osdtrace: trace replica subop writes in single mode

[taodd]

Summary

Stacked on #186. Single mode traced only PrimaryLogPG::log_op_stats, which fires solely for client ops on the primary — replica write subops (MSG_OSD_REPOP) were invisible, although on a replicated pool they are two thirds of the write work. This adds a second lightweight probe so -s prints one line per replica subop write, with the same in-kernel -l threshold behavior:

osd 0 pg 1.18 subop_w size 4483 client 4126 tid 500 object benchmark_data_..._object499 subop_lat 4886
osd 1 pg 1.18 subop_w size 4483 client 4126 tid 500 object benchmark_data_..._object499 subop_lat 5575
osd 2 pg 1.18 op_w   size 4096 client 4126 tid 500 object benchmark_data_..._object499 osd_ops [set-alloc-hint,write] op_lat 6396

Anchor choice: repop_commit, not log_subop_stats

Both fire at essentially the same instant — in current Ceph, log_subop_stats is a static function whose only call site is the last statement of repop_commit. But a static single-caller function is inlined at -O2: it is absent from every embedded release DWARF JSON (checked quincy 17.2.6/el8, reef 18.2.4/el9, squid 19.2.0/el9) and only exists out-of-line on debug builds. ReplicatedBackend::repop_commit is invoked through the C_OSD_RepModifyCommit callback, so it cannot be inlined away, and is present in all embedded release JSONs. It also only ever handles MSG_OSD_REPOP, so no message-type filter is needed.

Changes

• Varpaths: append recv_stamp and pg id (cast:MOSDRepOp → pgid.pgid.m_pool/m_seed, layout gdb-verified: MOSDRepOp::pgid declared directly at 0x1a0) to the repop_commit list — varids 175-177; 170-174 keep their positions so full mode's v1 program is untouched. 8/10 varid budget.
• BPF: uprobe_repop_commit_v2, a standalone sibling of uprobe_log_op_stats_v2 — recv_stamp first, in-kernel -l gate, then owner/tid/payload/object-name/pg into a ringbuf event; no map state, no printk. Appended as the last program in the file since func_progid indices are positional in skeleton declaration order.
• Userspace: attach-table entry (OP_SINGLE_PROBE, exact, v=2), print_single_op branch for subop_w lines (no osd_ops field — a repop carries a serialized ObjectStore transaction, not a decoded OSDOp vector), and subops are excluded from the size-range histograms (they measure client-visible latency and would count each client write once per replica).
• Docs: -s help text and man page updated.

Verification (3-OSD vstart cluster, size-3 pool, tentacle dev build)

• subop_w line count = exactly 2× the client op_w count (136,438 vs 68,218 over a bench run)
• for any tid, the primary's op_w pairs with two replica subop_w lines on the same pg/object, with subop_lat < op_lat
• -l 15: only above-threshold ops and subops print — including matched slow pairs (client op slow because its replica subop was slow)
• histograms unchanged: entry count equals the 4K writes only (34,109), subops and deletes excluded
• full-mode regression: 15 probes × 3 OSDs attach, identical output — now with live peers [(0, 7136), (2, 5906)] on the replicated cluster
• probe body cost via bpftool stats: ~3.5 µs per subop event (debug build, contended laptop); -l fast path structurally identical to the validated log_op_stats_v2 gate

🤖 Generated with Claude Code

https://claude.ai/code/session_01SQMVEJezhKCMbVMWFpNi5B