| Age | Commit message (Collapse) | Author |
|---|
|
PHP's IOInterface::select accepts an associative choices array whose keys
are the selectable values, but the port narrowed it to Vec<String>, making
key-based selection unrepresentable. Accept PhpMixed (List or Array) like
PHP's array $choices; ConsoleIO already branched on both shapes internally.
Also mirror PHP in the single-select array_search fallback for numeric-keyed
arrays. All call sites keep their previous list-based behavior.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
The PHP test implicitly requires the process cwd to be an ancestor of
the Fixtures dir (phpunit runs inside the composer checkout); cargo
runs tests from the crate manifest dir, which is not. Replicate the
phpunit precondition with a drop-restoring CwdGuard chdir'ing to the
__DIR__ equivalent, and serialize it (together with the only other
cwd-mutating test in the repository binary,
test_repository_writes_installed_php) via #[serial] so the
process-global cwd cannot race.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
Add the __add_installer test seam that registers an installer as a
pre-built Rc handle, so the test can reproduce PHP's object-identity
semantics (assertSame / removeInstaller) via Rc::ptr_eq; add_installer
cannot serve because Rc::from(Box) reallocates, losing the caller's
pointer identity.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
Re-verified the reason: besides the missing as_any seam on
EventInterface (a cross-cutting trait change over every event type),
the mocked dispatchScript call expectation is also inexpressible since
dispatch_script is a concrete method with no call-recording seam.
Record both blockers in the ignore string and TODO(phase-d) comment.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
The ignore reason had rotted: ConsoleOutputInterface::set_error_output
is already ported, so a BufferedOutput error sink can be injected into
a real ConsoleOutput and the error-routed write read back, matching the
file's convention of replacing PHPUnit mock expectations with
BufferedOutput assertions.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
The only missing seam was the PHP test's
ReflectionMethod(getPhpExecCommand) access; add the test-only
__get_php_exec_command wrapper and port the test on the existing
get_listeners override and process-executor mock infrastructure.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
PHP's BufferIO extends ConsoleIO, so $io instanceof ConsoleIO matches
it; the port models that inheritance as composition, making the plain
ConsoleIO downcast reject BufferIO and throw where PHP renders tables.
Also try a BufferIO downcast and unwrap its inner ConsoleIO, which
unblocks the two FORMAT_TABLE cases and un-ignores test_audit.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
The PHP instanceof WrappableOutputFormatterInterface check always holds
here: OutputFormatter is the sole OutputFormatterInterface implementor
in the port and it implements the wrappable interface, so the former
todo!() can return true for every representable formatter.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
Spell out why the existing __new_mock seam cannot serve: it replaces
execute() wholesale (skipping the download step, ref
InstallationManagerMock), while the PHP test runs the real execute() and
spies only on the three per-operation methods it dispatches to. A
per-method spy on the real path would be a production design change.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
The old reason blamed missing mock infrastructure, but a spy against
dyn AutoloadGeneratorInterface is perfectly writable. The actual blocker
is that make_autoloader (PHP makeAutoloader, called from doDispatch's
script branches, where setDevMode is invoked) is an intentional no-op in
the port, so set_dev_mode is never reached and a spy would observe
nothing.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
Beyond the missing CommandCompletionTester harness, the PHP test's
expected suggestions depend on the Composer dev checkout environment
(its own composer.json/lock, installed vendor packages) and live
Packagist queries, which the port cannot reproduce without a fixture
environment.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
The ignore reason went stale: the getListeners override seam and a real
ProcessExecutor wired to the IO already cover the PHP setup, and IOStub
is the PHPUnit IOInterface-mock equivalent. IOStub now records
writeError calls (writeRaw was already recorded) so the
expects(once)->with(...) spies on writeError/writeRaw can be reproduced
as call-list equality assertions.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
The ignore reason went stale: the FilesystemMock seam added for
FileDownloaderTest already intercepts removeDirectoryAsync, and the
sibling fossil/hg testRemove ports use it. The seam now records the
removed directories instead of a bare call count so the PHP
->with($this->equalTo($this->workingDir)) argument assertion can be
reproduced faithfully.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
Without a PHP runtime no XdebugHandler is ever constructed, so the
COMPOSER_ORIG_INIS lookup is skipped and php_ini_loaded_file() /
php_ini_scanned_files() are both false; PHP would return
[(string) false] = [""]. Model that directly instead of todo!().
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
The tests fail before any network access: DiagnoseCommand::check_platform
captures phpinfo() via ob_start()/ob_get_clean(), which are todo!() in
shirabe-php-shim, so the command panics before producing output. Record
that as the primary blocker alongside the live-network requirement.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
PHP's getRemoteContents calls file_get_contents unconditionally: the
same stream wrapper reads file:// URLs, plain local paths and network
schemes. The Rust port only handled the explicit "file" scheme, so a
scheme-less local packages.json repository failed with "file could not
be downloaded". Extend the local branch to empty schemes; the http(s)
stub is unchanged.
This clears the first blocker of the create-project functional test;
its ignore reason now documents the remaining ones (live network and
Glob::to_regex emitting PCRE lookaheads the regex crate cannot
compile), which need a user decision.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
The ignore reason went stale: HttpDownloader is reqwest-based now, so
building it for HgDriver no longer reaches curl_multi_init() (todo!() in
shirabe-php-shim::curl). Test content unchanged.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
The file:// branch of RemoteFilesystem::get_remote_contents is now
implemented, so the dist download reaches the ZipArchive "is not a zip
archive" path the test asserts. Test content unchanged.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
validate_array judged emptiness via as_array(), which only matches
PhpMixed::Array, so a non-empty PhpMixed::List (e.g. a funding array) was
misjudged as empty and dropped, diverging from PHP's !count() check.
Match both Array and List.
The stale ignore reason on test_fund_command no longer applies:
init_temp_composer injects packagist:false so no network is reached, and
the downloader stack is reqwest-based (no curl shim todo!()).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
support
ProcessExecutor::execute_async's mock branch was an unimplemented
todo!(), blocking every test whose code path calls it (feature-branch
git diffing, system-unzip/7z fallback extraction). Implement it by:
- Adding Process::__mock (mirroring the existing ZipArchive::__mock
pattern) so execute_async can resolve with a fabricated, already-
terminated Process instead of spawning a real subprocess.
- Extracting the sync mock's expectation-matching logic into a shared
ProcessExecutor::mock_match, wrapping the mock state in a RefCell so
it works from execute_async's &self/&mut self receivers.
- Setting error_output/capture_output from the mock branch, matching
PHP's ProcessExecutorMock::executeAsync sharing doExecute with the
sync path; execute_async now takes &mut self for this (safe, since
the borrow only needs to live through the synchronous setup, not
across the .await).
- Turning a strict-mode expectation mismatch from a panic!() into a
shirabe_php_shim::RuntimeException Err, mirroring PHPUnit's
AssertionFailedError extending \RuntimeException: PHP call sites
that catch (\RuntimeException $e) around a mocked git/hg/svn call
(e.g. Git::get_mirror_default_branch, GitDriver::supports) treat a
mismatch as an ordinary recoverable failure, and now so does the
port. The RefMut is dropped before firing an expectation's optional
callback so a re-entrant callback doesn't panic on double-borrow.
Also fixes two real bugs found while porting test_private_repository_
no_interaction: GitHub::authorize_oauth and GitLab::authorize_oauth
checked their domains config via PhpMixed::as_array(), which only
matches the Array (map) variant, but github-domains/gitlab-domains
default to PhpMixed::List, so the check always returned false and OAuth
token lookup was silently skipped. Use the in_array shim instead,
matching PHP's in_array() semantics. Also fix Git::run_command's
"capture credentials from git remote -v" call, which used the panic-
swallowing execute_args wrapper instead of a fallible execute(), so a
mock mismatch there couldn't reach get_mirror_default_branch's catch.
Un-ignores:
- zip_downloader_test::test_system_unzip_only_{good,failed}
- zip_downloader_test::test_non_windows_fallback_{good,failed}
- event_dispatcher_test::test_dispatcher_outputs_error_on_failed_command
- root_package_loader_test::test_feature_branch_pretty_version
- version_guesser_test::test_guess_version_reads_and_respects_non_feature_branches_configuration_for_arbitrary_naming{,_regex}
- version_guesser_test::test_remote_branches_are_selected
- github_driver_test::test_private_repository_no_interaction (also
adds the missing #[serial], since it seeds the shared Git::VERSION
static that vcs_repository_test::test_load_versions depends on for
real)
- init_command_test::test_get_git_config, made deterministic by
pointing HOME at a throwaway dir with its own .gitconfig instead of
depending on the host's global git config
Deduplicates the GitVersionGuard/RestoreEnv test-drop-guard idioms into
tests/common/test_case.rs instead of reimplementing them per file.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
|
|
port
extract_stability_flags ported PHP's `isset($stabilityFlags[$name]) &&
$stabilityFlags[$name] > $stability` as `unwrap_or(i64::MAX) > stability`,
so the check always short-circuited to "already more unstable" and no flag
(e.g. from an explicit `*@dev` requirement) was ever recorded. Fixed using
Option::is_some_and, a direct translation of PHP's isset() && ... check.
Also fixes tests/common/test_case.rs's shared installation_manager()
helper, which built a real InstallationManager::new instead of the
__new_mock constructor (mirroring PHP's FactoryMock::createInstallationManager()),
so it always had zero installers registered and wrote install-path: null
into fixture installed.json files.
Un-ignores test_reinstall_command, test_locally_modified_packages_from_source/
_from_dist, and test_package_still_present_error_when_no_install_flag_used —
the first three were already passing (their #[ignore] reasons were stale),
the last is fixed by the test_case.rs change above.
Updates the #[ignore] reasons on installer_test.rs's three fixture-driven
integration tests to reflect their current state: the install pipeline now
runs end-to-end, but the ~189-fixture installer/ set still hits several
independent, unrelated bugs/gaps that need case-by-case triage rather than
a single fix.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
|
|
repo bugs
Replace the PhpMixed-based `$showWarnings` hack in VersionSelector::
findBestCandidate with a typed ShowWarnings enum (Always / Predicate),
letting ShowCommand::findLatestPackage pass its real closure instead of
hardcoding `true`. Fix the --no-dev branch in ShowCommand::execute,
which built `repos` from an empty package list instead of sharing the
same InstalledRepository as `installed_repo`. Pass repository handles
instead of pre-borrowed `&dyn RepositoryInterface` refs into get_package/
generate_package_tree/add_tree to stop a RefCell double-borrow panic on
--all/--locked. Add the missing CompletePackage/RootPackage
set_release_date setter so the outdated sorting-by-age test can set
fixture dates. Resolve OutputFormatterStyleStack::pop's empty-style
todo!() via clone_box(), and fix FileDownloader's cache-GC log call to
pass the VERY_VERBOSE verbosity PHP uses.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
|
|
abandoned propagation
The json format branch ignored search results entirely and always wrote
null. Encode results into the same name/description/abandoned/url shape
Composer's array-backed repositories produce. Also fix
ComposerRepository's RepositoryInterface::search adapter, which dropped
the abandoned field from raw API results even when present.
|
|
AliasPackage wrongly delegated several methods (id, names, unique
name, pretty string, full pretty version, repository, __toString)
to aliasOf, but PHP's AliasPackage inherits these from BasePackage
unmodified, so they must use the alias's own state instead. This
collapsed alias and aliasOf into the same SAT literal id, breaking
the solver's alias-resolution rules; un-ignore the two solver tests
that exposed it.
|
|
The Preg delimiter bug the ignore reason described was already fixed.
|
|
PHP matches the SPDX license-expression grammar with a single recursive
PCRE pattern ((?(DEFINE) subpatterns plus (?&name) recursion), which the
regex crate cannot express, so port it as a hand-written recursive-descent
parser instead. licenseid/licenseexceptionid dictionary lookups use
longest-match-first, guarded against accepting a short entry that is only
a coincidental prefix of a longer identifier run (e.g. "DOC" prefixing
"DocumentRef-..."), which a length-sorted greedy match alone would
misparse.
Un-ignores the two validating_array_loader tests that were blocked on
this todo!().
|
|
test_update was skipped for a stale reason; running it uncovered three
distinct bugs it was actually catching:
- ApplicationTester::run never restored SHELL_VERBOSITY after
Application::configureIO mutates it, so one dataset's -vv verbosity
leaked into later runs sharing the process (Symfony's tester restores
it in a finally block; the port dropped that).
- Installer::do_install built its RepositorySet with a hardcoded empty
temporary_constraints map instead of self.temporary_constraints, so
--with never actually constrained the resolver.
- BumpCommand was missing a <warning> tag pair around one of its output
lines.
|
|
actually run
Without this, Application::do_run force-disabled interactivity whenever
stdin wasn't a tty (as under cargo test), so ApplicationTester runs with
set_inputs silently produced non-interactive default output instead of
consuming the answers, masking real behavior as several stale #[ignore]s
blaming already-implemented ProcessExecutor/Process todo!()s.
Port composer/tests/bootstrap.php's env setup into a bootstrap() helper
called from get_application_tester(), un-ignore the now-passing
init/update command tests, and update init_command_test's expected
schema-validation wording to match the jsonschema crate (already the
accepted wording per 541a8b4f, not an unported gap).
|
|
PHP's printTable builds a 5-column row for text output, with the raw
Link object (cast via __toString) as its own column separate from the
formatted description string. The port had collapsed both into one
column, dropping an empty column for successful checks and throwing
off the rendered column widths/spacing versus real Composer output.
|
|
|
|
executeBatch now builds one future per operation — the PHP promise
chain prepare -> install/update/uninstall -> cleanup -> repo->write,
including the '<Op> of <pkg> failed' rejection handler covering the
chain up to cleanup — and drives the whole batch through
waitOnPromises()/Loop::wait, so archive extraction (the unzip
subprocesses gated by ProcessExecutor's semaphore) finally overlaps
across packages. Alias operations stay synchronous in the collection
loop like PHP.
The shared repository is threaded through the chains as
RefCell<&mut dyn InstalledRepositoryInterface>: execute() wraps the
incoming &mut once, and InstallerInterface::install/update/uninstall
take the cell so implementations borrow it only in their synchronous
head/tail, never across an await. InstallationManager's own
install/update/uninstall/download/get_installer/get_install_path/
mark_for_notification move to &self (cache and notifiable_packages
behind RefCell) so every chain can capture &self.
WritableRepositoryInterface::write and the InstallationManagerInterface
get_install_path it relies on lose their &mut manager requirement —
the per-op repo->write inside the chains only reads install paths.
Warm-cache create-project laravel/laravel: the package-operations
phase (109 installs) drops from ~3.0-3.8s serial to ~1.9s, on par
with real Composer (~2.1s) measured back-to-back; the resulting
vendor tree, installed.json included, stays byte-identical to
Composer's (diff -rq clean).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
InstallationManager::downloadAndExecuteBatch now matches PHP: every
update/install operation's installer->download() promise is collected
and driven concurrently through waitOnPromises()/Loop::wait instead of
being awaited one package at a time. Concurrency caps stay where PHP
puts them (HttpDownloader 12, ProcessExecutor 10 via their semaphores).
Error semantics follow PHP too: all downloads settle before the first
rejection is rethrown, rather than aborting on the first failure.
To let the collected futures and the cleanup closures own their
installer beyond the loop iteration that created them, the installer
registry becomes Vec<Rc<dyn InstallerInterface>> and get_installer
hands out clones (PHP closures capture $installer the same way), with
InstallerInterface methods taking &self across the six implementors —
the only genuinely mutable state was LibraryInstaller.vendor_dir
(canonicalized in place), now behind a RefCell.
as_plugin_installer_mut/as_binary_presence_interface lose their &mut.
The cleanup_promises entries are now the real thing: the PHP closure
including the getInstallationSource() guard and the
installer->cleanup($opType, $package, $initialPackage) call, replacing
the no-op futures (drops one TODO(phase-b) and two TODO(phase-c)).
Verified against the real network: create-project laravel/laravel
produces a vendor tree byte-identical to real Composer's (diff -rq
clean across all 109 packages including vendor/composer).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
Concurrent package operations call into the same downloader instances
through Rc<RefCell<dyn DownloaderInterface>>; with &mut self methods
every call holds a RefMut across its awaits, which panics with
'already mutably borrowed' the moment two operations overlap. This is
groundwork for fanning out InstallationManager's download/install
loops (same rework HttpDownloader/CurlDownloader already got).
- DownloaderInterface/ChangeReportInterface/ArchiveDownloader/
VcsDownloader methods now take &self; as_change_report_interface
returns &dyn instead of &mut dyn.
- Implementors move their genuinely mutable state behind cells:
FileDownloader.additional_cleanup_paths, the archive downloaders'
cleanup_executed, ZipDownloader.zip_archive_object,
VcsDownloaderBase.has_cleaned_changes, GitDownloader's stash/discard/
cache maps and GitUtil, SvnDownloader.cache_credentials,
PerforceDownloader.perforce. FileDownloader.io gains a RefCell layer
so get_local_changes can keep PHP's NullIO swap under &self.
- ProcessExecutor::execute_async now returns a future that captures
everything up front instead of borrowing the executor, and call
sites build the future before awaiting, so no borrow on the shared
executor is held while a subprocess runs.
- Filesystem::remove_directory_async becomes remove_directory_async_via
taking the Rc handle: the Filesystem is only borrowed for the sync
head/tail, never across the rm subprocess await (sync borrow_mut
users like rename/ensure_directory_exists would otherwise collide).
- DownloadManager async call sites hold shared borrows only.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
execute_async was a serial pump: it queued a job, then drove it to
completion itself via wait_id()'s blocking usleep loop before
returning, so concurrent callers never overlapped even when polled
together through FuturesUnordered.
Rewrite it as a single &self async fn mirroring the CurlDownloader/
HttpDownloader rework: a tokio Semaphore sized by max_jobs
(COMPOSER_MAX_PARALLEL_PROCESSES, PHP parity) gates admission, the
child is started non-blocking, and an async 1ms sleep loop pumps
is_running()/check_timeout() while yielding to the reactor so sibling
jobs genuinely run in parallel. The Job table, STATUS_* lifecycle,
start_job/mark_job_done/count_active_jobs/wait/wait_id all had no
remaining callers and are removed.
&self also lets callers hold only a shared borrow across their awaits
(zip_downloader, version_guesser, filesystem via the new
get_process_handle), which would otherwise panic with 'already
mutably borrowed' once two async jobs overlap on the same
Rc<RefCell<ProcessExecutor>>.
The async mock branch no longer consumes the expectation before
hitting its todo!(): the panic made that bookkeeping unobservable.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
The earlier measurements documented in the perf notes were taken with
--no-audit to keep the security-advisories request out of the timings,
but the flag never landed in the committed script. Add it so future
runs are comparable.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
|
regex::Regex::clone() does not share the underlying meta engine's
search-cache pool, so every fresh clone pays a ~10us warmup cost on
its first use. Two changes together eliminate this across nearly all
preg_* call sites:
- A php_regex! macro resolves PHP-style patterns to a per-call-site
&'static regex::Regex (via regex-macro's LazyLock), applied at the
majority of call sites throughout the codebase.
- Call sites still passing dynamic pattern strings go through
PATTERN_CACHE, which now stores Arc<(Regex, bool)> and hands out
Arc::clone()s instead of cloning the Regex itself.
PregPattern::resolve() returns a ResolvedPattern enum (Arc or
'static reference) rather than an owned Regex, so neither path ever
clones the Regex proper.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
|
|
serde_json::from_reader issues one read() syscall per byte against an
unbuffered Reader. FileRetriever passed a raw File straight through,
so resolving the composer-schema.json $ref read its ~71KB contents
one byte at a time (twice per require, since schema validation runs
both before and during the update).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
|
|
security-advisories in a p2 provider response is a JSON array, which
decodes to PhpMixed::List rather than PhpMixed::Array. The check only
matched Array, so it always missed and silently fell through to the
api-url fallback instead of using the already-cached provider data.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
|
|
Lighter-weight companion to scripts/bench/create-project.sh: compares
shirabe vs composer on dependency resolution alone (require --no-install
--no-audit), skipping the download/install step that dominates
create-project's runtime.
|
|
The `regex` crate does not support negative lookahead, so the ported
`^dev-(?!main$|master$|trunk$|latest$)` pattern panicked at runtime on
any `require` invocation that reached version-selection. Replace it
with equivalent hand-written string logic per docs/dev/regex-porting.md.
|
|
CurlDownloader::auth_helper was never cloned out to another owner, so
Rc<RefCell<AuthHelper>> only needed the RefCell for interior
mutability (all methods take &self). Likewise HttpDownloader::dispatch
cloned self.rfs into a local binding it only ever used synchronously
(copy/get_contents don't await), so the clone bought nothing.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
|
|
PHP's handleRedirect() throws a bare TransportException when the
Location header is missing, and the caller's single catch block always
unlinks the `~` partial file via rejectJob(). The Rust decide() loop
splits each failure path into its own branch and had unlinked on every
other one, but missed this branch, leaking the partial file.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
|
|
Replace sync_executor::block_on's reactor-less busy-spin poller with
tokio::task::block_in_place + Handle::current().block_on(), riding a
single tokio Runtime entered once in main.rs (falling back to a
disposable one when no ambient runtime exists, e.g. in tests). This
lets HttpDownloader::dispatch await CurlDownloader::download directly
instead of bouncing through the separate curl_runtime() bridge, which
is now deleted.
Manual create-project verification against the real network caught a
concurrency bug this exposed: async_fetch_file held http_downloader's
RefMut across the await on add(), which only panics once downloads
genuinely overlap. add() only needs &self, so borrow() fixes it.
With everything now sharing one real reactor, the FuturesOrdered
fan-out added for ComposerRepository::get_security_advisories/
load_async_packages finally overlaps for real: fetching 8 packages'
metadata dropped from ~7-40s to a consistent ~3-4s in a before/after
comparison, with identical resulting lock files.
sync_executor::block_on's call sites are still synchronous rather
than async fn propagated up to Command::execute, which remains the
end goal (see the TODO(phase-e) in sync_executor.rs) - nested block_on
calls elsewhere don't get this same overlap, only prevented panics.
|
|
|
|
get_security_advisories/load_async_packages serialized every
start_cached_async_download call through sync_executor::block_on
per name, matching PHP's structure but not its promise-based
concurrency. Wrap the mutable state those downloads touch (cache,
fresh_metadata_urls, packages_not_found_cache, degraded_mode) in
RefCell/Cell so start_cached_async_download and its async_fetch_file
helpers can drop to &self, then fan out all downloads for a batch via
FuturesOrdered before processing responses sequentially in original
order, mirroring PHP's build-all-promises-then-wait shape.
Real I/O overlap still awaits item 7 (HttpDownloader::add currently
resolves through the curl_runtime()/sync_executor::block_on bridge
either way), so this is architectural groundwork, not a perf win yet.
|
|
11 downloader/installer integration-test files each redefined an
identical current_thread `run()` helper to block on async code. Extract
one shared multi_thread Runtime into tests/common/async_runtime.rs so
concurrent #[test] threads can all block_on it, matching the direction
item 7 (top-level Runtime) will take in production code.
|
|
Loop::wait already had the target signature and a TODO(phase-c-promise)
marker noting it drove promises serially; swap the for-loop for
FuturesUnordered so all promises are polled together instead of one at
a time, keeping the "remember only the first error" semantics.
This adds the first real use of the futures dependency (already present
in Cargo.toml/Cargo.lock from earlier prep work, now finally consumed),
so those lockfile/manifest changes land in this commit.
Real overlap still doesn't happen yet: each promise (HttpDownloader::add/
add_copy etc.) resolves through a blocking bridge (curl_runtime()/
sync_executor::block_on) that fully occupies the thread until it
settles, so this is groundwork for once a single top-level Runtime
replaces those bridges. Updated the TODO(phase-c-promise) comment to
reflect that.
|
|
Replaces Job/Request/JobHandle/id_gen/running_jobs/max_jobs with a
tokio::sync::Semaphore permit held for the duration of each request.
get/add/copy/add_copy are now &self (add/add_copy are also genuinely
async); a shared execute()/dispatch() core replaces
add_job/run_rfs_job/start_job/settle_job, returning the Response
directly instead of deferring to wait()/count_active_jobs()/
get_response() (all removed — confirmed zero callers, same for the
now-unused STATUS_* constants).
get()/copy() stay synchronous rather than becoming async wrappers
around add()/add_copy(), bridging via the existing sync_executor
instead of the curl_runtime() introduced for CurlDownloader: their
callers (~35 files reaching HttpDownloader) are mostly plain sync fns
with no async boundary anywhere in the call chain, and forcing that
propagation now would pull forward the dedicated async-propagation
task. curl-eligible requests still route through curl_runtime()
inside dispatch(), same as before — nesting sync_executor::block_on
(no real reactor) around curl_runtime().block_on() (a real, separate
Runtime) is safe; it's only nesting curl_runtime() inside itself that
would panic.
CurlDownloader no longer needs Rc<RefCell<>> wrapping despite the
original design sketch: since item 2 made all of its methods &self,
a plain Option<CurlDownloader> field works fine under HttpDownloader's
own &self methods.
get/add/copy/add_copy becoming &self (rather than &mut self) requires
no changes at any of their ~35 calling files: RefMut/Ref both deref to
a type that can call &self methods just fine.
Verified manually against real network I/O (sandbox disabled): `shirabe
show -a` (get()'s sync_executor-bridged path) and `shirabe
create-project` (add_copy()'s genuinely async path via file_downloader.rs)
both complete correctly with no hang.
|
|
Freeform notes about intentionally-unported production behavior are
easy to miss on a read-through and impossible to grep for later.
|
|
Replaces the Job-table + tick()-driven polling loop with one async
download() that sends, decides (retry/redirect/fail/succeed via a new
decide() extracted from the former run_job), and loops until it
resolves — no more resolve/reject callbacks. The client switches from
reqwest::blocking::Client to the non-blocking reqwest::Client, with
body streaming now via tokio::fs.
Because real async I/O needs a live tokio reactor and none runs yet at
the process level (sync_executor::block_on is a no-reactor busy-spin
executor that only works when awaited futures resolve synchronously),
HttpDownloader::start_job drives CurlDownloader::download() through a
dedicated temporary current_thread Runtime (curl_runtime(), marked
TODO(phase-e)) instead. This keeps concurrency characteristics
unchanged for now — start_job still resolves one job at a time — real
parallel I/O lands once HttpDownloader/Loop are rearchitected on top of
FuturesUnordered.
count_active_jobs' curl.tick() polling and the Job.settled/curl_id
plumbing are removed as dead weight now that start_job settles curl
jobs synchronously, same as the rfs path already did.
abort_request is dropped: it had no caller (the PHP Promise-cancellation
flow it backs was never ported), and the job table it operated on no
longer exists.
Verified manually against real network I/O (sandbox disabled): `shirabe
show -a` (JSON metadata, in-memory body) and `shirabe create-project`
(actual dist zip download + extraction) both complete correctly with no
hang. Two unrelated pre-existing bugs surfaced during manual testing
(an event-dispatcher subscriber wiring gap during `require`, and a
RefCell reentrancy panic in `diagnose`) reproduce identically on the
pre-change code and are out of scope here.
|
