• mount: support direct-io for loopback device

mount supports direct-io option for loop block device. In the original loop setup, the Direct flag doesn't work even if the syscall returns successfully. In the patch, I use ioctl to set the direct-io and return error if the old kernel doesn't support this.

• snapshots|pkg: umount without DETACH and nosync after umount

It's a cleanup for the snapshot test. The detach umount might cause unexpected issue. It's hard to debug in the github action. The umount without detach might be slow but it's easy to figure out the flaky cases.

• snapshots/blockfile: deflaky the testsuite

The patch is trying to deflake the blockfile snapshot test case. The view-type snapshot mounts the device with readonly. If the filesystem needs to recovery, the kernel will return the EROFS and the error in dmesg is like.

2023-05-25T05:31:07.6688425Z [Thu May 25 05:29:39 2023] EXT4-fs (loop10): INFO: recovery required on readonly filesystem
2023-05-25T05:31:07.6688980Z [Thu May 25 05:29:39 2023] EXT4-fs (loop10): write access unavailable, cannot proceed (try mounting with noload)

For this case, I introduce a viewHook helper for test. It's used to mount the snapshot with write permission, so that the kernel can handle recovery at first. Then the ro mount won't fail.

And using 8MiB is to reduce IO pressure. :)

• vendor: should sync when we apply vendor: update github.com/containerd/continuity

Even if we use direct-io for loop, we still have the unsync error, like

 ~ /onlyme/file-13(mode: 777, uid: 0, gid: 0) -> /onlyme/file-13(mode: 755, uid: 0, gid: 0)

So I change the fs/fstest to force sync for new file. Hopefully, containerd/continuity#228 it's acceptable.

#!/usr/bin/env bash

set -euo pipefail

for i in {1..101}
do
   echo "$i times"
   sudo -E PATH=$PATH make root-test EXTRA_TESTFLAGS="-failfast"
done

I run this branch for test in azure Standard D4s v4. It's happy. :)