osdtrace: print detailed per-op info in single mode

[taodd]

Summary

Single mode (-s) previously emitted only exit-time latency histograms. This PR extends the single PrimaryLogPG::log_op_stats probe (uprobe_log_op_stats_v2) so each op prints a detailed line — target pg id, object name, decoded osd op list (incl. cls calls), client id, tid, total latency — while keeping the histograms:

osd 0 pg 1.11 op_w size 4096 client 4446 tid 8 object benchmark_data_..._object7 osd_ops [set-alloc-hint,write] op_lat 1942
osd 0 pg 1.7  op_w size 0    client 4446 tid 55268 object benchmark_data_..._object27627 osd_ops [delete] op_lat 1730
osd 0 pg 1.1d op_r size 24   client 4446 tid 27640 object benchmark_last_metadata osd_ops [read] op_lat 220

No new probes are attached: everything is read at log_op_stats time from the message via new cast:MOSDOp varpaths, reusing the existing capture_decoded_osd_ops() machinery and rodata constants.

Changes

• Varpath list: drop the unused inb/outb entries (read from PT_REGS_PARM3/4 by both program variants) and add pg (MOSDOp::pgid), object name (MOSDOp::hobj.oid.name) and MOSDOp::ops vector-bounds chains — filling the 10-varid budget exactly.
• BPF: populate the new fields only after the -l threshold gate, so the in-kernel fast path is unchanged (~670 ns). New read_hprobe_varfield_opt degrades silently to zeroed fields when the varids are absent (DWARF JSON exported before this change) instead of dropping events or spamming bpf_printk.
• Classification: read/write is now derived from decoded opcode mode bits — zero-payload delete and omap-only class methods are correctly labeled op_w — with wb>0 as the fallback for degraded DWARF data.
• dwarf_parser: resolve_type_name (used by cast:) now strips typedef/cv wrappers and keeps scanning CUs for a complete definition. Needed because MOSDOp is now using MOSDOp = _mosdop::MOSDOp<std::vector<OSDOp>> — the class lives in a namespace the type cache does not descend into, so the first cached hit is a typedef that can lead to a declaration-only DIE.
• Hardening: fill_map_hprobes (osdtrace + radostrace) errors out on >10 varpaths per function or >10 fields per varpath instead of silently corrupting adjacent varids / overflowing VarField_Kernel.fields.

Overhead (vstart, tentacle dev build, kernel 6.17, via bpftool prog stats)

path	before	after
-s per-op body	~2.4 µs	~3.0 µs (391 ms / 130k ops)
-s -l <ms> fast path	~670 ns	unchanged (structurally untouched)
full mode (15 probes)	~18 µs/op	unchanged

Test plan

• rados bench 4K write + cleanup: op_w lines with [set-alloc-hint,write], [delete], correct pg/object/client/tid, op_lat consistent with bench-reported latency
• seq read bench: op_r [read] lines
• Ctrl-C still prints the size-range latency histograms
• full-mode regression: 15 probes attach, per-op output identical to before (varids 90/91 shared with v1 are unchanged)
• degraded-DWARF path covered by design: absent varids leave fields zeroed (pg 0.0, object -), events still emitted

🤖 Generated with Claude Code

https://claude.ai/code/session_01SQMVEJezhKCMbVMWFpNi5B