Codegraph Daemon (v0.9.7) — MCP Mode Main Thread Spin Loop

[MiracleYyds]

Codegraph Daemon (v0.9.7) — MCP Mode Main Thread Spin Loop

Summary

Codegraph daemon (v0.9.7) running with --mcp flag enters a busy-wait / spin-loop on its main thread when no MCP client is connected, consuming 100% of one CPU core.

---

Evidence

Single-thread CPU burn

• Main thread only (TID 1): utime=94433 ticks, stime=2238 — 100% user-space CPU
• 6 worker threads: all idle (utime ~0)

Dense "fake" syscall pattern

Metric	Value
Write syscalls (syscw)	13.4 million in ~15 min (~15,000/sec)
Actual bytes written (wchar)	22,806 bytes total — ~1.7 bytes per call
SQLite DB modification time	Not changed during the spin

Extremely high write syscall count with near-zero actual data — classic kernel-level spin signature.

Involuntary context switching

Metric	Value
Voluntary context switches	18,278
Involuntary context switches	101,259
Ratio	5.5:1 involuntary >> voluntary

Process is preempted by the kernel scheduler, not voluntarily sleeping — confirms CPU starvation pattern.

No clients connected — nothing to do

• Unix socket file is empty (0 bytes) — no MCP client connected
• File watcher disabled (CODEGRAPH_NO_WATCH=1) — not file-change driven
• Process state: R (running) — always runnable

---

Root Cause

MCP Daemon Main Loop
         │
         ▼
  io_uring / epoll wait?
         │
    No event → no client
         │
         ▼
  ⚠ BUG: devolves into busy-poll
  - io_uring non-blocking poll
  - epoll_wait(timeout=0)?
  - No sleep, no blocking
         │
         ▼
  ~15k syscalls/sec
  0 bytes real I/O
  100% single-core CPU

When no MCP client is connected, the main event loop degenerates into a non-blocking poll that retries immediately with zero sleep. Under normal operation with an active client, read/write operations naturally block the thread, hiding the bug.

---

Impact

Dimension	Assessment
Functionality	None (no clients connected, no requests processed)
CPU	100% of one core
Memory	Normal (~212 MB RSS)
Disk	Near-zero writes
Other processes	CPU contention if resources are tight

---

Workaround

kill <pid>

The daemon auto-restarts on demand. The bug only manifests during idle periods — active queries interrupt the spin naturally.

---

Environment

• codegraph version: v0.9.7
• Launch args: codegraph serve --mcp
• Platform: Linux x86_64
• SQLite mode: WAL, synchronous=FULL