Memory ownership and long-running PHP

z-engine writes into live engine structures, so it has to be explicit about who owns every byte it touches. This document describes the ownership model introduced by the memory-lifetime overhaul (#62) and how to run z-engine safely in long-running processes (worker loops, FPM with opcache preload).

The ownership model

Every value wrapper carries two orthogonal ownership bits:

Bit	Meaning	Cleanup
ownsContainer	z-engine allocated the container CData (eg the 16-byte zval box)	freed through the FFI allocator
ownsReference	the wrapper holds exactly one reference on the refcounted payload	dropped exactly once through the engine primitives (zval_ptr_dtor / rc_dtor_func)

The engine refcount stays the single source of truth. Because every owning wrapper holds its own reference, two wrappers aliasing the same pointer can never double-free.

Which constructor you use decides what you own:

Construction	Ownership	Lifetime behaviour
new ReflectionValue($x), new StringEntry($s), new ObjectEntry($o), new ResourceEntry($r), new ReferenceEntry($ref)	owning	addref on construction, automatic release on destruction (or explicit release())
StringEntry::fromString($s)	owning	fresh refcount-1 emalloc string, never aliases caller memory
StringEntry::persistent($s)	owning	fresh refcount-1 malloc string for sinks inside persistent engine structures
*::fromCData($ptr), ReflectionValue::fromValueEntry($ptr)	borrowed	no addref, release() is a no-op, the caller guarantees the pointer stays valid
ObjectEntry::weakFor($o)	borrowed + guarded	no addref, but every access after the object died throws instead of dereferencing a dangling pointer

Rules of thumb:

• release() is idempotent; after it, any access to an owning wrapper throws.
• Handing a pointer to an engine structure that will release it later (a class entry field, an AST node, a hashtable) goes through transferReferenceOwnership() — the wrapper keeps the pointer readable but no longer drops the reference; the engine sink does.
• Manual refcount surgery on engine-owned references uses releaseReference() — full engine semantics: interned/immutable payloads untouched, destruction at refcount zero via rc_dtor_func, persistent blocks never freed with the request allocator.
• Setters with engine sinks (ReflectionClass::setFileName(), ClosureEntry::setThis(), DeclarationNode::setName()/setDocComment(), ReflectionValue::setNativeValue()) release the previous value and store an owned replacement — callers no longer need to keep source values alive.
• ReflectionValue::initializeNativeValue() exists for uninitialized engine output slots (cast_object retval, do_operation result), where there is no previous value to release.

Parsed ASTs

Compiler::parseString() returns a tree whose arena and payload references are owned by an AstOwnership handle that travels with every node materialized from the tree. The whole tree is destroyed (zend_ast_destroy + arena free) when the last wrapper is collected. Two consequences:

• keep a node (any node) of the tree alive for as long as you read from it;
• do not graft nodes from a parsed tree into the live compilation AST (getAST() / AstProcessHook) — the detached tree will be destroyed independently. Build fresh nodes inside the AST process handler instead.

Hook lifecycle

install() replaces an engine function pointer with a libffi trampoline and registers the hook in a Core-level registry that keeps it alive. uninstall() restores the original pointer; only the most recently installed hook of a field may be uninstalled (out-of-order uninstalls throw). reinstall() mints a fresh trampoline.

Core::shutdown() — registered automatically via register_shutdown_function in Core::init() — unwinds all hook chains in reverse installation order. User shutdown functions run before object destructors and before ext/ffi frees the callback trampolines, so every hooked engine pointer is restored while writing it is still safe.

Invariant: no trampoline pointer survives Core::shutdown() in any structure that outlives the request. This is what makes FPM + opcache preload safe: persistent engine structures (class entries, handler blocks) never carry a pointer into the next request's freed trampolines. After shutdown, z-engine performs no engine writes at all — hooks are inactive during shutdown-phase object destructors, and installing a new hook throws.

Core::init() is idempotent per process: a repeated call — eg re-booting z-engine after a manual Core::shutdown() inside one live worker process — reuses the process-wide FFI binding, so every CData minted before the re-boot (module entries, hook state, heap anchors) stays valid. Minting a second binding would free the first one's type data together with the old FFI object and leave those CData dangling by the time the exit handlers touch them.

Runtime models

• Worker loops (RoadRunner, Swoole, ReactPHP, FrankenPHP worker mode): the whole worker is one PHP request. Install hooks once at boot; Core::shutdown() runs at worker exit.
• Classic FPM + opcache preload: call Core::preload() in the preload script, then Core::init() and hook installation happen per request; Core::shutdown() guarantees the per-request trampolines are gone from persistent structures before the request ends.
• SAPIs that cycle FFI callback state between handled requests: call Core::reinstallHooks() at the start of each handled request. Prefer a single hook per engine field in such setups — stacked chains keep intermediate proceed() targets from the previous cycle.

Module lifecycle callbacks

AbstractModule::register() wires FFI-closure trampolines into the module entry's module_startup_func / module_shutdown_func / request_startup_func / request_shutdown_func slots when the module class implements ModuleLifecycleInterface, and into info_func when it implements ModuleInfoInterface. Two hard rules shape the delivery semantics:

1. The trampolines follow the standard hook lifecycle. Core::shutdown() restores the NULL pointers while the trampolines are still alive, so the module entry — which the persistent module registry references directly since PHP 8.4 — never points into freed libffi memory (ext/ffi frees every callback trampoline at its own RSHUTDOWN). Every trampoline additionally checks Core::isShutdown() and no-ops after shutdown.
2. Callbacks never throw across the FFI boundary (#50): ext/ffi aborts the process on an escaping exception, and a FAILURE result from MINIT escalates to a fatal E_CORE_ERROR. Failures are contained and reported as E_USER_WARNING; the engine always sees SUCCESS.

Consequences — only request-phase callbacks are guaranteed:

Callback	Delivery
moduleStartup()	guaranteed: the engine calls the MINIT trampoline inside AbstractModule::startup()
requestStartup()	guaranteed: delivered directly by startup() for the current request (dl() parity — the engine only activates modules at request start)
requestShutdown()	guaranteed: delivered at request end by z-engine's own shutdown chain, immediately after Core::shutdown() (user shutdown functions run in registration order and Core::init() registered first). The engine's own RSHUTDOWN walk happens later, after the trampoline pointers were cleared. Engine writes are already forbidden inside this callback.
moduleShutdown()	best-effort only: real MSHUTDOWN runs after the FFI bridge teardown (for temporary modules in zend_post_deactivate_modules(), for persistent ones at process shutdown), where no PHP callback can be reached. It fires only if the engine destroys the module while the request is still alive.

Runtime-registered modules are wired for the current request only: after Core::shutdown() the entry keeps no callback pointers, so subsequent requests of the same process (FPM, workers) see the module without lifecycle callbacks unless it is registered again.

Module visibility and opcache constant folding

A runtime-registered module is a genuine module_registry entry, and every runtime lookup sees it: extension_loaded($name) with a runtime-built name, get_loaded_extensions(), new \ReflectionExtension(...), the phpinfo() module section. But opcache's optimizer pre-evaluates extension_loaded('<literal>') while the calling script is being compiled (zend_optimizer_eval_special_func_info()): a module that is not in the registry at compile time — with enable_dl off — folds the call to constant false, baked into the cached op array. A z-engine module registers at runtime, after the caller was compiled, so with opcache active such literal call sites keep answering false no matter how the registration went; dl()-loaded extensions face the same fold (#243). Ask ExtensionManager::has() / AbstractModule::isModuleRegistered() instead (both are fold-proof by construction), or pass the extension name through a variable.

Immortal-by-design allocations

The debug-build leak gate treats the following as expected, by design:

Allocation	Why it stays
Module entries, module name/globals buffers (AbstractModule)	since PHP 8.4 the engine module registry stores the registered zend_module_entry pointer directly (no copy), so the entry and its buffers must live for the process lifetime; all are malloc-backed
Persistent zend_module_dep[] arrays and their name/rel/version strings (AbstractModule::getModuleDependencies())	referenced from the registered module entry's deps field for the process lifetime; malloc-backed, bounded to one array per registered module
One zend_object_handlers block per hooked class entry	objects still dereference ->handlers after user shutdown functions ran, so freeing at shutdown would be a use-after-free; blocks are malloc-backed, keyed by class entry address, and bounded
One live libffi trampoline per installed hook	owned by ext/ffi, freed by its RSHUTDOWN
One zend_object_iterator_funcs vtable for the get-iterator bridge (IteratorBridge)	live engine iterators dereference ->funcs for their whole lifetime; the block is a malloc-backed process-wide singleton filled with libffi trampolines (trampolines themselves owned by ext/ffi). The bridge registers in the Core hook registry, so Core::shutdown() neutralizes surviving iterators (drops their cached current-value reference, swaps their handlers to std_object_handlers) while trampolines are still alive, and Core::reinstallHooks() re-mints the vtable for cycling SAPIs
Closures immortalized by ReflectionClass::addMethod()	the method table references the closure body for the rest of the request
Copied-out shared-memory function containers (FunctionLikeTrait::copyOutOfSharedMemory()) and the SHM originals they replace	opcache SHM is never written or freed; the writable zend_function container is malloc-backed (the engine's table destructor releases the body but never frees user function containers); bounded to one container per copied-out function
Added-method zend_function and added-constant zend_class_constant containers (ClassDelta::apply())	published table buckets reference them until the class dies; the engine releases the bodies/values they carry but never frees the containers themselves (it assumes arena storage); malloc-backed, bounded to one per added member
Removed-method structures and their bodies (ClassDelta::apply())	warmed-up inline caches and subclass method buckets may still point at the unpublished entry, so neither the structure nor its opcodes can be freed; bounded to one body per removed method
Previous function body of an in-flight redefine() (the entry was on the VM call stack during the swap)	the running frame still executes the old opcodes; the swap keeps the body allocated instead of freeing memory under the VM; bounded to one body per such self-redefinition
Final static-defaults duplicate of a redefined entry whose previous body outlives it (a live closure template holds the last body share)	the shared body arrays are freed by their last holder through its struct pointers, which do not know the entry's own defaults duplicate; bounded to one table per redefined entry (each swap releases the previous duplicate eagerly)
Engine-chained arena blocks for >32 KiB parses	allocated by the engine; z-engine never frees memory it did not allocate
Refcount-0 persistent strings	never freed with the request allocator; bounded, reclaimed at process end
Engine-original interface/trait buffers replaced by z-engine (including the trait alias/precedence lists replaced by addTraitAlias()/addTraitPrecedence() and their remove* counterparts)	possibly shared or in opcache SHM, never freed by z-engine; at most one per touched class
Persistent interned strings (StringEntry::persistentInterned)	interned-style (immutable, non-refcounted) blocks referenced by persistent tables and object properties; bounded by the number of persisted keys/values, reclaimed at process end
Persistent hashtables and their engine-grown data blocks (PersistentHashTable)	registries that must outlive the request by design; the engine resizes their data with the persistent allocator, so only PersistentHashTable::destroy() may release them (see below)
The shared uninitialized_bucket sentinel block	one uint32_t[2] per process backing every uninitialized persistent table, mirroring the engine's static
Persistent object clones (PersistentObjectFactory::persistentClone)	refcount-pinned malloc objects designed to survive the request boundary; detached from the object store before teardown so no engine path ever frees them. Clones managed by the persistent heap are the exception: PersistentHeap::remove() releases them exactly once through the graph inventory (see below)
The zengine module entry and its zval-sized globals anchor (ZEngineModule)	the single framework-wide module: the anchor is the one address a later request can use to rediscover the heap registry; covered by the module-entry rows above, bounded to one module per process
The persistent-heap root registry table (PersistentHeap)	maps heap keys to graph descriptors for the lifetime of the heap; released only by PersistentHeap::destroy(). Stored graphs themselves are droppable via remove(), not immortal (docs/persistent-heap.md)
Arena-mimicking blocks of classes copied out of opcache shared memory (ReflectionClass::copyOutOfSharedMemory())	the same copy model as a specialized class (row below): the copy is request memory reclaimed by the allocator at request end, while the shared-memory original it replaces in the class table is never written and never freed. Bounded to one copy per mutated shared-memory class per request
Arena-mimicking blocks of specialized classes (ClassSpecializer, see docs/class-specialization.md)	the class entry struct, property-info/class-constant blocks, properties_info_table, iterator/arrayaccess caches, duplicated arg_info blocks and duplicated type lists are structures the engine never frees for userland classes (they live in the compiler arena for a compiled class); the specializer allocates them as request memory reclaimed by the allocator at request end — request-lifetime by design, bounded per specialized class

Everything else is a bug: the test suite runs with report_memleaks=1 on a debug build and fails on any leak report.

Dismantling persistent data (cross-request free)

"Immortal by design" is the default for persistent allocations, not a life sentence: a long-running process that keeps persisting new state needs a way to drop old state again. That path is deliberately narrow, because releasing persistent memory is the one operation the engine cannot help with.

Primitive	What it releases
Core::untrackAndFree($ptr)	a block allocated in this same request, through the owning CData in the tracked-block registry (right allocator guaranteed, no-op for engine-original buffers)
Core::persistentFree($ptr)	any malloc-backed block, through libc free(), with no bookkeeping at all
PersistentHashTable::destroy()	one persistent table: zend_hash_destroy() for the engine-grown data block, then the struct itself
PersistentHeap::remove($key) / PersistentHeap::destroy()	one stored object graph (or the whole heap): every cloned object, minted string and persistent table of the graph, exactly once, driven by the per-key inventory (docs/persistent-heap.md)

Core::persistentFree() exists because the tracked-block registry is a PHP static: it dies with the request that filled it. A block persisted by request N can only be dropped by request N+k, where the registry no longer knows it — hence the raw form. It is a plain free(3): pass only blocks that are provably malloc-backed (z-engine persistent FFI allocations, or engine structures allocated with pemalloc(..., 1)), never request memory, never an interned zend_string (the engine's interned-string table references it), and never a block that is still reachable from a live engine structure. The two primitives stay orthogonal — persistentFree() does not touch the registry, so a caller dropping a block allocated in the current request must Core::untrack() it as well, or a recycled address could later be freed a second time through untrackAndFree().

PersistentHashTable::destroy() composes both halves for a table and is safe on sealed (markImmutable()) tables: it re-baselines the immutable refcount so the engine's debug-build "nobody else holds this array" assertion in zend_hash_destroy() holds. Stored payloads are never touched (persistent tables carry a NULL pDestructor by construction), so nested tables and buffers must be released by their owner before the container goes away.