[byrnedj]

DTO_COLLECT_STATS kept one global table that every thread updated with atomic adds. Collecting stats got slower as threads were added, so it couldn't be left on to profile a multithreaded workload. This PR changes how stats are collected and adds the output needed to follow a run over time and process it with tools.

Changes

• Per-thread counters. Each thread counts in its own block, mapped (demand-zero) when it first records a call. When the thread exits, the block is folded into a retired total and unmapped. Output sums the live blocks and the retired total. Increments are relaxed loads and stores to the thread's own block, with no locked instructions.
• Log-scaled histogram. The old histogram had 512 linear 4 KiB buckets, and everything from 2 MiB up landed in the last one. The new one has a bucket for calls under 64 B, then four buckets per power of two up to 4 GiB, then one for anything larger (106 in all).
• DTO_STATS_INTERVAL_MS=N. A sampler thread writes what happened in each N ms. The interval records add up to the summary.
• DTO_STATS_FORMAT=json. Output is JSON Lines: interval and summary records with calls/bytes/ns for each op (memset/memcpy/memmove/memcmp) and path (cpu/dsa/dsa_failed), plus failure reasons. The summary also has the histogram and the configuration at exit (wait method, the auto-tuned min_bytes and cpu_size_fraction, NUMA, number of WQs). Schema is in the README.
• DTO_STATS_FILE expands %p to the PID and appends instead of truncating, so a child process no longer overwrites its parent's output. Every record carries its pid.
• fork. The child starts with empty counts and reopens a %p file under its own PID. Locks are taken around fork, so a child can't inherit a lock held by the sampler or a printer. SIGINT output uses trylock.
• Removed dto_stats_log(). With DTO_LOG_FILE set it wrote each line twice, the second time from a va_list that had already been consumed. Also removed: turning stats on no longer forces the TRACE log level.

Overhead

Extra time per call with stats on, CPU path (DTO_USESTDC_CALLS=1), threads pinned, median of 5 runs, gnr-qual1:

size	threads	main	this PR
4 KiB	1	+55 ns	+44 ns
4 KiB	4	+357 ns	+46 ns
4 KiB	16	+1319 ns	+42 ns
4 KiB	32	+2612 ns	+37 ns
64 KiB	32	+1724 ns	+37 ns

What's left is the two clock_gettime calls. With stats off, nothing changed. On the DSA path (1 and 4 WQs), runs with stats on were 12–20% faster than main at 4 and 16 threads; 32 threads was too noisy to compare.

Testing

New ctest -R stats (tests/test_stats.py plus a workload of known calls, -fno-builtin). It needs no DSA hardware and checks:

• totals and the bucket each size lands in
• interval records summing to the totals
• calls from 200 short-lived threads all counted, with their blocks unmapped
• separate parent and child records after fork
• %p files
• the text tables

On gnr-qual1, stats passes as root (dsa0 wq0.0) and as a user (CPU path), and functional passes 10/10.

The second commit links libdto with pthread. The per-thread blocks use thread keys, and the sampler uses a thread and a condition variable; before glibc 2.34 these are in libpthread, not libc. Without the link, preloading DTO into a program that doesn't link libpthread itself (sh, true) failed on RHEL 8 with undefined symbol: pthread_key_create. The stats test now checks this case.

Merging with other open PRs

dto.c merges cleanly with #34 and #40. #34 and #38 conflict only in the lines each PR adds to CMakeLists.txt, TESTING.md and README.md.

A follow-up will add dtoctl --stats-run. It runs a program CPU-only and then with DSA, optionally under perf, and reports DSA speedup per operation with an Amdahl's-law estimate, using the JSON output added here.