Re: Syncer "giving up" on buffers

On Tue, 2 Sep 2003, Tim Robbins wrote:

> On Tue, Sep 02, 2003 at 05:10:48AM +1000, Bruce Evans wrote:
>
> > Apparently the bug fixed in ext2fs/fs.h revs 1.3, 1.4 and 1.6 (etc.)
> > was restored in rev.1.14.  I think this is because B_LOCKED buffers
> > were ignored in the sync() in boot() and flushed later when
> > vfs_unmountall() calls ext2fs_unmount(), but they are now seen in the
> > sync() so vfs_unmountall() is not called.
> >
> > Rev.1.3 of ext2fs/fs.h (etc.) abuses B_LOCKED to do little more than
> > make the sync() ignore ext2fs's private buffers (its complications are
> > mainly to handle the resulting B_LOCKED buffers).  It wants to brelse()
> > the buffers so that their BUF_REFCOUNT() is 0 and the sync() in boot()
> > is happy to handle them.  In the original fix, I think the buffers
> > could be B_DELWRI and then the sync() would fulush them, but setting
> > B_DELWRI was wrong and was changed (in rev.1.4) to setting the private
> > flag B_DIRTY instead.  Rev.1.13 esssentially removes the brelse() and
> > adds a new complication (BUF_KERNPROC()) and keeps the old ones.  I
> > think the BUF_KERNPROC() is less than useful -- without the brelse()'s,
> > the buffers are completely private to the file system.
>
> Is there any particular reason why ext2fs keeps these buffers locked instead
> of reading/writing them in when it needs to, or storing them in malloc'd
> memory and reading/writing them at mount/unmount time? Do we need to
> ensure that group descriptors & inode/block bitmaps are not written to
> disk until the filesystem gets unmounted, or is it merely to improve
> performance and simplify the code?

It is mostly to improve performance and complicate the code :-).  ext2fs
in Linux does it that way (with a few more optimizations) at least when
parts of the FreeBSD ext2fs code were obtained form Linux, and the port
to FreeBSD was fairly quick and dirty.

I fixed this in my version a few years ago but never got around to
committing the changes.  See sys.dif.gz in ~bde on freefall for an
old version.

Unchanged ext2fs has separate private caches for:
- superblocks.  These are easy to handle like summary info in ffs
  superblocks.  malloc()ing the space for this works well because the
  space is relatively tiny and syncing the info in the normal way is
  not very important (fsck can rebuild it after a crash).
- group descriptors.  These weren't so easy to handle like ffs does.
  Essentially the same handling works, but much more code is involved.

Bruce