1use std::collections::HashMap;
2use std::mem::size_of;
3use std::ops::Deref;
4use std::sync::atomic::{AtomicU8, AtomicU32, AtomicU64, Ordering};
5use std::sync::{Arc, LazyLock, Mutex, Weak};
6use std::task::{Context, Poll, Waker};
7
8use orbit_core::Fleet;
9#[cfg(unix)]
10use orbit_core::shm::{ShmRegion, ring_segment_name};
11
12use super::{ChunkDescriptor, ChunkGeometry, ChunkPlan, DEFAULT_CHUNK_BYTES};
13use crate::wake::{Doorstep, Driver};
14use crate::{Error, Incarnation, Result, lock_unpoisoned};
15
16const MAGIC: u32 = 0x4F_50_41_59; const VERSION: u16 = 4;
18
19const SLOT_FREE: u8 = 0;
20const SLOT_RESERVED: u8 = 1;
21const SLOT_LIVE: u8 = 2;
22const SLOT_READING: u8 = 3;
23const SLOT_RECLAIMING: u8 = 4;
24
25#[derive(Clone, Copy, Debug, Eq, Hash, PartialEq)]
27pub struct PayloadArenaSpec {
28 pub kind: u8,
29 pub slots_per_node: usize,
31 pub slot_size: usize
33}
34
35impl PayloadArenaSpec {
36 pub const fn new(
37 kind: u8,
38 slots_per_node: usize,
39 slot_size: usize
40 ) -> Self {
41 Self { kind, slots_per_node, slot_size }
42 }
43
44 pub const fn lane_bytes(self) -> usize {
45 self.slots_per_node * self.slot_size
46 }
47
48 pub const fn slots_for(
49 self,
50 payload_len: usize
51 ) -> usize {
52 payload_len.div_ceil(self.slot_size)
53 }
54
55 pub const fn chunk_geometry(self) -> Option<ChunkGeometry> {
58 ChunkGeometry::new(self.slot_size, self.slots_per_node)
59 }
60
61 pub const fn plan_chunks(
64 self,
65 data_bytes: usize
66 ) -> Option<ChunkPlan> {
67 let lane_bytes = self.lane_bytes();
68 let limit = if DEFAULT_CHUNK_BYTES < lane_bytes { DEFAULT_CHUNK_BYTES } else { lane_bytes };
69 ChunkPlan::for_limit(data_bytes, self.slot_size, limit)
70 }
71
72 pub const fn plan_chunks_with(
74 self,
75 data_bytes: usize,
76 slots_per_chunk: usize
77 ) -> Option<ChunkPlan> {
78 let geometry = match self.chunk_geometry() {
79 Some(geometry) => geometry,
80 None => return None
81 };
82 geometry.plan(data_bytes, slots_per_chunk)
83 }
84
85 fn validate(self) -> Result<()> {
86 if self.slots_per_node == 0
87 || self.slot_size == 0
88 || !self.slots_per_node.is_power_of_two()
89 || !self.slot_size.is_power_of_two()
90 || self.slots_per_node > u32::MAX as usize
91 || self.slot_size > u32::MAX as usize
92 {
93 return Err(Error::Malformed(format!(
94 "payload arena kind={} slots_per_node={} slot_size={}: both geometry values must be non-zero powers of two fitting u32",
95 self.kind, self.slots_per_node, self.slot_size
96 )));
97 }
98 Ok(())
99 }
100}
101
102#[repr(C, align(64))]
103struct Header {
104 magic: u32,
105 version: u16,
106 header_size: u16,
107 slot_meta_size: u32,
108 slot_size: u32,
109 slots_per_node: u32,
110 fleet_capacity: u16,
111 _reserved: [u8; 2],
112 next_generation: AtomicU64,
113 epoch: AtomicU64,
114 _reserved_credit: [u8; 24]
115}
116
117impl Header {
118 fn new(
119 fleet_capacity: u16,
120 spec: PayloadArenaSpec
121 ) -> Self {
122 Self {
123 magic: MAGIC,
124 version: VERSION,
125 header_size: size_of::<Self>() as u16,
126 slot_meta_size: size_of::<SlotMeta>() as u32,
127 slot_size: spec.slot_size as u32,
128 slots_per_node: spec.slots_per_node as u32,
129 fleet_capacity,
130 _reserved: [0; 2],
131 next_generation: AtomicU64::new(1),
132 epoch: AtomicU64::new(1),
133 _reserved_credit: [0; 24]
134 }
135 }
136
137 fn compatible(
138 &self,
139 fleet_capacity: u16,
140 spec: PayloadArenaSpec
141 ) -> bool {
142 self.magic == MAGIC
143 && self.version == VERSION
144 && self.header_size as usize == size_of::<Self>()
145 && self.slot_meta_size as usize == size_of::<SlotMeta>()
146 && self.slot_size as usize == spec.slot_size
147 && self.slots_per_node as usize == spec.slots_per_node
148 && self.fleet_capacity == fleet_capacity
149 }
150}
151
152#[repr(C, align(64))]
153struct CreditWord {
154 generation: AtomicU32,
155 waiters: AtomicU32,
156 _padding: [u8; 56]
157}
158
159#[repr(C, align(8))]
160struct SlotMeta {
161 state: AtomicU8,
162 _reserved: AtomicU8,
163 reader_node: std::sync::atomic::AtomicU16,
164 allocation_first: AtomicU64,
165 generation: AtomicU64,
166 producer_incarnation: AtomicU64,
167 reader_incarnation: AtomicU64
168}
169
170const _: () = assert!(size_of::<Header>() == 64);
171const _: () = assert!(size_of::<CreditWord>() == 64);
172const _: () = assert!(size_of::<SlotMeta>() == 40);
173
174#[derive(Clone, Copy)]
175struct Geometry {
176 slots_per_node: usize,
177 slot_size: usize,
178 total_slots: usize,
179 credits_offset: usize,
180 metadata_offset: usize,
181 payload_offset: usize,
182 segment_size: usize
183}
184
185impl Geometry {
186 fn new(
187 fleet_capacity: u16,
188 spec: PayloadArenaSpec
189 ) -> Self {
190 let total_slots = usize::from(fleet_capacity) * spec.slots_per_node;
191 let credits_offset = size_of::<Header>();
192 let metadata_offset =
193 credits_offset + usize::from(fleet_capacity) * size_of::<CreditWord>();
194 let payload_offset =
195 (metadata_offset + total_slots * size_of::<SlotMeta>()).next_multiple_of(64);
196 let segment_size = payload_offset + total_slots * spec.slot_size;
197 Self {
198 slots_per_node: spec.slots_per_node,
199 slot_size: spec.slot_size,
200 total_slots,
201 credits_offset,
202 metadata_offset,
203 payload_offset,
204 segment_size
205 }
206 }
207}
208
209enum Backing {
210 Memory(AlignedBytes),
211 #[cfg(unix)]
212 Shm(ShmRegion)
213}
214
215struct AlignedBytes {
216 ptr: *mut u8,
217 layout: std::alloc::Layout
218}
219
220impl AlignedBytes {
221 fn zeroed(size: usize) -> Self {
222 let layout = std::alloc::Layout::from_size_align(size, 64).expect("payload arena layout");
223 let ptr = unsafe { std::alloc::alloc_zeroed(layout) };
225 assert!(!ptr.is_null(), "payload arena allocation failed");
226 Self { ptr, layout }
227 }
228}
229
230impl Drop for AlignedBytes {
231 fn drop(&mut self) {
232 unsafe { std::alloc::dealloc(self.ptr, self.layout) }
234 }
235}
236
237unsafe impl Send for AlignedBytes {}
239unsafe impl Sync for AlignedBytes {}
240
241struct Arena {
242 _fleet: Arc<Fleet>,
243 backing: Backing,
244 geometry: Geometry,
245 kind: u8,
246 node: u16,
247 incarnation: Incarnation,
248 allocation: Mutex<usize>,
249 credit_wakers: Mutex<Vec<Waker>>,
250 driver: Mutex<Option<Driver>>
251}
252
253impl Arena {
254 fn base(&self) -> *mut u8 {
255 match &self.backing {
256 Backing::Memory(bytes) => bytes.ptr,
257 #[cfg(unix)]
258 Backing::Shm(region) => region.as_ptr()
259 }
260 }
261
262 fn header(&self) -> &Header {
263 unsafe { &*self.base().cast::<Header>() }
265 }
266
267 fn metadata(&self) -> &[SlotMeta] {
268 unsafe {
270 std::slice::from_raw_parts(
271 self.base().add(self.geometry.metadata_offset).cast::<SlotMeta>(),
272 self.geometry.total_slots
273 )
274 }
275 }
276
277 fn credits(&self) -> &[CreditWord] {
278 unsafe {
280 std::slice::from_raw_parts(
281 self.base().add(self.geometry.credits_offset).cast::<CreditWord>(),
282 self.geometry.total_slots / self.geometry.slots_per_node
283 )
284 }
285 }
286
287 fn local_credit(&self) -> &CreditWord {
288 &self.credits()[usize::from(self.node)]
289 }
290
291 fn payload(
292 &self,
293 absolute_slot: usize
294 ) -> *mut u8 {
295 unsafe {
297 self.base().add(self.geometry.payload_offset + absolute_slot * self.geometry.slot_size)
298 }
299 }
300
301 fn reserve(
302 self: &Arc<Self>,
303 payload_len: usize
304 ) -> Result<Publication> {
305 let count = self.required_slots(payload_len)?;
306
307 let mut hint = lock_unpoisoned(&self.allocation);
308 let lane_start = usize::from(self.node) * self.geometry.slots_per_node;
309 let candidate = (0..self.geometry.slots_per_node)
310 .map(|offset| (*hint + offset) & (self.geometry.slots_per_node - 1))
311 .find(|&local| {
312 local + count <= self.geometry.slots_per_node
313 && self.metadata()[lane_start + local..lane_start + local + count]
314 .iter()
315 .all(|slot| slot.state.load(Ordering::Acquire) == SLOT_FREE)
316 })
317 .ok_or(Error::PayloadFull { requested_slots: count })?;
318 let first = lane_start + candidate;
319 for slot in &self.metadata()[first..first + count] {
320 slot.state.store(SLOT_RESERVED, Ordering::Relaxed);
321 }
322 let mut generation = self.header().next_generation.fetch_add(1, Ordering::Relaxed);
323 if generation == 0 {
324 generation = self.header().next_generation.fetch_add(1, Ordering::Relaxed);
325 }
326 for slot in &self.metadata()[first..first + count] {
327 slot.allocation_first.store(first as u64, Ordering::Relaxed);
328 slot.generation.store(generation, Ordering::Relaxed);
329 slot.producer_incarnation.store(self.incarnation.get(), Ordering::Relaxed);
330 slot.reader_node.store(u16::MAX, Ordering::Relaxed);
331 slot.reader_incarnation.store(0, Ordering::Relaxed);
332 }
333 self.metadata()[first].state.store(SLOT_RESERVED, Ordering::Release);
334 *hint = (candidate + count) & (self.geometry.slots_per_node - 1);
335 Ok(Publication {
336 arena: Arc::clone(self),
337 first,
338 count,
339 payload_len,
340 generation,
341 published: false
342 })
343 }
344
345 fn release(
346 &self,
347 first: usize,
348 count: usize,
349 generation: u64
350 ) {
351 let slots = &self.metadata()[first..first + count];
352 if slots.iter().any(|slot| slot.generation.load(Ordering::Acquire) != generation) {
353 return;
354 }
355 for slot in slots.iter().skip(1) {
356 slot.state.store(SLOT_FREE, Ordering::Relaxed);
357 }
358 slots[0].state.store(SLOT_FREE, Ordering::Release);
359 let owner = first / self.geometry.slots_per_node;
360 let credit = &self.credits()[owner];
361 credit.generation.fetch_add(1, Ordering::SeqCst);
362 if credit.waiters.load(Ordering::SeqCst) > 0 {
363 crate::wake_on(&credit.generation);
364 }
365 }
366
367 fn required_slots(
368 &self,
369 payload_len: usize
370 ) -> Result<usize> {
371 if payload_len == 0 {
372 return Err(Error::Malformed("an empty payload needs no arena allocation".to_owned()));
373 }
374 let count = payload_len.div_ceil(self.geometry.slot_size);
375 if count > self.geometry.slots_per_node {
376 return Err(Error::PayloadTooLarge {
377 len: payload_len,
378 capacity: self.geometry.slots_per_node * self.geometry.slot_size
379 });
380 }
381 Ok(count)
382 }
383
384 fn has_run(
385 &self,
386 count: usize
387 ) -> bool {
388 let lane_start = usize::from(self.node) * self.geometry.slots_per_node;
389 (0..=self.geometry.slots_per_node - count).any(|local| {
390 self.metadata()[lane_start + local..lane_start + local + count]
391 .iter()
392 .all(|slot| slot.state.load(Ordering::Acquire) == SLOT_FREE)
393 })
394 }
395
396 fn wait_available(
397 &self,
398 count: usize
399 ) -> Result<()> {
400 let credit = self.local_credit();
401 loop {
402 if self.has_run(count) {
403 return Ok(());
404 }
405 credit.waiters.fetch_add(1, Ordering::SeqCst);
406 let seen = credit.generation.load(Ordering::SeqCst);
407 let outcome =
408 if self.has_run(count) { Ok(()) } else { crate::wait_on(&credit.generation, seen) };
409 credit.waiters.fetch_sub(1, Ordering::SeqCst);
410 outcome?;
411 }
412 }
413
414 fn poll_available(
415 self: &Arc<Self>,
416 count: usize,
417 cx: &mut Context<'_>
418 ) -> Poll<Result<()>> {
419 if self.has_run(count) {
420 return Poll::Ready(Ok(()));
421 }
422 {
423 let mut wakers = lock_unpoisoned(&self.credit_wakers);
424 if !wakers.iter().any(|waker| waker.will_wake(cx.waker())) {
425 wakers.push(cx.waker().clone());
426 }
427 }
428 if let Err(error) = self.ensure_driver() {
429 return Poll::Ready(Err(error));
430 }
431 if self.has_run(count) { Poll::Ready(Ok(())) } else { Poll::Pending }
432 }
433
434 fn ensure_driver(self: &Arc<Self>) -> Result<()> {
435 let mut driver = lock_unpoisoned(&self.driver);
436 if driver.is_none() {
437 *driver = Some(Driver::start(
438 Arc::as_ptr(self),
439 format!("orbit-payload-{}-{}-driver", self.kind, self.node)
440 )?);
441 }
442 Ok(())
443 }
444
445 fn extent_count(
446 &self,
447 first: usize,
448 generation: u64
449 ) -> usize {
450 let lane_end = (first / self.geometry.slots_per_node + 1) * self.geometry.slots_per_node;
451 self.metadata()[first..lane_end]
452 .iter()
453 .take_while(|slot| {
454 slot.allocation_first.load(Ordering::Acquire) as usize == first
455 && slot.generation.load(Ordering::Acquire) == generation
456 })
457 .count()
458 }
459
460 fn node_dead(
461 &self,
462 node: u16,
463 incarnation: Incarnation
464 ) {
465 let metadata = self.metadata();
466 for (first, slot) in metadata.iter().enumerate() {
467 let state = slot.state.load(Ordering::Acquire);
468 let producer_node = first / self.geometry.slots_per_node;
469 let allocation_start = slot.allocation_first.load(Ordering::Acquire) as usize == first;
470 let abandoned_publication = allocation_start
471 && matches!(state, SLOT_RESERVED | SLOT_LIVE)
472 && producer_node == usize::from(node)
473 && slot.producer_incarnation.load(Ordering::Acquire) == incarnation.get();
474 let abandoned_read = state == SLOT_READING
475 && slot.reader_node.load(Ordering::Acquire) == node
476 && slot.reader_incarnation.load(Ordering::Acquire) == incarnation.get();
477 if !abandoned_publication && !abandoned_read {
478 continue;
479 }
480 if slot
481 .state
482 .compare_exchange(state, SLOT_RECLAIMING, Ordering::SeqCst, Ordering::Acquire)
483 .is_err()
484 {
485 continue;
486 }
487 let generation = slot.generation.load(Ordering::Acquire);
488 let count = self.extent_count(first, generation);
489 if count != 0 {
490 self.release(first, count, generation);
491 }
492 }
493 }
494}
495
496impl Doorstep for Arena {
497 fn generation(&self) -> &AtomicU32 {
498 &self.local_credit().generation
499 }
500
501 fn listening(
502 &self,
503 delta: i32
504 ) {
505 if delta > 0 {
506 self.local_credit().waiters.fetch_add(1, Ordering::SeqCst);
507 } else {
508 self.local_credit().waiters.fetch_sub(1, Ordering::SeqCst);
509 }
510 }
511
512 fn drain(&self) {
513 let wakers = std::mem::take(&mut *lock_unpoisoned(&self.credit_wakers));
514 for waker in wakers {
515 waker.wake();
516 }
517 }
518}
519
520impl Drop for Arena {
521 fn drop(&mut self) {
522 if let Some(mut driver) = lock_unpoisoned(&self.driver).take() {
523 driver.stop(&self.local_credit().generation);
524 }
525 }
526}
527
528#[derive(Clone)]
531pub struct PayloadArena {
532 arena: Arc<Arena>
533}
534
535impl PayloadArena {
536 pub fn open(
537 fleet: Arc<Fleet>,
538 incarnation: Incarnation,
539 spec: PayloadArenaSpec
540 ) -> Result<Self> {
541 spec.validate()?;
542 Ok(Self { arena: open(&fleet, incarnation, spec)? })
543 }
544
545 pub fn kind(&self) -> u8 {
546 self.arena.kind
547 }
548
549 pub fn slot_size(&self) -> usize {
550 self.arena.geometry.slot_size
551 }
552
553 pub fn slots_per_node(&self) -> usize {
554 self.arena.geometry.slots_per_node
555 }
556
557 pub fn chunk_geometry(&self) -> ChunkGeometry {
558 ChunkGeometry::new(self.slot_size(), self.slots_per_node())
560 .expect("opened payload arena has valid chunk geometry")
561 }
562
563 pub fn plan_chunks(
564 &self,
565 data_bytes: usize
566 ) -> Result<ChunkPlan> {
567 PayloadArenaSpec::new(self.kind(), self.slots_per_node(), self.slot_size())
568 .plan_chunks(data_bytes)
569 .ok_or_else(|| {
570 Error::Malformed("payload arena cannot plan zero-sized geometry".to_owned())
571 })
572 }
573
574 pub fn plan_chunks_with(
575 &self,
576 data_bytes: usize,
577 slots_per_chunk: usize
578 ) -> Result<ChunkPlan> {
579 self.chunk_geometry().plan(data_bytes, slots_per_chunk).ok_or_else(|| {
580 Error::Malformed(format!(
581 "chunk slot count {slots_per_chunk} exceeds payload lane capacity {}",
582 self.slots_per_node()
583 ))
584 })
585 }
586 pub fn wait_available(
589 &self,
590 payload_len: usize
591 ) -> Result<()> {
592 let count = self.arena.required_slots(payload_len)?;
593 self.arena.wait_available(count)
594 }
595
596 pub fn is_available(
600 &self,
601 payload_len: usize
602 ) -> Result<bool> {
603 let count = self.arena.required_slots(payload_len)?;
604 Ok(self.arena.has_run(count))
605 }
606
607 pub fn poll_available(
610 &self,
611 payload_len: usize,
612 cx: &mut Context<'_>
613 ) -> Poll<Result<()>> {
614 let count = match self.arena.required_slots(payload_len) {
615 Ok(count) => count,
616 Err(error) => return Poll::Ready(Err(error))
617 };
618 self.arena.poll_available(count, cx)
619 }
620
621 pub fn reset_all(&self) {
624 let mut hint = lock_unpoisoned(&self.arena.allocation);
625 *hint = 0;
626 for slot in self.arena.metadata() {
627 slot.state.store(SLOT_FREE, Ordering::Release);
628 }
629 for credit in self.arena.credits() {
630 credit.generation.fetch_add(1, Ordering::SeqCst);
631 crate::wake_on(&credit.generation);
632 }
633 }
634
635 pub fn node_dead(
640 &self,
641 node: orbit_core::NodeId,
642 incarnation: Incarnation
643 ) {
644 self.arena.node_dead(node.get(), incarnation);
645 }
646
647 #[cfg(unix)]
649 pub fn unlink(&self) -> Result<()> {
650 match &self.arena.backing {
651 Backing::Memory(_) => {
652 self.reset_all();
653 Ok(())
654 }
655 Backing::Shm(region) => region.unlink().map_err(Error::Io)
656 }
657 }
658
659 #[cfg(test)]
660 pub(crate) fn publish(
661 &self,
662 payload: &[u8]
663 ) -> Result<Publication> {
664 let mut publication = self.reserve(payload.len())?;
665 publication.as_mut_slice().copy_from_slice(payload);
666 publication.make_live();
667 Ok(publication)
668 }
669
670 pub(crate) fn reserve(
671 &self,
672 payload_len: usize
673 ) -> Result<Publication> {
674 self.arena.reserve(payload_len)
675 }
676
677 pub(crate) fn read(
678 &self,
679 descriptor: ChunkDescriptor
680 ) -> Result<PayloadChunk> {
681 if descriptor.arena_kind() != self.kind() {
682 return Err(Error::Malformed(format!(
683 "chunk names payload arena kind {}, opened kind is {}",
684 descriptor.arena_kind(),
685 self.kind()
686 )));
687 }
688 let owner = usize::from(descriptor.owner_node());
689 let first_local = descriptor.first_slot() as usize;
690 let count = descriptor.slot_count() as usize;
691 let len = descriptor.payload_len() as usize;
692 if owner >= self.arena.geometry.total_slots / self.arena.geometry.slots_per_node
693 || count == 0
694 || first_local + count > self.arena.geometry.slots_per_node
695 || len == 0
696 || len > count * self.arena.geometry.slot_size
697 || (count > 1 && len <= (count - 1) * self.arena.geometry.slot_size)
698 {
699 return Err(Error::Malformed(
700 "chunk descriptor is outside its payload lane".to_owned()
701 ));
702 }
703 let first = owner * self.arena.geometry.slots_per_node + first_local;
704 let slots = &self.arena.metadata()[first..first + count];
705 let generation = descriptor.allocation_generation();
706 if slots.iter().any(|slot| slot.generation.load(Ordering::Acquire) != generation)
707 || slots.iter().skip(1).any(|slot| slot.state.load(Ordering::Acquire) != SLOT_RESERVED)
708 || slots[0]
709 .state
710 .compare_exchange(SLOT_LIVE, SLOT_READING, Ordering::Acquire, Ordering::Acquire)
711 .is_err()
712 {
713 return Err(Error::Malformed(
714 "chunk allocation is stale or already consumed".to_owned()
715 ));
716 }
717 slots[0].reader_incarnation.store(self.arena.incarnation.get(), Ordering::Release);
718 slots[0].reader_node.store(self.arena.node, Ordering::Release);
719 Ok(PayloadChunk {
720 arena: Arc::clone(&self.arena),
721 first,
722 count,
723 len,
724 generation,
725 descriptor
726 })
727 }
728}
729
730pub(crate) struct Publication {
731 arena: Arc<Arena>,
732 first: usize,
733 count: usize,
734 payload_len: usize,
735 generation: u64,
736 published: bool
737}
738
739impl Publication {
740 pub(crate) fn as_slice(&self) -> &[u8] {
741 unsafe { std::slice::from_raw_parts(self.arena.payload(self.first), self.payload_len) }
744 }
745
746 pub(crate) fn as_mut_slice(&mut self) -> &mut [u8] {
747 unsafe { std::slice::from_raw_parts_mut(self.arena.payload(self.first), self.payload_len) }
750 }
751
752 pub(crate) fn coordinates(&self) -> (u64, u32, u32, u32, u8, u16) {
753 let first_local = self.first % self.arena.geometry.slots_per_node;
754 (
755 self.generation,
756 first_local as u32,
757 self.count as u32,
758 self.payload_len as u32,
759 self.arena.kind,
760 self.arena.node
761 )
762 }
763
764 pub(crate) fn truncate(
765 &mut self,
766 payload_len: usize
767 ) -> Result<()> {
768 if payload_len == 0 || payload_len > self.payload_len {
769 return Err(Error::Malformed(format!(
770 "committed payload length {payload_len} is outside reserved capacity {}",
771 self.payload_len
772 )));
773 }
774 let count = self.arena.required_slots(payload_len)?;
775 if count < self.count {
776 for slot in &self.arena.metadata()[self.first + count..self.first + self.count] {
777 slot.state.store(SLOT_FREE, Ordering::Release);
778 }
779 self.count = count;
780 let credit = self.arena.local_credit();
781 credit.generation.fetch_add(1, Ordering::SeqCst);
782 if credit.waiters.load(Ordering::SeqCst) > 0 {
783 crate::wake_on(&credit.generation);
784 }
785 }
786 self.payload_len = payload_len;
787 Ok(())
788 }
789
790 pub(crate) fn make_live(&self) {
791 self.arena.metadata()[self.first].state.store(SLOT_LIVE, Ordering::Release);
792 }
793
794 pub(crate) fn mark_published(mut self) {
795 self.published = true;
796 }
797}
798
799impl Drop for Publication {
800 fn drop(&mut self) {
801 if !self.published {
802 self.arena.release(self.first, self.count, self.generation);
803 }
804 }
805}
806
807pub struct PayloadChunk {
810 arena: Arc<Arena>,
811 first: usize,
812 count: usize,
813 len: usize,
814 generation: u64,
815 descriptor: ChunkDescriptor
816}
817
818impl PayloadChunk {
819 pub const fn descriptor(&self) -> ChunkDescriptor {
820 self.descriptor
821 }
822
823 pub fn as_slice(&self) -> &[u8] {
824 unsafe { std::slice::from_raw_parts(self.arena.payload(self.first), self.len) }
827 }
828}
829
830impl Deref for PayloadChunk {
831 type Target = [u8];
832
833 fn deref(&self) -> &Self::Target {
834 self.as_slice()
835 }
836}
837
838impl Drop for PayloadChunk {
839 fn drop(&mut self) {
840 self.arena.release(self.first, self.count, self.generation);
841 }
842}
843
844#[derive(Hash, Eq, PartialEq)]
845enum Key {
846 Memory(usize, u8),
847 #[cfg(unix)]
848 Shm(String, u16)
849}
850
851static ARENAS: LazyLock<Mutex<HashMap<Key, Weak<Arena>>>> =
852 LazyLock::new(|| Mutex::new(HashMap::new()));
853
854fn open(
855 fleet: &Arc<Fleet>,
856 incarnation: Incarnation,
857 spec: PayloadArenaSpec
858) -> Result<Arc<Arena>> {
859 if !crate::waits_supported() {
860 return Err(Error::Io(std::io::Error::new(
861 std::io::ErrorKind::Unsupported,
862 "payload arenas need a platform that can wait on a shared word"
863 )));
864 }
865 let key = if fleet.is_shm() {
866 #[cfg(unix)]
867 {
868 Key::Shm(ring_segment_name(fleet.name(), spec.kind), fleet.node_id().get())
869 }
870 #[cfg(not(unix))]
871 unreachable!("non-Unix fleets cannot use POSIX SHM")
872 } else {
873 Key::Memory(Arc::as_ptr(fleet) as usize, spec.kind)
874 };
875 let mut arenas = lock_unpoisoned(&ARENAS);
876 arenas.retain(|_, arena| arena.strong_count() > 0);
877 if let Some(arena) = arenas.get(&key).and_then(Weak::upgrade) {
878 if arena.incarnation != incarnation {
879 return Err(Error::Malformed(format!(
880 "this process already opened payload kind {} as incarnation {}",
881 spec.kind,
882 arena.incarnation.get()
883 )));
884 }
885 if arena.geometry.slots_per_node != spec.slots_per_node
886 || arena.geometry.slot_size != spec.slot_size
887 {
888 return Err(Error::Malformed(format!(
889 "this process already opened payload kind {} with slots_per_node={} slot_size={}",
890 spec.kind, arena.geometry.slots_per_node, arena.geometry.slot_size
891 )));
892 }
893 return Ok(arena);
894 }
895
896 let geometry = Geometry::new(fleet.fleet_capacity(), spec);
897 let backing = match &key {
898 Key::Memory(..) => {
899 let bytes = AlignedBytes::zeroed(geometry.segment_size);
900 unsafe {
902 std::ptr::write(
903 bytes.ptr.cast::<Header>(),
904 Header::new(fleet.fleet_capacity(), spec)
905 );
906 }
907 Backing::Memory(bytes)
908 }
909 #[cfg(unix)]
910 Key::Shm(name, _) => Backing::Shm(open_shm(name, fleet.fleet_capacity(), spec, geometry)?)
911 };
912 let arena = Arc::new(Arena {
913 _fleet: Arc::clone(fleet),
914 backing,
915 geometry,
916 kind: spec.kind,
917 node: fleet.node_id().get(),
918 incarnation,
919 allocation: Mutex::new(0),
920 credit_wakers: Mutex::new(Vec::new()),
921 driver: Mutex::new(None)
922 });
923 arenas.insert(key, Arc::downgrade(&arena));
924 Ok(arena)
925}
926
927#[cfg(unix)]
928fn open_shm(
929 name: &str,
930 fleet_capacity: u16,
931 spec: PayloadArenaSpec,
932 geometry: Geometry
933) -> Result<ShmRegion> {
934 use std::io;
935
936 let (region, _initialization_lock) =
937 ShmRegion::open_or_create_locked(name, geometry.segment_size)?;
938 if region.created() {
939 unsafe {
941 std::ptr::write(region.as_ptr().cast::<Header>(), Header::new(fleet_capacity, spec));
942 }
943 } else {
944 let header = unsafe { &*region.as_ptr().cast::<Header>() };
946 if header.magic != MAGIC {
947 return Err(Error::Io(io::Error::new(
948 io::ErrorKind::InvalidData,
949 format!("SHM segment {name} has the wrong payload-arena magic")
950 )));
951 }
952 if !header.compatible(fleet_capacity, spec) {
953 return Err(Error::Io(io::Error::new(
954 io::ErrorKind::InvalidData,
955 format!("SHM segment {name} has an incompatible payload-arena layout")
956 )));
957 }
958 }
959 Ok(region)
960}
961
962pub fn segment_size_for(
963 fleet_capacity: u16,
964 spec: PayloadArenaSpec
965) -> usize {
966 Geometry::new(fleet_capacity, spec).segment_size
967}
968
969#[cfg(test)]
970mod tests {
971 use super::*;
972 use crate::exchange::{ExchangeId, Flow};
973
974 fn fleet() -> Arc<Fleet> {
975 Arc::new(Fleet::join("payload-arena-test", 2).expect("fleet"))
976 }
977
978 fn descriptor(
979 publication: &Publication,
980 chunk_id: u64
981 ) -> ChunkDescriptor {
982 let (generation, first, count, len, kind, owner) = publication.coordinates();
983 ChunkDescriptor::new(
984 ExchangeId::from_stream_id(crate::StreamId::from_net_id(orbit_core::NetId64::make(
985 240, 0, 1
986 ))),
987 Flow::AtoB,
988 chunk_id,
989 generation,
990 first,
991 count,
992 len,
993 kind,
994 owner
995 )
996 }
997
998 #[test]
999 fn one_chunk_spans_the_number_of_slots_its_payload_needs() {
1000 let arena =
1001 PayloadArena::open(fleet(), Incarnation::new(1), PayloadArenaSpec::new(241, 8, 256))
1002 .expect("arena");
1003 let bytes = vec![0x5a; 777];
1004 let publication = arena.publish(&bytes).expect("publication");
1005 let descriptor = descriptor(&publication, 1);
1006 assert_eq!(descriptor.slot_count(), 4);
1007 publication.mark_published();
1008
1009 let chunk = arena.read(descriptor).expect("chunk");
1010 assert_eq!(&*chunk, bytes);
1011 }
1012
1013 #[test]
1014 fn dropping_a_chunk_returns_the_whole_extent_as_credit() {
1015 let arena =
1016 PayloadArena::open(fleet(), Incarnation::new(1), PayloadArenaSpec::new(242, 4, 64))
1017 .expect("arena");
1018 let first = arena.publish(&[1; 256]).expect("fills lane");
1019 let descriptor = descriptor(&first, 1);
1020 first.mark_published();
1021 assert!(matches!(arena.publish(&[2]), Err(Error::PayloadFull { .. })));
1022 drop(arena.read(descriptor).expect("consume"));
1023 assert!(arena.publish(&[3]).is_ok());
1024 }
1025
1026 #[test]
1027 fn a_cancelled_publication_never_spends_credit() {
1028 let arena =
1029 PayloadArena::open(fleet(), Incarnation::new(1), PayloadArenaSpec::new(243, 2, 64))
1030 .expect("arena");
1031 drop(arena.publish(&[1; 128]).expect("reserved publication"));
1032 assert!(arena.publish(&[2; 128]).is_ok());
1033 }
1034
1035 #[test]
1036 fn returned_credit_wakes_a_parked_producer() {
1037 let arena =
1038 PayloadArena::open(fleet(), Incarnation::new(1), PayloadArenaSpec::new(244, 2, 64))
1039 .expect("arena");
1040 let publication = arena.publish(&[1; 128]).expect("fills lane");
1041 let descriptor = descriptor(&publication, 1);
1042 publication.mark_published();
1043 let chunk = arena.read(descriptor).expect("held chunk");
1044
1045 let waiting = arena.clone();
1046 let waiter = std::thread::spawn(move || waiting.wait_available(1));
1047 std::thread::sleep(std::time::Duration::from_millis(20));
1048 drop(chunk);
1049 waiter.join().expect("waiter").expect("credit wake");
1050 }
1051
1052 #[cfg(unix)]
1053 #[test]
1054 fn returned_credit_wakes_only_the_owning_producer_node() {
1055 let name: &'static str = Box::leak(format!("pc{:x}", std::process::id()).into_boxed_str());
1056 let spec = PayloadArenaSpec::new(248, 1, 64);
1057 let producer = PayloadArena::open(
1058 Arc::new(
1059 Fleet::join_shm_as(name, 2, orbit_core::NodeId::ZERO).expect("producer fleet")
1060 ),
1061 Incarnation::new(1),
1062 spec
1063 )
1064 .expect("producer arena");
1065 producer.reset_all();
1066 let consumer = PayloadArena::open(
1067 Arc::new(
1068 Fleet::join_shm_as(name, 2, orbit_core::NodeId::new(1)).expect("consumer fleet")
1069 ),
1070 Incarnation::new(2),
1071 spec
1072 )
1073 .expect("consumer arena");
1074
1075 let publication = producer.publish(&[7; 64]).expect("fill producer lane");
1076 let descriptor = descriptor(&publication, 1);
1077 publication.mark_published();
1078 let chunk = consumer.read(descriptor).expect("consume producer chunk");
1079
1080 let waiting = producer.clone();
1081 let waiter = std::thread::spawn(move || waiting.wait_available(64));
1082 let deadline = std::time::Instant::now() + std::time::Duration::from_secs(1);
1083 while producer.arena.local_credit().waiters.load(Ordering::SeqCst) == 0 {
1084 assert!(std::time::Instant::now() < deadline, "producer did not park");
1085 std::thread::yield_now();
1086 }
1087 assert_eq!(
1088 consumer.arena.local_credit().waiters.load(Ordering::SeqCst),
1089 0,
1090 "an unrelated node must not share the producer's credit wait word",
1091 );
1092
1093 drop(chunk);
1094 waiter.join().expect("waiter thread").expect("producer wake");
1095 producer.unlink().expect("unlink payload arena");
1096 }
1097
1098 #[cfg(feature = "tokio")]
1099 #[tokio::test]
1100 async fn returned_credit_wakes_a_pending_task() {
1101 use std::future::poll_fn;
1102
1103 let arena =
1104 PayloadArena::open(fleet(), Incarnation::new(1), PayloadArenaSpec::new(247, 2, 64))
1105 .expect("arena");
1106 let publication = arena.publish(&[1; 128]).expect("fills lane");
1107 let descriptor = descriptor(&publication, 1);
1108 publication.mark_published();
1109 let chunk = arena.read(descriptor).expect("held chunk");
1110
1111 let waiting = arena.clone();
1112 let waiter = tokio::spawn(async move { poll_fn(|cx| waiting.poll_available(1, cx)).await });
1113 tokio::task::yield_now().await;
1114 drop(chunk);
1115 tokio::time::timeout(std::time::Duration::from_secs(2), waiter)
1116 .await
1117 .expect("credit task stayed pending")
1118 .expect("credit task")
1119 .expect("credit wake");
1120 }
1121
1122 #[test]
1123 fn death_reclaims_only_unpublished_or_held_allocations() {
1124 let arena =
1125 PayloadArena::open(fleet(), Incarnation::new(7), PayloadArenaSpec::new(245, 2, 64))
1126 .expect("arena");
1127
1128 let abandoned = arena.arena.reserve(128).expect("reserved by producer");
1129 arena.node_dead(orbit_core::NodeId::ZERO, Incarnation::new(7));
1130 let replacement = arena.publish(&[2; 128]).expect("producer credit reclaimed");
1131 drop(abandoned);
1132 let replacement_descriptor = descriptor(&replacement, 2);
1133 replacement.mark_published();
1134 let held = arena.read(replacement_descriptor).expect("held by reader");
1135
1136 arena.node_dead(orbit_core::NodeId::ZERO, Incarnation::new(6));
1137 assert!(matches!(arena.publish(&[3]), Err(Error::PayloadFull { .. })));
1138 arena.node_dead(orbit_core::NodeId::ZERO, Incarnation::new(7));
1139 let after_reader_death = arena.publish(&[4; 128]).expect("reader credit reclaimed");
1140 drop(held);
1141 let after_descriptor = descriptor(&after_reader_death, 3);
1142 after_reader_death.mark_published();
1143 assert!(arena.read(after_descriptor).is_ok());
1144 }
1145
1146 #[test]
1147 fn committed_unread_payload_is_reclaimed_with_its_dead_producer() {
1148 let arena =
1149 PayloadArena::open(fleet(), Incarnation::new(8), PayloadArenaSpec::new(246, 2, 64))
1150 .expect("arena");
1151 let publication = arena.publish(b"survives").expect("publication");
1152 let descriptor = descriptor(&publication, 1);
1153 publication.mark_published();
1154
1155 arena.node_dead(orbit_core::NodeId::ZERO, Incarnation::new(8));
1156 assert!(arena.read(descriptor).is_err());
1157 assert!(arena.publish(&[9; 128]).is_ok());
1158 }
1159}