LATX, fix: Preserve queued RCU callbacks across fork

[LaurenIsACoder]

Scope and invariant

Based on upstream bfe17ae7c1. Pending RCU callbacks must stay represented by both complete queue links and the pending count when fork snapshots the process.

Previously the consumer subtracted a batch before waiting for its grace period. A fork during that wait inherited queued nodes without the matching count; the child could leave them unprocessed or block a later drain. Concurrent producers could also be captured between queue-link updates.

Serialize enqueue/dequeue/count transitions with fork, decrement only when a node is actually dequeued, and preserve the child's queued work. The queue mutex is never held over a callback, grace-period wait or worker creation. Keep the existing lazy/deferred worker policy so namespace setup does not gain an unexpected helper thread.

Boundary: callbacks already dequeued into a vanished thread's stack are not replayed. Inherited queued callbacks start at the existing first-call/drain/explicit-deferred-start safe points; this is not a promise of immediate cleanup in an otherwise idle child.

Author self-review and combined validation

The actual published diffs were reviewed individually and checked against the local candidates. Follow-up corrections were folded into their existing commits; author/sign-off metadata updates preserved the per-PR source trees. The series remains six commits across #472, #473, #474, #475 and #476.

All five branches merge without conflicts on bfe17ae7c1. The combined source tree 80aaaaf13688387640f528d1c0bccb6a800f8d6e builds and passes 33/33 lat-pr-fast tests. On that combined binary, the 500-cycle KZT unload and retained-handle checks, 520-cycle SMC/fork checks, and 512 successful TSYNC operations alongside thread creation also complete successfully. A tests-disabled configuration registers no tests.

The integration suite still reports 27 skips, including after selecting the available guest sysroot with LATX_X86_64_SYSROOT: 25 require clang and two require the x86 cross compiler. These are not passes. This author self-review does not substitute for the upstream CI matrix, maintainer review or real-application acceptance.

Validation

LoongArch64, GCC 15.3, x86_64-linux-user, LAT O1, KZT compiled in, release:

./configure --target-list=x86_64-linux-user --enable-latx --optimize-O1 \
  --extra-ldflags=-ldl --enable-kzt --enable-tests --disable-docs \
  --disable-werror --meson=meson
ninja -C build -j2 all meson-test-prereq
env -u LAT_LD_PREFIX meson test -C build --no-rebuild \
  --suite lat-pr-fast --print-errorlogs --num-processes 2
env -u LAT_LD_PREFIX meson test -C build --no-rebuild \
  --suite latx-integration --print-errorlogs --num-processes 2

• Isolated branch: 25/25 lat-pr-fast passed.
• Maintained native regression uses the real public RCU API, pthreads and fork: a reader-held grace period, lazy/deferred child startup, empty queues, callbacks that queue callbacks, and four concurrent producers across repeated forks. It verifies exact-once callback delivery within its covered queue boundary and uses bounded timeouts.
• The reader-held fork regression fails against the old queue accounting; both the new regression and the pre-existing lazy-worker test pass with this change.
• The held-reader scenario is registered for LAT/KZT, whose existing atfork path does not wait for the parent grace-period lock. It is not evidence for every build configuration.
• Integration suite: 27 skipped (exit 77) because the clang/guest-building toolchain is unavailable; skipped tests are not passes.

This is an independent infrastructure fix. The separate KZT unloaded-header reclamation PR must be merged after this one to preserve its queued retirements across fork. No runtime installation or application acceptance is included.