semaphore-handshake-audit

name	semaphore-handshake-audit
description	Audit LLK inter-thread synchronization (Tensix semaphores + ATGETM/ATRELM mutexes) for races/deadlock — SEMINIT correctness vs usage, post/get balance, wait-direction, RISC-MMIO-vs-Tensix ordering, and mutex acquire/release balance. Use after touching any t6_semaphore_*/semaphore_post/semaphore_get/SEMINIT/SEMWAIT/SEMPOST/SEMGET/t6_mutex_* or any math↔pack / unpack↔math handshake (MATH_PACK, UNPACK_TO_DEST, UNPACK_SYNC, MATH_DONE, FPU_SFPU).
user_invocable	true

/semaphore-handshake-audit — Inter-thread semaphore & mutex correctness

Ground-truth precedence: the live ISA doc (tt-isa-docs MCP, fetched each run) outranks every rule, table, and example baked into this skill — treat those as dated illustrations. If the live ISA doc contradicts a baked rule here, do NOT silently proceed: surface the conflict to the user and ask whether the baked rule should be overwritten, discarded, or kept. Default to the ISA doc.

MANDATORY — before any verdict, read the shared grounding policy. The per-architecture source ladder (which docs to consult), the ground-or-abstain rule, and the Source preflight (list the sources you'll consult with their reachability + hierarchy, then PAUSE for the user) are defined once in race-audit-all → .claude/skills/race-audit-all/SKILL.md. Your FIRST action is to Read that file and follow its "Ground-truth source ladder", "Ground-or-abstain", and "Source preflight" sections — they are load-bearing: a verdict produced without them is ungrounded and MUST NOT be reported. If that file genuinely cannot be read, say so and abstain rather than proceed ungrounded. (If you were spawned by a race-audit-all sweep — your prompt already lists the confirmed sources — skip the Source preflight and do not pause; the orchestrator ran it once.)

Coverage — floor, not ceiling. The grep patterns and site lists in this skill are a seed, not an exhaustive enumeration. After running them, widen the search with full reasoning. The techniques here are illustrative examples, not the allowed set — use any approach your reasoning suggests, including ones not listed: e.g. semantic search (by behavior/effect, not just token), resolving macros / wrappers / typedefs / indirection the literal pattern can't match, following the call graph to callers and callees, and diffing the WH/BH/QSR variants to catch a site present in one arch and missing in another. If you can find a hazard, primitive, or site the encoded patterns don't cover — by any means — pursue and report it; do not clamp a stronger analysis to this list or to these techniques. State any residual coverage gaps explicitly (no silent caps).

Execution — parallel by default. When enumeration yields more than a few sites/files, fan out concurrent Agent calls by default (one per file/subsystem, a fresh context each), saturating the available concurrency (~10–16 at once); go inline only for a trivial set. The per-file fan-out described under Thoroughness is the default, not an exhaustive-only option. The cross-referencing/synthesis of results stays sequential (it must follow the per-unit findings). The heavyweight Workflow tool still requires explicit multi-agent opt-in — it is the opt-in exhaustive tier, not the default. Don't over-spawn a tiny diff.

Persisting results — single writer, incremental. Agents only return their findings; they never write a shared file (no concurrent-write clobbering). If findings are persisted to a file, the orchestrator/caller is the sole writer and appends each wave's returns as they arrive — incremental, never only-at-the-end — so an interrupt preserves every completed wave's findings.

The bug class (precise)

The three Tensix threads coordinate through 8 hardware semaphores (Semaphores[0..7], each a 4-bit Value 0–15 plus a Max) and two ATGETM/ATRELM mutexes. Unlike config-register races (covered by cfg-word-overlap-audit, reconfig-stall-audit, mmio-race-audit), these bugs are in the handshake protocol: a mis-initialized, unbalanced, wrong-direction, or wrongly-ordered post/wait/get → lost synchronization (silent data corruption) or deadlock (TENSIX TIMED OUT).

Ground-truth HW semantics (confirm via tt-isa-docs MCP: SEMINIT.md, SEMWAIT.md, SEMPOST.md, SyncUnit.md)

• SEMPOST → Value++ capped at 15 (HW). SEMGET → Value-- floored at 0. Both silently no-op at the limit.
• SEMWAIT(block_mask, sem_mask, cond): stalls the thread's instruction issue while the condition holds. STALL_ON_MAX = block while Value == Max (wait for a consumer to drain). STALL_ON_ZERO = block while Value == 0 (wait for a producer to post).
• SEMINIT(NewMax, NewValue, sem_mask) sets both Value and Max atomically. Crucial: Max is used ONLY by SEMWAIT; SEMPOST ignores it and still caps at 15. So Max is purely the wait threshold — it does not bound posting. Backpressure exists only if the producer calls wait_on_max before each post.
• LLK wrappers (ckernel.h): t6_semaphore_init(idx, min, max) → TTI_SEMINIT(max, min, …) (so wrapper min = initial Value, max = Max). t6_semaphore_post/get are Tensix SEMPOST/SEMGET (in-stream, ordered). semaphore_post/get (no t6_) are RISC MMIO writes to pc_buf_base[PC_BUF_SEMAPHORE_BASE+idx] — asynchronous to the Tensix stream. The LLK_ASSERT(value<MAX) / (value>0) guards are debug-only → release builds silently over/underflow.

Semaphore map (WH/BH — ckernel_structs.h)

#	Name	Roles
0	FPU_SFPU	fpu↔sfpu; also reused (aliased SFPU_FPU) in some experimental kernels
1	MATH_PACK	math↔pack on dest register (the main double-buffer)
2	UNPACK_TO_DEST	unpack↔math, unpack-direct-to-dest
3	UNPACK_OPERAND_SYNC	unpack↔pack/math operand get/release
4	PACK_DONE	pack iteration begin/end (perf)
5	UNPACK_SYNC	RISC↔unpack, config-context acquire/release
6	UNPACK_MATH_DONE	(see name reuse note below)
7	MATH_DONE	math-done barrier for unpack-to-dest

What to check — and the rules

1. INIT correctness vs usage (do this FIRST — it's the most-missed)

• Every semaphore used with wait_on_max/STALL_ON_MAX MUST be SEMINIT'd with Max = the intended buffer depth, by one thread, before any thread uses it. If it relies on the HW-reset default Max, the producer's backpressure gate is wrong → over-post (silent corruption) or premature/never block.
  • In-tree, only MATH_PACK is SEMINIT'd inside tt-llk: _llk_math_pack_sync_init_ (llk_math_common.h) sets Max=1 (DstSync::SyncFull) or Max=2 (SyncHalf), Value=0. UNPACK_TO_DEST, MATH_DONE, PACK_DONE have no SEMINIT in tt-llk — they are boot-initialized by BRISC firmware, not the compute-API startup. Ground truth (verified): brisc.cc → c_tensix_core::initialize_tensix_semaphores() (tt_metal/hw/inc/internal/.../c_tensix_core.h) runs once before any TRISC kernel and issues ex_sem_init(sem, max, value) = Max=1, Value=0 for MATH_PACK, UNPACK_TO_DEST, MATH_DONE, PACK_DONE. So all are depth-1 by default and init-ordered before use; MATH_PACK is then re-SEMINIT'd per DstSync mode by _llk_math_pack_sync_init_. The LLK test harness mirrors this in tests/helpers/include/boot.h. When auditing, confirm any NEW wait-on-max semaphore is added to initialize_tensix_semaphores (or re-init'd in LLK); flag any wait-on-max semaphore with no reachable SEMINIT in BRISC firmware or tt-llk.
• Initial Value must match the empty/full convention — producer/consumer semaphores start at Value=0 (empty). A stale non-zero value (no re-init across kernels, or across a DstSync mode switch) desyncs from cycle one.
• The issuing thread + ordering: semaphores are thread-agnostic (single bank), but SEMINIT executes in one thread's stream. It must land before any thread's first wait/post/get. Verify the init barrier (e.g. _llk_math_pack_sync_init_ does tensix_sync() + while(semaphore_read(MATH_PACK)>0){} to drain prior packs before re-SEMINIT). Missing barrier → init races a peer still using the old value.
• Mode-switch re-init: SyncFull↔SyncHalf need Max 1↔2. Changing dest-sync mode without re-SEMINIT (and a dest drain) leaves Max mismatched to the real buffer depth.
• Max ≤ 15 and ≥ the deepest outstanding count the producer can post between waits. Since SEMPOST ignores Max, ALSO confirm usage rule #2.

2. Post/get (producer/consumer) balance

• Exactly one SEMGET per SEMPOST over any complete iteration. Walk every branch: a post inside an if whose get is unconditional (or vice-versa), or an early return/continue between them, drifts Value → producer eventually wedges at Max forever (deadlock) or consumer drains a slot that was never filled.
• Because SEMPOST ignores Max, every producer post must be preceded (in that thread's program order) by the wait_on_max gate — that gate is the only thing preventing overshoot past the intended depth. A post on a path that skips the wait → silent overflow toward 15.
• Cross-layer handshakes (don't false-flag a deadlock): for experimental/custom ops, the producer and consumer halves can live in different layers — an LLK function may provide only one half (e.g. a wait_on_zero+get) while the matching post is issued by the compute kernel / model layer (ttnn/cpp/.../kernels/..., models/.../kernel_includes/..., tt_metal/hw/inc/api/compute/...), exactly like MATH_PACK's halves sit in cmath/cpack but are assembled by the compute API. So a wait_on_zero with no producer in tt-llk is NOT automatically a deadlock — chase the producer repo-wide (grep -rIn '<SEM>' ttnn models tt_metal/hw/inc/api) before flagging. Only call it a deadlock if no reachable layer posts it.

3. Wait-direction

• Producer side waits STALL_ON_MAX (block while full); consumer side waits STALL_ON_ZERO (block while empty), then SEMGET. Swapped condition → either no synchronization (proceeds immediately) or permanent stall. Canonical good pair: math wait_on_max→compute→post(MATH_PACK); pack wait_on_zero→pack→get(MATH_PACK).

4. RISC-MMIO vs Tensix-instruction ordering

• semaphore_post/semaphore_get (no t6_) are RISC stores that execute asynchronously to the Tensix stream (overlaps /mmio-race-audit). When a RISC semaphore_post is meant to gate a Tensix consumer, there must be a SEMWAIT/TTI_SEMGET that actually stalls on it (the UNPACK_SYNC context-acquire idiom: RISC semaphore_post → STALLWAIT(STALL_UNPACK, TRISC_CFG) → MOP → t6_semaphore_get). A bare MMIO post/get with no in-stream wait can land early/late. On Quasar, HW AutoTTSync changes these RISC↔Tensix ordering rules (read Confluence 1340276980 at audit for what it actually orders; see /mmio-race-audit) — don't apply WH/BH manual-ordering verdicts blindly.
• Ordering vs a paired blocking mailbox_write: when a producer releases the consumer with a semaphore_post AND hands data via a mailbox_write in the same path, the semaphore_post must come first — a full mailbox FIFO can stall the producer before it posts, deadlocking against the waiting consumer (→ mailbox-sync-audit, signal-before-blocking-write).

5. Mutex (ATGETM/ATRELM) balance & deadlock

• t6_mutex_acquire(idx)/t6_mutex_release(idx) (mutex::REG_RMW=0, mutex::SFPU=4) must be balanced on every path — an early return between acquire and release holds the mutex and deadlocks every other thread that acquires it. The mutex only helps parties that take it (REG_RMW is not taken by the math thread — see cfg-word-overlap-audit).
• mutex::SFPU is declared but currently never acquired in tt-llk. It exists because SFPU instructions can issue from both T1 and T2; if a kernel ever issues SFPU from pack concurrently with math without taking it → race. Flag any new cross-thread SFPU issue that doesn't.

6. Cross-thread deadlock cycle

• Build the wait-for graph: math waits MATH_PACK-max (needs pack get) and UNPACK_TO_DEST-zero (needs unpack post); pack waits MATH_PACK-zero (needs math post); unpack waits UNPACK_TO_DEST-max (needs math get). Any cycle with no thread able to make progress = deadlock. Most often introduced by reordering wait-before-work, or an op that posts on one thread whose consumer is conditionally skipped, or mismatched init/uninit pairing across threads.

Method

1. Enumerate every site, tagged by thread via filename (*unpack*/cunpack=T0, *math*/cmath=T1, *pack*/cpack=T2):

   cd tt_metal/tt-llk
   grep -rInE "t6_semaphore_(post|get|wait_on_max|wait_on_zero|init)|semaphore_(post|get)\(|TTI_SEM(POST|GET|WAIT|INIT)|t6_mutex_(acquire|release)|TTI_ATGETM|TTI_ATRELM" \
     tt_llk_* --include=*.h | grep -v /tests/
   # init sites (incl. above tt-llk — kernels/firmware do some SEMINIT):
   grep -rInE "SEMINIT|t6_semaphore_init" tt_metal/hw tt_metal/tt-llk --include=*.h
   

2. Per semaphore, list every (thread, op, condition, file:line). Classify ops: SEMINIT(Max,Value) / SEMPOST(producer) / SEMGET(consumer) / SEMWAIT(max|zero) / RISC-MMIO post|get.
3. Run the rules: init present + correct Max/Value (#1) → balance over all branches (#2) → directions (#3) → MMIO ordering (#4) → mutex balance (#5) → deadlock graph (#6).
4. Confirm semantics against tt-isa-docs (WH/BH) when unsure what a SEMWAIT condition or Max does.

Verdict

• Init present (right thread, right Max/Value, ordered before use) + balanced post/get on all paths + correct directions + MMIO posts gated by an in-stream wait + mutexes balanced + no wait cycle → SAFE.
• Wait-on-max semaphore with no reachable SEMINIT, or Max ≠ buffer depth, or stale initial Value → INIT-BUG (corruption or deadlock depending on default).
• Post/get imbalance or skipped wait-before-post on a reachable branch → RACE/DEADLOCK (real).
• Window exists only on an unused/experimental path, or value-invariant → LATENT — say so; let the maintainer decide.

Architecture note

WH/BH share this model. Quasar adds HW AutoTTSync affecting RISC↔Tensix ordering (read Confluence 1340276980 at audit for its guarantees), so rule #4 verdicts differ; the semaphore-protocol rules (#1–#3, #5–#6) still apply. Verify the semaphore map/wrappers resolve for Quasar before concluding.

Verified non-bugs (don't re-flag)

• deepseek_compute_kernel_init (tt_metal/hw/inc/api/compute/experimental/deepseek_compute_kernel_hw_startup.h): MATH inits FPU_SFPU (sem 0), PACK inits SFPU_FPU (= UNPACK_MATH_DONE, sem 6). Different semaphores on purpose — a two-semaphore bidirectional FPU↔SFPU handshake (documented in its @note), not a typo.
• UNPACK_TO_DEST / MATH_DONE having no SEMINIT inside tt-llk is not a missing-init bug: BRISC firmware c_tensix_core::initialize_tensix_semaphores() boot-inits them (and MATH_PACK, PACK_DONE) to Max=1, Value=0 before any TRISC kernel. In particular MATH_DONE's wait_on_max is a real depth-1 gate (Max=1), not a no-op — do not assume a default Max of 15.
• FPU_SFPU (sem 0) in the experimental llk_unpack_AB_reduce_custom*.h is consumer-only (wait_on_zero+get, "wait for blocked_matmul_and_pack to signal") — not a missing-producer deadlock. The matching post is issued by the compute-kernel layer (e.g. sdpa.h/custom_tilize.h MATH post, llk_math_sdpa_custom_mm.h, driven by blocked_matmul_and_pack<> in sparse_sdpa_compute.cpp); init via the deepseek startup (Max=1). Classic cross-layer handshake.

Thoroughness (optional, full sweep)

Best done as: deterministic enumeration (grep → per-semaphore table) → one agent per semaphore to run rules #1–#6 → adversarial verify: for each SAFE verdict, try to construct a branch/order that breaks balance or deadlocks; for each wait-on-max semaphore, try to prove its SEMINIT is unreachable. Only run a Workflow if the user opts into multi-agent orchestration. Never upgrade "balanced in the common path" to SAFE without checking early-return/conditional paths.

Output

For each semaphore/mutex: the per-thread op table (file:line), SEMINIT Max/Value + issuing thread + whether ordered before use, balance result, wait-directions, MMIO-ordering, and verdict (SAFE / INIT-BUG / RACE-DEADLOCK / LATENT) with one-line reasoning + fix. End with a wait-for-graph note (any cycle) and totals per verdict per arch.