Files
ESP32_Serial_Swiss_Army_Knife/docs/phase8d16_implementation.md
T

97 lines
17 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# Phase 8D.16 — Broker visibility and confirmed writer assignment
## Status and scope (2026-09-13)
User-authorized implementation complete; focused host regressions, broader regressions and production build pass. **Independent review complete with no actionable findings; target validation and user phase sign-off remain pending.** Reviewer independently reran broker management/wrap, cookie Broker and Display, dispatcher, lifecycle25, UI119+C/HTML/CSP, throughput and diff checks, all PASS. Parent final `pio run` confirmation PASS6.93s,100,300 B RAM/1,782,613 B flash; diff check PASS. Host doubles do not establish real multicore timing or target margins.
Settings → Broker is admin-only and uses the existing Serial/Display label/value and action styles. It shows up to eight connected clients with ID, type, name, writer/observer role, pending output bytes, output high-water and dropped bytes. Selection is explicit and Refresh clears it. Only Assign writer followed by confirmation submits a mutation. Viewing, selecting, refreshing, switching settings/terminals and reading results do not change the lease or close either terminal.
The initial Git status was clean. Preserve the previously signed-off combined binary WebSocket send and CPU160 configuration; neither was edited. No 8D.15 diagnostics UI (diagnostics remain shell-only), 8D.17/18 popovers/icon changes, disconnect/revoke-only UI, serial persistence/configuration change, new generic runner/task/timer/mutex allocation, queue depth/item or stack-size increase, generated assets, SDK patches, upload, erase, branch or commit. No full M3 or resource-reserve acceptance claim.
## Production files and ownership
- `src/session_broker.{c,h}` owns the atomic non-consuming compact management projection, conditional writer assignment and non-reused client IDs. All serial mediation, event queues and data buffers remain broker-owned.
- `src/web_broker_settings.{c,h}` owns HTTP authorization/parsing/encoding, one originating-login-bound operation/result slot and typed execution adapter.
- `src/admin_ssh_console.{c,h}` queues only a numeric operation ID on the existing four-entry request queue. The same dispatcher executes the broker API outside its critical sections, never through a constructed command string.
- `src/web_server.c`, `src/CMakeLists.txt` compose three optional routes and the new source.
- `src/web_ui.c` supplies authored HTML/JS, native select/confirmation, bounded result polling and uncertainty recovery. Loader/CSP and generated/vendored assets are unchanged.
## Authoritative broker contract and wrap analysis
`session_broker_get_management_snapshot()` takes the existing mutex with **zero wait**, copying writer ID, lease generation and all active compact rows in one lock acquisition. It never reads UART payloads or pops events. Busy/uninitialized snapshots are unavailable, not empty-success snapshots. Pending/HWM/drop fields describe broker output, not peer receipt; counters cover the connection or last counter clear. Disconnected clients disappear; global historical accounting remains unchanged.
`session_broker_assign_writer_current(target, generation)` requires a nonzero target and generation. It shares the existing force-writer implementation, with the generation comparison and connected-target lookup **inside the same broker mutex, before any lease/counter/event effects**. Stale generations return `ESP_ERR_INVALID_STATE`; absent/reused targets return `ESP_ERR_NOT_FOUND` when the generation is otherwise current. Both map to a typed `conflict` result. Assignment to the already-current target is a no-op only after validation; UI disables this redundant action. There is no compare-unlock-force race.
The existing unconditional force/release APIs and transport request/release policy remain available. At most one connected client owns the lease; assignment makes the former writer an observer. Already-accepted serial TX bytes are not recalled or discarded. Events remain advisory: transport reconciliation still uses authoritative broker state.
Two independent stale-reference problems required small broker changes:
1. **Client IDs:** three slot bits plus 29 generation bits; generations previously wrapped from `536870911` to `1`. Free slots at their maximum generation are now skipped/retired until reboot. The final generation is still usable, including ID `UINT32_MAX` in slot 7. Exhausting all eight slots returns the existing `ESP_ERR_NO_MEM` without DTR/client publication. No previously-issued ID is reused within a boot; no wider transport protocol ID or per-client allocation is introduced. Reboot resets broker IDs but invalidates all originating web sessions, so an old authorized confirmation cannot span boots.
2. **Lease confirmation:** a separate 32-bit generation starts at 1, survives counter clear and advances under the mutex for each granted/released/revoked event, **before attempting advisory delivery**. All current lease transitions emit these events; denied requests and same-writer no-ops do not advance it. A forced transfer can advance twice (revoke plus grant), so this is an opaque version, not a transition count. Release/reacquire by the same writer invalidates old confirmation (ABA). At `UINT32_MAX` it saturates permanently; management snapshots remain readable but conditional assignment rejects and UI disables assignment. Ordinary transport requests, release/disconnect, console force and UART0/USB recovery remain available. Counter/event-sequence wrap cannot resurrect a confirmation because neither is used as its token.
New observer connections/disconnections do not invalidate unrelated confirmations. The selected target's non-reused ID still prevents reassignment to its replacement; any writer disconnect changes the lease generation. The central grant/release/revoke emission invariant must be preserved by future lease mutations.
## HTTP and operation contract
All routes require current cookie/principal authentication and administrator role. Normal users cannot read management details/results or transfer. GET retains the established bodyless/queryless/header/Origin policy. POST additionally requires canonical same-origin Origin, CSRF and exact supported JSON Content-Type, with no query or transfer encoding. Responses use no-store/nosniff/no-referrer; scratch cleanup and unread-body close behavior follow Display.
| Route | Method | Bound / response |
|---|---|---|
| `/api/settings/broker` | GET | 2,048-byte response buffer; exactly `generation`, `writer`, `clients`. 503 when broker unavailable/busy. |
| `/api/settings/broker-operation` | POST | Nonempty, at most 256 bytes/four receive calls. Exactly `action:"assign"`, `generation` (14294967294), `target` (14294967295). 202 means admitted, not executed. |
| `/api/settings/broker-operation` | GET | 96-byte response buffer; exactly `id`, `action`, `state`, latest result for this originating login only. |
Each client row has exactly `id`, `type`, `name_hex`, `pending`, `high_water`, `dropped`. Type is the existing enum (0 Console, 1 USB, 2 Web, 3 SSH, 4 Internal). Names are at most 23 bytes encoded as at most 46 lowercase hex characters; browser decoding and DOM `textContent` avoid JSON/HTML injection. Invalid UTF-8 displays replacement characters, while IDs remain the authoritative identity. `dropped` is an unsigned decimal **string**, preserving all 64 bits in JavaScript. Writer/observer role is derived from the same snapshot's writer ID. No passwords, tickets, verifiers, private keys or UART bytes are included.
The narrow flat parser rejects duplicates, missing/unknown fields, escaped names/actions, strings in numeric fields, nested values, leading zeros, negatives, fractions, exponents, integer overflow, embedded NUL and trailing garbage. Exact 256-byte whitespace-padded requests are supported. Excessive fragmentation/receive errors fail rather than retry/drain indefinitely.
One static slot is shared across administrators, but results are isolated by non-reused **web-session ID**, not username. Another login sees `{id:0,action:"none",state:"idle"}`. Pending/full/not-ready dispatcher or exhausted operation IDs returns 503/Retry-After 1. IDs never wrap; completed results are replaceable, not durable history or idempotency records.
The dispatcher checks original session/principal/admin and a **30-second dequeue admission deadline** before calling the conditional broker API. No timer or hard cancellation is added: a blocked dispatcher can retain the bounded pending record beyond the deadline until dequeue. Work admitted before logout/expiry/stop may finish; no delivery guarantee is made. Auth stop/restart invalidates queued originating identities. Principal material is wiped at completion and local copies are wiped on exit.
States: `idle` (no retained result for this login, not proof of cancellation), `pending` (queued/executing), `ok` (assignment/no-op completed at execution time, not a future lease guarantee), `conflict` (stale/exhausted generation, absent target or uninitialized broker; no change by this operation), `failed` (other owner failure), `cancelled` (currentness/dequeue deadline denied before owner admission).
Optional registration orders snapshot → result GET → mutation POST. Failed snapshot/result registration prevents later routes; failed POST removes result GET, unless unregister itself fails, leaving reads only. Other settings, authentication, serial/admin transports and recovery survive optional failures. Budget: **30 → 33 URI handlers**, six sockets/no LRU unchanged.
## Browser behavior
- Admin-only full-page Settings navigation; no empty placeholder, contextual popover, icon redesign or new ordinary-user status detail.
- No default selected target; no transfer on opening/selecting. Refresh always clears selection rather than silently renewing a previous confirmation.
- Confirmation identifies target ID/type/name and snapshot writer ID, explains observer transfer and already-queued UART bytes. The exact selected target/generation is captured before asynchronous session revalidation and submission; concurrent changes are rejected at the broker lock.
- One explicit POST, no automatic mutation retry. At most ten one-second result GETs and a 15-second overall automatic-check deadline, including delayed session work. Manual Check Result remains available after uncertainty.
- Lost acknowledgement, another-tab result replacement, invalid response and navigation preserve explicit uncertain-outcome messaging. Known terminal results refresh the snapshot once; stale/conflict requires fresh selection/confirmation.
- Navigation, pagehide, logout and expiry abort/fence stale browser work and clear selected identities/rows. This is not backend cancellation. Both terminals retain their existing hidden-output draining and selected-input policy.
## Validation actually run
All commands below passed in this implementation session with bounded terminal runtimes:
- `python3 tests/session_broker_diagnostics/run.py`: existing seven diagnostic groups plus two management/wrap groups, actual broker/console under deterministic RTOS/serial doubles. Covers atomic/non-consuming projection, zero-wait contention, USB/SSH/Web interleavings, exactly-one-writer transfer, stale/reused targets, release/reacquire ABA, counter clear, local force-release, writer disconnect, saturation and final client IDs.
- `python3 tests/web_cookie_auth/run.py --broker`: six new groups plus shared auth/store/installed-IDF boundary tests; actual HTTP policy/parser/operation module with broker dependency double. Bounds, safe eight-row encoding, session-isolated results, queue failure/replay, stale-owner result mapping, currentness/deadline/expiry/revocation/stop-restart and lost acknowledgement.
- `python3 tests/admin_console_boundary/run.py`: actual dispatcher, including new Broker ID routing, zero/full/not-ready admission, unchanged four-slot capacity and no command-runner invocation; existing certificate and SSH adapters also pass.
- `python3 tests/web_admin_transport/server_lifecycle.py`: 25 groups, including six Broker descriptor/name allocation positions, failed result unregister, failed stop/retry/restart and unrelated route isolation. Orchestration uses HTTPD/TLS doubles, not actual TLS allocation faults.
- `python3 tests/web_ui_session/run.py`: **119 browser groups** (eight new Broker groups) plus actual C renderer/HTML structure/shared styles/no-store/loader-CSP checks. Snapshot validation, safe names/64-bit counters, no view/selection mutation, confirmation/cancel, stale completion, polling bounds, lost/replaced results, navigation/401/pagehide/logout fences. No real browser geometry or target visual pass claimed.
- Cookie variants `--display`, `--network`, `--accounts`, `--serial-settings`, `--settings`, `--admin`; canonical `tests/web_network_settings/run.py`; `tests/web_auth_parse/run.py` (294 cases); `tests/admin_console_boundary/accounts.py`, `lifecycle.py`; `tests/admin_ssh_policy/run.py`; `tests/web_serial_performance/run.py` (including 513 installed-SDK wire comparisons); `tests/web_httpd_idle/run.py`; `tests/web_admin_transport/run.py --tickets`; `tests/web_session_store/run.py --serial`; `tests/web_diagnostics/run.py`; `tests/web_login_ui/run.py`.
- `git --no-pager diff --check` passed.
Tests initially caught and fixed a six-field Broker row incorrectly checked as seven fields. Harness integration corrections: static helper collision with extracted console code, HTTP test output bound, and lifecycle route-count expectations. No remaining known focused test failures. Self-review inspected broker transitions/ID generation, HTTP/session policy, optional registration and browser confirmation/result flow; subsequent independent review found no actionable findings (see status above).
### Build/resource evidence
Actual pre-edit `pio run`: PASS **7.14 s**, **100,196 B RAM / 1,765,233 B flash**. Post-integration build: PASS82.93s. Final confirmation after formatting/review: PASS21.20s, unchanged sizes. Final substantive-source build: PASS **20.54 s**, **100,300 B RAM / 1,782,613 B flash**: **+104 B RAM / +17,380 B flash** versus baseline, within the existing 4 MiB application partition. Authored response sizes: HTML28,929 bytes; app JS109,731 bytes after copy-text cleanup (not generated assets). Existing nonfatal SDK `FATFS_PRINT_FLOAT` bool-default warning occurred in the integration build; no SDK/config workaround applied.
CPU160 verified in `sdkconfig.defaults`, active `sdkconfig.esp32-s3-devkitc-1-n16r8` and generated `config/sdkconfig.h`; PlatformIO's generic board banner still says240MHz and is not the configured frequency. No CPU/scheduling/throughput tuning was performed.
Static image delta excludes runtime URI allocation overhead and peak nested stack usage. Snapshot handler uses bounded 2,048-byte response plus compact eight-row snapshot and auth locals on the existing HTTPD stack; operation admission uses256-byte request/96-byte result buffers and one slot, no payload heap allocation. HTTPD/dispatcher live margins, full-mix internal/DMA/PSRAM minima and scheduling overhead are **unmeasured**. A build pass does not establish safe runtime reserves.
## Pending target checklist — not executed
- [ ] Compare admin Broker rows with UART0 broker snapshots/counters for zero, one and full supported client mix; verify ID/type/name/role/pending/HWM/drop semantics without consuming serial data. Normal user direct GET/POST requests must be denied.
- [ ] Open Settings/Broker, select/cancel/Refresh, switch terminals/views and confirm no lease/socket change; exercise keyboard/touch selection and native confirmation on narrow/wide screens.
- [ ] Assign among USB, role-user SSH and both web serial clients while both admin routes remain open. Verify exactly one writer, observers still receive binary output, former writer input is rejected, no in-band controls and already-accepted TX semantics.
- [ ] Hold confirmation while target disconnects/reconnects; confirm old target fails without changing the current writer. Repeat with competing USB/SSH/browser writer requests, admin-shell force, local expected-writer release and same-writer release/reacquire. Refresh/reselect/confirm explicitly after conflict.
- [ ] Test two independent admin logins and two tabs on one login: pending capacity, isolated/latest replaced results, lost POST/result response, bounded polling/manual recovery and no duplicate transfer. Logout/expiry/role revoke/HTTPS stop-restart while queued must reject stale admission; already-admitted work may finish.
- [ ] With full mix at signed-off160MHz/230400 baud, repeat settled non-consuming throughput captures and verify binary transparency/drop isolation. Measure boot/full-mix/operation internal/DMA/PSRAM free/minimum/largest blocks and HTTPD/dispatcher stack minima, including maximum rows and concurrent HTTPS load.
- [ ] Validate optional route allocation failures and failed stop/restart where a safe target fault-injection mechanism is available; unavailable settings must not remove UART0 or native USB recovery. Host failure injection is not target proof.
- [ ] Record exact revision/build/client mix, observations and limits; independent review is complete, but explicit user 8D.16 sign-off remains required. No erase or identity/configuration migration is required.