fix(ckpt): open loop device in direct-io mode

[qiwu575757]

The btrfs-loop backend attached its image with losetup --find --show, which leaves the loop device in buffered mode (DIO=0). Any O_DIRECT issued by btrfs on top of the loop is then silently downgraded to buffered IO against the host image file: data crosses two page caches (the loop filesystem's and the host ext4's) and every request is serialized through the single loop kernel thread. The loop backend therefore had no end-to-end O_DIRECT at all, and its data path became the bottleneck under concurrent IO.

Measured on 5.10.134-19.6.3, ws-ckpt 0.4.2, 40 GiB image on a 100 GiB ESSD PL1 host, comparing the same workloads before and after:

checkpoint under fio pressure 305.10 ms -> 62.55 ms (5.3x -> 1.08x
of btrfs-base; p50 42 vs 41 ms and
p99 237 vs 237 ms now match base)
checkpoint under dual pressure 327.98 ms -> 139.75 ms
10h soak throughput 705 -> 1355 cycles/h
soak checkpoint / rollback 291.07 -> 120.58 ms / 332.13 -> 107.84 ms

Add --direct-io=on to all four losetup --find --show call sites, so bootstrap, post-rename remount, legacy rollback reattach and image-grow reattach all open the device in direct mode. No plain attach remains in this file, so a fallback path cannot silently reintroduce buffered mode.

Validated on the same host with the full ws-ckpt suite after the change: 421075 operations, zero non-zero exits and zero SHA256 integrity mismatches, covering daemon kill -9 recovery, host filesystem exhaustion, image size cap, 4-process and 4-workspace concurrency, a 10h soak and 32k snapshot accumulation.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

Qi Wu seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 1dafa3672d

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".

[yummypeng]

LGTM