net/ffi/mod.rs
1//! C FFI bindings for cross-language integration.
2//!
3//! This module provides a C-compatible API for using Net from
4//! other languages (Python, Node.js, Go, etc.).
5//!
6//! # Safety
7//!
8//! All public FFI functions in this module accept raw pointers from C code.
9//! Each is declared `pub unsafe extern "C" fn` so the unsafety is
10//! explicit at the type level; the module-wide contract callers
11//! must uphold is:
12//! - Pointers are valid and properly aligned
13//! - Opaque handle pointers (`*mut T`) were produced by this crate's
14//! matching constructor (`Box::into_raw` inside the FFI surface).
15//! Foreign-allocated pointers, even if valid and aligned, will UB
16//! when consumed by `Box::from_raw` in the corresponding `_free`.
17//! - String pointers point to valid UTF-8 data
18//! - Buffer sizes are accurate
19//! - Handles are not used after `net_shutdown`
20//!
21//! The per-function `# Safety` rustdoc is intentionally suppressed
22//! at the module level — every entry point shares the same contract
23//! and the module doc-comment above (plus `include/README.md`) is
24//! the source of truth. Adding individual `# Safety` blocks would
25//! duplicate the same wording 200 times without adding signal.
26//!
27//! # Thread Safety
28//!
29//! All FFI functions are thread-safe. The event bus handle can be shared
30//! across threads.
31//!
32#![allow(clippy::missing_safety_doc)]
33// The cross-cutting C-side safety contract for every `unsafe` block in
34// this module is documented in the `# Safety` section above:
35// caller-validated pointer / length / lifetime / handle-not-after-shutdown
36// invariants documented in `include/net.h`. Inlining `// SAFETY:` on each
37// block would add ~200 identical "see module preamble" comments without
38// adding any signal beyond what the preamble already says.
39#![expect(
40 clippy::undocumented_unsafe_blocks,
41 reason = "module-wide FFI safety contract documented in the # Safety preamble above"
42)]
43#![expect(
44 clippy::multiple_unsafe_ops_per_block,
45 reason = "FFI entry points routinely deref + write to multiple out-parameter fields under the same caller contract; splitting per-op would obscure the single boundary-cross"
46)]
47
48//! # Tokio runtime restriction
49//!
50//! Internal FFI ops (`net_poll`, `net_flush`, `net_shutdown`,
51//! `net_redex_*`, `net_mesh_new`, the cortex FFI, the mesh FFI)
52//! drive the bus's tokio runtime via `Runtime::block_on`. That
53//! function panics with "Cannot start a runtime from within a
54//! runtime" if the calling thread is already inside a tokio
55//! runtime context. The functions are `extern "C"`, so a panic
56//! unwinds across the FFI boundary into C / Go-cgo / Python /
57//! NAPI — undefined behavior.
58//!
59//! **The common-case C / Go / Python caller has no Rust tokio
60//! runtime, so this is unreachable for them.** The narrow path is:
61//!
62//! - A **Rust** caller loads the cdylib and calls these
63//! functions from inside its own `#[tokio::main]` (or any
64//! thread that has called `Runtime::enter()`).
65//! - A non-Rust caller embeds a Rust library that runs its own
66//! tokio runtime and forwards calls into this cdylib on the
67//! same thread.
68//!
69//! Both forms are unusual but reachable. **Do not call any FFI
70//! op from a thread that already holds a tokio runtime
71//! context.** If you must, spawn the FFI call on a fresh OS
72//! thread that doesn't carry a runtime guard, or wrap the call
73//! with `tokio::task::spawn_blocking(|| net_xxx(...))` to escape
74//! the worker pool.
75//!
76//! `net_init` (`mod.rs:284-316`) hardens against this for runtime
77//! *construction*; the steady-state ops do not, since the cost
78//! of a `Handle::try_current()` check on every poll would be
79//! measurable for the common path that doesn't hit the bug.
80//!
81//! # `catch_unwind` + caller-held locks
82//!
83//! Several FFI entries (`net_blob_publish`, `net_blob_resolve`,
84//! `net_*_wait_for_token`) wrap their body in
85//! `std::panic::catch_unwind(AssertUnwindSafe(...))` so a panic
86//! during the call returns a typed `NET_ERR_BLOB_PANIC` /
87//! `NET_ERR_PANIC` code rather than unwinding across the FFI
88//! boundary. That stops the substrate-side undefined behavior,
89//! but it does NOT make the wrapped code transparently panic-safe
90//! from the caller's perspective.
91//!
92//! **If the caller invokes an FFI op while holding an OS-level
93//! lock, a `sync.Mutex` (Go), `threading.Lock` (Python), or any
94//! other mutex with poisoning semantics, and the FFI body panics,
95//! the mutex is left in a poisoned state.** Subsequent acquires
96//! on the same mutex by the caller observe the poisoning and
97//! either error (Rust `parking_lot` with `poison_on_unwind`) or
98//! deadlock (Go's `sync.Mutex` doesn't poison; the caller has
99//! observed a return value that may not reflect the state of
100//! the FFI op).
101//!
102//! Recommended caller pattern: **do not hold a caller-side lock
103//! across an FFI call**. Acquire the lock, prepare the inputs,
104//! release the lock, then call the FFI. Re-acquire if you need
105//! to update caller state with the result.
106//!
107//! The hazard is documented per-binding in:
108//! - Python: `bindings/python/README.md` (caller-mutex notes)
109//! - Node: `bindings/node/README.md`
110//! - Go: `bindings/go/net/redex.go` lifecycle docs
111//! - C: `include/net.h` (every wait-family declaration)
112//!
113//! # Memory Management
114//!
115//! - Handles returned by `net_init` must be freed with `net_shutdown`
116//! - String buffers passed to `net_poll` are owned by the caller
117//! - Error codes are returned as integers (0 = success, negative = error)
118//!
119//! # Example (C)
120//!
121//! ```c
122//! #include "net.h"
123//!
124//! int main() {
125//! // Initialize with default config
126//! void* bus = net_init("{\"num_shards\": 4}");
127//! if (!bus) return 1;
128//!
129//! // Ingest an event
130//! int result = net_ingest(bus, "{\"token\": \"hello\"}", 19);
131//! if (result < 0) { /* handle error */ }
132//!
133//! // Poll events
134//! char buffer[65536];
135//! result = net_poll(bus, "{\"limit\": 100}", buffer, sizeof(buffer));
136//!
137//! // Shutdown
138//! net_shutdown(bus);
139//! return 0;
140//! }
141//! ```
142
143// FFI functions accept raw pointers but are not marked `unsafe` to maintain
144// C ABI compatibility. Safety is documented in the module-level docs.
145#![allow(clippy::not_unsafe_ptr_arg_deref)]
146
147use std::ffi::CStr;
148use std::os::raw::{c_char, c_int};
149use std::ptr;
150
151use tokio::runtime::Runtime;
152
153use crate::bus::EventBus;
154use crate::config::EventBusConfig;
155use crate::consumer::ConsumeRequest;
156use crate::event::{Event, RawEvent};
157
158/// C FFI for CortEX / NetDb / RedexFile. Requires `netdb` (for the
159/// unified facade) and `redex-disk` (for persistent storage paths on
160/// `Redex` / `RedexFile`). Go / cgo consumers target this surface.
161///
162/// `missing_docs` is suppressed on this module: these are extern "C"
163/// shims over already-documented Rust adapters, and the per-function
164/// contract is documented in the binding-side READMEs (Go / TS / Py).
165/// Re-documenting each shim would duplicate with drift risk.
166/// Per-FFI-handle quiescing protocol shared by cortex / mesh
167/// handles to close the audit-#23/#24/#25 use-after-free hazards
168/// when a `_free` races a concurrent op. See module docs for the
169/// soundness story (intentional box leak) and the per-handle
170/// recipe.
171#[cfg(any(
172 all(feature = "netdb", feature = "redex-disk"),
173 feature = "net",
174 feature = "redis",
175))]
176pub mod handle_guard;
177
178#[cfg(all(feature = "netdb", feature = "redex-disk"))]
179#[allow(missing_docs)]
180pub mod cortex;
181
182/// C FFI for the Dataforts Phase 3 blob surface. Exposes the
183/// BlobRef wire codec, the global adapter registry, and the
184/// `publish_blob` / `resolve_payload` helpers for cgo / native
185/// consumers.
186#[cfg(feature = "dataforts")]
187#[allow(missing_docs)]
188pub mod blob;
189
190/// Stub definitions for the `net_mesh_blob_adapter_*` symbols
191/// when the `dataforts / netdb / redex-disk` feature triple is
192/// off. cgo / dlsym consumers link these symbols unconditionally
193/// (see `bindings/go/blob.go`), so a libnet built without the
194/// triple must still satisfy them — each stub returns
195/// `NET_ERR_FEATURE_NOT_BUILT` (or null) so Go programs route to
196/// a clean error rather than fail at program load. The module is
197/// empty when the feature triple is on (the real impls in
198/// `ffi::blob` cover the same symbol names).
199#[allow(missing_docs)]
200pub mod blob_stubs;
201
202/// C FFI for the encrypted-UDP mesh transport + channels. Requires
203/// the `net` feature (which brings in the crypto + transport). Go /
204/// cgo consumers target this surface alongside `ffi::cortex`. See
205/// the `ffi::cortex` note for why `missing_docs` is suppressed here.
206#[cfg(feature = "net")]
207#[allow(missing_docs)]
208pub mod mesh;
209
210/// C FFI for the `aggregator.registry` RPC client + channel
211/// visibility setter. Stage 5 of `SDK_AGGREGATOR_SUBNET_PLAN.md`.
212/// Rides the `net` feature alongside `ffi::mesh` because every
213/// op needs a `MeshNodeHandle`, and `cortex` because the
214/// underlying `behavior::aggregator` module's RPC surface is
215/// cortex-only (`mesh_rpc`, `cortex::rpc`, `postcard`).
216#[cfg(all(feature = "net", feature = "cortex"))]
217#[allow(missing_docs)]
218pub mod aggregator;
219
220/// C FFI for stateless predicate evaluation (Phase 9c of
221/// `CAPABILITY_SYSTEM_SDK_PLAN.md`). Pure helpers — no handles,
222/// no state. Mirrors the SDK-layer `evaluatePredicate` /
223/// `evaluate_predicate` surface every binding ships, exposed at
224/// the C ABI for raw consumers (C / C++ / Zig / Swift / etc.).
225#[cfg(feature = "net")]
226pub mod predicate;
227
228/// C FFI for stateless capability-set validation (Phase 9a of
229/// `CAPABILITY_SYSTEM_SDK_PLAN.md`). Pure helper — `caps_json`
230/// in, `report_json` out. Mirrors the SDK-layer
231/// `validate_capabilities` surface, exposed at the C ABI for raw
232/// consumers.
233#[cfg(feature = "net")]
234pub mod schema;
235
236/// C FFI for predicate debug-session helpers (Phase 9d of
237/// `CAPABILITY_SYSTEM_SDK_PLAN.md`). Pure helpers — single-eval
238/// `evaluate_with_trace`, corpus-wide
239/// `aggregate_debug_report`, and host-side
240/// `redact_metadata_keys`. Mirror what every other binding
241/// ships at the SDK layer; exposed at the C ABI for raw
242/// consumers.
243#[cfg(feature = "net")]
244pub mod predicate_debug;
245
246/// C FFI for the Redis Streams consumer-side dedup helper. Mirrors
247/// the Rust `net::adapter::RedisStreamDedup` surface for Go / C / Zig
248/// consumers. See `ffi::redis_dedup` module docs for the wire
249/// shape and the dedup contract.
250#[cfg(feature = "redis")]
251pub mod redis_dedup;
252
253#[cfg(feature = "net")]
254use crate::adapter::net::{NetAdapterConfig, ReliabilityConfig, StaticKeypair};
255#[cfg(any(feature = "redis", feature = "jetstream", feature = "net"))]
256use crate::config::AdapterConfig;
257#[cfg(feature = "jetstream")]
258use crate::config::JetStreamAdapterConfig;
259#[cfg(feature = "redis")]
260use crate::config::RedisAdapterConfig;
261#[cfg(feature = "net")]
262use std::ffi::CString;
263
264/// Opaque handle to an event bus instance.
265///
266/// This wraps the EventBus along with a Tokio runtime for async operations.
267///
268/// # Lifetime / soundness
269///
270/// The handle storage is *intentionally leaked* on `net_shutdown` rather
271/// than freed via `Box::from_raw`. Reasoning: every FFI entry point
272/// dereferences the C-side `*mut NetHandle` to access the atomics that
273/// gate shutdown. The previous Dekker-style SeqCst handshake between
274/// `FfiOpGuard::try_enter` (which calls `fetch_add` on `active_ops`) and
275/// `net_shutdown` (which loads `active_ops` then `Box::from_raw`s the
276/// handle) was unsound: SeqCst orders the atomic operations only — the
277/// non-atomic `Box::from_raw` could deallocate the storage between
278/// shutdown's load and a concurrent FFI op's `fetch_add`, producing a
279/// use-after-free on the freed atomic. By never freeing the box, the
280/// atomic memory backing the handle is always valid; concurrent FFI ops
281/// observe `shutting_down=true` after shutdown signals it and bail
282/// before touching `bus`/`runtime`.
283///
284/// `bus` and `runtime` are stored in `ManuallyDrop` so that
285/// `net_shutdown` can `take` them out (via `ptr::read`) in order to
286/// call `bus.shutdown().await`. Because `shutting_down` is set first
287/// and shutdown waits for `active_ops` to drop to zero before reading
288/// these fields, no FFI op can be racing the read. If the wait times
289/// out, the `ptr::read` is skipped and both fields are leaked along
290/// with the box.
291pub struct NetHandle {
292 /// Owned `EventBus`. Read out via `ManuallyDrop::take` during
293 /// shutdown once `active_ops` has drained to zero. After that
294 /// point, `shutting_down` is `true` and no FFI op may access this
295 /// field.
296 bus: std::mem::ManuallyDrop<EventBus>,
297 /// Owned tokio runtime. Same lifetime contract as `bus`.
298 runtime: std::mem::ManuallyDrop<Runtime>,
299 /// Set to `true` once `net_shutdown` begins. All other FFI
300 /// functions check this flag and return `ShuttingDown` before
301 /// touching `bus` / `runtime`.
302 shutting_down: std::sync::atomic::AtomicBool,
303 /// Number of in-flight FFI operations (excluding shutdown itself).
304 /// `net_shutdown` spins until this drops to zero (with a deadline)
305 /// before reading `bus` / `runtime` to call shutdown.
306 active_ops: std::sync::atomic::AtomicU32,
307 /// Set to `true` after `net_shutdown` has consumed `bus` /
308 /// `runtime` via `ManuallyDrop::take`. A second `net_shutdown`
309 /// call observes this and returns `Success` without re-taking
310 /// (which would be UB). FFI ops also check this before touching
311 /// `bus` / `runtime`, defending against a contract-violating
312 /// caller that races a post-shutdown call.
313 bus_taken: std::sync::atomic::AtomicBool,
314 /// Set to `true` after `bus.shutdown()` returns from the
315 /// first `net_shutdown` call. A second/third concurrent
316 /// `net_shutdown` caller spins until this flips before
317 /// returning success — without this gate the second caller
318 /// observed `bus_taken == true` and returned `Success` while
319 /// the first caller was still mid-`block_on(bus.shutdown())`,
320 /// falsely signaling completion of an in-progress shutdown.
321 shutdown_completed: std::sync::atomic::AtomicBool,
322}
323
324/// Maximum time `net_shutdown` will wait for in-flight FFI operations
325/// to complete before giving up. If the deadline expires, the bus is
326/// leaked rather than read out — leaking is correct (the box is
327/// already leaked permanently for soundness reasons) but means the
328/// adapter's `flush()` / `shutdown()` won't run.
329const FFI_SHUTDOWN_DEADLINE: std::time::Duration = std::time::Duration::from_secs(5);
330
331/// RAII guard that increments `active_ops` on creation and decrements on drop.
332struct FfiOpGuard<'a> {
333 handle: &'a NetHandle,
334}
335
336impl<'a> FfiOpGuard<'a> {
337 /// Try to enter an FFI operation. Returns `None` if the handle is
338 /// shutting down or if `bus` / `runtime` have already been taken.
339 ///
340 /// Soundness rests on the fact that the box backing `handle` is
341 /// never freed (see `NetHandle` doc). The `fetch_add` is therefore
342 /// always on valid memory regardless of whether shutdown is in
343 /// progress. The subsequent loads decide whether the op is allowed
344 /// to proceed; if shutdown was signaled or `bus_taken` flipped
345 /// before our increment was visible, we bail without touching
346 /// `bus` / `runtime`. The `bus_taken` check defends against a
347 /// contract-violating caller that races a post-shutdown call: even
348 /// if `shutting_down` was reset somehow, an op that would touch the
349 /// already-taken `ManuallyDrop` fields is rejected.
350 fn try_enter(handle: &'a NetHandle) -> Option<Self> {
351 handle
352 .active_ops
353 .fetch_add(1, std::sync::atomic::Ordering::SeqCst);
354 if handle
355 .shutting_down
356 .load(std::sync::atomic::Ordering::SeqCst)
357 || handle.bus_taken.load(std::sync::atomic::Ordering::SeqCst)
358 {
359 handle
360 .active_ops
361 .fetch_sub(1, std::sync::atomic::Ordering::AcqRel);
362 None
363 } else {
364 Some(Self { handle })
365 }
366 }
367}
368
369impl Drop for FfiOpGuard<'_> {
370 fn drop(&mut self) {
371 self.handle
372 .active_ops
373 .fetch_sub(1, std::sync::atomic::Ordering::AcqRel);
374 }
375}
376
377/// Returns `true` when `handle` is non-null and aligned for
378/// `NetHandle`. Every `extern "C"` entry point that derefs the
379/// raw handle must gate on this — a misaligned pointer produced
380/// by an over-eager `void *` cast in a foreign caller would be
381/// immediate UB on `&*handle`, even before the `is_null` check.
382#[inline]
383fn handle_is_valid(handle: *const NetHandle) -> bool {
384 !handle.is_null() && (handle as usize).is_multiple_of(std::mem::align_of::<NetHandle>())
385}
386
387/// Error codes returned by FFI functions.
388#[repr(C)]
389pub enum NetError {
390 /// Success (no error).
391 Success = 0,
392 /// Null pointer passed.
393 NullPointer = -1,
394 /// Invalid UTF-8 string.
395 InvalidUtf8 = -2,
396 /// Invalid JSON.
397 InvalidJson = -3,
398 /// Initialization failed.
399 InitFailed = -4,
400 /// Ingestion failed (backpressure).
401 IngestionFailed = -5,
402 /// Poll failed.
403 PollFailed = -6,
404 /// Buffer too small.
405 BufferTooSmall = -7,
406 /// Shutting down.
407 ShuttingDown = -8,
408 /// Integer overflow: result does not fit in `c_int`.
409 IntOverflow = -9,
410 /// Stream handle does not belong to the supplied node handle.
411 /// Previously the send-family FFIs accepted any (stream, node)
412 /// pair without verifying they were created from the same node,
413 /// allowing silent cross-session traffic.
414 MismatchedHandles = -10,
415 /// `CString::new` failure: the input bytes are valid UTF-8 by
416 /// Rust's `String` invariant but contain an interior NUL byte
417 /// — and the C ABI cannot represent that, since C strings are
418 /// NUL-terminated. Pre-fix this was reported as
419 /// `InvalidUtf8`, which was wrong: the input is UTF-8-valid;
420 /// it just has a NUL where C expects it not to. A binding
421 /// reading the typed error and seeing "invalid UTF-8" would
422 /// chase the wrong cause.
423 InteriorNul = -11,
424 /// Unknown error.
425 Unknown = -99,
426}
427
428impl From<NetError> for c_int {
429 fn from(e: NetError) -> Self {
430 e as c_int
431 }
432}
433
434/// Enter an FFI operation with lifetime protection. Returns an `FfiOpGuard`
435/// that prevents `net_shutdown` from deallocating the handle until the guard
436/// is dropped. Returns `Err` with the error code if shutdown is in progress.
437#[inline]
438fn enter_ffi_op(handle: &NetHandle) -> Result<FfiOpGuard<'_>, c_int> {
439 FfiOpGuard::try_enter(handle).ok_or(NetError::ShuttingDown.into())
440}
441
442/// Initialize a new event bus.
443///
444/// # Parameters
445///
446/// - `config_json`: JSON configuration string (UTF-8, null-terminated).
447/// Pass NULL or empty string for default configuration.
448///
449/// # Returns
450///
451/// Opaque handle to the event bus, or NULL on failure.
452/// The handle must be freed with `net_shutdown`.
453///
454/// # Example Configuration
455///
456/// ```json
457/// {
458/// "num_shards": 8,
459/// "ring_buffer_capacity": 1048576,
460/// "backpressure_mode": "DropOldest",
461/// "batch": {
462/// "min_size": 1000,
463/// "max_size": 10000,
464/// "max_delay_ms": 10
465/// }
466/// }
467/// ```
468#[unsafe(no_mangle)]
469pub unsafe extern "C" fn net_init(config_json: *const c_char) -> *mut NetHandle {
470 // Parse and validate the config BEFORE constructing the tokio
471 // runtime. Building the runtime first would let any subsequent
472 // early-return path (`CStr::to_str` Err, `parse_config_json`
473 // returning None, `EventBus::new` returning Err) drop the
474 // local `Runtime` on function return. Dropping a multi-thread
475 // tokio runtime from inside ANOTHER tokio runtime's worker
476 // thread panics with "Cannot drop a runtime in a context where
477 // blocking is not allowed", unwinding across this `extern "C"`
478 // boundary into a Python / Go-cgo / NAPI / PyO3 caller —
479 // undefined behaviour. By validating inputs first, the runtime
480 // is only built once we know it will be installed into the
481 // `NetHandle` and survive the call.
482 let config = if config_json.is_null() {
483 EventBusConfig::default()
484 } else {
485 let config_str = match unsafe { CStr::from_ptr(config_json) }.to_str() {
486 Ok("") => EventBusConfig::default(),
487 Ok(s) => match parse_config_json(s) {
488 Some(cfg) => cfg,
489 None => return ptr::null_mut(),
490 },
491 Err(_) => return ptr::null_mut(),
492 };
493 config_str
494 };
495
496 // Now construct the runtime — its lifetime is tied to the
497 // returned `NetHandle` (via `create_with_config`), so the only
498 // remaining drop is on `net_shutdown`, which already handles
499 // it via `runtime.block_on(...)` (see #74) outside any other
500 // tokio context.
501 let runtime = match Runtime::new() {
502 Ok(rt) => rt,
503 Err(_) => return ptr::null_mut(),
504 };
505
506 create_with_config(runtime, config)
507}
508
509/// Parse JSON configuration into EventBusConfig.
510///
511/// Supports:
512/// - `num_shards`: number of shards
513/// - `ring_buffer_capacity`: ring buffer size per shard
514/// - `backpressure_mode`: "DropNewest", "DropOldest", "FailProducer"
515fn parse_config_json(json_str: &str) -> Option<EventBusConfig> {
516 let value: serde_json::Value = serde_json::from_str(json_str).ok()?;
517
518 let mut builder = EventBusConfig::builder();
519
520 if let Some(num_shards) = value.get("num_shards").and_then(|v| v.as_u64()) {
521 let num_shards = u16::try_from(num_shards).ok()?;
522 builder = builder.num_shards(num_shards);
523 }
524
525 if let Some(capacity) = value.get("ring_buffer_capacity").and_then(|v| v.as_u64()) {
526 let capacity = usize::try_from(capacity).ok()?;
527 builder = builder.ring_buffer_capacity(capacity);
528 }
529
530 if let Some(bp_value) = value.get("backpressure_mode") {
531 let bp_mode = if let Some(mode) = bp_value.as_str() {
532 match mode {
533 "DropNewest" | "drop_newest" => crate::config::BackpressureMode::DropNewest,
534 "DropOldest" | "drop_oldest" => crate::config::BackpressureMode::DropOldest,
535 "FailProducer" | "fail_producer" => crate::config::BackpressureMode::FailProducer,
536 // Pre-fix every other string silently fell back to
537 // `DropNewest`. A typo (`"DropOldset"`) thus
538 // changed durability profile at deploy time with
539 // no error. Reject unknowns to match the contract
540 // already enforced by `parse_poll_request_json`.
541 _ => return None,
542 }
543 } else if let Some(obj) = bp_value.as_object() {
544 // Object form: `{"Sample": {"rate": N}}` for the
545 // sampling mode that has an associated value.
546 if let Some(sample) = obj.get("Sample").or_else(|| obj.get("sample")) {
547 let rate = sample.get("rate").and_then(|v| v.as_u64())?;
548 let rate = u32::try_from(rate).ok()?;
549 if rate == 0 {
550 // Validated again by `EventBusConfig::validate`,
551 // but reject earlier so the parser surface
552 // matches the validator surface.
553 return None;
554 }
555 crate::config::BackpressureMode::Sample { rate }
556 } else {
557 return None;
558 }
559 } else {
560 return None;
561 };
562 builder = builder.backpressure_mode(bp_mode);
563 }
564
565 // Parse Redis config
566 #[cfg(feature = "redis")]
567 if let Some(redis) = value.get("redis") {
568 if let Some(url) = redis.get("url").and_then(|v| v.as_str()) {
569 let mut redis_config = RedisAdapterConfig::new(url);
570
571 if let Some(prefix) = redis.get("prefix").and_then(|v| v.as_str()) {
572 redis_config = redis_config.with_prefix(prefix);
573 }
574 if let Some(max_len) = redis.get("max_stream_len").and_then(|v| v.as_u64()) {
575 let max_len = usize::try_from(max_len).ok()?;
576 redis_config = redis_config.with_max_stream_len(max_len);
577 }
578 if let Some(pipeline_size) = redis.get("pipeline_size").and_then(|v| v.as_u64()) {
579 let pipeline_size = usize::try_from(pipeline_size).ok()?;
580 redis_config = redis_config.with_pipeline_size(pipeline_size);
581 }
582
583 builder = builder.adapter(AdapterConfig::Redis(redis_config));
584 }
585 }
586
587 // Parse JetStream config
588 #[cfg(feature = "jetstream")]
589 if let Some(jetstream) = value.get("jetstream") {
590 if let Some(url) = jetstream.get("url").and_then(|v| v.as_str()) {
591 let mut js_config = JetStreamAdapterConfig::new(url);
592
593 if let Some(prefix) = jetstream.get("prefix").and_then(|v| v.as_str()) {
594 js_config = js_config.with_prefix(prefix);
595 }
596 if let Some(max_messages) = jetstream.get("max_messages").and_then(|v| v.as_i64()) {
597 js_config = js_config.with_max_messages(max_messages);
598 }
599 if let Some(replicas) = jetstream.get("replicas").and_then(|v| v.as_u64()) {
600 let replicas = usize::try_from(replicas).ok()?;
601 js_config = js_config.with_replicas(replicas);
602 }
603
604 builder = builder.adapter(AdapterConfig::JetStream(js_config));
605 }
606 }
607
608 // Parse Net config
609 #[cfg(feature = "net")]
610 if let Some(net) = value.get("net") {
611 let bind_addr: std::net::SocketAddr = net
612 .get("bind_addr")
613 .and_then(|v| v.as_str())
614 .and_then(|s| s.parse().ok())?;
615
616 let peer_addr: std::net::SocketAddr = net
617 .get("peer_addr")
618 .and_then(|v| v.as_str())
619 .and_then(|s| s.parse().ok())?;
620
621 let psk: [u8; 32] = net
622 .get("psk")
623 .and_then(|v| v.as_str())
624 .and_then(|s| hex::decode(s).ok())
625 .and_then(|v| v.try_into().ok())?;
626
627 let role = net
628 .get("role")
629 .and_then(|v| v.as_str())
630 .unwrap_or("initiator");
631
632 let mut net_config = match role {
633 "initiator" => {
634 let peer_pubkey: [u8; 32] = net
635 .get("peer_public_key")
636 .and_then(|v| v.as_str())
637 .and_then(|s| hex::decode(s).ok())
638 .and_then(|v| v.try_into().ok())?;
639 NetAdapterConfig::initiator(bind_addr, peer_addr, psk, peer_pubkey)
640 }
641 "responder" => {
642 let secret_key: [u8; 32] = net
643 .get("secret_key")
644 .and_then(|v| v.as_str())
645 .and_then(|s| hex::decode(s).ok())
646 .and_then(|v| v.try_into().ok())?;
647 let public_key: [u8; 32] = net
648 .get("public_key")
649 .and_then(|v| v.as_str())
650 .and_then(|s| hex::decode(s).ok())
651 .and_then(|v| v.try_into().ok())?;
652 let keypair = StaticKeypair::from_keys(secret_key, public_key);
653 NetAdapterConfig::responder(bind_addr, peer_addr, psk, keypair)
654 }
655 _ => return None,
656 };
657
658 // Apply optional settings
659 if let Some(reliability) = net.get("reliability").and_then(|v| v.as_str()) {
660 net_config = net_config.with_reliability(match reliability {
661 "light" => ReliabilityConfig::Light,
662 "full" => ReliabilityConfig::Full,
663 _ => ReliabilityConfig::None,
664 });
665 }
666
667 if let Some(pool_size) = net.get("packet_pool_size").and_then(|v| v.as_u64()) {
668 if let Ok(size) = usize::try_from(pool_size) {
669 net_config = net_config.with_pool_size(size);
670 }
671 }
672
673 // Reject `0` for `heartbeat_interval_ms` and
674 // `session_timeout_ms`. `EventBusConfig::validate` rejects
675 // zero `Duration`s for `cooldown`, `metrics_window`, etc.,
676 // but the Net adapter's JSON parser had no equivalent guard
677 // — a `0` here flowed through to `Duration::from_millis(0)`,
678 // which on the heartbeat path busy-loops the heartbeat task
679 // and saturates a CPU. Treat zero as a misconfig and refuse
680 // to build the bus, surfacing as `InvalidJson` so the FFI
681 // caller sees a typed failure rather than a hung daemon.
682 if let Some(interval_ms) = net.get("heartbeat_interval_ms").and_then(|v| v.as_u64()) {
683 if interval_ms == 0 {
684 return None;
685 }
686 net_config =
687 net_config.with_heartbeat_interval(std::time::Duration::from_millis(interval_ms));
688 }
689
690 if let Some(timeout_ms) = net.get("session_timeout_ms").and_then(|v| v.as_u64()) {
691 if timeout_ms == 0 {
692 return None;
693 }
694 net_config =
695 net_config.with_session_timeout(std::time::Duration::from_millis(timeout_ms));
696 }
697
698 if let Some(batched) = net.get("batched_io").and_then(|v| v.as_bool()) {
699 net_config = net_config.with_batched_io(batched);
700 }
701
702 builder = builder.adapter(AdapterConfig::Net(Box::new(net_config)));
703 }
704
705 builder.build().ok()
706}
707
708fn create_with_config(runtime: Runtime, config: EventBusConfig) -> *mut NetHandle {
709 let bus = match runtime.block_on(EventBus::new(config)) {
710 Ok(bus) => bus,
711 Err(_) => {
712 // Send the runtime off to a fresh OS thread for
713 // dropping. Dropping a multi-thread tokio `Runtime`
714 // from inside another tokio runtime's worker thread
715 // panics ("Cannot drop a runtime in a context where
716 // blocking is not allowed"); a panic here would unwind
717 // across this `extern "C"` frame. The fresh thread
718 // guarantees a non-tokio context, so the drop is sound
719 // regardless of the caller's runtime environment. We
720 // don't `join()` the thread — the drop completes on
721 // its own and the caller has already been told
722 // `net_init` failed (returning null).
723 std::thread::spawn(move || drop(runtime));
724 return ptr::null_mut();
725 }
726 };
727
728 let handle = Box::new(NetHandle {
729 bus: std::mem::ManuallyDrop::new(bus),
730 runtime: std::mem::ManuallyDrop::new(runtime),
731 shutting_down: std::sync::atomic::AtomicBool::new(false),
732 active_ops: std::sync::atomic::AtomicU32::new(0),
733 bus_taken: std::sync::atomic::AtomicBool::new(false),
734 shutdown_completed: std::sync::atomic::AtomicBool::new(false),
735 });
736
737 Box::into_raw(handle)
738}
739
740/// Ingest a single event.
741///
742/// # Parameters
743///
744/// - `handle`: Event bus handle from `net_init`.
745/// - `event_json`: JSON event string (UTF-8).
746/// - `len`: Length of the event string in bytes.
747///
748/// # Returns
749///
750/// - `0` on success
751/// - Negative error code on failure
752#[unsafe(no_mangle)]
753pub unsafe extern "C" fn net_ingest(
754 handle: *mut NetHandle,
755 event_json: *const c_char,
756 len: usize,
757) -> c_int {
758 if !handle_is_valid(handle) || event_json.is_null() {
759 return NetError::NullPointer.into();
760 }
761
762 let handle = unsafe { &*handle };
763 let _guard = match enter_ffi_op(handle) {
764 Ok(g) => g,
765 Err(err) => return err,
766 };
767
768 // `slice::from_raw_parts` requires `len <= isize::MAX`. A
769 // C caller passing a sign-extended `-1` (or any
770 // `len > isize::MAX as usize`) triggers immediate UB before
771 // any other validation runs. Reject such inputs explicitly
772 // — caller should never see this in practice; surfacing a
773 // typed error is safer than UB.
774 if len > isize::MAX as usize {
775 return NetError::InvalidJson.into();
776 }
777 // Parse event JSON
778 let json_bytes = unsafe { std::slice::from_raw_parts(event_json as *const u8, len) };
779 let json_str = match std::str::from_utf8(json_bytes) {
780 Ok(s) => s,
781 Err(_) => return NetError::InvalidUtf8.into(),
782 };
783
784 let event = match Event::from_str(json_str) {
785 Ok(e) => e,
786 Err(_) => return NetError::InvalidJson.into(),
787 };
788
789 // Ingest
790 match handle.bus.ingest(event) {
791 Ok(_) => NetError::Success.into(),
792 Err(_) => NetError::IngestionFailed.into(),
793 }
794}
795
796/// Ingest a raw JSON string (fastest path).
797///
798/// The JSON string is stored directly without parsing.
799/// This is the recommended method for high-throughput ingestion.
800///
801/// # Parameters
802///
803/// - `handle`: Event bus handle from `net_init`.
804/// - `json`: JSON string (UTF-8).
805/// - `len`: Length of the JSON string in bytes.
806///
807/// # Returns
808///
809/// - `0` on success
810/// - Negative error code on failure
811#[unsafe(no_mangle)]
812pub unsafe extern "C" fn net_ingest_raw(
813 handle: *mut NetHandle,
814 json: *const c_char,
815 len: usize,
816) -> c_int {
817 if !handle_is_valid(handle) || json.is_null() {
818 return NetError::NullPointer.into();
819 }
820
821 let handle = unsafe { &*handle };
822 let _guard = match enter_ffi_op(handle) {
823 Ok(g) => g,
824 Err(err) => return err,
825 };
826
827 // `slice::from_raw_parts` requires `len <= isize::MAX`.
828 if len > isize::MAX as usize {
829 return NetError::InvalidJson.into();
830 }
831 let json_bytes = unsafe { std::slice::from_raw_parts(json as *const u8, len) };
832 let json_str = match std::str::from_utf8(json_bytes) {
833 Ok(s) => s,
834 Err(_) => return NetError::InvalidUtf8.into(),
835 };
836
837 let raw = RawEvent::from_str(json_str);
838
839 match handle.bus.ingest_raw(raw) {
840 Ok(_) => NetError::Success.into(),
841 Err(_) => NetError::IngestionFailed.into(),
842 }
843}
844
845/// Ingest multiple raw JSON strings (fastest batch path).
846///
847/// # Parameters
848///
849/// - `handle`: Event bus handle.
850/// - `jsons`: Array of pointers to JSON strings.
851/// - `lens`: Array of lengths for each JSON string.
852/// - `count`: Number of events in the arrays.
853///
854/// # Returns
855///
856/// Number of successfully ingested events, or negative error code.
857#[unsafe(no_mangle)]
858pub unsafe extern "C" fn net_ingest_raw_batch(
859 handle: *mut NetHandle,
860 jsons: *const *const c_char,
861 lens: *const usize,
862 count: usize,
863) -> c_int {
864 if !handle_is_valid(handle) || jsons.is_null() || lens.is_null() {
865 return NetError::NullPointer.into();
866 }
867 if count == 0 {
868 return 0;
869 }
870
871 let handle = unsafe { &*handle };
872 let _guard = match enter_ffi_op(handle) {
873 Ok(g) => g,
874 Err(err) => return err,
875 };
876 let mut events = Vec::with_capacity(count);
877 // Track per-entry drops so the caller's accounting can
878 // reconcile the returned count against the input count.
879 // Pre-fix per-entry rejects (null pointer, oversized length,
880 // invalid UTF-8) were silently `continue`-d and the caller
881 // saw `count - drops` accepted events without any signal as
882 // to which input indices were dropped. A binding that
883 // attributed the drop to back-pressure and retried got the
884 // wrong indices and double-published the good ones.
885 //
886 // The C-API contract is "returns count of accepted events";
887 // expanding it to take an out-param of dropped indices is
888 // an API addition, not a fix-in-place. Emit `tracing::warn!`
889 // with the offending index AND reason so operators
890 // observing the bus can correlate drop counts to specific
891 // inputs without changing the C surface. For high-volume
892 // bindings this should still be sized at one log line per
893 // dropped entry; if that ever matters in practice the
894 // `*_ex` follow-up can return the indices structurally.
895 let mut dropped_null = 0usize;
896 let mut dropped_oversize = 0usize;
897 let mut dropped_invalid_utf8 = 0usize;
898
899 for i in 0..count {
900 let json_ptr = unsafe { *jsons.add(i) };
901 let len = unsafe { *lens.add(i) };
902
903 if json_ptr.is_null() {
904 tracing::warn!(
905 index = i,
906 "net_ingest_raw_batch: dropping entry with null pointer"
907 );
908 dropped_null += 1;
909 continue;
910 }
911
912 // `slice::from_raw_parts` requires `len <= isize::MAX`.
913 // Skip pathological per-entry lengths rather than UB.
914 if len > isize::MAX as usize {
915 tracing::warn!(
916 index = i,
917 len,
918 "net_ingest_raw_batch: dropping entry with len > isize::MAX"
919 );
920 dropped_oversize += 1;
921 continue;
922 }
923 let json_bytes = unsafe { std::slice::from_raw_parts(json_ptr as *const u8, len) };
924 match std::str::from_utf8(json_bytes) {
925 Ok(json_str) => events.push(RawEvent::from_str(json_str)),
926 Err(_) => {
927 tracing::warn!(
928 index = i,
929 "net_ingest_raw_batch: dropping entry with invalid UTF-8"
930 );
931 dropped_invalid_utf8 += 1;
932 }
933 }
934 }
935 let total_dropped = dropped_null + dropped_oversize + dropped_invalid_utf8;
936 if total_dropped > 0 {
937 // Aggregate summary for log-pipeline filters that fold
938 // per-index lines.
939 tracing::warn!(
940 input_count = count,
941 dropped_null,
942 dropped_oversize,
943 dropped_invalid_utf8,
944 "net_ingest_raw_batch: {} of {} entries dropped before ingest",
945 total_dropped,
946 count,
947 );
948 }
949
950 let count = handle.bus.ingest_raw_batch(events);
951 // Returning `c_int::MAX` on overflow would be ambiguous with a real
952 // `INT_MAX` ingest. Signal overflow explicitly so callers doing
953 // accounting in high-throughput paths do not silently miscount.
954 c_int::try_from(count).unwrap_or_else(|_| NetError::IntOverflow.into())
955}
956
957/// Ingest multiple events.
958///
959/// # Parameters
960///
961/// - `handle`: Event bus handle.
962/// - `events_json`: JSON array of events (UTF-8, null-terminated).
963///
964/// # Returns
965///
966/// Number of successfully ingested events, or negative error code.
967#[unsafe(no_mangle)]
968pub unsafe extern "C" fn net_ingest_batch(
969 handle: *mut NetHandle,
970 events_json: *const c_char,
971) -> c_int {
972 if !handle_is_valid(handle) || events_json.is_null() {
973 return NetError::NullPointer.into();
974 }
975
976 let handle = unsafe { &*handle };
977 let _guard = match enter_ffi_op(handle) {
978 Ok(g) => g,
979 Err(err) => return err,
980 };
981
982 let json_str = match unsafe { CStr::from_ptr(events_json) }.to_str() {
983 Ok(s) => s,
984 Err(_) => return NetError::InvalidUtf8.into(),
985 };
986
987 // Parse as JSON array
988 let array: Vec<serde_json::Value> = match serde_json::from_str(json_str) {
989 Ok(a) => a,
990 Err(_) => return NetError::InvalidJson.into(),
991 };
992
993 let events: Vec<Event> = array.into_iter().map(Event::new).collect();
994 let count = handle.bus.ingest_batch(events);
995
996 // Returning `c_int::MAX` on overflow would be ambiguous with a real
997 // `INT_MAX` ingest. Signal overflow explicitly — matches the
998 // `net_ingest_raw_batch` contract.
999 c_int::try_from(count).unwrap_or_else(|_| NetError::IntOverflow.into())
1000}
1001
1002/// Parse the JSON request body passed to `net_poll` into a
1003/// `ConsumeRequest`. Returns the negative `NetError` code on parse
1004/// failure so the caller can surface it back across FFI. Both `limit`
1005/// and `cursor` are optional, but if either key is present with the
1006/// wrong JSON type it is an explicit error — silently falling back to
1007/// the default would hide caller bugs (e.g. the Go binding that
1008/// previously serialized `cursor` but had it dropped server-side).
1009fn parse_poll_request_json(json_str: &str) -> Result<ConsumeRequest, c_int> {
1010 let value: serde_json::Value =
1011 serde_json::from_str(json_str).map_err(|_| c_int::from(NetError::InvalidJson))?;
1012
1013 let limit = match value.get("limit") {
1014 None | Some(serde_json::Value::Null) => 100usize,
1015 Some(v) => match v.as_u64() {
1016 // `as usize` would silently truncate on 32-bit targets for
1017 // values above `usize::MAX`. Reject such inputs explicitly
1018 // so a caller asking for e.g. 2^33 events on a wasm32
1019 // build gets `InvalidJson` instead of a tiny wrap-around.
1020 Some(n) => usize::try_from(n).map_err(|_| c_int::from(NetError::InvalidJson))?,
1021 None => return Err(NetError::InvalidJson.into()),
1022 },
1023 };
1024 let cursor = match value.get("cursor") {
1025 None | Some(serde_json::Value::Null) => None,
1026 Some(v) => match v.as_str() {
1027 Some(s) => Some(s.to_owned()),
1028 None => return Err(NetError::InvalidJson.into()),
1029 },
1030 };
1031 let mut req = ConsumeRequest::new(limit);
1032 req.from_id = cursor;
1033 Ok(req)
1034}
1035
1036/// Poll events from the bus.
1037///
1038/// # Parameters
1039///
1040/// - `handle`: Event bus handle.
1041/// - `request_json`: JSON request string (UTF-8, null-terminated).
1042/// Example: `{"limit": 100, "ordering": "InsertionTs"}`
1043/// - `out_buffer`: Output buffer for JSON response.
1044/// - `buffer_len`: Size of the output buffer.
1045///
1046/// # Returns
1047///
1048/// - Number of bytes written to buffer on success
1049/// - Negative error code on failure
1050#[unsafe(no_mangle)]
1051pub unsafe extern "C" fn net_poll(
1052 handle: *mut NetHandle,
1053 request_json: *const c_char,
1054 out_buffer: *mut c_char,
1055 buffer_len: usize,
1056) -> c_int {
1057 if !handle_is_valid(handle) || out_buffer.is_null() {
1058 return NetError::NullPointer.into();
1059 }
1060
1061 let handle = unsafe { &*handle };
1062 let _guard = match enter_ffi_op(handle) {
1063 Ok(g) => g,
1064 Err(err) => return err,
1065 };
1066
1067 // Parse request
1068 let request = if request_json.is_null() {
1069 ConsumeRequest::new(100)
1070 } else {
1071 let json_str = match unsafe { CStr::from_ptr(request_json) }.to_str() {
1072 Ok(s) => s,
1073 Err(_) => return NetError::InvalidUtf8.into(),
1074 };
1075 match parse_poll_request_json(json_str) {
1076 Ok(req) => req,
1077 Err(code) => return code,
1078 }
1079 };
1080
1081 // Reject buffers too small to even hold an empty-response
1082 // JSON envelope. This catches the degenerate "tiny buffer"
1083 // case before we hit the adapter — `BufferTooSmall` returned
1084 // here means "no work was done, caller's cursor is unchanged."
1085 // 256 bytes comfortably fits the empty-response JSON below
1086 // even with a long echoed `next_id` cursor.
1087 const MIN_RESPONSE_BUFFER: usize = 256;
1088 if buffer_len < MIN_RESPONSE_BUFFER {
1089 return NetError::BufferTooSmall.into();
1090 }
1091
1092 // Stash the cursor before moving `request` into `poll()` so
1093 // the post-poll fallback can echo it back to the caller. On
1094 // overflow we write a minimal "no events delivered, cursor
1095 // unchanged" response so the caller's next poll re-fetches
1096 // the same range — events are not lost on idempotent
1097 // adapters (Redis XRANGE, JetStream direct_get).
1098 let cursor_snapshot = request.from_id.clone();
1099
1100 // Poll
1101 let response = match handle.runtime.block_on(handle.bus.poll(request)) {
1102 Ok(r) => r,
1103 Err(_) => return NetError::PollFailed.into(),
1104 };
1105
1106 // Serialize response. Events that fail to parse are included as raw
1107 // strings so the caller can see all events and detect parse failures.
1108 let total_events = response.events.len();
1109 let mut parsed_events: Vec<serde_json::Value> = Vec::with_capacity(total_events);
1110 let mut parse_errors: usize = 0;
1111 for e in &response.events {
1112 match e.parse() {
1113 Ok(v) => parsed_events.push(v),
1114 Err(_) => {
1115 parse_errors += 1;
1116 // Include the raw bytes as a string so the caller doesn't silently lose events
1117 if let Ok(raw) = e.raw_str() {
1118 parsed_events.push(serde_json::Value::String(raw.to_string()));
1119 }
1120 }
1121 }
1122 }
1123 let response_json = match serde_json::to_string(&serde_json::json!({
1124 "events": parsed_events,
1125 "next_id": response.next_id,
1126 "has_more": response.has_more,
1127 "count": parsed_events.len(),
1128 "parse_errors": parse_errors,
1129 })) {
1130 Ok(s) => s,
1131 Err(_) => return NetError::Unknown.into(),
1132 };
1133
1134 // Buffer overflow: emit a minimal fallback response that echoes
1135 // the caller's original cursor as `next_id`. The caller's next
1136 // poll runs against the same range and re-delivers the events
1137 // (idempotent on Redis XRANGE / JetStream direct_get). Without
1138 // this, a caller that trusts `next_id` blindly would advance
1139 // past the unread batch.
1140 if response_json.len() + 1 > buffer_len {
1141 let fallback = serde_json::to_string(&serde_json::json!({
1142 "events": [],
1143 "next_id": cursor_snapshot,
1144 "has_more": true,
1145 "count": 0,
1146 "parse_errors": 0,
1147 "buffer_too_small": true,
1148 "events_dropped": total_events,
1149 }))
1150 .unwrap_or_else(|_| String::from(
1151 r#"{"events":[],"next_id":null,"has_more":true,"count":0,"parse_errors":0,"buffer_too_small":true}"#
1152 ));
1153 if fallback.len() < buffer_len {
1154 unsafe {
1155 ptr::copy_nonoverlapping(
1156 fallback.as_ptr() as *const c_char,
1157 out_buffer,
1158 fallback.len(),
1159 );
1160 *out_buffer.add(fallback.len()) = 0;
1161 }
1162 }
1163 return NetError::BufferTooSmall.into();
1164 }
1165
1166 // Copy to output buffer
1167 unsafe {
1168 ptr::copy_nonoverlapping(
1169 response_json.as_ptr() as *const c_char,
1170 out_buffer,
1171 response_json.len(),
1172 );
1173 *out_buffer.add(response_json.len()) = 0; // Null terminate
1174 }
1175
1176 // Data was already copied into the caller's buffer; a
1177 // `c_int` overflow here means the byte count exceeds c_int's
1178 // range, NOT that the buffer was too small. Returning
1179 // `BufferTooSmall` would tell the caller to "resize and retry"
1180 // when retrying can't fix the actual condition. `IntOverflow`
1181 // is the documented variant for this case.
1182 match c_int::try_from(response_json.len()) {
1183 Ok(n) => n,
1184 Err(_) => NetError::IntOverflow.into(),
1185 }
1186}
1187
1188/// Get event bus statistics.
1189///
1190/// # Parameters
1191///
1192/// - `handle`: Event bus handle.
1193/// - `out_buffer`: Output buffer for JSON statistics.
1194/// - `buffer_len`: Size of the output buffer.
1195///
1196/// # Returns
1197///
1198/// Number of bytes written, or negative error code.
1199#[unsafe(no_mangle)]
1200pub unsafe extern "C" fn net_stats(
1201 handle: *mut NetHandle,
1202 out_buffer: *mut c_char,
1203 buffer_len: usize,
1204) -> c_int {
1205 if !handle_is_valid(handle) || out_buffer.is_null() {
1206 return NetError::NullPointer.into();
1207 }
1208
1209 let handle = unsafe { &*handle };
1210 let _guard = match enter_ffi_op(handle) {
1211 Ok(g) => g,
1212 Err(err) => return err,
1213 };
1214 let stats = handle.bus.stats();
1215 let shard_stats = handle.bus.shard_stats();
1216
1217 let stats_json = match serde_json::to_string(&serde_json::json!({
1218 "events_ingested": stats.events_ingested.load(std::sync::atomic::Ordering::Relaxed),
1219 "events_dropped": stats.events_dropped.load(std::sync::atomic::Ordering::Relaxed),
1220 "batches_dispatched": stats.batches_dispatched.load(std::sync::atomic::Ordering::Relaxed),
1221 "shard_events_ingested": shard_stats.events_ingested,
1222 "shard_events_dropped": shard_stats.events_dropped,
1223 "shard_batches_dispatched": shard_stats.batches_dispatched,
1224 })) {
1225 Ok(s) => s,
1226 Err(_) => return NetError::Unknown.into(),
1227 };
1228
1229 if stats_json.len() + 1 > buffer_len {
1230 return NetError::BufferTooSmall.into();
1231 }
1232
1233 unsafe {
1234 ptr::copy_nonoverlapping(
1235 stats_json.as_ptr() as *const c_char,
1236 out_buffer,
1237 stats_json.len(),
1238 );
1239 *out_buffer.add(stats_json.len()) = 0;
1240 }
1241
1242 // See net_poll above — the data was already copied, so an
1243 // overflowing length is `IntOverflow`, not `BufferTooSmall`.
1244 match c_int::try_from(stats_json.len()) {
1245 Ok(n) => n,
1246 Err(_) => NetError::IntOverflow.into(),
1247 }
1248}
1249
1250/// Flush all pending batches to the adapter.
1251///
1252/// # Parameters
1253///
1254/// - `handle`: Event bus handle.
1255///
1256/// # Returns
1257///
1258/// - `0` on success
1259/// - Negative error code on failure
1260#[unsafe(no_mangle)]
1261pub unsafe extern "C" fn net_flush(handle: *mut NetHandle) -> c_int {
1262 if !handle_is_valid(handle) {
1263 return NetError::NullPointer.into();
1264 }
1265
1266 let handle = unsafe { &*handle };
1267 let _guard = match enter_ffi_op(handle) {
1268 Ok(g) => g,
1269 Err(err) => return err,
1270 };
1271
1272 match handle.runtime.block_on(handle.bus.flush()) {
1273 Ok(_) => NetError::Success.into(),
1274 Err(_) => NetError::Unknown.into(),
1275 }
1276}
1277
1278/// Shut down the event bus and free resources.
1279///
1280/// # Parameters
1281///
1282/// - `handle`: Event bus handle. After this call, the handle is invalid.
1283///
1284/// # Returns
1285///
1286/// - `0` on success
1287/// - Negative error code on failure (including `Unknown` if the
1288/// bounded wait for in-flight FFI operations expired before the bus
1289/// could be shut down cleanly)
1290///
1291/// # Notes
1292///
1293/// The handle's storage is intentionally leaked: the box is never
1294/// returned to the allocator. See `NetHandle`'s docs for why. This is
1295/// a one-time cost per shutdown — typically per-process, since most C
1296/// callers initialize the bus once and shut down once.
1297#[unsafe(no_mangle)]
1298pub unsafe extern "C" fn net_shutdown(handle: *mut NetHandle) -> c_int {
1299 if !handle_is_valid(handle) {
1300 return NetError::NullPointer.into();
1301 }
1302
1303 // Scope the `&NetHandle` borrow into an inner block so it is
1304 // verifiably out of scope before the
1305 // `ManuallyDrop::take(&mut (*handle).bus)` calls below.
1306 // Holding an `&NetHandle` in scope for the whole function
1307 // while taking a raw `&mut (*handle).bus` later would rely on
1308 // NLL ending the immutable borrow before the mutable take —
1309 // a pattern fragile under stacked/tree borrow models. The
1310 // block-scoped borrow makes the lifetime constraint explicit
1311 // and obvious to both the compiler and any future maintainer.
1312 let drained_and_taken = {
1313 // SAFETY: The C contract guarantees `handle` is valid here and that
1314 // `net_shutdown` is not called concurrently with itself. Future
1315 // dereferences of the box from concurrent FFI ops on other threads
1316 // are also sound because we never free the box (see below).
1317 let handle_ref = unsafe { &*handle };
1318
1319 // Signal shutdown so concurrent FFI calls bail before touching
1320 // `bus`/`runtime`. SeqCst pairs with `FfiOpGuard::try_enter`.
1321 handle_ref
1322 .shutting_down
1323 .store(true, std::sync::atomic::Ordering::SeqCst);
1324
1325 // Bounded wait for in-flight ops to drain. Without a deadline, a
1326 // hung concurrent operation (e.g. `net_flush` against a stalled
1327 // adapter) would pin a CPU at 100% inside this loop forever.
1328 //
1329 // `std::hint::spin_loop()` is a CPU pause hint, not a yield. On
1330 // a single-threaded executor (or any configuration where the FFI
1331 // caller's thread is the same one that needs to make progress on
1332 // the in-flight async work) the tight spin starves the very tokio
1333 // worker we're waiting for, *causing* the deadline to expire when
1334 // it otherwise wouldn't. `thread::yield_now` lets the OS schedule
1335 // whatever's blocked, and a 1ms `thread::sleep` between yields
1336 // prevents the loop from saturating a CPU on platforms where
1337 // `yield_now` is a
1338 // near-no-op under low contention. The drain we expect to take
1339 // milliseconds, so a millisecond-granularity poll is fine.
1340 let deadline = std::time::Instant::now() + FFI_SHUTDOWN_DEADLINE;
1341 let mut drained = false;
1342 loop {
1343 if handle_ref
1344 .active_ops
1345 .load(std::sync::atomic::Ordering::SeqCst)
1346 == 0
1347 {
1348 drained = true;
1349 break;
1350 }
1351 if std::time::Instant::now() >= deadline {
1352 break;
1353 }
1354 std::thread::yield_now();
1355 std::thread::sleep(std::time::Duration::from_millis(1));
1356 }
1357
1358 if !drained {
1359 // In-flight ops may still be reading `bus`/`runtime`; reading
1360 // them out via `ManuallyDrop::take` would race those readers.
1361 // Leak both fields along with the box. Future ops still see
1362 // `shutting_down=true` and bail before touching either field,
1363 // so the leaked memory is never read again.
1364 return NetError::Unknown.into();
1365 }
1366
1367 // Idempotent shutdown: if a previous `net_shutdown` already
1368 // moved out the bus/runtime, do not call `ManuallyDrop::take`
1369 // a second time (that would be UB). The first call may still
1370 // be inside `runtime.block_on(bus.shutdown())` though — pre-
1371 // fix the second caller observed `bus_taken == true` and
1372 // returned `Success` immediately, falsely signaling
1373 // completion of an in-progress shutdown. Spin on
1374 // `shutdown_completed` (set by the first caller AFTER
1375 // `bus.shutdown()` returns) so subsequent callers wait for
1376 // the actual completion.
1377 if handle_ref
1378 .bus_taken
1379 .swap(true, std::sync::atomic::Ordering::SeqCst)
1380 {
1381 // Wait for the first caller to actually finish.
1382 // Bounded by the same FFI_SHUTDOWN_DEADLINE as the
1383 // `active_ops` drain — if the first caller is wedged
1384 // longer than that, we surface a Transient error rather
1385 // than block forever.
1386 let inner_deadline = std::time::Instant::now() + FFI_SHUTDOWN_DEADLINE;
1387 while !handle_ref
1388 .shutdown_completed
1389 .load(std::sync::atomic::Ordering::Acquire)
1390 {
1391 if std::time::Instant::now() >= inner_deadline {
1392 return NetError::Unknown.into();
1393 }
1394 std::thread::yield_now();
1395 std::thread::sleep(std::time::Duration::from_millis(1));
1396 }
1397 return NetError::Success.into();
1398 }
1399 drained
1400 };
1401 let _ = drained_and_taken;
1402
1403 // SAFETY: `active_ops` reached zero with `shutting_down=true`, so:
1404 // - Every FFI op that started before shutdown has fully
1405 // completed (decremented `active_ops` on guard drop).
1406 // - Any future FFI op will observe `shutting_down=true` and
1407 // bail in `try_enter` before touching `bus` / `runtime`.
1408 // Plus, `bus_taken` was just CAS'd from false → true, so no other
1409 // shutdown is concurrently moving the same fields out. The
1410 // immutable `handle_ref` borrow above has been dropped (block
1411 // scope ended), so the `&mut`-via-raw-pointer below is the
1412 // only live access — no stacked/tree-borrow race.
1413 //
1414 // We deliberately do NOT call `Box::from_raw` here. The box's
1415 // `shutting_down` / `active_ops` / `bus_taken` atomics must remain
1416 // valid memory because future FFI ops still dereference the
1417 // C-side pointer to check them. Leaking the box is the
1418 // correctness fix for the previous use-after-free; the per-handle
1419 // storage cost is a one-time overhead.
1420 let bus = unsafe { std::mem::ManuallyDrop::take(&mut (*handle).bus) };
1421 let runtime = unsafe { std::mem::ManuallyDrop::take(&mut (*handle).runtime) };
1422
1423 // Flush pending batches and gracefully shut down the adapter
1424 // before dropping the runtime. Without this, pending events in
1425 // ring buffers and batch workers would be silently lost.
1426 let result = runtime.block_on(bus.shutdown());
1427
1428 // `bus` and `runtime` go out of scope here and are dropped.
1429 // The leaked box keeps the atomics alive for any straggler ops.
1430
1431 // Signal completion to any second/third caller spinning on
1432 // `shutdown_completed` in the idempotent path above. Done
1433 // AFTER `bus.shutdown()` returns and AFTER the bus / runtime
1434 // drop, so subsequent callers can rely on this flag as a
1435 // hard "shutdown is fully done" barrier.
1436 unsafe { &*handle }
1437 .shutdown_completed
1438 .store(true, std::sync::atomic::Ordering::Release);
1439
1440 match result {
1441 Ok(()) => NetError::Success.into(),
1442 Err(_) => NetError::Unknown.into(),
1443 }
1444}
1445
1446/// Get the number of shards.
1447///
1448/// # Parameters
1449///
1450/// - `handle`: Event bus handle.
1451///
1452/// # Returns
1453///
1454/// Number of shards, or 0 if handle is null.
1455#[unsafe(no_mangle)]
1456pub unsafe extern "C" fn net_num_shards(handle: *mut NetHandle) -> u16 {
1457 if !handle_is_valid(handle) {
1458 return 0;
1459 }
1460 let handle = unsafe { &*handle };
1461 let _guard = match enter_ffi_op(handle) {
1462 Ok(g) => g,
1463 Err(_) => return 0,
1464 };
1465 handle.bus.num_shards()
1466}
1467
1468/// Get the library version.
1469///
1470/// # Returns
1471///
1472/// Version string (static, do not free).
1473#[unsafe(no_mangle)]
1474pub unsafe extern "C" fn net_version() -> *const c_char {
1475 static VERSION: &[u8] = b"0.8.0\0";
1476 VERSION.as_ptr() as *const c_char
1477}
1478
1479/// Generate a new Net keypair.
1480///
1481/// # Returns
1482///
1483/// JSON string with hex-encoded public_key and secret_key.
1484/// The caller must free the returned string with `net_free_string`.
1485/// Returns NULL if Net feature is not enabled.
1486#[cfg(feature = "net")]
1487#[unsafe(no_mangle)]
1488pub unsafe extern "C" fn net_generate_keypair() -> *mut c_char {
1489 let keypair = StaticKeypair::generate();
1490 let json = serde_json::json!({
1491 "public_key": hex::encode(keypair.public_key()),
1492 "secret_key": hex::encode(keypair.secret_key()),
1493 });
1494
1495 match CString::new(json.to_string()) {
1496 Ok(s) => s.into_raw(),
1497 Err(_) => ptr::null_mut(),
1498 }
1499}
1500
1501/// Free a string returned by Net functions.
1502///
1503/// # Parameters
1504///
1505/// - `s`: String pointer returned by `net_generate_keypair` or similar.
1506#[cfg(feature = "net")]
1507#[unsafe(no_mangle)]
1508pub unsafe extern "C" fn net_free_string(s: *mut c_char) {
1509 if !s.is_null() {
1510 unsafe {
1511 drop(CString::from_raw(s));
1512 }
1513 }
1514}
1515
1516// `net.h` declares both `net_generate_keypair` and
1517// `net_free_string` unconditionally — a consumer linking against
1518// a cdylib built without the `net` feature would otherwise hit
1519// a load-time missing-symbol error despite the header advertising
1520// the symbol. Provide always-empty stubs so the symbol is
1521// resolvable on every build configuration. Mirrors the
1522// `nat-traversal` cfg pattern in `mesh.rs`.
1523
1524/// Stub for builds without the `net` feature.
1525///
1526/// `net.h` declares `net_generate_keypair` unconditionally, so
1527/// the symbol must be resolvable on every build configuration.
1528/// Returns NULL since keypair generation requires the net feature.
1529#[cfg(not(feature = "net"))]
1530#[unsafe(no_mangle)]
1531pub unsafe extern "C" fn net_generate_keypair() -> *mut c_char {
1532 ptr::null_mut()
1533}
1534
1535/// Stub for builds without the `net` feature.
1536///
1537/// Mirrors the always-on signature in `net.h`. Reclaims a
1538/// CString-allocated pointer if non-null.
1539#[cfg(not(feature = "net"))]
1540#[unsafe(no_mangle)]
1541pub unsafe extern "C" fn net_free_string(s: *mut c_char) {
1542 if !s.is_null() {
1543 unsafe {
1544 drop(std::ffi::CString::from_raw(s));
1545 }
1546 }
1547}
1548
1549// =========================================================================
1550// Structured (non-JSON) API — _ex variants
1551// =========================================================================
1552
1553/// Ingestion receipt for C consumers.
1554#[repr(C)]
1555pub struct NetReceipt {
1556 /// Shard the event was assigned to.
1557 pub shard_id: u16,
1558 /// Insertion timestamp (nanoseconds).
1559 pub timestamp: u64,
1560}
1561
1562// Pin layout invariants for `NetReceipt`. `#[repr(C)]` already
1563// gives C ABI compatibility per platform, but doesn't catch a
1564// future field-reorder or field-add — both would silently break
1565// any C/Go/Python binding that hard-codes the struct layout.
1566// Static asserts on 64-bit targets (the production deployment
1567// shape) trip CI before such a change reaches a binary release.
1568//
1569// 64-bit: `u16 (2) + 6 pad + u64 (8)` = 16 bytes, alignment 8.
1570#[cfg(target_pointer_width = "64")]
1571const _: () = assert!(
1572 std::mem::size_of::<NetReceipt>() == 16,
1573 "NetReceipt size changed on 64-bit; bindings hard-code 16. \
1574 If the change is intentional, bump the binding versions and \
1575 update this assertion."
1576);
1577#[cfg(target_pointer_width = "64")]
1578const _: () = assert!(
1579 std::mem::align_of::<NetReceipt>() == 8,
1580 "NetReceipt alignment changed on 64-bit; bindings expect 8."
1581);
1582
1583/// A single stored event for C consumers.
1584///
1585/// # Safety contract for callers
1586///
1587/// `id`/`id_len` and `raw`/`raw_len` are produced by Rust as a
1588/// `Box<[u8]>` whose fat-pointer length is reconstructed at free
1589/// time from `id_len` / `raw_len`. The fields are `pub` because
1590/// `#[repr(C)]` exposes them to C, **but they must be treated as
1591/// read-only** between the `net_poll_*` call that produced them
1592/// and the `net_free_poll_result` that consumes them.
1593///
1594/// Mutating `id_len` or `raw_len` (or copying the struct, replacing
1595/// the pointer, and then freeing) causes
1596/// `Box::from_raw(slice_from_raw_parts_mut(ptr, wrong_len))` to be
1597/// undefined behavior on free — the allocator records the
1598/// allocation size and any mismatch is UB.
1599#[repr(C)]
1600pub struct NetEvent {
1601 /// Event ID (not null-terminated, use `id_len`).
1602 /// Read-only after `net_poll_*`; do not mutate.
1603 pub id: *const c_char,
1604 /// Length of the event ID. Read-only after `net_poll_*`; do not
1605 /// mutate (mutation causes UB on free).
1606 pub id_len: usize,
1607 /// Raw JSON payload (not null-terminated, use `raw_len`).
1608 /// Read-only after `net_poll_*`; do not mutate.
1609 pub raw: *const c_char,
1610 /// Length of the raw JSON payload. Read-only after
1611 /// `net_poll_*`; do not mutate (mutation causes UB on free).
1612 pub raw_len: usize,
1613 /// Insertion timestamp (nanoseconds).
1614 pub insertion_ts: u64,
1615 /// Shard ID.
1616 pub shard_id: u16,
1617}
1618
1619// Pin layout invariants for `NetEvent`. See `NetReceipt`'s
1620// asserts for rationale. Bindings (C, Go, Python, Node) hard-
1621// code 48 bytes on 64-bit; an accidental reorder or new field
1622// would silently shift every offset.
1623//
1624// 64-bit: `4 × 8 (ptrs/usize) + u64 (8) + u16 (2) + 6 trail` = 48.
1625#[cfg(target_pointer_width = "64")]
1626const _: () = assert!(
1627 std::mem::size_of::<NetEvent>() == 48,
1628 "NetEvent size changed on 64-bit; bindings hard-code 48. \
1629 If the change is intentional, bump the binding versions and \
1630 update this assertion."
1631);
1632#[cfg(target_pointer_width = "64")]
1633const _: () = assert!(
1634 std::mem::align_of::<NetEvent>() == 8,
1635 "NetEvent alignment changed on 64-bit; bindings expect 8."
1636);
1637
1638/// Poll result for C consumers.
1639#[repr(C)]
1640pub struct NetPollResult {
1641 /// Array of events. Free with `net_free_poll_result`.
1642 pub events: *mut NetEvent,
1643 /// Number of events in the array.
1644 pub count: usize,
1645 /// Cursor for the next poll (null-terminated). NULL if no more.
1646 pub next_id: *mut c_char,
1647 /// 1 if more events are available, 0 otherwise.
1648 pub has_more: c_int,
1649}
1650
1651/// Stats for C consumers.
1652#[repr(C)]
1653pub struct NetStats {
1654 /// Total events ingested.
1655 pub events_ingested: u64,
1656 /// Events dropped due to backpressure.
1657 pub events_dropped: u64,
1658 /// Batches dispatched to the adapter.
1659 pub batches_dispatched: u64,
1660}
1661
1662/// Ingest raw JSON with structured receipt.
1663#[unsafe(no_mangle)]
1664pub unsafe extern "C" fn net_ingest_raw_ex(
1665 handle: *mut NetHandle,
1666 json: *const c_char,
1667 len: usize,
1668 out: *mut NetReceipt,
1669) -> c_int {
1670 if !handle_is_valid(handle) || json.is_null() {
1671 return NetError::NullPointer.into();
1672 }
1673
1674 let handle = unsafe { &*handle };
1675 let _guard = match enter_ffi_op(handle) {
1676 Ok(g) => g,
1677 Err(err) => return err,
1678 };
1679
1680 // `slice::from_raw_parts` requires `len <= isize::MAX`.
1681 if len > isize::MAX as usize {
1682 return NetError::InvalidJson.into();
1683 }
1684 let json_bytes = unsafe { std::slice::from_raw_parts(json as *const u8, len) };
1685 let json_str = match std::str::from_utf8(json_bytes) {
1686 Ok(s) => s,
1687 Err(_) => return NetError::InvalidUtf8.into(),
1688 };
1689
1690 let raw = RawEvent::from_str(json_str);
1691
1692 match handle.bus.ingest_raw(raw) {
1693 Ok((shard_id, timestamp)) => {
1694 if !out.is_null() {
1695 unsafe {
1696 (*out).shard_id = shard_id;
1697 (*out).timestamp = timestamp;
1698 }
1699 }
1700 NetError::Success.into()
1701 }
1702 Err(_) => NetError::IngestionFailed.into(),
1703 }
1704}
1705
1706/// Poll events with structured result (no JSON overhead).
1707///
1708/// The caller must free the result with `net_free_poll_result`.
1709#[unsafe(no_mangle)]
1710pub unsafe extern "C" fn net_poll_ex(
1711 handle: *mut NetHandle,
1712 limit: usize,
1713 cursor: *const c_char,
1714 out: *mut NetPollResult,
1715) -> c_int {
1716 if !handle_is_valid(handle) || out.is_null() {
1717 return NetError::NullPointer.into();
1718 }
1719
1720 // Pre-validate `limit` BEFORE calling `bus.poll` — the bus
1721 // advances the consumer cursor before returning, so any
1722 // post-poll allocation failure (e.g. `Layout::array::<NetEvent>`
1723 // overflow on a pathological `count`, or `std::alloc::alloc`
1724 // returning null under memory pressure) would drop the response
1725 // and lose every event the cursor just stepped past. Reject
1726 // requests whose `count * size_of::<NetEvent>` would overflow
1727 // `isize::MAX` (the `Layout::array` cap) up front, so the
1728 // failure happens before the cursor moves.
1729 if limit > 0
1730 && (std::mem::size_of::<NetEvent>())
1731 .checked_mul(limit)
1732 .is_none_or(|v| v > isize::MAX as usize)
1733 {
1734 return NetError::IntOverflow.into();
1735 }
1736
1737 let handle = unsafe { &*handle };
1738 let _guard = match enter_ffi_op(handle) {
1739 Ok(g) => g,
1740 Err(err) => return err,
1741 };
1742
1743 let mut request = ConsumeRequest::new(limit);
1744 if !cursor.is_null() {
1745 if let Ok(s) = unsafe { CStr::from_ptr(cursor) }.to_str() {
1746 if !s.is_empty() {
1747 request = request.from(s);
1748 }
1749 }
1750 }
1751
1752 let response = match handle.runtime.block_on(handle.bus.poll(request)) {
1753 Ok(r) => r,
1754 Err(_) => return NetError::PollFailed.into(),
1755 };
1756
1757 let count = response.events.len();
1758
1759 // Allocate events array.
1760 //
1761 // Each iteration allocates two boxed byte slices via
1762 // `Vec::to_vec().into_boxed_slice()`, which panic on OOM in
1763 // the global allocator. A panic across this `extern "C"`
1764 // body is UB — under the cgo/N-API/cffi unwind model the
1765 // panic propagates into a frame that doesn't expect it. Wrap
1766 // the per-event build in `catch_unwind`, track how many
1767 // events we've fully written, and on panic / mid-loop
1768 // failure free the partial array via `free_events_array`
1769 // so neither UB nor the partial allocations leak.
1770 let events_ptr = if count > 0 {
1771 let layout = match std::alloc::Layout::array::<NetEvent>(count) {
1772 Ok(l) => l,
1773 Err(_) => return NetError::Unknown.into(),
1774 };
1775 let ptr = unsafe { std::alloc::alloc(layout) as *mut NetEvent };
1776 if ptr.is_null() {
1777 return NetError::Unknown.into();
1778 }
1779
1780 // Shared counter so the outer scope can clean up partial
1781 // writes if any iteration panics.
1782 let completed = std::cell::Cell::new(0usize);
1783 let build_result = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
1784 for (i, event) in response.events.iter().enumerate() {
1785 let id_bytes = event.id.as_bytes().to_vec().into_boxed_slice();
1786 let id_len = id_bytes.len();
1787 let id_ptr = Box::into_raw(id_bytes) as *const c_char;
1788
1789 let raw_bytes = event.raw.to_vec().into_boxed_slice();
1790 let raw_len = raw_bytes.len();
1791 let raw_ptr = Box::into_raw(raw_bytes) as *const c_char;
1792
1793 unsafe {
1794 ptr.add(i).write(NetEvent {
1795 id: id_ptr,
1796 id_len,
1797 raw: raw_ptr,
1798 raw_len,
1799 insertion_ts: event.insertion_ts,
1800 shard_id: event.shard_id,
1801 });
1802 }
1803 completed.set(i + 1);
1804 }
1805 }));
1806 if build_result.is_err() {
1807 // A panic landed mid-loop. Free fully-written events
1808 // (those past `completed.get()` were never written, so
1809 // the inner `id`/`raw` pointers aren't valid). The
1810 // events array was allocated for `count` NetEvent
1811 // slots, so the dealloc must use that same layout.
1812 free_events_array_partial(ptr, completed.get(), count);
1813 return NetError::Unknown.into();
1814 }
1815 ptr
1816 } else {
1817 ptr::null_mut()
1818 };
1819
1820 // Leak next_id if present.
1821 let next_id_ptr = match response.next_id {
1822 Some(ref s) => match std::ffi::CString::new(s.as_str()) {
1823 Ok(c) => c.into_raw(),
1824 Err(_) => {
1825 // Free already-allocated events before returning
1826 // error. `s.as_str()` is valid UTF-8 by `String`
1827 // invariant, so this is the interior-NUL path —
1828 // an upstream cursor id that contains `\0` cannot
1829 // round-trip through a C string. Pre-fix this
1830 // returned `InvalidUtf8`, which mis-described
1831 // the cause; bindings now see the more accurate
1832 // `InteriorNul`.
1833 free_events_array(events_ptr, count);
1834 return NetError::InteriorNul.into();
1835 }
1836 },
1837 None => ptr::null_mut(),
1838 };
1839
1840 unsafe {
1841 (*out).events = events_ptr;
1842 (*out).count = count;
1843 (*out).next_id = next_id_ptr;
1844 (*out).has_more = if response.has_more { 1 } else { 0 };
1845 }
1846
1847 NetError::Success.into()
1848}
1849
1850/// Free an events array and all its id/raw allocations.
1851///
1852/// `count` is the number of fully-written events (those whose
1853/// inner `id` / `raw` boxed slices were initialized). It must
1854/// also match the `Layout::array::<NetEvent>` used at allocation
1855/// time — every existing caller writes exactly `count` events
1856/// before invoking this function. For partial-cleanup paths
1857/// (e.g. panic mid-build), use [`free_events_array_partial`].
1858fn free_events_array(events: *mut NetEvent, count: usize) {
1859 free_events_array_partial(events, count, count);
1860}
1861
1862/// Free an events array where only `walk_count` entries have
1863/// fully-initialized `id`/`raw` allocations, but the array
1864/// itself was allocated for `alloc_count` slots. Per-event
1865/// boxes are freed for `0..walk_count`; the array is then
1866/// deallocated with the original `Layout::array::<NetEvent>(alloc_count)`
1867/// to match the allocation. Used by `net_poll_ex`'s panic-mid-loop
1868/// recovery path.
1869fn free_events_array_partial(events: *mut NetEvent, walk_count: usize, alloc_count: usize) {
1870 if events.is_null() || alloc_count == 0 {
1871 return;
1872 }
1873 for i in 0..walk_count {
1874 let event = unsafe { &*events.add(i) };
1875 if !event.id.is_null() {
1876 unsafe {
1877 let _ = Box::from_raw(std::ptr::slice_from_raw_parts_mut(
1878 event.id as *mut u8,
1879 event.id_len,
1880 ));
1881 }
1882 }
1883 if !event.raw.is_null() {
1884 unsafe {
1885 let _ = Box::from_raw(std::ptr::slice_from_raw_parts_mut(
1886 event.raw as *mut u8,
1887 event.raw_len,
1888 ));
1889 }
1890 }
1891 }
1892 if let Ok(layout) = std::alloc::Layout::array::<NetEvent>(alloc_count) {
1893 unsafe {
1894 std::alloc::dealloc(events as *mut u8, layout);
1895 }
1896 }
1897}
1898
1899/// Free the internal allocations of a poll result returned by `net_poll_ex`.
1900///
1901/// This frees the events array (including each event's `id` and `raw` buffers)
1902/// and the `next_id` string. It does **not** free the `NetPollResult` struct
1903/// itself, which is caller-provided (typically stack-allocated or managed by
1904/// the caller).
1905#[unsafe(no_mangle)]
1906pub unsafe extern "C" fn net_free_poll_result(result: *mut NetPollResult) {
1907 if result.is_null() {
1908 return;
1909 }
1910
1911 let result = unsafe { &mut *result };
1912
1913 // Free events array and all id/raw allocations.
1914 free_events_array(result.events, result.count);
1915
1916 // Free next_id.
1917 if !result.next_id.is_null() {
1918 unsafe {
1919 drop(std::ffi::CString::from_raw(result.next_id));
1920 }
1921 }
1922
1923 // Null the fields so a second `net_free_poll_result` on the
1924 // same struct is a safe no-op rather than a double-free. The
1925 // C header's contract just says "free a poll result"; without
1926 // this clear, a defensive caller calling free twice (or two
1927 // wrappers each calling free in their destructor) would
1928 // re-`Box::from_raw` an already-freed pointer.
1929 result.events = std::ptr::null_mut();
1930 result.count = 0;
1931 result.next_id = std::ptr::null_mut();
1932 result.has_more = 0;
1933}
1934
1935/// Get stats without JSON serialization.
1936#[unsafe(no_mangle)]
1937pub unsafe extern "C" fn net_stats_ex(handle: *mut NetHandle, out: *mut NetStats) -> c_int {
1938 if !handle_is_valid(handle) || out.is_null() {
1939 return NetError::NullPointer.into();
1940 }
1941
1942 let handle = unsafe { &*handle };
1943 let _guard = match enter_ffi_op(handle) {
1944 Ok(g) => g,
1945 Err(err) => return err,
1946 };
1947 let stats = handle.bus.stats();
1948
1949 unsafe {
1950 (*out).events_ingested = stats
1951 .events_ingested
1952 .load(std::sync::atomic::Ordering::Relaxed);
1953 (*out).events_dropped = stats
1954 .events_dropped
1955 .load(std::sync::atomic::Ordering::Relaxed);
1956 (*out).batches_dispatched = stats
1957 .batches_dispatched
1958 .load(std::sync::atomic::Ordering::Relaxed);
1959 }
1960
1961 NetError::Success.into()
1962}
1963
1964#[cfg(test)]
1965mod tests {
1966 use super::*;
1967
1968 #[test]
1969 fn test_parse_config_valid() {
1970 let config = parse_config_json(r#"{"num_shards": 8}"#);
1971 assert!(config.is_some());
1972 }
1973
1974 #[test]
1975 fn test_parse_config_num_shards_overflow() {
1976 // u16::MAX is 65535, so 65536 should fail
1977 let config = parse_config_json(r#"{"num_shards": 65536}"#);
1978 assert!(
1979 config.is_none(),
1980 "num_shards exceeding u16::MAX should fail"
1981 );
1982
1983 // Much larger value should also fail
1984 let config = parse_config_json(r#"{"num_shards": 100000}"#);
1985 assert!(
1986 config.is_none(),
1987 "num_shards exceeding u16::MAX should fail"
1988 );
1989 }
1990
1991 #[test]
1992 fn test_parse_config_num_shards_max_valid() {
1993 // u16::MAX (65535) should be valid
1994 let config = parse_config_json(r#"{"num_shards": 65535}"#);
1995 assert!(config.is_some(), "num_shards at u16::MAX should be valid");
1996 }
1997
1998 #[test]
1999 fn test_parse_config_invalid_json() {
2000 let config = parse_config_json(r#"{"num_shards": invalid}"#);
2001 assert!(config.is_none());
2002 }
2003
2004 #[test]
2005 fn test_parse_config_empty() {
2006 let config = parse_config_json(r#"{}"#);
2007 assert!(config.is_some(), "empty config should use defaults");
2008 }
2009
2010 /// Pin: known `backpressure_mode` strings round-trip; an
2011 /// unknown value (typo) is rejected with `None`, not silently
2012 /// downgraded to `DropNewest`. Pre-fix a deploy-time typo
2013 /// like `"DropOldset"` swapped the operator's intended
2014 /// durability for `DropNewest` with no diagnostic.
2015 #[test]
2016 fn parse_config_rejects_unknown_backpressure_mode() {
2017 // Known values still parse.
2018 for s in [
2019 "DropNewest",
2020 "drop_newest",
2021 "DropOldest",
2022 "drop_oldest",
2023 "FailProducer",
2024 "fail_producer",
2025 ] {
2026 let cfg = parse_config_json(&format!(r#"{{"backpressure_mode": "{}"}}"#, s));
2027 assert!(cfg.is_some(), "known mode `{}` must parse", s);
2028 }
2029
2030 // Typos must fail.
2031 for s in ["DropOldset", "FailProduce", "drop_oldst", "garbage", ""] {
2032 let cfg = parse_config_json(&format!(r#"{{"backpressure_mode": "{}"}}"#, s));
2033 assert!(
2034 cfg.is_none(),
2035 "unknown mode `{}` must reject (pre-fix this silently \
2036 fell through to DropNewest)",
2037 s,
2038 );
2039 }
2040
2041 // Wrong JSON type also fails — pre-fix this hit the
2042 // `and_then(|v| v.as_str())` short-circuit and was
2043 // ignored entirely.
2044 let cfg = parse_config_json(r#"{"backpressure_mode": 42}"#);
2045 assert!(
2046 cfg.is_none(),
2047 "non-string non-object backpressure_mode must reject"
2048 );
2049 let cfg = parse_config_json(r#"{"backpressure_mode": true}"#);
2050 assert!(cfg.is_none(), "boolean backpressure_mode must reject");
2051 }
2052
2053 /// Pin: the `Sample { rate }` mode is reachable from JSON
2054 /// via `{"backpressure_mode": {"Sample": {"rate": N}}}`,
2055 /// and a zero rate is rejected (validator already rejects
2056 /// it; the parser must too, so the surface is consistent).
2057 #[test]
2058 fn parse_config_supports_sample_mode_with_validation() {
2059 let cfg = parse_config_json(r#"{"backpressure_mode": {"Sample": {"rate": 10}}}"#);
2060 assert!(cfg.is_some(), "Sample with non-zero rate must parse");
2061
2062 let cfg = parse_config_json(r#"{"backpressure_mode": {"Sample": {"rate": 0}}}"#);
2063 assert!(cfg.is_none(), "Sample with rate=0 must reject");
2064
2065 let cfg = parse_config_json(r#"{"backpressure_mode": {"Sample": {}}}"#);
2066 assert!(cfg.is_none(), "Sample missing rate must reject");
2067 }
2068
2069 // Regression: the Go binding's `Poll(limit, cursor)` serializes a
2070 // `"cursor"` field that the FFI JSON path previously ignored —
2071 // cross-shard pagination silently broke. `parse_poll_request_json`
2072 // must round-trip the cursor into `ConsumeRequest.from_id`.
2073 #[test]
2074 fn test_parse_poll_request_preserves_cursor() {
2075 let req = parse_poll_request_json(r#"{"limit": 50, "cursor": "abc:123"}"#).unwrap();
2076 assert_eq!(req.limit, 50);
2077 assert_eq!(req.from_id.as_deref(), Some("abc:123"));
2078 }
2079
2080 #[test]
2081 fn test_parse_poll_request_no_cursor_defaults_to_none() {
2082 let req = parse_poll_request_json(r#"{"limit": 10}"#).unwrap();
2083 assert_eq!(req.limit, 10);
2084 assert_eq!(req.from_id, None);
2085 }
2086
2087 #[test]
2088 fn test_parse_poll_request_empty_uses_default_limit() {
2089 let req = parse_poll_request_json(r#"{}"#).unwrap();
2090 assert_eq!(req.limit, 100);
2091 assert_eq!(req.from_id, None);
2092 }
2093
2094 // Regression: a wrong-typed `"limit"` previously hit
2095 // `.as_u64().unwrap_or(100)` and silently defaulted. Caller bugs
2096 // (e.g. sending a string or a negative number) must surface as
2097 // `InvalidJson` instead.
2098 #[test]
2099 fn test_parse_poll_request_wrong_type_limit_errors() {
2100 let err = parse_poll_request_json(r#"{"limit": "50"}"#).unwrap_err();
2101 assert_eq!(err, c_int::from(NetError::InvalidJson));
2102 let err = parse_poll_request_json(r#"{"limit": -1}"#).unwrap_err();
2103 assert_eq!(err, c_int::from(NetError::InvalidJson));
2104 }
2105
2106 #[test]
2107 fn test_parse_poll_request_wrong_type_cursor_errors() {
2108 let err = parse_poll_request_json(r#"{"cursor": 123}"#).unwrap_err();
2109 assert_eq!(err, c_int::from(NetError::InvalidJson));
2110 }
2111
2112 #[test]
2113 fn test_parse_poll_request_null_fields_use_defaults() {
2114 let req = parse_poll_request_json(r#"{"limit": null, "cursor": null}"#).unwrap();
2115 assert_eq!(req.limit, 100);
2116 assert_eq!(req.from_id, None);
2117 }
2118
2119 /// `usize::MAX` is always a valid usize regardless of target
2120 /// pointer width, so it must parse successfully on both 32- and
2121 /// 64-bit builds. This pins the boundary case.
2122 #[test]
2123 fn test_parse_poll_request_limit_at_usize_max() {
2124 let json = format!(r#"{{"limit": {}}}"#, usize::MAX);
2125 let req = parse_poll_request_json(&json).unwrap();
2126 assert_eq!(req.limit, usize::MAX);
2127 }
2128
2129 /// Regression: `as usize` silently truncates on 32-bit targets
2130 /// for `u64` values above `usize::MAX`. The parser must return
2131 /// `InvalidJson` instead of wrapping. We only run this on 32-bit
2132 /// targets because on 64-bit `usize::MAX == u64::MAX`, leaving
2133 /// nothing that fits in u64 but not usize.
2134 #[cfg(target_pointer_width = "32")]
2135 #[test]
2136 fn test_parse_poll_request_limit_overflows_usize_on_32bit() {
2137 // 2^33 — fits in u64, but exceeds usize::MAX on a 32-bit build.
2138 let err = parse_poll_request_json(r#"{"limit": 8589934592}"#).unwrap_err();
2139 assert_eq!(err, c_int::from(NetError::InvalidJson));
2140 }
2141
2142 /// CR-22: pin parity between the Rust-side `NetError` enum and
2143 /// the two C-header copies. The Rust enum is the source of
2144 /// truth; C / Go consumers `errors.Is` against the named codes.
2145 /// Pre-CR-22 the headers were missing `-9` (IntOverflow) and
2146 /// `-10` (MismatchedHandles); a consumer receiving those values
2147 /// would fall into the unknown-code branch and lose actionable
2148 /// distinction.
2149 ///
2150 /// We extract every integer literal that appears as the
2151 /// right-hand side of an `= ` token in the file and check
2152 /// that each Rust-side value is present in BOTH headers. The
2153 /// test does NOT verify symbolic names; the sealing
2154 /// constraint is the numeric value alone.
2155 ///
2156 /// Both `include_str!` paths point inside `net/crates/net/`.
2157 /// `include/net.go.h` is a manually-synced mirror of the
2158 /// repo-root `go/net.h`. Reaching outside the crate root
2159 /// (`include_str!("../../../../../go/net.h")`) breaks
2160 /// `cargo publish` and any out-of-repo vendoring of this
2161 /// crate, so the in-crate copy is the supported source. A
2162 /// drift between the two surfaces here as a parity-test
2163 /// failure: one of them will be missing the new value.
2164 #[test]
2165 fn cr22_c_header_parity_with_rust_neterror() {
2166 let primary = include_str!("../../include/net.h");
2167 let go_copy = include_str!("../../include/net.go.h");
2168
2169 // The Rust enum's full set of values (mirrors `pub enum
2170 // NetError` above). When a new variant is added in the
2171 // Rust source, this list — AND both headers — must be
2172 // updated together. The asserts that follow then catch a
2173 // missing header update at the next CI run.
2174 let rust_values: &[i32] = &[0, -1, -2, -3, -4, -5, -6, -7, -8, -9, -10, -11, -99];
2175
2176 // Pull every numeric literal that looks like an enum-value
2177 // assignment (`= <number>` followed by `,` or whitespace).
2178 // Whitespace-tolerant: skips `= 0`, `= 0`, `= -10`, etc.
2179 fn extract_assigned_values(src: &str) -> Vec<i32> {
2180 let mut out = Vec::new();
2181 let mut chars = src.chars().peekable();
2182 while let Some(c) = chars.next() {
2183 if c != '=' {
2184 continue;
2185 }
2186 // Skip whitespace.
2187 while let Some(&peek) = chars.peek() {
2188 if peek == ' ' || peek == '\t' {
2189 chars.next();
2190 } else {
2191 break;
2192 }
2193 }
2194 // Optional sign.
2195 let mut buf = String::new();
2196 if let Some(&peek) = chars.peek() {
2197 if peek == '-' || peek == '+' {
2198 buf.push(peek);
2199 chars.next();
2200 }
2201 }
2202 // Digits.
2203 let mut had_digit = false;
2204 while let Some(&peek) = chars.peek() {
2205 if peek.is_ascii_digit() {
2206 buf.push(peek);
2207 chars.next();
2208 had_digit = true;
2209 } else {
2210 break;
2211 }
2212 }
2213 if had_digit {
2214 if let Ok(v) = buf.parse::<i32>() {
2215 out.push(v);
2216 }
2217 }
2218 }
2219 out
2220 }
2221
2222 let primary_vals = extract_assigned_values(primary);
2223 let go_vals = extract_assigned_values(go_copy);
2224
2225 for &v in rust_values {
2226 assert!(
2227 primary_vals.contains(&v),
2228 "CR-22 regression: include/net.h is missing the value {} \
2229 (Rust NetError defines it). Add the matching `NET_ERR_*` \
2230 enumerator before merging.",
2231 v
2232 );
2233 assert!(
2234 go_vals.contains(&v),
2235 "CR-22 regression: bindings/go/net/net.h is missing the value {} \
2236 (Rust NetError defines it).",
2237 v
2238 );
2239 }
2240 }
2241
2242 /// CR-5: pin that `examples/capability.c` does not double-include
2243 /// `net.h` and `net.go.h`. Both files use the `NET_SDK_H` include
2244 /// guard, so when both are included in one TU the second is
2245 /// silently skipped — every `net_validate_capabilities` /
2246 /// `net_predicate_*` call the example makes becomes an
2247 /// implicit-declaration error on GCC 14+/Clang 16+, and a silent
2248 /// `int`-return miscompile on older toolchains. The deeper fix
2249 /// (renaming one guard so they compose cleanly) is tracked as
2250 /// CR-28; this test catches the example-level regression.
2251 #[test]
2252 fn cr5_example_does_not_double_include_net_headers() {
2253 let example = include_str!("../../examples/capability.c");
2254 let net_h_included = example.contains("#include \"../include/net.h\"");
2255 let net_go_h_included = example.contains("#include \"../include/net.go.h\"");
2256 assert!(
2257 net_go_h_included,
2258 "examples/capability.c must include net.go.h to declare \
2259 net_validate_capabilities + net_predicate_* symbols"
2260 );
2261 assert!(
2262 !net_h_included,
2263 "examples/capability.c must NOT also include net.h: \
2264 both headers share the NET_SDK_H guard, so the second \
2265 include is silently skipped, leaving the example's \
2266 net_predicate_* calls implicitly declared. Drop the \
2267 redundant include — net.go.h is a superset."
2268 );
2269 }
2270
2271 /// `handle_is_valid` rejects null and any pointer not aligned for
2272 /// `NetHandle`. A foreign caller producing a misaligned pointer
2273 /// (e.g. via an over-eager `void *` cast on a packed struct) hits
2274 /// `&*handle` UB before any other check fires; this gate is the
2275 /// pre-deref discriminator.
2276 #[test]
2277 fn handle_is_valid_rejects_null_and_misaligned() {
2278 // Null is rejected.
2279 assert!(
2280 !handle_is_valid(std::ptr::null::<NetHandle>()),
2281 "null pointer must not be considered a valid handle"
2282 );
2283
2284 // Aligned but non-null is accepted (we use a small backing
2285 // buffer to materialize a pointer without dereferencing it).
2286 // `align_of::<NetHandle>()` is the alignment we must match.
2287 let align = std::mem::align_of::<NetHandle>();
2288 let buf = vec![0u8; align * 2];
2289 let base = buf.as_ptr() as usize;
2290 let aligned = (base + align - 1) & !(align - 1);
2291 let aligned_ptr = aligned as *const NetHandle;
2292 assert!(
2293 handle_is_valid(aligned_ptr),
2294 "aligned non-null pointer must validate (align={align}, ptr={aligned_ptr:p})"
2295 );
2296
2297 // A pointer one byte past `aligned_ptr` is misaligned for any
2298 // type with align > 1, and `NetHandle` (containing `AtomicU32`,
2299 // `AtomicBool`, ManuallyDrop'd EventBus + Runtime) easily
2300 // exceeds 1.
2301 if align > 1 {
2302 let misaligned_ptr = (aligned + 1) as *const NetHandle;
2303 assert!(
2304 !handle_is_valid(misaligned_ptr),
2305 "misaligned pointer must be rejected (align={align}, ptr={misaligned_ptr:p})"
2306 );
2307 }
2308 }
2309
2310 /// Pin: zero values for `heartbeat_interval_ms` and
2311 /// `session_timeout_ms` must reject the entire config (parser
2312 /// returns `None`). Pre-fix the parser threaded `0` through
2313 /// to `Duration::from_millis(0)`, which on the Net adapter's
2314 /// heartbeat path results in a busy-loop that pegs a CPU and
2315 /// produces no diagnostic — the FFI caller saw a successful
2316 /// `net_init` followed by a hung daemon. The validator-level
2317 /// guard for cooldown / metrics_window has no equivalent on
2318 /// the Net-adapter side, so the parser is the only place that
2319 /// can refuse the build.
2320 #[cfg(feature = "net")]
2321 #[test]
2322 fn parse_config_rejects_zero_heartbeat_and_session_timeout() {
2323 // 32-byte hex strings (64 chars) so `hex::decode` produces
2324 // exactly the [u8; 32] the parser requires for `psk` and
2325 // `peer_public_key`.
2326 let psk = "0".repeat(64);
2327 let peer_pk = "1".repeat(64);
2328
2329 // Sanity: a config with both fields *non-zero* must parse
2330 // successfully — proves the rejection in the negative
2331 // cases below is caused by the zero, not a missing
2332 // required field on the surrounding `net` block.
2333 let baseline = format!(
2334 r#"{{"net":{{"bind_addr":"127.0.0.1:9000","peer_addr":"127.0.0.1:9001",
2335 "psk":"{psk}","peer_public_key":"{peer_pk}",
2336 "heartbeat_interval_ms":1000,"session_timeout_ms":30000}}}}"#
2337 );
2338 assert!(
2339 parse_config_json(&baseline).is_some(),
2340 "baseline net config with non-zero heartbeat/session_timeout must parse"
2341 );
2342
2343 // heartbeat_interval_ms = 0 → reject.
2344 let zero_hb = format!(
2345 r#"{{"net":{{"bind_addr":"127.0.0.1:9000","peer_addr":"127.0.0.1:9001",
2346 "psk":"{psk}","peer_public_key":"{peer_pk}",
2347 "heartbeat_interval_ms":0,"session_timeout_ms":30000}}}}"#
2348 );
2349 assert!(
2350 parse_config_json(&zero_hb).is_none(),
2351 "heartbeat_interval_ms=0 must reject (pre-fix this produced a CPU-pegging busy loop)"
2352 );
2353
2354 // session_timeout_ms = 0 → reject.
2355 let zero_to = format!(
2356 r#"{{"net":{{"bind_addr":"127.0.0.1:9000","peer_addr":"127.0.0.1:9001",
2357 "psk":"{psk}","peer_public_key":"{peer_pk}",
2358 "heartbeat_interval_ms":1000,"session_timeout_ms":0}}}}"#
2359 );
2360 assert!(
2361 parse_config_json(&zero_to).is_none(),
2362 "session_timeout_ms=0 must reject"
2363 );
2364 }
2365}