[FEATURE] Forward permission evidence through original exact metadata endpoints #5

Closed
opened 2026-09-29 12:34:58 +02:00 by erikinkinen · 0 comments
Owner

Problem and motivation

Native filesystem permission mediation needs explicit ownership, mode and ACL evidence from the original retained filesystem object. The private providers expose that evidence, but Vfsd exact metadata endpoints currently offer no route to it.

Proposed behavior and scope

Only original job-lifetime READ_METADATA directory endpoints serve selector-free snapshots and bounded ACL pages. Vfsd uses the retained mount/provider handle and fresh private correlation, checks framing, handle, kind, revision and range, then replaces private identifiers with caller request correlation and the live VFS record ID. It does not reopen paths, synthesize ACL absence or broaden another endpoint's operation ceiling. Unknown evidence stays unknown, refused staging is cleared, and original owned completion custody prevents refetch/replay after refusal. Actual buffer geometry bounds pages without a new complete-ACL limit; entries remain in existing staging while the fixed prefix is rewritten.

Eighteen controls are authored and unrun: five wire, five provider transport, seven real-dispatch and one owned-completion. The original transport interface is shared and native path codecs/resolver fixtures are split; changed code remains below 1,000 lines (maximum 898), with inline rustdoc and missing_docs retained. Component contracts and the technical manual are updated. Tests, builds, lint checks, rustdoc, VM/manual/profiler runs and CI observation remain deferred by user instruction. No pass, warning-free build or measured speedup is claimed.

Native permission enforcement remains open: trusted credentials, full ACL acquisition and actual operation-specific effects still need to be connected. Value-inode ACL payloads and ACL-aware persistent mutation also remain unfinished. Both complete EriX guest-build generations and the Rust/LLVM guest rebuild remain required and unproved. Phase completion is 16.22% weighted, 4.81% whole-item, with 20/460 parents and 435/895 sub-items complete.

Add selector-free exact VFS permission snapshots and revision-bound ACL pages. The exact receiver chooses the original retained directory handle. Only job-lifetime READ_METADATA directory endpoints accept these operations; read, write, enumerate and entry-mutation endpoints retain their existing ceilings. The snapshot describes this directory, never a caller-selected descendant.

Forward requests through the original mounted provider route with fresh internal request correlation. Validate snapshot handle and node kind and every ACL page's request, handle, revision, ordinal and bound before returning evidence. Publish Vfsd's monotonic exact-object ID in place of the private provider handle. This number remains descriptive and cannot select a different record. ACL page requests must match the endpoint's current object ID; each page is tied to its original provider revision. Consumers still validate complete count, ordering and mode correspondence across all pages.

Use shared checked codecs with disjoint VFS operation IDs and fixed framing. Bound page transfers by the existing response storage, not a new total-ACL limit. Clear refused provider staging. Unsupported providers explicitly refuse; never synthesize ownership, ACL absence or a policy verdict from legacy stat. Native request custody and one-shot completion remain in the existing owned exact dispatcher.

Authority, security and reliability

No path, mount selector, credential, provider endpoint or capability is accepted from the requester. Vfsd uses the exact record's retained provider handle and original mount. The native receiver proves route possession before operation admission. Snapshot revisions do not authorize effects or outlive object retirement. No metadata operation opens, re-resolves, truncates or grants an object. Compatibility deny-only policy and operation-specific conditional effects remain unfinished and must be wired separately.

Acceptance criteria

Checklist
  • Define disjoint exact metadata codecs with selector-free snapshot input and complete page correlation.
  • Implement the actual provider transport and exact dispatcher with original route custody, metadata-only ceilings, refusal cleanup and no private handle disclosure.
  • Author malformed, cross-route, retired-record, wrong-operation, short-buffer, page and original-provider tests; execution is waived by user.
  • Update contracts, manual, original signed dependency/catalog adoption, linked reviews and phase tracking.

Alternatives and tradeoffs

Broadening a write or read endpoint with metadata authority would violate its exact operation contract. Reopening a path would lose retained-object identity. Using legacy stat would mistake unsupported ownership or ACLs for known evidence. The bridge preserves the existing directory metadata scope and paged provider contract. It adds no metadata cache: providers retain their freshness checks and consumers retain complete-ACL obligations.

Tracking and rollout

Continue native permission mediation, building on ACL page transport. Coordinate shared ABI, Vfsd, manual, catalog adoption and phase completion. Existing provider and ordinary exact operations keep their encodings. Full native policy and EriX/toolchain guest builds remain required.

Related records and contracts
Repository Links
vfsd issue 5
## Problem and motivation Native filesystem permission mediation needs explicit ownership, mode and ACL evidence from the original retained filesystem object. The private providers expose that evidence, but Vfsd exact metadata endpoints currently offer no route to it. ## Proposed behavior and scope Only original job-lifetime READ_METADATA directory endpoints serve selector-free snapshots and bounded ACL pages. Vfsd uses the retained mount/provider handle and fresh private correlation, checks framing, handle, kind, revision and range, then replaces private identifiers with caller request correlation and the live VFS record ID. It does not reopen paths, synthesize ACL absence or broaden another endpoint's operation ceiling. Unknown evidence stays unknown, refused staging is cleared, and original owned completion custody prevents refetch/replay after refusal. Actual buffer geometry bounds pages without a new complete-ACL limit; entries remain in existing staging while the fixed prefix is rewritten. Eighteen controls are authored and unrun: five wire, five provider transport, seven real-dispatch and one owned-completion. The original transport interface is shared and native path codecs/resolver fixtures are split; changed code remains below 1,000 lines (maximum 898), with inline rustdoc and missing_docs retained. Component contracts and the technical manual are updated. Tests, builds, lint checks, rustdoc, VM/manual/profiler runs and CI observation remain deferred by user instruction. No pass, warning-free build or measured speedup is claimed. [Native permission enforcement](https://git.erikinkinen.fi/erix/posixd/issues/10) remains open: trusted credentials, full ACL acquisition and actual operation-specific effects still need to be connected. Value-inode ACL payloads and ACL-aware persistent mutation also remain unfinished. Both complete EriX guest-build generations and the Rust/LLVM guest rebuild remain required and unproved. [Phase completion](https://git.erikinkinen.fi/erix/integration/issues/65) is **16.22% weighted**, **4.81% whole-item**, with **20/460 parents** and **435/895 sub-items** complete. Add selector-free exact VFS permission snapshots and revision-bound ACL pages. The exact receiver chooses the original retained directory handle. Only job-lifetime READ_METADATA directory endpoints accept these operations; read, write, enumerate and entry-mutation endpoints retain their existing ceilings. The snapshot describes this directory, never a caller-selected descendant. Forward requests through the original mounted provider route with fresh internal request correlation. Validate snapshot handle and node kind and every ACL page's request, handle, revision, ordinal and bound before returning evidence. Publish Vfsd's monotonic exact-object ID in place of the private provider handle. This number remains descriptive and cannot select a different record. ACL page requests must match the endpoint's current object ID; each page is tied to its original provider revision. Consumers still validate complete count, ordering and mode correspondence across all pages. Use shared checked codecs with disjoint VFS operation IDs and fixed framing. Bound page transfers by the existing response storage, not a new total-ACL limit. Clear refused provider staging. Unsupported providers explicitly refuse; never synthesize ownership, ACL absence or a policy verdict from legacy stat. Native request custody and one-shot completion remain in the existing owned exact dispatcher. ## Authority, security and reliability No path, mount selector, credential, provider endpoint or capability is accepted from the requester. Vfsd uses the exact record's retained provider handle and original mount. The native receiver proves route possession before operation admission. Snapshot revisions do not authorize effects or outlive object retirement. No metadata operation opens, re-resolves, truncates or grants an object. Compatibility deny-only policy and operation-specific conditional effects remain unfinished and must be wired separately. ## Acceptance criteria <details> <summary>Checklist</summary> - [x] Define disjoint exact metadata codecs with selector-free snapshot input and complete page correlation. - [x] Implement the actual provider transport and exact dispatcher with original route custody, metadata-only ceilings, refusal cleanup and no private handle disclosure. - [x] Author malformed, cross-route, retired-record, wrong-operation, short-buffer, page and original-provider tests; execution is waived by user. - [x] Update contracts, manual, original signed dependency/catalog adoption, linked reviews and phase tracking. </details> ## Alternatives and tradeoffs Broadening a write or read endpoint with metadata authority would violate its exact operation contract. Reopening a path would lose retained-object identity. Using legacy stat would mistake unsupported ownership or ACLs for known evidence. The bridge preserves the existing directory metadata scope and paged provider contract. It adds no metadata cache: providers retain their freshness checks and consumers retain complete-ACL obligations. ## Tracking and rollout Continue [native permission mediation](https://git.erikinkinen.fi/erix/posixd/issues/10), building on [ACL page transport](https://git.erikinkinen.fi/erix/lib-ipc/issues/3). Coordinate [shared ABI](https://git.erikinkinen.fi/erix/lib-ipc/pulls/2), [Vfsd](https://git.erikinkinen.fi/erix/vfsd/pulls/3), [manual](https://git.erikinkinen.fi/erix/docs/pulls/4), [catalog adoption](https://git.erikinkinen.fi/erix/integration/pulls/12) and [phase completion](https://git.erikinkinen.fi/erix/integration/issues/65). Existing provider and ordinary exact operations keep their encodings. Full native policy and EriX/toolchain guest builds remain required. <details> <summary>Related records and contracts</summary> | Repository | Links | | --- | --- | | vfsd | [issue 5](https://git.erikinkinen.fi/erix/vfsd/issues/5) | </details>
erikinkinen added this to the Alpha 1 milestone 2026-10-30 09:35:16 +01:00
Sign in to join this conversation.
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
erix/vfsd#5
No description provided.