Add a flow view of the osd, pg, and object

The trace lines already carried the mapping and the parser threw it away.
A radostrace line names the object, its placement group, and the acting
set, so the panel opened with m shows all three levels without a new
remote command. osdtrace names no object, so a run with only that source
collapses to two levels and the panel title says which one it reached.

o orders every branch by op count or by latency and s reverses it. Rows
carry the mean size the op declared, which separates a slow heavy request
from one that is slow for the little it asks for. A read reports the
length it requested rather than the bytes returned.

Two keys behaved oddly once the panel existed. r and t switched the trace
source before starting it, which in a view with no selected source looked
like a key that did nothing, and x cleared osdtrace alone while the panel
drew from every source.

Signed-off-by: Seyeong Kim <me@seyeong.kim>

Author: xtrusia

CHANGELOG.md

@@ -8,6 +8,14 @@ Notable changes are documented here. The format follows
 
 ### Added
 
+- Added a flow view, opened with `m`, that shows the OSD, placement group, and
+  object behind the ops in the trace buffer. radostrace lines already name all
+  three, so the view needs no extra remote command; with only osdtrace running
+  it collapses to OSD and placement group. `o` orders it by op count or by
+  latency and `s` reverses the order. Each row also carries the mean op size,
+  which separates a slow heavy request from one that is slow for the little it
+  asks for. A read reports the length it requested rather than bytes returned,
+  so a client that always asks for 4MiB reports 4MiB whatever the object holds.
 - Insights now name the checks behind a `HEALTH_WARN` or `HEALTH_ERR` instead of
   telling the operator to go run `ceph health detail`. `ceph -s` already carried
   them, so this costs no extra remote command.
@@ -20,6 +28,7 @@ Notable changes are documented here. The format follows
 
 ### Fixed
 
+- `x` now clears every captured trace source rather than osdtrace alone.
 - Node readiness no longer breaks on hosts without `ceph-osd` processes. The
   remote OSD count fell back through `pgrep -c ... || echo 0`, which emitted two
   lines because `pgrep -c` prints `0` and exits 1 on no match. The extra line

README.md

@@ -55,6 +55,11 @@ In a source clone, use `cargo run -- init-config`, edit the generated file, then
 - Per-OSD commit and apply latency from `ceph osd perf` next to the eBPF trace
   numbers, and health check names read straight out of `ceph -s`.
 - osdtrace eBPF latency tracing with per-OSD and per-PG breakdown of queue, BlueStore, and KV-commit latency.
+- A flow view of the OSD, placement group, and object behind the observed
+  ops, built from the trace lines already streaming. radostrace names the
+  object so the view has three levels; with only osdtrace it collapses to
+  OSD and placement group. Rows carry the mean op size next to the latency; a
+  read reports the length it requested, not the bytes returned.
 - No standing agent: no permanent daemon on the nodes; the osdtrace runner script removes itself on stop, quit, or TTL expiry. (The cephtrace tracer binaries you deploy do persist under `~/.cephlens/bin/`.)
 - Edit hosts and trace settings live in the TUI; changes apply to open SSH streams immediately.
 - Export recorded sessions as Markdown reports with the same diagnostic rules used by the TUI.
@@ -286,6 +291,9 @@ p          run a probe readiness check
 c          edit config
 t/f/r      view osdtrace / kfstrace / radostrace; press again to start or stop (confirmed)
 a          start or stop all trace sources (confirmed)
+m          flow view: the osd -> pg -> object mapping behind the live ops
+o          flow view: order by op count or by latency
+s          flow view: reverse the order
 i          install osdtrace
 x          clear captured trace events
 ?          toggle the help overlay

src/app.rs

@@ -19,6 +19,7 @@ use serde_json::Value;
 use crate::{
     collect::{parse_cluster_summary, parse_osds, run_probe},
     editor::ConfigEditor,
+    flow::{FlowMetric, FlowSort},
     kfstrace::{KfsEvent, parse_kfs_event},
     model::{NodeSummary, Snapshot},
     radostrace::{RadosEvent, parse_rados_event},
@@ -177,6 +178,10 @@ pub(crate) struct App {
     pub(crate) trace_following: bool,
     pub(crate) trace_session: Option<String>,
     pub(crate) trace_source: TraceSource,
+    pub(crate) flow_view: bool,
+    pub(crate) flow_metric: FlowMetric,
+    pub(crate) flow_sort: FlowSort,
+    pub(crate) flow_scroll: usize,
     pub(crate) kfstrace_events: Vec<KfsEvent>,
     pub(crate) kfstrace_active: usize,
     pub(crate) kfstrace_stop: Arc<AtomicBool>,