RFC - [content-store] Commit should check first fsync() error

[fuweid]

Description

In the #10128, the content-store's Commit API fsync dirty-pages before enter bbolt transaction to avoid long mutex contention. However, the commit doesn't check the error from fsync.

containerd/core/metadata/content.go

Lines 575 to 582 in 2ec82c4

	var innerErr error
	// We pre-sync the in-flight writes to the disk. This avoids the [subsequent fp.Sync() call]
	// (https://github.com/containerd/containerd/blob/c4c3c6ea568ce0cfbcf754863abadeea37d77c8f/plugins/content/local/writer.go#L95)
	// from taking too long (10s+) while holding the metadata database lock as in the following
	// `update` transaction. We intentionally ignore any error on Sync() because it will be
	// handled by the subsequent `fp.Sync` anyway.
	nw.Sync()

Checked https://wiki.postgresql.org/wiki/Fsync_Errors and I think we should check it before we enter bbolt transaction just in case that we lost the error, even if we call fsync again in the transaction.

Steps to reproduce the issue

No Idea about how to reproduce fsync error.

We might introduce dmflaky or something like - https://github.com/kdave/xfstests/blob/0e5c12dfd008efc2848c98108c9237487e91ef35/tests/generic/108#L5-L12

Describe the results you received and expected

Should fast-fail

What version of containerd are you using?

4167416

Any other relevant information

No response

Show configuration if it is related to CRI plugin.

No response

[fuweid]

ping @xinyangge-db @dmcgowan @AkihiroSuda