- The round-2 report guessed the apply thread's ~48% mutex time was the backend
pendingMutex; the symbolized call graph at 56a77348 says it is the idle poll in
ServerLoop::ApplyThreadMain taking m_controlMutex twice per iteration
(ControlIsPending in the ready lambda, PumpControlRequest), with DrainRing at
29% of the thread. Every package below is cut from that graph, not from self%.
- T (transport): the control mailbox gets an atomic shadow bit m_controlPosted so
the idle poll takes no lock (handshake, C2 critical section and the
NOT_INITIALIZED answer unchanged); Doorbell::Wait spins on a once-calibrated
iteration budget and reads no clock in steady state (the park phase alone
computes deadlines); client/server wait and park counts ride the stats line
as wait[srv= srvpark= cli= clipark=] and the wire ledger as cliwait=/clipark=.
- D (role guards): OnApplyThread() is inline - one relaxed atomic key against the
caller's thread pointer (aarch64 __builtin_thread_pointer, else
std::this_thread::get_id) - and m_applyThreadId, the function-static guard and
the pthread_equal call are gone; the three scope depth counters stop being
thread_local (their only reader is the apply-thread guard, so they count only
when OnApplyThread() is true; a GL-thread scope no longer exempts the apply
thread, pinned by a new case); the guard chains are ordered cheapest-first.
- B (persistent maps): the shadow allocator rounds the SIZE to a page multiple, so
TrackWriteMap aligns OUTWARD when the shadow is page-granular - a sub-page map
is one protected page, edges and their per-draw XXH3 disappear; the inward arm
stays as the fallback. Review found a hole the edge hashes had hidden since
round 1: a map that faulted but was not consumed by the walk that advanced the
epoch was never pushed. The walk now drains every marked member before the
binding walk, registration bumps the epoch so a fresh map ships once on the
draw path, and a non-4 KB kernel page refuses the mprotect arm by name.
- C (client per verb): an image-unit high-water mark (push build only, fed by
glBindImageTexture) replaces the 192-unit sweep per draw; the residual fill's
7-term supplied conjunction is a memo keyed on its four inputs; the server's
63-field stamp per verb is a constexpr per-class mask table.
- generate_mipmap no longer skips its wait under MOBILEGL_IPC_BATCH_WAITS: its sink
is the backend GenerateMipmap, which reads MGB_CTX->GetActiveTextureUnit(), and
a batched publish let the next validate move the serial under it -
Fatal{UnmigratedPipeInput, "GetActiveTextureUnit@ReadPixels"} on the gate tree.
The rule is written at the site: a record may skip its wait only if its apply
reads nothing the residual fill writes. The texture-death scenario gets the
framebuffer case's fence before it reads the object_death tally.
- Redmi 2f7cbe2e, FCL + Minecraft 26.3-rc-3, VD12, CPU pinned: inproc 64.3 fps
(client 15.0 ms/frame, apply 12.4) -> 103-114 fps (client 8.7-9.2 ms/frame on a
mid core, apply 7.0-7.4); monolith 115 (120 Hz vsync-capped runs) / 206 (one
uncapped run, 5.0 ms/frame on the big core). Gates: unit 2203/2203,
integration-split 111/111 (the two formerly flaky cases 6/6 on repeat),
generators, include closure, dirty surface and doc citations green.
VD12 Minecraft on the Redmi (FCL, DirectGLES): the split's in-world frame rate
13 -> 60-66 fps against monolith's 110-114, and persistent-map wire bytes
25 MB/frame -> ~0.23 MB/frame. The residual split-specific costs are named at
the end, not hidden: the R-1 serialization (WaitForApplied ~12%, P11's
versioned-inputs design) and a split-only mutex complex still under
investigation.
Persistent-map push, three arms picked per buffer at map time:
- mprotect dirty tracker: the mapped range's page-aligned interior is
protected PROT_READ; the app's next write faults, the SIGSEGV handler marks
the page in the slot's bitmap and unprotects, and the push ships one 64 KB
block per block containing a faulted page. Ownership is three tests -
SEGV_ACCERR only (a userspace raise and an unmapped page chain verbatim),
inside a live tracked range, and not a refault on a still-set bit (a write
we unprotected succeeds, so that is a non-write access and it crashes
honestly). The alignment is INWARD: an outward-aligned edge page holds
foreign bytes, and a signal-blocking worker thread faulting on one is a
process kill - that was the ctest-only SmallRing segfault. Edge spans
(at most two per map) are hashed per push and ship their containing block
only on change (unconditional edges measured 112 MB/frame). Re-arm is one
mprotect per contiguous run. The first push ships everything once.
- hash-suppress arm (MOBILEGL_IPC_PERSISTENT_HASH_SUPPRESS, default 1):
per-block XXH3 against the last push's hashes; untracked buffers only.
- PushDrawConsumers: only the buffers the upcoming draw/dispatch can read
(VAO attribs/element, indexed binding points, indirect/parameter slots,
TexBuffer backing - the last also retires the per-frame-cadence cloud
twitch), plus the fault-epoch skip: an unmoved epoch proves no interior
byte changed, tracked edges are hashed in place, and any hash-arm member
vetoes. The membership map caches the tracker slot pointer (a 64-slot
rescan per buffer per draw was measured and removed).
MOBILEGL_IPC_BATCH_WAITS (default 1; MOBILEGL_PIPE_VERIFY forces 0): a
value-class record (kCtxState/kCtxCso/kCtxObject with no reply slot) is
published without waiting for its own apply; the barrier is taken at the next
pull-reading verb, the only place BARRIER-PULLED fields are read, so the R-1
fence posts are untouched and only the round trips between them change. Two
consequences are pinned rather than argued: kCtxObject verbs the tests read
synchronously get explicit fences (GenerateMipmapFieldsCross,
FramebufferDeathCrosses), and a frontend object whose last SharedPtr drops on
the apply thread - impossible before batching - now enqueues its
(kind, lifetimeId) on the deferred destroy queue and the GL thread replays
the DestroyAndFree helper at the verb hooks (CopyImageLayeredScenario aborted
Fatal{RoleViolation, MGPipeSlots} without it).
Doorbell::Wait's spin reads the clock once per 64 yields: on the workload
device a steady_clock read is a real syscall (no vDSO) and the per-iteration
reads were ~20% of the whole process at ~1000 waits/frame.
Gates: gen_pipe --check/--self-test green; unit 2188/2188;
integration-split 110/111, the one red being GenerateMipmapDepthPixels whose
DirectGLES and DirectVulkan MONOLITH siblings flake identically with every
code path of this package provably inert (PushIsArmed false, no ClientSession,
deferral's monolith early-out).