kern/45676: ffs/wapbl problems

[yamt%NetBSD.org@localhost]

>Number:         45676
>Category:       kern
>Synopsis:       ffs/wapbl problems
>Confidential:   no
>Severity:       serious
>Priority:       medium
>Responsible:    kern-bug-people
>State:          open
>Class:          sw-bug
>Submitter-Id:   net
>Arrival-Date:   Fri Dec 02 12:15:00 +0000 2011
>Originator:     YAMAMOTO Takashi
>Release:        NetBSD current
>Organization:
        
>Environment:
        
>Description:
        1. ffs_full_fsync has code to skip wapbl_flush if vp->v_dirtyblkhd is
        empty.  i don't think it's a safe optimization because the buffer
        which contains dinode would hang on devvp, not this vp.  probably
        another per-vnode/inode field is necessary to remember if the log is
        involved for the file.

        2. ffs_full_fsync and ffs_fsync skip wapbl_flush for FSYNC_DATAONLY.
        it seems wrong because block pointers should be written out.

        3. ffs_full_fsync and ffs_fsync does not obey FSYNC_CACHE when wapbl
        is used but wapbl_flush is not necessary.

        4. ffs_full_fsync and ffs_fsync should be merged as now softdep has
        gone.

        5. wapbl_flush issues two DIOCCACHESYNC calls which are typically
        very expensive.  while i'm not sure why _only_ this part of code
        needs this level of guarantee, at least it should be reduced to
        a single call, probably using a checksum as the comment in
        wapbl_write_commit suggests.

        6. wapbl should have underlying device requirements documented
        somewhere.  (torn writes, disk cache behaviour, etc)

>How-To-Repeat:
        code inspection.
>Fix:

>Unformatted: