vmm: one-shot mode allocates CIDs outside the IdPool and is invisible to the VMM

从 vmm/src/one_shot.rs 开始，重点查看 run_one_shot 和 CID 扫描，然后跟踪 app.rs reload_vms/reload_vms_sync 以及 crates/dstackup/src/cid.rs。比较提出的协调方案，并确定应如何在现有分配机制中表示 one-shot CID。完成的标准是：并发的 one-shot 分配和主服务分配不能选择同一个 CID。

Summary

One-shot mode (vmm/src/one_shot.rs) and the main VMM service allocate vsock CIDs from the same configured range through two mechanisms that do not know about each other. Nothing prevents them from picking the same CID.

The two allocators

Main service — IdPool over [cid_start, cid_start + cid_pool_size), rebuilt on reload from the supervisor's process list (app.rs, reload_vms / reload_vms_sync).

One-shot — one_shot.rs:24-56:

// scan `ps aux` for qemu-system-x86_64 ... guest-cid=<n>
let mut one_shot_cid = config.cvm.cid_start;
while existing_cids.contains(&one_shot_cid) {
    one_shot_cid += 1;
    ...
}

It starts at cid_start, avoids collisions by scraping ps aux, and never touches the pool.

Why they can collide

One-shot launches QEMU directly (cmd.status() at the end of run_one_shot) rather than registering the process with the supervisor. occupied_cids in both reload paths is built from supervisor.list(), so a one-shot VM's CID is invisible to the main service and never gets occupied in the pool.

The blindness is one-directional:

	sees the other's CIDs?	via
one-shot → main service	yes	ps aux finds the qemu processes
main service → one-shot	no	one-shot never reaches the supervisor

So the main service can allocate a CID that a running one-shot VM already holds.

Two secondary issues in the same code path:

• TOCTOU — the ps aux scan and the QEMU launch are not atomic; a concurrent allocation in the window collides regardless.
• Parsing — CIDs are recovered by string-splitting ps aux output on guest-cid=, which is sensitive to how QEMU arguments are formatted.

Note on #907

Before #907, IdPool::allocate() had an off-by-one that made it skip cid_start entirely, while one-shot starts at cid_start. That incidentally kept the two apart. #907 fixed the off-by-one (correctly — one_shot.rs:48 and crates/dstackup/src/cid.rs both already treat the window as [start, start+size)), which removes the accidental separation.

This is not a regression introduced by #907. The protection only ever held for exactly one one-shot VM: a second one takes cid_start + 1, which was already inside the main pool's allocation range. The underlying problem is that the two allocators were never coordinated.

Possible directions

1. Register one-shot processes with the supervisor so the existing pool machinery covers them.
2. Reserve a dedicated range for one-shot outside [cid_start, cid_start + cid_pool_size).
3. Have one-shot allocate through IdPool rather than ps aux.

(1) seems most consistent with how the rest of the system tracks VMs, but one-shot is deliberately a lighter path, so (2) may be the cheaper fix.

Confidence

The code paths are confirmed by reading: one-shot starts at cid_start, does not register with the supervisor, and both reload paths source occupied_cids from supervisor.list() only. Not verified on hardware — I have not observed an actual vsock CID collision, and I do not know how much one-shot mode is used in practice, which bounds how much this matters.

Found while reviewing #907.