, as gh api markdown renders it: keywords and targets, other repositories, code and <pre>, attributes, links, the ends of blocks and table cells, and inline elements; the verdicts and the writes through a fake gh api, the render request, gh's input and output in UTF-8, the bound on lookups, an API error at each read, a comment that quotes the marker or has no author, and the workflow's triggers, permissions, checkout, concurrency, and inputs. validate-nvx runs it on Linux and Windows. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
doc/contribute.md asks every pull request to reference an issue on the first line of its description, but nothing checks it: pull requests such as #1675, whose only references sit in the template's HTML comment, and #1516, whose Refs #1509 names a pull request, reference no issue, and a reviewer has to notice. Add the Issue link workflow, which pull_request_target runs, from the default branch, when a pull request is opened, reopened, pushed to, edited, labelled, or unlabelled. Its script, check_issue_link.py, reads the pull request through gh api, in UTF-8 whatever the locale, and passes it when GitHub links it to an issue of the repository (closingIssuesReferences) or when its description holds close, fix, resolve (in any form), Part of, or Refs, with an optional colon, then #N, microsoft/nvx#N, or the issue's URL, where N is an issue. A pull request's number, a missing number, a bare #N, and another repository's issue do not count. The script reads the description as GitHub renders it, through the Markdown API in the repository's context, and only the text that GitHub shows counts: not an HTML comment, an attribute such as a link's title, or code, inline, fenced, indented, or in <pre>, nor a keyword that code or the end of a block, such as a list item or a table cell, separates from its number. A link counts by where it goes, and of its text only a keyword that starts it: a link to an issue or a pull request, absolute or root-relative, with nothing after the number other than a slash, a query, or a fragment, as that keyword and OWNER/REPO#N, which no text after the link extends, and any other link as code does. An issue's URL that GitHub shows as text follows the same rule, but for the punctuation that ends a sentence. GitHub's own renderer keeps the check in step with what the pull request's page shows, which a parser of CommonMark, GFM tables, and raw HTML written here kept missing. It looks up at most the first 10 numbers, so a description cannot make it call the API without bound. Pull requests by dependabot[bot] or github-actions[bot], or labelled no-issue, are exempt, and closed ones are left alone. An unlinked pull request gets the needs-issue label and one marker comment that says how to reference an issue, and the check fails; editing the description reruns it, which then removes the label and rewrites the comment. Only a comment of github-actions[bot] that starts with the marker is the script's. The script does every read before its first write, so an API error fails the check and changes nothing, and its summary names the verdict and the numbers, never the description. The job has no condition, since a skipped run would replace the previous verdict, runs one check a pull request at a time, keeps no checkout credentials, and gets only the pull request's number. Its token reads contents and writes pull requests; reading an issue of a public repository needs no issues permission, so the nightly report and issue labelling stay the only workflows that write issues. Evidence: dry runs against the live repository pass #1687, #1686, and #1682, which GitHub links to issues, and #1679, whose description references #1676, and flag #1675, #1516, #1510, and #1509, as reading their descriptions confirms. A 64 KiB description renders in about a second, and its text reads in at most 60 ms. test_issue_link.py covers the reading of GitHub's HTML, as gh api markdown renders it: keywords and targets, other repositories, code and <pre>, attributes, links, the ends of blocks and table cells, and inline elements; the verdicts and the writes through a fake gh api, the render request, gh's input and output in UTF-8, the bound on lookups, an API error at each read, a comment that quotes the marker or has no author, and the workflow's triggers, permissions, checkout, concurrency, and inputs. validate-nvx runs it on Linux and Windows. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
doc/contribute.md asks every pull request to reference an issue on the first line of its description, but nothing checks it: pull requests such as #1675, whose only references sit in the template's HTML comment, and #1516, whose Refs #1509 names a pull request, reference no issue, and a reviewer has to notice. Add the Issue link workflow, which pull_request_target runs, from the default branch, when a pull request is opened, reopened, pushed to, edited, labelled, or unlabelled. Its script, check_issue_link.py, reads the pull request through gh api, in UTF-8 whatever the locale, and passes it when GitHub links it to an issue of the repository (closingIssuesReferences) or when its description holds close, fix, resolve (in any form), Part of, or Refs, with an optional colon, then #N, microsoft/nvx#N, or the issue's URL, where N is an issue. A pull request's number, a missing number, a bare #N, and another repository's issue do not count. The script reads the description as GitHub renders it, through the Markdown API in the repository's context, and only the text that GitHub shows counts: not an HTML comment, an attribute such as a link's title, or code, inline, fenced, indented, or in <pre>, nor a keyword that code or the end of a block, such as a list item or a table cell, separates from its number. A link counts by where it goes, and of its text only a keyword that starts it: a link to an issue or a pull request, absolute or root-relative, with nothing after the number other than a slash, a query, or a fragment, as that keyword and OWNER/REPO#N, which no text after the link extends, and any other link as code does. An issue's URL that GitHub shows as text follows the same rule, but for the punctuation that ends a sentence. GitHub's own renderer keeps the check in step with what the pull request's page shows, which a parser of CommonMark, GFM tables, and raw HTML written here kept missing. It looks up at most the first 10 numbers, so a description cannot make it call the API without bound. Pull requests by dependabot[bot] or github-actions[bot], or labelled no-issue, are exempt, and closed ones are left alone. An unlinked pull request gets the needs-issue label and one marker comment that says how to reference an issue, and the check fails; editing the description reruns it, which then removes the label and rewrites the comment. Only a comment of github-actions[bot] that starts with the marker is the script's. The script does every read before its first write, so an API error fails the check and changes nothing, and its summary names the verdict and the numbers, never the description. The job has no condition, since a skipped run would replace the previous verdict, runs one check a pull request at a time, keeps no checkout credentials, and gets only the pull request's number. Its token reads contents and writes pull requests; reading an issue of a public repository needs no issues permission, so the nightly report and issue labelling stay the only workflows that write issues. Evidence: dry runs against the live repository pass #1687, #1686, and #1682, which GitHub links to issues, and #1679, whose description references #1676, and flag #1675, #1516, #1510, and #1509, as reading their descriptions confirms. A 64 KiB description renders in about a second, and its text reads in at most 60 ms. test_issue_link.py covers the reading of GitHub's HTML, as gh api markdown renders it: keywords and targets, other repositories, code and <pre>, attributes, links, the ends of blocks and table cells, and inline elements; the verdicts and the writes through a fake gh api, the render request, gh's input and output in UTF-8, the bound on lookups, an API error at each read, a comment that quotes the marker or has no author, and the workflow's triggers, permissions, checkout, concurrency, and inputs. validate-nvx runs it on Linux and Windows. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>


Refs #1509
What and why
Current captured-snapshot status: actual Draft, base
3103f24c292f82907c77dda5cd9d8c105b608971, headaacccd99e132233f4ae6e440b30aeadf941031f4, tree14d9c35296ddc44615a009ec4816a5850027e182. This layer is four single-parentcommits on the captured NVX
devbaselinec9817ae73263a0c32498dd43ebeb089164af8cbe(the merged v6.1/1567 baseline),not latest dev. Later observed
1ad73759...and its09cd771c...pinare not included or validated here.
The lower inherited OpenVMM pin is
73cf403574a8907ee94dab0ab9d648152b832b05, not historical5d656d85...orPhase 3's feature pin
b09c5884.... All four Phase 2 patches replayedidentically (
range-diff =and matching stable patch IDs), preservingpublic-path admission, short-TEMP refusal, ARM debug ISA/config admission,
recovery hints, lower restore readers, and upstream fixture fixes.
The exact layer audit remains 60 existing-test changes, 21 additions.
The upper PR's published
24bdced0...selectsb09c5884...; the whole stackhas 16 single-parent commits above captured
c981and no new merge commits.Its integrated runtime evidence is applicable by preserved production-input
identity, not a claim of reexecution at each final head. Overall
qualification is not green: the Windows privilege and host clock/service
failures and the proof limits below remain.
Historical initial-publication record: the frozen identities and
publication-only statements below describe the original snapshot, not the
current head or new qualification.
Historical captured-C5/413586/5d status (superseded): actual Draft, base
f925fcd32110a280ae89b515af4da51521c3b02d, head41358684863a511d63f747c50f6368b4b6e214c0(four single-parent commits).The baseline is captured
devc5fdc8b850c63dcb0dddb1a1a6ede5a42a20d692,not latest dev; later
24fa7233...and its PIT pin were not taken in.This lower layer inherits OpenVMM
5d656d854a8daca0f23833666b0b71229fe2ab73, not Phase 3's separate6036d7ee...feature pin.
The current review fixes carry selected ISA through ARM
Image-debugtimechecks, validate its matching config with the existing watchdog guard before
scenarios start, and recommend the ARM distro producer plus namespaced
scratch output. X86 behavior and lower restore-option detection are preserved.
The exact layer audit is 60 existing-test changes, 21 additions; the
latest two-file config fix updates the valid ARM fixture and adds positive
and each-missing-detector refusal coverage.
Phase 2 of the three-phase ARM64 stack depends on
Phase 1 and enables the existing
managed sandbox with native local artifacts. This is publication of the frozen commit
cb017cf8aee13e5e7276236d1b0ab328e1b52e7f, not a rewrite or a rebase ontocurrent
dev. The review delta is only791efdd345d0f888f3db8e63d9398ce60e382f92..cb017cf8aee13e5e7276236d1b0ab328e1b52e7f(40 files), not the combined Phase 1 and Phase 2 delta.
adapter; validate the ARM kernel Image, matching initramfs architecture
manifest, and Windows OpenVMM PE identity before launch. Preserve the legacy
artifact API's x86 default.
architecture-qualified paths, package inputs, and manifests. Export them
through the ARM Docker artifact target.
directory-hash selection with the same e2fsprogs package and recompute its
checksums, leaving already-canonical x86 bytes and the pinned hash unchanged.
fixtures for native artifacts. Public sandbox options, saved formats,
control protocols, and the one-VP CLI lifecycle contract do not change.
ARM snapshot modes and release-catalog/download qualification remain excluded.
The sole parent is
791efdd345d0f888f3db8e63d9398ce60e382f92; the exact treeis
61959019583b0f95d30f30188273fe342a57b832. The historical OpenVMM gitlinkremains
61cdd63978ea58ce045d159208ed4ca22f30cbb8, unchanged in this layer.The eventual Phase 3 PR owns combined ARM pin/CI integration, including the
advance to
38016f6637f0c61fb55c9ee31033ceec428e346c; this lower historicalslice is not being presented as independently qualified with that newer VMM.
Validation
Current bounded evidence
This refreshed Phase 2 source has the same four patches as the preserved
layer. The local replay stage ran only the affected Python and static checks
below; no new Phase 2 Rust artifact tests or native VM tests were run.
Earlier 25-test Rust/product results remain historical receipts, not new
results for
aacccd99. This description update ran no tests or builds.python scripts\nvx.py verify73cf4035...submodule.python -m unittest scripts.test_time_abi.MonitorTests scripts.test_microvm_tests.GuestConsoleEchoTests scripts.test_microvm_tests.MicrovmTests.test_arm_debug_kernel_requires_each_detector_before_running_scenarios scripts.test_microvm_tests.MicrovmTests.test_arm_debug_boot_passes_selected_architecture_to_the_shared_monitor scripts.test_microvm_tests.MicrovmTests.test_debug_kernel_runs_the_same_host_restore_scenarios_on_vmlinux_debug scripts.test_microvm_tests.MicrovmTests.test_arm_missing_sandbox_inputs_recommend_matching_producer_outputs scripts.test_microvm_tests.MicrovmTests.test_invalid_script_counter_architecture_fails_before_process_start scripts.test_microvm_tests.MicrovmTests.test_public_sandbox_scenarios_need_their_artifacts_before_any_runs scripts.test_guest_binaries.GuestBinaryArchitectureTests.test_distro_cli_selects_architecture_qualified_work_and_output scripts.test_guest_binaries.GuestBinaryArchitectureTests.test_arm_distro_producer_uses_matching_root_packages_and_manifest scripts.test_guest_binaries.GuestBinaryArchitectureTests.test_arm_public_scenarios_require_namespaced_matching_layer_inputs scripts.test_guest_binaries.GuestBinaryArchitectureTests.test_arm_lifecycle_identity_fixture_uses_matching_provenance scripts.test_guest_binaries.GuestBinaryArchitectureTests.test_arm_microvm_runner_selects_image_and_requires_explicit_scenarios -vformat --checkandcheckon the 15 affected Python files in the retained local handoff--pythonplatform Linuxand--pythonplatform Windows, with explicit--pythonpathto the existing maintained interpreterpython scripts\check_test_layout.pyandpython scripts\check_markdown_links.pypython scripts\check_test_changes.py --base 3103f24c292f82907c77dda5cd9d8c105b608971 --head aacccd99e132233f4ae6e440b30aeadf941031f4The local source and command receipts are retained in
review-fixes\baseline-refresh-20261113-c981, includingfinal-local-source-proof.jsonandfinal-lightweight-handoff.json.Integrated Phase 3 execution: applicability and remaining failures
The independently supplied sealed handoff
C:\Users\gautamahuja\aci\phase3-refresh-20261113\evidence\final-phase3-refresh-qualification-handoff.jsonwas read and its exact SHA-256 verified:
4877476782c70b72978e7060b41f249429341c49246e32e1e5cb05d2c4d9f83a.These are the validator owner's results, not additional tests run by this
Phase 2 description update.
Actual runtime execution used NVX
2000aa078713467add77565acdae688a6d6ecfcfwith OpenVMMb09c58843ae310dfff9ae68e9e822f9c5846b914, normal native ARM64openvmm 0.5.0, executable SHA-256f171f776a76c7aeef5faae7221e7a508d4d3de48f7e62f227bc7fbc372eb7653.Its measured build profile was opt-level 0, not inferred from a directory name.
records and 128 repeat commands/135 records; 115 adapter cases across
nine targets; 833 real guest tests with the exact 12 ignored names retained.
The 46 adapter snapshot exclusions are not passes.
passed, covering the selected 1/2/4/8-VP SMP/sample/warp scope. Scratch and
RAM each completed ten Start/Stop cycles under one provision. These remain
the exact performed cells, not an unrestricted platform matrix.
24bdced073a24418c1ebbf1d0775bb5ba42266bais one test-only appendto
2000, changing two files/three methods. All 1032 production-input rowsare identical, SHA-256
73ae1805271658bcb25d876f5bd113afee42a70a05830713901ae56abdd0da28.Runtime applicability uses that entire input identity and retained bridge
receipt SHA-256
7787029762c5560e2420c41a8986bc9c2839d27078cd5ab484c395d58cc4c3b0;it is not new-head runtime reexecution.
24bdfull Python run counted 2255 tests: zero assertion-failedmethods; three Windows 1314 setup-error records in two baseline Ubuntu
methods; 107 skip records (104 whole-method skips and three subcase skips);
2147 complete-success methods and two partially skipped methods. The
corrected six targeted methods and current static/language checks passed.
The original
2000assertion failures and corrected accounting receiptsare retained; skipped subcases are not covered.
Overall qualification remains NOT GREEN. The Windows adapter host
library had 57 passes and one baseline failure (1314 mapped as Uncategorized
instead of PermissionDenied); the two Ubuntu setup methods above remain
blocked by Windows symlink privilege. Full doctor exited 1: H4 failed with
UTC free-running and H7 failed with W32Time off. H1/H2/H3/H5/H6 passed on the
full 13-sample/10-second schedule; H5 measured 26 ns offset/156 ns uncertainty
over 66 pairs, and H6's actual eight-VP/128-MiB idle 0.1/1/5/1 scope measured
52 ns offset/391 ns uncertainty with no backward time. No host, service,
security, or threshold changes waived these failures.
Normal-product ARM capture refusal was demonstrated with the cold-only
error. Normal-product restore refusal was not proved (the retained
attempt failed before manifest admission), and deterministic bounded
VP0/AP-latch overlap was not proved. Snapshots remain unsupported.
Native ARM NVX CI, trusted upstream OpenVMM CI, physical x86,
Linux/KVM/MSHV, NetworkService, and synchronized-UTC qualification are not
claimed. All PRs remain actual Draft; no trial trust bypass, Ready, CI
dispatch, or merge was enabled.
Historical 413586 bounded evidence (superseded)
The latest
41358684follow-up changed Python only. Its retained red/greenreceipts show the old ARM branch did not read the config or refuse either
missing detector; the fixed branch refuses both before scenarios run.
No new test campaign was run for this metadata update.
python -m unittest scripts.test_microvm_tests.MicrovmTests.test_arm_debug_kernel_requires_each_detector_before_running_scenarios scripts.test_microvm_tests.MicrovmTests.test_arm_debug_boot_passes_selected_architecture_to_the_shared_monitor scripts.test_microvm_tests.MicrovmTests.test_debug_kernel_runs_the_same_host_restore_scenarios_on_vmlinux_debug scripts.test_microvm_tests.MicrovmTests.test_arm_missing_sandbox_inputs_recommend_matching_producer_outputs scripts.test_microvm_tests.MicrovmTests.test_invalid_script_counter_architecture_fails_before_process_start scripts.test_time_abi.MonitorTests -vpython scripts\nvx.py verifyandpython scripts\check_test_layout.pypython -m ruff format --check scripts\nvx_tools\microvm_tests.py scripts\test_microvm_tests.pyandpython -m ruff check scripts\nvx_tools\microvm_tests.py scripts\test_microvm_tests.py, using the existing maintained Python environmentpython -m pyright --pythonplatform Linux scripts\nvx_tools\microvm_tests.py scripts\test_microvm_tests.pyand the same selection with--pythonplatform Windows, in that environmentpython scripts\check_test_changes.py --base f925fcd32110a280ae89b515af4da51521c3b02d --head HEADat413586844bef5ffa...: artifact unit selector (25 tests), batch-refusal selector under short and normal TEMP (1 each), Clippy, rustdoc, and rustfmt41358684.No native ARM debug-kernel qualification, full matrix, or passing CI is
claimed. Earlier Win1314/emulation limitations remain; the separate new-host
UTC H4 failure remains unqualified, and normal OpenVMM trust gates are not
waived. All three PRs remain actual Draft; no CI/Ready/merge was enabled.
Historical publication evidence
This session performs publication only. No source edits, builds, test
execution, VMs, Docker runs, package installation, host preparation, CI
dispatch, or merge actions were performed.
git show -s --format='%H%n%P%n%T%n%s%n%b' cb017cf8aee13e5e7276236d1b0ab328e1b52e7fgit ls-tree cb017cf8aee13e5e7276236d1b0ab328e1b52e7f -- openvmm61cdd63978ea58ce045d159208ed4ca22f30cbb8.git merge --ff-only --no-edit cb017cf8aee13e5e7276236d1b0ab328e1b52e7fgit status --porcelain=v1gh api repos/microsoft/nvx/git/ref/heads/gautamahuja-microsoft-nvx-phase-2-pr --jq .object.shacb017cf8aee13e5e7276236d1b0ab328e1b52e7f.python scripts\check_test_changes.py --base 791efdd345d0f888f3db8e63d9398ce60e382f92 --head cb017cf8aee13e5e7276236d1b0ab328e1b52e7forigin/dev.Historical qualification supplied by the coordinator includes representative
native run116: 818 passed, 0 failed, 12 ignored. It was not rerun here,
and its exact invocation was not independently reconstructed in this
publication session. This is not a blanket all-platform-green claim.
The previously recorded Windows Win1314 and actual-x86-under-emulation
limitations remain unresolved. Fresh ARM producers and final combined
qualification belong to the separate Phase 3 validation owner.
The documented quality, Rust/runtime-adapter, and CLI gates in
.github/actions/check-quality/action.yml,.github/actions/check-nvx-rs/action.yml, and.github/actions/validate-nvx/action.ymlwere inspected, not rerun.python scripts\nvx.py verify,python scripts\local_gates.py --base origin/dev --run,python scripts\check_test_layout.py, Rust formatting/lint/test/API checks,guest production, and backend runtime suites are not run by this
publication session. The requested CI exercise is on the Ready Phase 3 PR;
this Phase 2 PR stays an actual draft.
Existing-test disclosure (
tests-modified):nvx-rs/src/host/artifacts/architecture/tests.rsnvx-rs/tests/cli/harness.rsnvx-rs/tests/cli.rsnvx-rs/tests/cli/download.rsnvx-rs/tests/cli/dry_run.rsnvx-rs/tests/cli/formats.rsnvx-rs/tests/cli/lifecycle.rsnvx-rs/tests/cli/one_shot.rsnvx-rs/tests/cli/openvmm.rsnvx-rs/tests/cli/validation.rsnvx-rs/tests/managed_lifecycle.rsscripts/nvx_tools/managed_exec_tests.pyscripts/nvx_tools/microvm_tests.pyscripts/nvx_tools/sandbox_lifecycle_tests.pyscripts/nvx_tools/sandbox_proxy_tests.pyscripts/nvx_tools/sandbox_ram_upper_tests.pyscripts/test_microvm_tests.pyscripts/test_nvx_tools.pyguest/common/nvx-sandbox-smoketests/runtime-adapter/tests/common/mod.rstests/runtime-adapter/tests/guest_session.rstests/runtime-adapter/tests/layered_launch.rsscripts/test_guest_binaries.pyandtests/runtime-adapter/tests/runtime.rsThe test-change audit found no CI test-wiring changes in this exact layer.
Historical review fixes and guarded stack sync (2026-11-12)
This section records the earlier
f865578bpublication and then-Ready state.The current Draft status, captured baseline, head, and bounded evidence are
stated above; the original receipts below are retained without requalification.
The earlier frozen-slice and publication-only statements above record the
initial publication. They are retained as historical disclosures, not as
claims about this update's head, validation, or draft state. This PR is now
Ready (
isDraft=false); this update preserves that existing state and thetests-modifiedlabel.Following the authorized lower-layer fixes, the current Phase 1 base is
d06bf47f89eddf2fa68edeb99caf3fd697d35f45. Only the original Phase 2 commit wasreplayed onto it, as
c517c2382fa80a748eb793e88ee0ef08cd73ae94, followed bythe focused review fix.
The current head is
f865578bdd7a4767e4b208f67d99792cb9215ac2, tree4015eb54220da30fee92225e3a109fc89f3c6aa0. The current review range isd06bf47f89eddf2fa68edeb99caf3fd697d35f45..f865578bdd7a4767e4b208f67d99792cb9215ac2(two linear commits, 41 files), not the combined stack or current
dev.The historical OpenVMM gitlink remains
61cdd63978ea58ce045d159208ed4ca22f30cbb8.The original and replayed Phase 2 diffs have the same zero-context stable
patch ID,
e0ef031933b888a440744e86a47cb1b9c01496a8. The range-diff shows onlyinherited import context from the Phase 1 source-export fix, plus the new
focused commit. Named local backup refs preserve the original tip and the
exact remote lease. The push used
--force-with-lease=refs/heads/gautamahuja-microsoft-nvx-phase-2-pr:cb017cf8aee13e5e7276236d1b0ab328e1b52e7fafter rechecking the existing stack order, Ready states, refs, and absence of
auto-merge/queue entries. No
devintake, source merge, pin promotion, stackrecreation, or mutation of another owning branch was performed.
Focused changes and test disclosure
Artifacts::probe()now rejects divergence between public paths and aremembered ARM artifact directory before applying that directory's native
checks. It reuses
runtime_identity()'s private path comparison withoutadding provenance reads. The probe rustdoc states the invariant.
path overrides while original ARM artifacts remain valid, verifies
missing-file errors, and verifies that probing needs no runtime-identity
metadata. A new test in
nvx-rs/src/host/artifacts/tests.rsretainsexplicit-path and x86 override compatibility.
batch_openvmminnvx-rs/tests/cli.rspads its script with an inert REMline containing 64 fixed bytes. Short Windows TEMP paths therefore reach
the expected non-MZ refusal. The exact ARM error and no-execution marker
assertions remain; the production PE parser is unchanged.
Fresh targeted validation on the rebased slice
These are native ARM64 Windows checks, not blanket all-platform qualification.
The maintained resource guard admitted the serialized Rust stages with
CARGO_BUILD_JOBS=1;NVX_BUNDLE=skipavoided bundle downloads and is notevidence that artifact bundling passed. No OpenVMM build, VM, Docker workload,
package installation, manual CI dispatch, or merge operation was run here.
python scripts\nvx.py verify61cdd639...submodule, without changing its gitlink or upstream URL.cargo test --manifest-path nvx-rs\Cargo.toml --locked --lib --features cli,testing host::artifacts::cargo test --manifest-path nvx-rs\Cargo.toml --locked --features cli,testing --test cli a_run_refuses_a_batch_file_as_openvmm -- --exactwithTEMP=TMP=C:\tTEMP/TMPcargo clippy --manifest-path nvx-rs\Cargo.toml --locked --all-targets --all-features -- -D warningscargo doc --manifest-path nvx-rs\Cargo.toml --locked --no-deps --all-featureswithRUSTDOCFLAGS=-D warningscargo fmt --manifest-path nvx-rs\Cargo.toml --all --checkpython scripts\check_test_layout.pypython scripts\check_test_changes.py --base d06bf47f89eddf2fa68edeb99caf3fd697d35f45 --head f865578bdd7a4767e4b208f67d99792cb9215ac2python scripts\check_test_changes.py --base c517c2382fa80a748eb793e88ee0ef08cd73ae94 --head f865578bdd7a4767e4b208f67d99792cb9215ac2null, not zero running containers. Heavy/native lane explicitly released afterward.The historical run116, Win1314 and actual-x86-under-emulation limitations
above remain disclosed. This update did not rerun Linux, x86-under-emulation,
full Rust/Python suites, artifact bundling, or backend VM qualification; those
are not claimed green by these focused checks. Phase 3 still owns the
combined newer OpenVMM pin and CI integration.