Threads

Guest threads, their synchronization and how the host schedules them. Each entry: symptom, cause, fix, the games it was found on. The symptom index lists every symptom.

Thread creation and cores

Publishing a new thread's scheduler fields

• Symptom: TSan reports thread_main writing host_work while thr_runnable reads it during a same-core yield (t_notify_sched, issue #411).
• Cause: CreateThread had already linked the object into g_all, but its pthread initialized host_work, the Mach port and QoS without the scheduler lock G. StartThread could make the object runnable before those writes completed.
• Fix: initialize those shared fields and read the starting priority under G. CreateThread publishes the pthread handle and registry link together after pthread_create returns, rather than letting that call write a published field. The initial main thread obtains its QoS before publication; the priority getter/setter and deferred QoS updates also hold G. Readers keep the same lock, with no added startup wait, timeout or game rule.
• Check: the real start/exit and native-wait paths in t_notify_sched, both preemption modes, run under TSan with the thr_runnable suppression removed. The unsuppressed original reported the startup race. Changing priority before start and consuming the deferred update at a real gate also exposed the unlocked priority setter; the extended test covers that path with DM_QOS_MAP=1.
• Found on: host unit tests; no particular game's hang was attributed to this race.

Starting a higher-priority thread on the caller's core

• Symptom: a parent returns from StartThread before a newly runnable higher-priority child pinned to its core finishes its work (deterministic regression, not attributed to a game failure).
• Cause: StartThread signalled the child's pthread but did not request the priority handoff already used when waking a waiter. Atmosphère's KThread::Run makes the child runnable; scheduler priority rules then apply. It does not promise that an equal-priority or other-core child has entered its function before the call returns.
• Fix: note_runnable records the same handoff for starts and wakes (hle_thread.c). No unconditional child-entry handshake or timeout.
• Check: make -C dm/host/tests GAME=010038B015560000 preempt delays the child's host dispatch by 1 ms and 20 ms. With DM_PREEMPT=0, both delayed starts failed before the fix. Afterward both DM_PREEMPT=0 and 1 pass; the broader mode already supplied the handoff before this correction. Equal-priority, other-core, migratable, host-helper, suspended and host-GPU-work children do not hold up the caller.
• Found on: the threading audit of FINAL FANTASY TACTICS; its affected workers are lower priority and initially migratable, so this correction does not remove its local wake cut.

Core mask changes move the ideal core into the new mask

• Symptom: per-core job queues never drained.
• Cause: a SetThreadCoreMask that pins a thread to other cores (ideal core −3 = keep) left the ideal core outside the new mask, so GetCurrentProcessorNumber reported a core the thread cannot run on.
• Fix: the host moves the ideal core into the new mask, as Horizon does, so GetCurrentProcessorNumber reports a core the thread can run on.

Ideal core −2 resolves to the process default core

• Symptom: crash indexing a per-core table with a huge offset (index −2).
• Cause: threads created with ideal core −2 (use process default) were recorded as core −2, and GetCurrentProcessorNumber returned −2.
• Fix: −2 resolves to the process default core (0), and a negative core is never returned (hle_thread.c).
• Found on: Super Mario Bros. Wonder, in its font engine (main+0x107dbc0).

Free thread count (svcGetInfo type 24)

• Symptom: the main thread waits forever on an event right after nnMain starts, with only three or four threads created; DM_THR=1 shows no CreateThread for the thread it waits on.
• Cause: Unity's thread start (main+0x3e1a50) creates a thread only when nn::os::GetFreeThreadCount() is at least 1. That is svcGetInfo type 24 (FreeThreadCount, Horizon 11 and later), which the host answered 0 like every type it did not know, so the thread was never asked for. The kernel answers the thread resource limit less the threads alive (Atmosphère kern_svc_info.cpp); pm gives an application 96 threads (pm_spec.cpp, 256 with a boost).
• Fix: the host answers 96 less the game's threads that have not exited (thr_limit_used; guest threads the host spawns for itself are left out).
• Found on: Monster Train 2 (Unity 2022.3.54f1 on SDK 18.3.5: main+0x3e7288 starts a thread and waits for it to signal main+0x539cf80). With the fix the game creates 40 threads and reaches its title.
• Note: of the 20 games extracted on this machine, only this main imports the function.

Pausing threads and reading their context for garbage collectors

• Symptom: in garbage-collected titles (Unity / IL2CPP), live objects freed, with their finalizers releasing native resources.
• Cause: the Boehm collector pauses each thread, reads its registers for roots and scans its stack. SetThreadActivity (0x32) and GetThreadContext3 (0x33) were placeholders that only reported success: threads kept running and the context was empty, so roots held only in registers were missed.
• Fix: both really pause and report registers (hle_thread.c). A thread inside an SVC counts as paused at once, with the gate frame as its context (thr_svc_enter / thr_svc_leave around every gate SVC; it waits at the SVC exit while paused). A thread in guest code gets SIGUSR2, and the handler saves the ucontext and waits there. A thread in host code is retried every 100 µs. The context is Horizon's 0x320-byte ThreadContext.
• Found on: Unity / IL2CPP titles in general.

SVCs leave the stack below sp untouched

• Symptom: in a garbage-collected title whose collector stops threads cooperatively, live objects freed while their thread sleeps, and their memory reused: a crash in the collected runtime's code on a pointer made of another object's bytes (Streets of Rage 4: CRASH sig=11 at main+0x46a4, BRUTE's SyncBlock::GetSyncBlock, on 0x5b005b00000002, a string's length 2 and "[["), in about two boots of three, during the first seconds of loading.
• Cause: Horizon's kernel never writes a thread's stack during an SVC. The host's SVC gate opened its 800-byte register frame below the guest's sp and ran the SVC's host code there. BRUTE's Boehm-style collector parks a thread entering a blocking call (main+0x2ff4c80: setjmp, then the sp, published) and scans that stack from the recorded sp up. In Thread.Sleep the SDK path to svc 0xB is all tail calls, so the parking function's frame, which holds the caller's x20 (a loop's only reference to an array), is popped but not overwritten. The scan finds the array there on the Switch; the gate's frame overwrote that word (with the saved fpcr), the collector freed the array, and another thread's string[] took its memory.
• Fix: the gate runs on a stack of the thread's own (entry.S dm_gate, thread.c guest_gate_stack): each thread's KGATE slot points to a static stub that loads its gate-stack top, and the register frame and the host's SVC code go there. A guest thread's gate stack is the idle rest of its own pthread stack, so it outlives the thread's exit from inside ExitThread; the guest main thread, whose guest code runs on its pthread's stack, gets a mapped one. Below the guest's sp only the rewriter's gate block writes, its 16-byte push of x29 / x30. The frame records the block's sp at regs[99], where the thread dump, GetThreadContext3's sp and guest_caller read the guest's stack and return address. No rewriter or game change.
• Found on: Streets of Rage 4 (MonoGame through Sickhead's BRUTE, SDK 9.3.1): after the change, six boots in six reached the frame loop (before, two in three crashed).

Same-core wake handoffs

• Symptom: crashes around movie changes in the game's own player threads, about one run in two: SignalEvent → MultiWaitImpl::NotifyAndWakeupThread on a null MultiWait at address 0x20; WaitAny walking freed holders; a worker reading a freed object at main+0x225338. The other way round, a handoff that waits too widely: a black screen at boot that never loads a texture (packaged textures: first uploads 0), or a load that never ends behind an animated loading screen, with the thread dump's same-core handoffs N, total T ms larger than the run itself.
• Cause: Horizon schedules the highest-priority runnable thread on a core (Mesosphere's scheduler). A lower-priority thread that wakes a higher-priority thread pinned to its core cannot continue until that work blocks. The host's pthreads run in parallel. The earlier approximation also confused host dispatch state with guest readiness: wake() cleared the guest wait before the pthread resumed, but thread_info could still report that pthread waiting. The caller then continued too soon. Its 5 ms timeout opened the same hole under a longer host delay; the recorded yield_pending flag was unused unless the profile enabled checks after every thread SVC. The first unbounded handoff then waited for every runnable higher-priority thread pinned to the caller's core, not only the ones its wake made runnable. The host runs a lower-priority thread beside such threads at all other times, and a core's busy loaders, job workers and pollers are seldom all blocked at the same moment, so a wake could hold its caller for minutes.
• Fix: every game's thread SVC honours yield_pending. Before returning, it waits for the threads its wakes made runnable that are higher priority and pinned to the caller's core, and for those such a thread makes runnable in turn while the handoff lasts (Dread's dispatcher wakes the worker), until each blocks, sleeps, yields to any thread (SleepThread(-2), which takes it off its core so that lower-priority threads run there), exits, is suspended, or enters explicit host GPU work (svc_thread / dm_yield_to_core_peers in hle_thread.c). A wake records the thread in the waker's current handoff and in those the waker belongs to (note_runnable, at most four); blocking ends that (hand_leave), and a handoff ends when its caller's thread SVC returns (hand_epoch). Guest runnable state is authoritative, including the wake-to-pthread-resume interval; a host dispatch snapshot cannot release the handoff, and there is no timeout that lets unfinished work through. A higher-priority peer the wake did not make runnable does not hold the caller, as at every other point on the host. Host-created helpers and threads that can run on another core are excluded.
• Check: make -C dm/host/tests GAME=010093801237C000 preempt delays a woken pthread's dispatch by 1 ms and 20 ms through real event, mutex and condition-variable SVCs. The worker's payload and exit must precede the signaller's return. The original implementation failed both delays; changing readiness alone fixed 1 ms but still failed 20 ms. Equal-priority, other-core, migratable, host-helper, suspended and host-GPU-work peers do not block the caller. Its handoff-scope cases add a busy higher-priority peer that nobody woke: the woken worker must still finish first and the signaller must return (the whole-core wait never returns), a worker woken by the woken dispatcher must finish before the signaller returns (fails without the recorded chain), and a woken thread that yields to any thread releases the signaller (never returns without that rule). The native-mutex case wakes a higher-priority pinned peer while holding the mutex it needs; the signal must return at that native block, and the peer finishes after unlock (hangs without the wait annotation).
• Found on: Metroid Dread, whose per-build movie request-ring gates and core_preempt on setting are now removed; the whole-core wait on The Legend of Zelda: Tears of the Kingdom (issue #247) and Donkey Kong Country: Tropical Freeze (issue #259, section 6; no profile here).
• Native helper joins: dm_host_thread_join also releases the guest core during a native helper's lifetime, preserving an enclosing host-work scope. Xenoblade Chronicles 3 exposed an audio-unit setup join in AKEventManager during a later stall. That sample establishes the missing scheduler annotation, but does not establish an unbounded Apple audio-unit construction deadlock. The preemption regression gates a native helper, verifies that a same-core signaller returns while the peer joins it, then checks completion, the join result and nested host work.
• Scope: these are cooperative thread-SVC handoffs, not a complete asynchronous guest scheduler. The optional core_preempt profile key (DM_PREEMPT=0 / 1 overrides it) additionally checks already-runnable peers after every thread SVC; it defaults off. Historically that broader policy moved Super Mario Bros. Wonder's first frame from 0.6 s to 2.9 s by waiting for busy decompression threads. Wake-triggered handoffs apply with either setting.

A same-core handoff holds only a waker pinned to that core

• Symptom: the picture stops for good while sound goes on; the main thread spins in the game's own wait loop (a sample of the process shows it in guest code, no SVC), the other guest threads wait, and the job workers it waits for sit in svc_thread → dm_yield_to_core_peers → nanosleep.
• Cause: the same-core handoff held the waker until every higher-priority thread its wake made runnable on its core had blocked, whatever the waker's own affinity mask. Batman: Arkham Knight's job workers (priority 43, mask 0x5) finish a job and signal the main thread (priority 29, pinned to core 0), which right away hands them two more jobs and spins on a counter until both are done (main+0x12f8010, the counter main+0x511c0f0); it never blocks, so the worker that signalled it never came back, and as New Story loaded neither job ran. Horizon's scheduler moves a thread preempted on one core to another core of its mask that runs nothing of higher priority (Mesosphere's KScheduler migrates suggested threads), so there the worker goes on on core 2. The host also gave the guest main thread an affinity mask of 7, where Horizon creates every thread, the main thread too, with its ideal core alone (KThread::Initialize).
• Fix: dm_yield_to_core_peers returns at once for a caller whose mask allows another core, and such a waker is not marked for the handoff (note_runnable, hle_thread.c); the main thread's mask is 1 (core 0). A pinned waker keeps the handoff as before. make -C dm/host/tests GAME=<id> preempt has the case (exclusion 7: the signaller's mask is 0x5, and it returns before the woken worker's payload).
• Found on: Batman: Arkham Knight: every route run froze at New Story (frame 7,800 in one, 11,700 in another); with the fix the opening cutscene plays.

Lock-free queues published without barriers

• Symptom: the picture stops and stays (sound goes on) at random places, about one run in two: every task worker waits on its own event, and the task a waiter waits for counts a subtask that is marked queued but is in no queue (the ring's head equals its tail).
• Cause: a game's task ring publishes an item with two plain stores, the item and then the new tail, with no lock on that side, and its pop reads the tail and then the item with no barrier between. ARMv8 lets another core see those stores, and satisfy those loads, out of order; the Switch's Cortex-A57 cores practically never show it, Apple's cores do: a worker sees the new tail with the slot's old null, takes the empty slot (head + 1, the slot cleared) and the item is lost.
• Fix: the game's own code is wrong, so a per-build cut (its patches.py) puts a dmb ishst between the push's two stores and a dmb ishld between the pop's two loads, in the place of reloads of the tail and the head whose values the instructions around them take from registers instead.
• Found on: Batman: Arkham Knight (push main+0xc5adbc, pop main+0xc5ab30), by pull request #340, whose runs froze in 2 of 3 before the cuts and in none of 5 after; another engine's queue of the same shape can drop items the same way.

A host lock held across a same-core handoff

• Symptom: the game freezes when a movie starts playing beside the game (a guide movie in a corner of the screen): the last frame stays, the process idles at a few percent of a core, and the thread dump shows a lower-priority job thread running/not waiting with its backtrace in the movie player (ReleaseOutputBufferIndex), the main thread waiting for it, and the movie's input thread running/not waiting too. A sample of the process shows the job thread in md_pump → md_event → dm_yield_to_core_peers, and the input thread in h_movie_SendIn → md_pump waiting for a pthread mutex.
• Cause: the host's movie decoder holds its decode lock (dmu) across guest code: the "input available" and "output ready" events are signalled with the SDK's SignalEvent from inside md_pump. The signal woke the movie's higher-priority threads pinned to the signaller's core, so the signaller's same-core wake handoff waited for them to block before its SVC returned. One of them, the input thread, had called SendInputBufferForDecode and was waiting for dmu in host code. A thread waiting for a host lock still counted as runnable, so the handoff never ended and the lock was never released. On Horizon that lock is guest-visible (the SDK's decoder holds an nn::os mutex there), and a thread waiting for it is blocked.
• Fix: dm_host_lock (hle_thread.c) takes a host lock that HLE code holds while it runs guest code. A guest thread that has to wait for it is in a native HLE wait (dm_thread_native_wait) while it waits: blocked for thr_runnable, and out of the handoff that made it runnable, so the handoff does not wait for it. The movie decoder takes dmu and the deferred-event lock g_dfmu with it (hle_movie.c). Releasing dmu before signalling instead would let a signal land after Stop has returned, on events the game is tearing down (the reason md_defer_body signals under its lock).
• Check: make -C dm/host/tests GAME=010093801237C000 preempt, handoff scope "woken thread waiting for the signaller's host lock": the signaller holds a host lock across its signal and the woken higher-priority worker waits for it with dm_host_lock. The signaller must return before the worker finishes; without the native wait around the lock the test never returns.
• Found on: Kirby and the Forgotten Land: the first guide movie ("A Jump", 480×270 VP9) froze the game on the first walk from the beach, every time. The job worker (priority 46) and the movie threads (priority 44) are all pinned to core 0.
• Found on: Xenoblade Chronicles 3, whose 140-second opening VP9 movie stopped at its first output release. A live stack sample identified the input thread in md_pump waiting on dmu, and the output thread holding it inside SignalEvent → dm_yield_to_core_peers.

Simultaneous crashes of several threads

• Symptom: several threads crash at once and the dumps interleave, then the process hangs instead of exiting.
• Cause: Unity's job threads tend to crash together.
• Fix: crash.c lets only the first crashing thread print (crash_gate); the others stop in pause(), and a 15 s alarm ends the process if the dump itself stalls.
• Found on: Unity titles in general.

Synchronization and address arbitration

SignalToAddress type 2 with count ≤ 0

• Symptom: none seen; found while investigating Fire Emblem Engage's demo crash, which it did not change (that was the address-space shortage, Size-aligned regions exhaust the alias address space).
• Cause: for SignalToAddress type 2 (SignalAndModifyByWaitingCountIfEqual) with count ≤ 0 and waiters present, the new value is value − 2, not value − 1 (mesosphere KAddressArbiter; Ryujinx and yuzu agree). The host had −1.
• Fix: the host writes value − 2, the specified behaviour.

WaitForAddress type 3: 64-bit wait if equal

• Symptom: a 64-bit wait could return at once or block when it should not (not met in a game).
• Cause: Horizon 18 added a 64-bit "wait if equal" arbitration type, which SDK 19 exports as nn::os::WaitForAddress(const long*, long); the host used to fall into its 32-bit "wait if less than" test.
• Fix: type 3 compares the 64-bit value (hle_thread.c).
• Found on: Pokémon Legends: Z-A (SDK 19.3.11): Z-A does not call it, but the SDK has it.

Unity's stream buffer handshake loses a wake-up on Apple cores

• Symptom: boot stops before or soon after the first frame with the CPU at 0% and every guest thread waiting on a condition variable: the main thread in nn::os::AcquireSemaphore after posting a command to UnityGfxDeviceWorker, the worker in AcquireSemaphore too, one condition wait after it (wait # one higher); 3 of 5 Mac runs, at random points.
• Cause: a memory-order race in the game's code, not a host bug. Unity 2018.4's ThreadedStreamBuffer (the ring buffer every Unity worker thread is fed through) wakes a sleeping side with a Dekker handshake: the reader that finds nothing sets "reader sleeping" with a store-release and loads the write position again before sleeping on its semaphore; a writer stores the write position with a store-release and then checks that flag (and the mirror pair for a writer short of space). The second access of each side is a plain ldr, and AArch64 lets a plain load complete before an earlier store-release is visible, so both sides can read the old value and neither signals. The Switch's Cortex-A57 allows it too, but Apple cores hit it within seconds.
• Fix: a per-build cut turns each of those loads into a load-acquire (ldar), which a store-release before it orders after the store (eight sites in that build's patches.py); 14 of 14 runs boot and keep presenting.
• Found on: FINAL FANTASY CRYSTAL CHRONICLES Remastered Edition 1.0.0 and 1.0.3 (Unity 2018.4.2f1, SDK 8.2.1; its patches.py: 1.0.0 main+0x1e425b0 and the stream functions after it, the same sites +0x25380 in 1.0.3).
• Check: another Unity title that stops this way: look for stlr followed by a plain ldr of the other word in its ThreadedStreamBuffer (the functions around the AcquireSemaphore the worker waits in).

Lock acquire without acquire ordering runs a queue item twice

• Symptom: a game that runs on the Switch crashes on the Mac 10 to 60 s into a scene, at a different place each run, with garbage in a list the game just built (Breath of the Wild 1.6.0: h_nvnBufferGetAddress called with a row's metadata word as the buffer pointer, main+0x14ef6e0; also faults in main+0x12970dc, +0x13615f4, +0x1445d48, +0x1472470). A trace of the draw-list builders shows two threads inside the same view at once, running the same job descriptor.
• Cause: the game's own spin lock takes the lock with ldxr w9,[x23]; cbnz w9,<ldxr>; stxr w9,w8,[x23]; cbnz w9,<ldxr>: an exclusive load with no acquire ordering, and no barrier after the loop. The architecture lets the loads that follow (the queue's count, head and ring) be satisfied before the lock is held; the game ran on the Switch's Cortex-A57 without that showing (why is not known), Apple cores do it, so two workers read the same head, both pop it and both run the job. The unlock is ordered (dmb ish before the store), and Nintendo's SDK locks use ldaxr. The same loop followed by dmb ish is ordered and harmless: Tears of the Kingdom's main has 32 such, Super Mario Bros. Wonder's 33, all fenced. In Breath of the Wild 1.6.0 the job queue pop (main+0x13a0784) and submit (main+0x13a0068) have the bare form, 412 loops of this shape in the module (counted without the barrier check; 1.9.0 has none of either form); Super Mario 3D All-Stars' main has 7.
• Fix: patch_nso.py makes every such acquire an ldaxr (acquire_lock_sites), in every module of every game, found by the instruction shape (an exclusive load, a cbnz back to it, a store-exclusive of a non-zero value to the same address, a cbnz back to the load), unless a dmb ish or dmb ishld follows the loop. No address list. dm/tools/tests/test_acquire_lock.py checks the shape on hand-assembled words.
• Found on: The Legend of Zelda: Breath of the Wild 1.6.0 (issue #251): the same job descriptor ran on two threads 8 us apart.
• Check: another game that fails like this: trace which job a worker takes (a branch slot at the job entry, a lock-free per-thread ring); the same descriptor on two threads at once is this bug. Count the unfenced bare ldxr acquires (acquire_lock in the rewriter's statistics line).

Late PhysX task destructor after the unlock

• Symptom: a PhysX worker crashes in a task's removeReference with a null task manager, right as the game switches to 30 fps (pc=main+0x10ccf28, lr=main+0x13c236c); on the iPhone and in 2 of 5 Mac runs, in every Mac run with 12 busy loops loading the CPU.
• Cause: the game publishes completion before task teardown ends, then reuses the same task. Its main thread rebuilds each world's completion task before simulate; the worker sets done, notifies and unlocks inside run, then calls release, whose destructor clears mTm. Another core may rebuild it in between. This is a guest lifetime race: an unlock cannot protect subsequent destructor writes. PhysX's task contract places recycling in release, after run, and its PxLightCpuTask destructor clears mTm. There is no measured Switch timing guarantee that makes the ordering safe. The game's offsets, affinities and replay evidence are in its note.
• Fix: a per-build cut makes that one release a ret (its patches.py); the destructor frees nothing and the step rewrites every field.
• Check: dm/tools/check_thread_cuts.py replays the original construction, release and removeReference instructions with a forced late release: the original faults at the recorded null-manager load; normal release-before-rebuild succeeds; the registered cut preserves the new task. This controlled replay has no host mutex or graphics implementation in its failure path.
• Found on: The Legend of Zelda: Link's Awakening. How it was found is in 05.

The thread ID at ThreadLocalRegion+0x110 (SDK 22)

• Symptom: game called nn::diag abort: Result 0, call site sdk+0x153c34 during main's static initialisers, bt: sdk+0x153c34 sdk+0x137eb4 sdk+0x13dc50 sdk+0x13fa90 sdk+0xa4d64 main+0x134 run_module_inits+0xa8 guest_main+0x15c main+0xfc start+0x1a20, and no game code on the stack at all.
• Cause: those frames are pthread_mutex_lock+0xa4, nn::nlibsdk::heap::TlsHeapCentral::LockMutexSlow, nn::nlibsdk::heap::CentralHeap::MakeCache and nn::mem::StandardAllocator::Allocate, called out of __nnmusl_init_dso: the first heap lock of the SDK's own allocator. pthread_mutex_lock takes a slow path when InternalCriticalSection::IsLockedByCurrentThread reports the mutex is the caller's, and that check is mrs tpidrro_el0; ldr w8,[x8,#0x110]; ldr w9,[x0]; and w9,w9,#0xbfffffff; cmp w8,w9; cset w0,eq — the owner is the 32-bit thread id at ThreadLocalRegion+0x110, which SDK 22's InternalCriticalSection::Enter (sdk+0x14b560) writes from there. The host wrote only +0x1F8 (the thread object pointer), which is what SDK 16's mutexes compared, so +0x110 stayed 0, which is the owner word of an unlocked mutex: every unlocked mutex read as the caller's, and the abort fired on the first one.
• Fix: the host stores each thread's kernel ID, the one svcGetThreadId answers (thr_t id, from 0x51 up, never 0), at ThreadLocalRegion+0x110 as well as the thread object at +0x1F8: guest_thread_set_id in thread.c, called by hle_thread.c when a guest thread starts and for the main thread before any guest code runs. SDK 22 still wants +0x1F8 for the current thread: ThreadManager::CreateThread (sdk+0x149a78) and its thread init (sdk+0x149ff0) both str x0,[tpidrro_el0+0x1f8].
• Check: during the bring-up the thread's ThreadLocalRegion showed +0x110 at 0 with +0x1F8 set; with the fix the allocator's first lock takes the fast path and the boot reaches nnMain.
• Found on: Well Dweller (SDK 22.2.0). Any SDK 22+ game that locks a mutex before nnMain would hit this; a game whose mutexes are all taken by a thread that already owns them would not notice.

Yields and spin waits

The yield back-off (DM_YIELD_BACKOFF) and the rewriter's spin waits (patch_nso.py "Spin waits", dm_spin_wait in hle_thread.c), with their measurements, are in development.md ("Yield loops", "Spin loops").

Timer slack in guest sleeps (sleep_precise)

• Symptom: a 30 Hz game thread runs at 27 Hz, and whatever it paces comes out slow (cutscenes at 36 ms per frame instead of 33.3; gameplay was fine at 60).
• Cause: macOS timer slack: nanosleep and mach_wait_until overshoot by about 10% at any QoS (measured: 33 ms → 36.4 ms, 5 ms → 6.1 ms, 1 ms → 1.26 ms). Dread's sound thread (main+0xf95b0) reads a millisecond clock, updates, then sleeps "33 ms minus this round's work", so on the host its period was 36.9 ms. The cutscene loop renders a frame whenever the audio play position, which that thread refreshes, moves (main+0x119a3d0; the position is played samples ÷ sample rate of the cutscene's stream), so cutscenes ran at about 27 fps.
• Fix: SleepThread sleeps three quarters of the remaining time per step until within 20 µs (sleep_precise in hle_thread.c).
• Found on: Metroid Dread: cutscenes at 33.3 ms, the sound thread at 33.2 ms. Super Mario Bros. Wonder's frame-time p95 also went from 19.7 to 18.4 ms.
• Note: timed waits (WaitSynchronization, condition variables) still use the host's own timeouts.

Sleeps under 20 µs take the yield path

• Symptom: a guest thread keeps a core at 100% with billions of SleepThread calls.
• Cause: sleep_precise returns at once for anything under 20 µs (its nanosleep would overshoot), so a loop of such sleeps never sleeps.
• Fix: sleeps under 20 µs take the yield path and its back-off (plain yields for the first 2 ms of a run, then short sleeps).
• Found on: Super Mario Galaxy: Hagi polls with SleepThread(100 ns) (main+0x517c), 1.4 billion calls in 200 s, and the render thread nearly as busy. The two threads went from 199 s and 171 s of CPU per 200 s to 5 s and 19 s per 120 s, at the same frame rate.

Short timed waits are polls

• Symptom: a thread polls at full speed although the power line's yields (if any) show no back-off sleeps; a sample of the host shows it in svc_thread → block → pthread_cond_wait with timeouts of nanoseconds, between plain cthread_yields.
• Cause: a guest loop that waits with a timeout of a few hundred nanoseconds (a condition variable, an event) until another thread signals. Each wait times out at once, as it would on Horizon, so the loop spins. When it also sleeps under 20 µs between its waits, those sleeps take the yield path (Sleeps under 20 µs take the yield path), but each wait was another SVC between two yields, so every yield began a new run and the back-off never started; a loop of waits alone had no yield at all.
• Fix: a wait with a timeout under 20 µs (WaitProcessWideKeyAtomic, WaitSynchronization, WaitForAddress) is a step of the run of yields (poll_wait_begin / poll_wait_end in hle_thread.c): it continues the run when the SVC before it was one of the run's steps and under SPIN_GAP_NS (20 µs) of guest work came after it, and once the run has lasted the back-off's start (2 ms) its timeout becomes the back-off's sleep. A lengthened wait still ends at once when it is signalled. A wait that is signalled ends the run. The yields' register test does not apply to waits, since the SDK's own frames hold the time left at the SVC (SDK 3.5.1's InternalConditionVariableImplByHorizon::TimedWait keeps it in x21); a loop that works more than 20 µs between its waits starts a new run at each one.
• Found on: Hyrule Warriors: Definitive Edition (KTGL, SDK 3.5.1): its main thread waits for its frame's event with TimedWaitEvent(30 ns) and SleepThread(1 µs) in turn, its update thread with TimedWaitEvent(500 ns) alone. In its first battle at 60 fps on the Mac: CPU 102% → 27% of a core, CPU energy 4.6 → 1.5 W, wakeups 358,000 → 72,000 a second, frame time p95 17.5 → 16.9 ms.
• Check: make -C dm/host/tests GAME=<id> build/t_pollwait && dm/host/tests/build/t_pollwait: loops of waits alone and of waits between short sleeps (condition variable and WaitSynchronization) back off after 2 ms, a loop with 30 µs of work between its waits does not, and a lengthened wait returns at the signal.

Yields that make progress return at once (thr_yield)

• Symptom: a loading screen sits for a minute with the CPU nearly idle; one thread yields thousands of times a second, every yield a back-off sleep (the power line's top yielder, with as many back-off sleeps).
• Cause: the thread was working, not waiting: its loop calls nn::os::YieldThread once per item, and the back-off made each yield a 250 µs sleep.
• Fix: a yield whose caller's x19–x28 changed since the last one is progress (thr_yield in hle_thread.c; the SDK reaches the SVC through tail branches, so those registers are the loop's own): it returns at once with no host call, and the run starts again. Polls keep those registers (Three Houses' limiter, Galaxy 2's Hagi sleep poll at main+0x11ec); counters move them.
• Found on: Animal Crossing: New Horizons: a black screen with the plane icon for 77 s after the opening movie, SaveProc at about 2,600 yields a second, its grid walks over the new island (main+0x157d020) yielding once per cell; also about 40 s of the same at boot. The island now shows about 10 s after the movie, as with DM_YIELD_BACKOFF=0. Polls that keep their registers: Fire Emblem: Three Houses, Super Mario Galaxy 2.
• Check: check such a change on The Legend of Zelda: Breath of the Wild outdoors: its PlacementMgr and RadarMgr turned out to yield per item too, and a sched_yield on each of their 200,000 yields a second cost a third of a core more than the back-off had; with no host call, CPU energy fell instead (numbers in development.md "Yield loops").
• Note: a temporary log of which of x19–x28 move at each yield, per thread, is what told the two kinds apart.

A wait at two yields is not progress

• Symptom: one guest thread keeps a host core at 100% with about a billion yields a minute (SVC 0xb in DM_HLECOUNT), none of them a back-off sleep, while the game itself plays normally.
• Cause: the loop waits at two yields whose callee-saved registers differ, so each yield looked as if the loop had moved on since the last one, and returned at once. Hades' NetworkPollThread (main+0x27aaa0) yields between its passes with w23 = 0, and in each pass yields while the nifm request is on hold with w23 = 2,000, its retry count.
• Fix: a yield is progress only when x19–x28 differ from both of the last two yields (thr_yield in hle_thread.c): a loop that comes back to the registers of the yield before last is waiting, and the run goes on to the back-off. Counters and cursors (Animal Crossing's grid walk, Breath of the Wild's batches) never come back, so they still return at once.
• Found on: Hades, with Network requests stay on hold before they settle. Neither change alone was enough: in 60 s of its smoke route the process used 117 s of CPU on main, 119 s with only the nifm change (627 million submits, a pass now ended at the second query) and 122 s with only this one (1.38 billion queries, the 50 ms hold still timed out). With both it used 62 s, as with the poll thread's yields turned into sleeps: 216,000 yields, most of them back-off sleeps.
• Check: a temporary per-thread count of the yields only this rule keeps in a run, and of the back-off sleeps those runs then took, on boots and play of other games: Tears of the Kingdom's main thread and ModuleSystem workers about 0.1% of their yields (761 of 926,116 on the main thread in 120 s) and no back-off sleep; Breath of the Wild outdoors at most 7 per thread in 150 s, no back-off sleep; Three Houses up to 95 per thread, Donkey Kong Country: Tropical Freeze about 1,000 on the main thread, neither with a back-off sleep from them; Kirby at most 1; Celeste, BALL x PIT, Bayonetta and Ghost Trick none. Their frame rates matched main in runs that kept the window's focus (a run whose window loses focus to another app drops to about 30 fps, on main as well).

Events and notifications

Service-event producers follow the binding lifetime

• Symptom: a host helper thread aborts when signalling a destroyed SystemEvent; repeated vi::GetDisplayVsyncEvent calls double the event's tick rate. Rapid replacement can abort with too many vsync events, or a retired vsync worker can signal another service's event at reused storage.
• Cause: lazy binding calls the SDK's CreateSystemEvent, which requires a guest thread. Checking only the guest address also mistakes a replacement binding for the original event, even if its kernel handle number is checked: both can be reused before the next tick. An eight-slot worker table counts destroyed events until their workers next run.
• Fix: hle_signal_system_event retains lazy binding on guest threads; host threads skip an unbound event and log once. Each service-event binding gets a monotonically increasing identity. hle_signal_system_event_binding matches that identity and signals under the same lock as CloseHandle, so a stale producer cannot signal a replacement. The vi registry keeps one worker per binding identity in an allocated list; a worker whose binding closed removes only its own entry and exits. A new vi binding at reused storage gets its own worker, while repeated requests for a live binding share the existing one.
• Found on: synthetic host regression cases in t_system_event.c, using real binding and event SVCs with worker startup held until after destruction and handle reuse; no game failure is claimed.
• Check: duplicate vi registration, nine rapid replacements before the first tick, storage and handle reuse by a different service, vi rebinding at reused storage, host signalling after destruction, and guest lazy binding. The capacity and unrelated-service cases fail on the original PR #202 and pass with lifetime identities and the allocated registry.

One auto-clear event shared by a pool's workers

• Symptom: the main thread waits forever on a job's done event at boot while every worker sleeps, two of them on one event (main+0x1426c waiting on Job::m_doneEvent); 4 of 6 Mac runs.
• Cause: the engine puts a job in worker k's private queue but signals one shared SDK auto-clear event. Its condition wake chooses by priority, then FIFO; it has no target-worker identity. The wrong worker checks only its own queue and sleeps again. This is a guest assumption about waiter order, exposed by host timing. Host pick / wseq agrees with Atmosphère's condition-tree comparator and Signal, including the historic 0.19.0 implementation. This describes the SDK's condition-based auto-clear event; kernel readable events instead notify all synchronization waiters. Game offsets and observations belong in its note.
• Fix: the game's own wait wrapper (main+0x12500, {event, pending}) signals the event once more when it takes a nonzero pending word after waking, and nothing set it: a build cut makes the signal set it first (its patches.py), so a worker woken in another's place passes the wake on; 6 of 6 runs boot.
• Found on: FINAL FANTASY TACTICS - The Ivalice Chronicles.
• Check: t_preempt checks priority/FIFO selection and mutex reacquisition through real condition SVCs. dm/tools/check_thread_cuts.py runs the game's actual signal and wait wrappers with both worker arrival orders: only the original wrong-worker cases strand a job. The game's note records the Mac boot comparison with the corrected host.
• Note: an unconditional StartThread child-entry handshake changed the timing but did not eliminate the hang and is not a kernel guarantee. The separate same-core start handoff applies only when priority and affinity require it.

The oe notification event and its queue

• Symptom: abort in os::WaitSystemEvent under oe::PopNotificationMessage (The Legend of Zelda: Link's Awakening); one guest thread at about 1000 ms/s on the iPhone only, the phone hot, CPU energy 5–7 W (Monster Hunter Rise's MessageSuperviseThread, priority 32, in the [nvn] threads lines).
• Cause: the host keeps the notifications in a queue of its own. The blocking pop waits on the system event oe::Initialize would have created, and the host's Initialize is a no-op. Monster Hunter Rise's thread waits on oe::GetNotificationMessageEvent() with os::WaitAny, which does not clear an event, then pops notifications until the queue is empty. The host only ever signalled its notification event, so after the first focus change (iOS changes focus at launch; the Mac does not) it stayed set and the loop never blocked again.
• Fix: the host now implements the blocking pop on its notification queue (a condition variable, then the TryPop path), and the event follows the queue as the am service's does: set while a notification is queued, cleared when the last one is popped (hle_oe.c mq_ev_sync, hle_clear_system_event).
• Found on: The Legend of Zelda: Link's Awakening (SDK 8.2.0, the blocking pop); Monster Hunter Rise (the event).
• Check: Mac repro DM_OE_SCRIPT=focus=2@5,focus=1@6 (the script also starts at the game's first GetNotificationMessageEvent, since such a game pops only once the event fires): before 995 ms/s, CPU 109%, 6.97 W; after, the thread is idle and CPU energy is 0.05 W.

Native HLE waits release the guest core

• Symptom: Paper Mario: The Thousand-Year Door 1.0.1 stays black before its first frame. The main thread stops at main+0x32796c, after starting SYS_Notification; the priority-28 notifier is in oe::PopNotificationMessage, while the priority-44 main thread waits for its same-core wake handoff.
• Cause: the notification HLE blocks in a host pthread_cond_wait, outside the guest kernel's wait bookkeeping. The notifier remains classified as runnable and retains its inherited handoff, so its lower-priority waker cannot continue even though the notifier has no work.
• Fix: dm_thread_native_wait marks the blocked native-wait interval under the scheduler lock and drops the inherited handoff. PopNotificationMessage enters that state while holding the queue mutex, waits, and restores runnable state after reacquiring the mutex. Ordinary native execution is still runnable.
• Found on: Paper Mario: The Thousand-Year Door.
• Check: the original PR #291 reproducer on upstream 1402ded2 blocks until its watchdog exits with code 99. make -C dm/host/tests GAME=0100ECD018EBE000 notify-scheduler now uses real CreateThread, StartThread, wait-for-exit and close SVCs; only the assembly transition to a guest stack is replaced by a C test entry. The original notification HLE strands StartThread until the watchdog exits with code 99. The fixed HLE passes 100 runs in each scheduling mode, including a wake without a queued message, runnable-state restoration and delivery of focus notification 15. The full host test suite (make -C dm/host/tests GAME=0100ECD018EBE000 run) also passes with CoreAudio access. The game's note records its separate Mac startup comparison and iPhone report.