mirror of
https://github.com/sabnzbd/sabnzbd.git
synced 2026-09-14 22:51:34 -04:00
develop
1
Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
|
|
35d5355f49 |
Replace CherryPy with Uvicorn and Starlette (#3550)
* Migrate web interface from CherryPy to Uvicorn/Starlette
Squashed rebase of feature/uvicorn (34 commits) onto develop, reconciled
with ~3 months of intervening develop changes.
Replaces the CherryPy webserver and request handling with Uvicorn/Starlette
across the API, web interface, RSS, config pages and related modules.
Reconciliation with develop during the rebase:
- api.py: kept develop's security/behaviour fixes (orphan path-traversal
guard, expanded log redaction incl. host_whitelist and
remote_label_replacement, get_dconfig single-return, get_retryable_jobs,
connections default, translated NNTP test errors) on top of the Starlette
request/response rewrite.
- interface.py: ported the RSS route handlers to develop's DB-backed
RSSRepository API (process_feed, rss_repository / find_job_by_url /
clear_feed / clear_downloaded / flag_downloaded).
- misc.py: kept develop's hachoir-based get_media_duration.
- requirements.txt: dropped the CherryPy stack, adopted develop's newer pins.
Also applied ruff --fix (PEP 604 unions, builtin generics) to align with
develop's lint config.
Verified: ruff check, black --check, and the affected test suites
(9413 passed, 1 skipped) all pass.
* Update starlette/uvicorn versions
* Fix race issues in global rss state
* Fix test race in server shutdown
The uvicorn migration turned /shutdown (and the shutdown API) into fire-and-forget: it spawned shutdown_program() in a background thread and replied immediately, whereas develop ran it synchronously and only replied once halt() had persisted all state. Because the module-scoped test teardown doesn't wait for the process to exit, the next module's clean_cache_dir wiped the shared cache dir (and reused the fixed port) while the previous instance was still saving state and holding the port — producing the three intermittent failures (deleted sabnzbd.log → "File log disabled or not found"; un-persisted [sorters] → KeyError; stale instance → missing wizard .quoteBlock).
* Fix robots and description, add favicon
* Remove remains of http basic auth
* Setup Starlette once configuration is available, fix static file relative cwd and url_base config
* abort_and_show_error when webserver fails to start
* Guard stopping webserver that never started
* Delegate XFF handling to ProxyHeadersMiddleware
* Merged params at request.state.params instead of modifying private apis
* Both shutdown routes share implementation and do not block event loop
* Run sync handlers via run_in_threadpool and facilitate eventual migration to async
* Pool database connections
* Online backup of database due to WAL changes
* Fix exception on None request.client (test clients or unix sockets)
* Fix flakey tests due to process not fully shutting down
* Restore X-Frame-Options behaviour via middleware
* Fix set_config_default with multiple keywords
* Remove broken logging call
* Restore api logging functionality
* Cache-Control: no-store
* Login only via POST
* Remove 401 (basic-auth) and add 404 handling via redirect
* Fix crash when shutdown not an int
* Use BaseRedirectResponse helper
* Remove trailing slashes from wizard routes
* URL helper, absolute URLs everywhere, fixes issues with nested navigation
* Fix scheduler adding multiple daysofweek
* Restore CherryPy api behaviour merging body with query params (body wins)
* Clearer documentation of get_request_params and request_params
* First stage supporting gradual api async
* Fix rss ajax consuming flash
* Restore access log functionality
* Hostname check in middleware
* Request logging in middleware
* Param parsing in middleware
* Security checks in middleware
* secured_expose is now purely route registration
* Lookup api handler once per request
* Fix flakey alert dialogs
* Trigger restart via BackgroundTask
* Restore CherryPy first param wins and get/post consistency
* Remove dead code
* Secure cookies based on protocol the client used
* Fix various issues with port_is_free
1. port_is_free answered the wrong question. It connect-probed ("is something answering?") rather than bind-probed ("can I bind?"). A port could report free and then kill startup at uvicorn's bind().
2. The bind-all remap crossed address families. :: was mapped to 127.0.0.1, probing IPv4 for an IPv6 bind — a regression against portend, which maps :: → ::1.
3. The call sites passed the wrong host. browserhost is a client-reachable address; the thing that has to be bindable is web_host.
4. Errors were swallowed. A bare except OSError hid gaierror, so an unresolvable host reported "free".
5. find_free_port had a port-0 trap. Under a bind-probe, currentport=0 always succeeds and returned 0 — the old failure sentinel. Now guarded, and None instead of 0.
6. Ports 80/443 were misdiagnosed. EACCES was folded into "occupied", producing ten futile probes and a panic claiming another program held the port. PermissionError now propagates to a dedicated panic explaining the actual remedies.
7. The tests were largely tautological. Three tests covering one branch, an IPv6 test with no IPv6 in it, a timeout test that never engaged the timeout, TOCTOU-prone fixed-range probes, no SO_REUSEADDR on the helper listener, and nothing asserting the property that matters — that "free" implies bindable.
8. A portability bug I introduced, then fixed. I'd baked Linux SO_REUSEADDR overlap semantics into four assertions; macOS differs. Now platform-aware, with the IPv6 regression re-covered by checking the socket family directly.
* Claim the bind address for uvicorn on startup, resolves "49" in err handling from cherrypy
* Rename function BaseRedirectResponse to base_redirect_response
* Restore error response on change web directory
* Add missing typings
* Fix return type of retry job for future types
* A better fix for xdist compatibility - test overwrote db_path
* Secure session cookies (rss flash)
* Inline or remove some functions
* Retry job futuretype behaviour
* Sneak a worksteal fix in
* Test and fix retry_job futuretype behaviour
|