fix(ckpt): reap finished tasks in accept loop - #2554
Conversation
The accept loop spawned every connection into a `JoinSet` but only reaped it at shutdown, so the set kept one entry — task handle plus output slot — for each connection that had already finished. Every ws-ckpt CLI invocation opens one socket connection, so the daemon grew monotonically for its whole lifetime and only released the memory on exit. Measured on 5.10.134-19.6.3 with ws-ckpt 0.4.2, the growth was linear in connection count and independent of both the storage backend and the kind of operation, which is what identified the connection path rather than snapshot state as the source: btrfs-base 10h soak 7.8-7.9 KB per operation btrfs-loop 10h soak 7.9-8.0 KB per operation loop, direct-io, 10h 8.0 KB per operation isolated 60s sample 8.4 KB per operation VmData tracked VmRSS almost exactly (5418 MB of 5366 MB resident), the file descriptor count stayed flat, and the snapshot pool was held constant during the soaks, so the growth was heap held by the daemon rather than page cache, descriptors, or index state. A `kill -9` returned RSS from 5.7 GB to 46 MB. Under a checkpoint-heavy agent workload this reaches roughly 370 MB/h, and about 1.7 GB/h under the functional suite, so a long-lived daemon eventually OOMs. Drain completed tasks with `try_join_next` after each select iteration. It is synchronous, so reaping cannot delay the next accept, and it avoids the second mutable borrow of `join_set` that a `join_next` branch inside the same `select!` would need while the accept arm still spawns into the set. Shutdown still drains with `join_next` under the existing timeout, so in-flight connections keep their grace period. After the change, 20000 operations grow RSS by 2.6 MB (0.13 KB per operation, down from ~8 KB), and a further 40000 operations add 2.0 MB in total with no linear trend, i.e. what remains is allocator headroom rather than accumulation. Descriptors stay at 14. Signed-off-by: Qi Wu <shijing.wq@alibaba-inc.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: e4040e7e90
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| // awaits, so reaping cannot delay the next accept; a join_next branch in | ||
| // the select! above would instead need a second mutable borrow of | ||
| // join_set while the accept arm still spawns into it. | ||
| while join_set.try_join_next().is_some() {} |
There was a problem hiding this comment.
When a burst of clients is accepted and then no new client connects, this reap runs only immediately after an accept()/cancel select iteration. Any connection task that completes after the last accept does not wake this loop, so its JoinSet entry remains until another client connects or the daemon shuts down; with many concurrent checkpoint requests this can still retain one task allocation per request. Drive reaping from task completion itself, such as a join_next/poll branch or another wakeup source, rather than only after accepts.
Useful? React with 👍 / 👎.
1 participant
The accept loop spawned every connection into a
JoinSetbut only reaped it at shutdown, so the set kept one entry — task handle plus output slot — for each connection that had already finished. Every ws-ckpt CLI invocation opens one socket connection, so the daemon grew monotonically for its whole lifetime and only released the memory on exit.Measured on 5.10.134-19.6.3 with ws-ckpt 0.4.2, the growth was linear in connection count and independent of both the storage backend and the kind of operation, which is what identified the connection path rather than snapshot state as the source:
btrfs-base 10h soak 7.8-7.9 KB per operation
btrfs-loop 10h soak 7.9-8.0 KB per operation
loop, direct-io, 10h 8.0 KB per operation
isolated 60s sample 8.4 KB per operation
VmData tracked VmRSS almost exactly (5418 MB of 5366 MB resident), the file descriptor count stayed flat, and the snapshot pool was held constant during the soaks, so the growth was heap held by the daemon rather than page cache, descriptors, or index state. A
kill -9returned RSS from 5.7 GB to 46 MB. Under a checkpoint-heavy agent workload this reaches roughly 370 MB/h, and about 1.7 GB/h under the functional suite, so a long-lived daemon eventually OOMs.Drain completed tasks with
try_join_nextafter each select iteration. It is synchronous, so reaping cannot delay the next accept, and it avoids the second mutable borrow ofjoin_setthat ajoin_nextbranch inside the sameselect!would need while the accept arm still spawns into the set. Shutdown still drains withjoin_nextunder the existing timeout, so in-flight connections keep their grace period.After the change, 20000 operations grow RSS by 2.6 MB (0.13 KB per operation, down from ~8 KB), and a further 40000 operations add 2.0 MB in total with no linear trend, i.e. what remains is allocator headroom rather than accumulation. Descriptors stay at 14.