Audit and benchmark non-file hot paths

[siraben]

Summary

• Adds focused C benchmarks for non-file hot paths: syscall identity, time syscalls, futex EAGAIN, eventfd, epoll readiness, mmap faults, fd control, fork/wait, signal masks, and sched_yield.
• Defers poll notify-pipe setup until a poll actually blocks, avoiding pipe/fcntl/kevent setup for immediate-ready epoll/poll paths.
• Avoids host timezone work for gettimeofday when the guest passes a null tz pointer.
• Tightens fcntl(F_GETFD/F_SETFD) to update/read cloexec under the fd table lock and avoids bit-setting an invalid F_DUPFD_CLOEXEC result.

Benchmark Notes

Final focused run:

CLANG=/opt/homebrew/opt/llvm/bin/clang RUNS=5 WARMUPS=1 PREPARE_APK_INDEXES=0 \
  BENCH_FILTER='^(syscall_identity|identity_extended|time_syscalls|futex_eagain|eventfd_loop|epoll_eventfd|mmap_faults|fd_control|process_wait|signal_mask|yield_loop|pipe_throughput|dev_urandom_stream)$' \
  OUT=/tmp/ish-nonfile-final \
  I386_BASE_FS=e2e_out/bench-base/i386 \
  RV64_BASE_FS=e2e_out/bench-base/riscv64 \
  ./benchmarks/hot_paths/compare.sh

Benchmark	i386 best	RV64 best	RV64 vs i386
dev_urandom_stream	0.007796s / 2052.33 MiB/s	0.007492s / 2135.61 MiB/s	1.041x
epoll_eventfd	0.226719s	0.150893s	1.503x
eventfd_loop	0.200460s	0.202724s	0.989x
fd_control	0.787822s	0.866376s	0.909x
futex_eagain	0.241512s	0.268871s	0.898x
identity_extended	1.055663s	1.086358s	0.972x
mmap_faults	0.012561s / 20380.54 MiB/s	0.012897s / 19849.58 MiB/s	0.974x
pipe_throughput	0.010779s / 2968.74 MiB/s	0.010400s / 3076.92 MiB/s	1.036x
process_wait	0.014454s	0.014640s	0.987x
signal_mask	0.398589s	0.405305s	0.983x
syscall_identity	1.921237s	2.194061s	0.876x
time_syscalls	0.501590s	0.556313s	0.902x
yield_loop	0.052319s	0.052326s	1.000x

Earlier focused before/after for the poll change showed RV64 epoll_eventfd moving from 0.207266s to 0.136769s, a 1.515x speedup on that immediate-ready workload. The final broader run still shows epoll_eventfd at 1.503x vs i386.

Validation

• ninja -C build-bench-rv64 ish
• python3 -m py_compile benchmarks/hot_paths/compare.py benchmarks/hot_paths/run.py
• git diff --check
• Focused cross-arch benchmark run above

[siraben]

Follow-up hot-path pass pushed through ecc7ec12.

Implemented and kept:

Area	Benchmark	Before best	After best	Speedup	Notes
fakefs negative metadata cache	stat_misses	1.470629s	0.813193s	1.81x	5 runs, RV64 old build vs current branch
poll grouping / callback fanout	poll_many_ready	0.136954s	0.034819s	3.93x	5 runs, RV64 old build vs current branch

Current focused RV64/i386 comparison after the changes:

Benchmark	i386 best	RV64 best	RV64 speedup
recursive_ls_deep	0.160000s	0.080000s	2.00x
find_stat_deep	3.800000s	5.060000s	0.75x
fstat_loop	0.058137s	0.023193s	2.51x
stat_misses	3.130487s	0.813193s	3.85x
eventfd_loop	0.177551s	0.195855s	0.91x
epoll_eventfd	0.223901s	0.144718s	1.55x
poll_many_ready	0.144021s	0.034819s	4.14x
fd_high_churn	0.728030s	0.278294s	2.62x

Discarded after measurement:

Experiment	Result
fakefs inode-stat cache	No measurable gain beyond the negative path cache; removed before commit.
fdtable next_free allocation hint	Slowed fd_high_churn on RV64 (0.263854s baseline vs 0.275629s with the hint in the earlier run); removed before commit.
fakefs stat cache size matrix	2048 remains the best tradeoff. 8192 did not improve the realistic metadata workloads.

Validation run before pushing:

• ninja -C build-bench-rv64 ish
• python3 -m py_compile benchmarks/hot_paths/compare.py benchmarks/hot_paths/run.py
• git diff --check