perf(btree): Stage 1b — lock-free root snapshot for read descents

[gburd]

Stage 1b: lock-free root snapshot for read descents

Lift the multi-core read-scaling ceiling by removing the per-operation
contention on the B-tree root page. Every descent fetches the root via
__memp_fget, which does atomic read-modify-writes on shared cache lines
(bucket-latch share count, buffer pin/refcount, buffer-latch share count).
Because every thread touches the root on every operation, that line is a true
serialization point: in-cache reads peak around 8 threads and then negatively
scale.

Approach

Each DB handle keeps a private, immutable copy of the root page taken at a
known root LSN, plus a wired (non-evictable) pointer to the live root buffer. A
plain read descent picks its first child out of the private copy — a pure read,
no pin, no latch, no shared write — and validates the copy against the live
root LSN before and after fetching the child. On the rare root change
(split/merge) it falls back to the normal pinned/latched descent and refreshes
the copy.

Wiring the root sidesteps the need for general epoch reclamation: the single
hot frame is simply never freed, so a lock-free reader can never dereference
reclaimed memory.

What's in the branch

• BH_WIRED/wired buffer mark + __memp_wire/__memp_unwire, evictor skips
  wired frames, capped at MPOOL_WIRED_MAX_PCT of the region so wiring cannot
  starve the cache. Only the one common tree root is wired.
• BAM_RSNAP per-handle root snapshot; __bam_rsnap_refresh /
  __bam_rsnap_child; superseded copies retired to a free list and freed at
  handle close (avoids a reader/free race without epoch reclamation).
• Fast path gated to plain btree read lookups: not write/stack/parent/next/
  del/min/max, not OPD, not recno/recnum, multiversion == 0, logging on,
  durable. A plain read descent acquires no root page lock on the normal path
  either (BAM_GET_ROOT locks the root only for WRITE/recno/recnum/downrev),
  so the fast path skips no lock it should have taken.

Correctness

• Every structural change bumps the page LSN under the page's exclusive latch;
  the reader's pre/post LSN check brackets the child selection and never acts
  on an unvalidated result.
• Eviction's refcount == 0 + exclusive-mtx_buf handshake is unchanged.
• Root page number is stable in BDB (splits push content down, root pgno fixed);
  compaction is the one runtime root move and now disarms the cached snapshot.

A use-after-free found in review is fixed in bb2e78e: __memp_wire now
reports whether the frame is actually wired, and the snapshot is cached only
when wiring took (otherwise the descent falls back); compaction invalidates the
cached root pointer; the wired-cap arithmetic no longer truncates to zero for
small caches.

Validation

• Clean debug + release builds, no new warnings.
• TCL: lock001, txn001, test001 (btree), ssi001, ssi002, test003, test011
  (duplicates/OPD), test026 (cursor delete), test111 (compaction incl.
  -revsplitoff) all pass.
• Multi-core scaling measurement on the 24-core box is tracked separately (see
  scaling-findings); headline metric is whether the ~8-thread ceiling lifts and
  the futex/atomic self-time drops at 8–24 threads.

Maps to ROADMAP #2 (latch-free buffer lookup). Stage 1c (lock-free descent of
deeper internal pages) is intentionally out of scope — profiling shows the next
bottleneck is elsewhere and a correct version needs epoch reclamation.