NetBSD/sys/ufs/lfs/CHANGES

#   $NetBSD: CHANGES,v 1.5 2005/12/11 12:25:26 christos Exp $

kernel:

- Instead of blindly continuing when it encounters an Inode that is
  locked by another process, lfs_markv will process the rest of the
  inodes passed to it and then return EAGAIN.  The cleaner will
  recognize this and not mark the segment clean.  When the cleaner runs
  again, the segment containg the (formerly) locked inode will sort high
  for cleaning, since it is now almost entirely empty.

- A beginning has been made to test keeping atime information in the
  Ifile, instead of on the inodes.  This should make read-mostly
  filesystems significantly faster, since the inodes will then remain
  close to the data blocks on disk; but of course the ifile will be
  somewhat larger.  This code is not enabled, as it makes the format of
  IFILEs change.

- The superblock has been broken into two components: an on-disk
  superblock using fixed-size types, exactly 512 bytes regardless of
  architecture (or could be enlarged in multiples of the media block
  size up to LFS_SBPAD); and an in-memory superblock containing the
  information only useful to a running LFS, including segment pointers,
  etc.  The superblock checksumming code has been modified to make
  future changes to the superblock format easier.

- Because of the way that lfs_writeseg works, buffers are freed before
  they are really written to disk: their contents are copied into large
  buffers which are written async.  Because the buffer cache does not
  serve to throttle these writes, and malloced memory is used to hold them,
  there is a danger of running out of kmem_map.  To avoid this, a new
  compile-time parameter, LFS_THROTTLE, is used as an upper bound for the
  number of partial-segments allowed to be in progress writing at any
  given time.

- If the system crashes between the point that a checkpoint is scheduled
  for writing and the time that the write completes, the filesystem
  could be left in an inconsistent state (no valid checkpoints on
  disk).  To avoid this, we toggle between the first two superblocks
  when checkpointing, and (if it is indicated that no roll-forward agent
  exists) do not allow one checkpoint to occur before the last one has
  completed.  When the filesystem is mounted, it uses the *older* of the
  first two superblocks.

- DIROPs:

  The design of the LFS includes segregating vnodes used in directory
  operations, so that they can be written at the same time during a
  checkpoint, avoiding filesystem inconsistency after a crash.  Code for
  this was partially written for BSD4.4, but was not complete or enabled.

  In particular, vnodes marked VDIROP could be flushed by getnewvnode at
  any time, negating the usefulness of marking a vnode VDIROP, since if
  the filesystem then crashed it would be inconsistent.  Now, when a
  vnode is first marked VDIROP it is also referenced.  To avoid running
  out of vnodes, an attempt to mark more than LFS_MAXDIROP vnodes wth
  VDIROP will sleep, and trigger a partial-segment write when no dirops
  are active.

- LFS maintains a linked list of free inode numbers in the Ifile;
  accesses to this list are now protected by a simple lock.

- lfs_vfree is not allowed to run while an inode has blocks scheduled
  for writing, since that could trigger a miscounting in lfs_truncate.

- lfs_balloc now correctly extends fragments, if a block is written
  beyond the current end-of-file.

- Blocks which have already been gathered into a partial-segment are not
  allowed to be extended, since if they were, any blocks following them
  would either be written in the wrong place, or overwrite other blocks.

- The LFS buffer-header accounting, which triggers a partial-segment
  write if too many buffer-headers are in use by the LFS subystem, has
  been expanded to include *bytes* used in LFS buffers as well.

- Reads of the Ifile, which almost always come from the cleaner, can no
  longer trigger a partial-segment write, since this could cause a
  deadlock.

- Support has been added (but not tested, and currently disabled by
  default) for true read-only filesystems.  Currently, if a filesystem
  is mounted read-only the cleaner can still operate on it, but this
  obviously would not be true for read-only media.  (I think the
  original plan was for the roll-forward agent to operate using this
  "feature"?)

- If a fake buffer is created by lfs_markv and another process draws the
  same block in and changes it, the fake buffer is now discarded and
  replaced by the "real" buffer containing the new data.

- An inode which has blocks gathered no longer has IN_MODIFIED set, but
  still does in fact have dirty blocks attached.  lfs_update will now
  wait for such an inode's writes to complete before it runs,
  suppressing a panic in vinvalbuf.

- Many filesystem operations now update the Ifile's mtime, allowing the
  cleaner to detect when the filesystem is idle, and clean more
  vigorously during such times (cf. Blackwell et al., 1995).

- When writing a partial-segment, make sure that the current segment is
  still marked ACTIVE afterward (otherwise the cleaner might try to
  clean it, since it might well be mostly empty).

- Don't trust the cleaner so much.  Sort the blocks during gathering,
  even if they came from the cleaner; verify the location of on-disk
  inodes, even if the cleaner says it knows where they came from.

- The cleaning code (lfs_markv in particular) has been entirely
  rewritten, and the partial-segment writing code changed to match.
  Lfs_markv no longer uses its own implementation of lfs_segwrite, but
  marks inodes with IN_CLEANING to differentiate them from the
  non-cleaning inodes.  This change fixes numerous problems with the old
  cleaner, including a buffer overrun, and lost extensions in active
  fragments.  lfs_bmapv looks up and returns the addresses of inode
  blocks, so the cleaner can do something intelligent with them.

  If IN_CLEANING is set on an inode during partial-segment write, only fake
  buffers will be written, and IN_MODIFIED will not be cleared, saving
  us from a panic in vinvalbuf.  The addition of IN_CLEANING also allows
  dirops to be active while cleaning is in progress; since otherwise
  buffers engaged in active dirops might be written ahead of schedule,
  and cause an inconsistent checkpoint to be written to disk.

  (XXX - even now, DIROP blocks can sometimes be written to disk, if we
  are cleaning the same blocks as are active?  Grr, I don't see a good
  solution for this!)

- Added sysctl entries for LFS.  In particular, `writeindir' controls
  whether indirect blocks are written during non-checkpoint writes.
  (Since there is no roll-forward agent as yet, there is no penalty in
  not writing indirect blocks.)

- Wake up the cleaner at fs-unmount time, so it can die (if we unmount
  and then remount, we could conceivably get more than one cleaner
  operating at once).

newfs_lfs:

- The ifile inode is now created with the schg flag set, since nothing
  ever modifies it.  This could be a pain for the roll-forward agent,
  but since that should really run *before* the filesystem is mounted,
  I don't care.

- For large disks, it may be necessary to write one or more indirect
  blocks when the ifile inode is created.  Newlfs has been changed to
  write the first indirect block, if necessary.  It should instead just
  build a set of inodes and blocks, and then use the partial-segment
  writing routine mentioned above to write an ifile of whatever size is
  desired.

lfs_cleanerd:

- Now writes information to the syslog.

- Can now deal properly with fragments.

- Sometimes, the cleaner can die.  (Why?)  If this happens and we don't
  notice, we're screwed, since the fs will overfill.  So, the invoked
  cleaner now spawns itself repeatedly, a la init(8), to ensure that a
  cleaner is always present to clean the fs.

- Added a flag to clean more actively, not on low load average but
  filesystem inactivity; a la Blackwell et al., 1995.

fsck_lfs:

- Exists, although it currently cannot actually fix anything (it is a
  diagnostic tool only at this point).