When we start taking a reference to the per-ag for every cached
buffer in the system, kernel lockstat profiling on an 8-way create
workload shows the mp->m_perag_lock has higher acquisition rates
than the inode lock and has significantly more contention. That is,
it becomes the highest contended lock in the system.

The perag lookup is trivial to convert to lock-less RCU lookups
because perag structures never go away. Hence the only thing we need
to protect against is tree structure changes during a grow. This can
be done simply by replacing the locking in xfs_perag_get() with RCU
read locking. This removes the mp->m_perag_lock completely from this
path.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NAlex Elder <aelder@sgi.com>

When we start taking references per cached buffer to the the perag
it is cached on, it will blow the current debug maximum reference
count assert out of the water. The assert has never caught a bug,
and we have tracing to track changes if there ever is a problem,
so just remove it.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NAlex Elder <aelder@sgi.com>

Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <david@fromorbit.com>

Dmapi support was never merged upstream, but we still have a lot of hooks
bloating XFS for it, all over the fast pathes of the filesystem.

This patch drops over 700 lines of dmapi overhead.  If we'll ever get HSM
support in mainline at least the namespace events can be done much saner
in the VFS instead of the individual filesystem, so it's not like this
is much help for future work.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <dchinner@redhat.com>

The block number comes from bulkstat based inode lookups to shortcut
the mapping calculations. We ar enot able to trust anything from
bulkstat, so drop the block number as well so that the correct
lookups and mappings are always done.
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>

If a filesystem is mounted without the inode64 mount option we
should still be able to access inodes not fitting into 32 bits, just
not created new ones.  For this to work we need to make sure the
inode cache radix tree is initialized for all allocation groups, not
just those we plan to allocate inodes from.  This patch makes sure
we initialize the inode cache radix tree for all allocation groups,
and also cleans xfs_initialize_perag up a bit to separate the
inode32 logical from the general perag structure setup.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NAlex Elder <aelder@sgi.com>

The use of radix_tree_preload() only works if the radix tree was
initialised without the __GFP_WAIT flag. The per-ag tree uses
GFP_NOFS, so does not trigger allocation of new tree nodes from the
preloaded array. Hence it enters the allocator with a spinlock held
and triggers the might_sleep() warnings.

Reported-by; Chris Mason <chris.mason@oracle.com>
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NAlex Elder <aelder@sgi.com>

Many places in the xfs code return E2BIG when they really mean
EFBIG; trying to grow past 16T on a 32 bit machine, for example,
says "Argument list too long" rather than "File too large" which is
not particularly helpful.

Some of these don't make perfect sense as EFBIG either, but still
better than E2BIG IMHO.
Signed-off-by: NEric Sandeen <sandeen@sandeen.net>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NAlex Elder <aelder@sgi.com>

This can't be enabled through the build system and has been dead for
ages.  Note that the CRC patches add back log checksumming, but the
code is quite different from the version removed here anyway.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <david@fromorbit.com>

The current default size of the reserved blocks pool is easy to deplete
with certain workloads, in particular workloads that do lots of concurrent
delayed allocation extent conversions.  If enough transactions are running
in parallel and the entire pool is consumed then subsequent calls to
xfs_trans_reserve() will fail with ENOSPC.  Also add a rate limited
warning so we know if this starts happening again.

This is an updated version of an old patch from Lachlan McIlroy.
Signed-off-by: NDave Chinner <david@fromorbit.com>
Signed-off-by: NAlex Elder <aelder@sgi.com>

Currently the code to implement the file operations is split over
two small files.  Merge the content of xfs_lrw.c into xfs_file.c to
have it in one place.  Note that I haven't done various cleanups
that are possible after this yet, they will follow in the next
patch.  Also the function xfs_dev_is_read_only which was in
xfs_lrw.c before really doesn't fit in here at all and was moved to
xfs_mount.c.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <david@fromorbit.com>
Signed-off-by: NAlex Elder <aelder@sgi.com>

We currently do background inode flush asynchronously, resulting in
inodes being written in whatever order the background writeback
issues them. Not only that, there are also blocking and non-blocking
asynchronous inode flushes, depending on where the flush comes from.

This patch completely removes asynchronous inode writeback. It
removes all the strange writeback modes and replaces them with
either a synchronous flush or a non-blocking delayed write flush.
That is, inode flushes will only issue IO directly if they are
synchronous, and background flushing may do nothing if the operation
would block (e.g. on a pinned inode or buffer lock).

Delayed write flushes will now result in the inode buffer sitting in
the delwri queue of the buffer cache to be flushed by either an AIL
push or by the xfsbufd timing out the buffer. This will allow
accumulation of dirty inode buffers in memory and allow optimisation
of inode cluster writeback at the xfsbufd level where we have much
greater queue depths than the block layer elevators. We will also
get adjacent inode cluster buffer IO merging for free when a later
patch in the series allows sorting of the delayed write buffers
before dispatch.

This effectively means that any inode that is written back by
background writeback will be seen as flush locked during AIL
pushing, and will result in the buffers being pushed from there.
This writeback path is currently non-optimal, but the next patch
in the series will fix that problem.

A side effect of this delayed write mechanism is that background
inode reclaim will no longer directly flush inodes, nor can it wait
on the flush lock. The result is that inode reclaim must leave the
inode in the reclaimable state until it is clean. Hence attempts to
reclaim a dirty inode in the background will simply skip the inode
until it is clean and this allows other mechanisms (i.e. xfsbufd) to
do more optimal writeback of the dirty buffers. As a result, the
inode reclaim code has been rewritten so that it no longer relies on
the ambiguous return values of xfs_iflush() to determine whether it
is safe to reclaim an inode.

Portions of this patch are derived from patches by Christoph
Hellwig.

Version 2:
- cleanup reclaim code as suggested by Christoph
- log background reclaim inode flush errors
- just pass sync flags to xfs_iflush
Signed-off-by: NDave Chinner <david@fromorbit.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>

This mangles the reserved blocks counts a little more.

1) add a helper function for the default reserved count
2) add helper functions to save/restore counts on ro/rw
3) save/restore reserved blocks on freeze/thaw
4) disallow changing reserved count while readonly

V2: changed field name to match Dave's changes
Signed-off-by: NEric Sandeen <sandeen@sandeen.net>
Signed-off-by: NAlex Elder <aelder@sgi.com>

Remove the XFS_LOG_FORCE argument which was always set, and the
XFS_LOG_URGE define, which was never used.

Split xfs_log_force into a two helpers - xfs_log_force which forces
the whole log, and xfs_log_force_lsn which forces up to the
specified LSN.  The underlying implementations already were entirely
separate, as were the users.

Also re-indent the new _xfs_log_force/_xfs_log_force which
previously had a weird coding style.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NAlex Elder <aelder@sgi.com>

Currently we define aliases for the buffer flags in various
namespaces, which only adds confusion.  Remove all but the XBF_
flags to clean this up a bit.

Note that we still abuse XFS_B_ASYNC/XBF_ASYNC for some non-buffer
uses, but I'll clean that up later.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <david@fromorbit.com>
Signed-off-by: NAlex Elder <aelder@sgi.com>

gcc warns of an array subscript out of bounds in xfs_mod_sb().
The code is written in such a way that if the array subscript is
out of bounds, then it will assert fail. Rearrange the code to
avoid the bounds check warning.
Signed-off-by: NDave Chinner <david@fromorbit.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>

Now that the perag structure is allocated memory rather than held in
an array, we don't need to have the busy extent array external to
the structure. Embed it into the perag structure to avoid needing an
extra allocation when setting up.
Signed-off-by: NDave Chinner <david@fromorbit.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NAlex Elder <aelder@sgi.com>

Add proper error handling in case an error occurs while initializing
new perag structures for a mount point.  The mount structure is
restored to its previous state by deleting and freeing any perag
structures added during the call.
Signed-off-by: NDave Chinner <david@fromorbit.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NAlex Elder <aelder@sgi.com>

Uninline xfs_perag_{get,put} so that tracepoints can be inserted
into them to speed debugging of reference count problems.
Signed-off-by: NDave Chinner <david@fromorbit.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NAlex Elder <aelder@sgi.com>

Reference count the per-ag structures to ensure that we keep get/put
pairs balanced. Assert that the reference counts are zero at unmount
time to catch leaks. In future, reference counts will enable us to
safely remove perag structures by allowing us to detect when they
are no longer in use.
Signed-off-by: NDave Chinner <david@fromorbit.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NAlex Elder <aelder@sgi.com>

The use of an array for the per-ag structures requires reallocation
of the array when growing the filesystem. This requires locking
access to the array to avoid use after free situations, and the
locking is difficult to get right. To avoid needing to reallocate an
array, change the per-ag structures to an allocated object per ag
and index them using a tree structure.

The AGs are always densely indexed (hence the use of an array), but
the number supported is 2^32 and lookups tend to be random and hence
indexing needs to scale. A simple choice is a radix tree - it works
well with this sort of index.  This change also removes another
large contiguous allocation from the mount/growfs path in XFS.

The growing process now needs to change to only initialise the new
AGs required for the extra space, and as such only needs to
exclusively lock the tree for inserts. The rest of the code only
needs to lock the tree while doing lookups, and hence this will
remove all the deadlocks that currently occur on the m_perag_lock as
it is now an innermost lock. The lock is also changed to a spinlock
from a read/write lock as the hold time is now extremely short.

To complete the picture, the per-ag structures will need to be
reference counted to ensure that we don't free/modify them while
they are still in use.  This will be done in subsequent patch.
Signed-off-by: NDave Chinner <david@fromorbit.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NAlex Elder <aelder@sgi.com>

Convert the remaining direct lookups of the per ag structures to use
get/put accesses. Ensure that the loops across AGs and prior users
of the interface balance gets and puts correctly.
Signed-off-by: NDave Chinner <david@fromorbit.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NAlex Elder <aelder@sgi.com>

Convert the old xfs tracing support that could only be used with the
out of tree kdb and xfsidbg patches to use the generic event tracer.

To use it make sure CONFIG_EVENT_TRACING is enabled and then enable
all xfs trace channels by:

   echo 1 > /sys/kernel/debug/tracing/events/xfs/enable

or alternatively enable single events by just doing the same in one
event subdirectory, e.g.

   echo 1 > /sys/kernel/debug/tracing/events/xfs/xfs_ihold/enable

or set more complex filters, etc. In Documentation/trace/events.txt
all this is desctribed in more detail.  To reads the events do a

   cat /sys/kernel/debug/tracing/trace

Compared to the last posting this patch converts the tracing mostly to
the one tracepoint per callsite model that other users of the new
tracing facility also employ.  This allows a very fine-grained control
of the tracing, a cleaner output of the traces and also enables the
perf tool to use each tracepoint as a virtual performance counter,
     allowing us to e.g. count how often certain workloads git various
     spots in XFS.  Take a look at

    http://lwn.net/Articles/346470/

for some examples.

Also the btree tracing isn't included at all yet, as it will require
additional core tracing features not in mainline yet, I plan to
deliver it later.

And the really nice thing about this patch is that it actually removes
many lines of code while adding this nice functionality:

 fs/xfs/Makefile                |    8
 fs/xfs/linux-2.6/xfs_acl.c     |    1
 fs/xfs/linux-2.6/xfs_aops.c    |   52 -
 fs/xfs/linux-2.6/xfs_aops.h    |    2
 fs/xfs/linux-2.6/xfs_buf.c     |  117 +--
 fs/xfs/linux-2.6/xfs_buf.h     |   33
 fs/xfs/linux-2.6/xfs_fs_subr.c |    3
 fs/xfs/linux-2.6/xfs_ioctl.c   |    1
 fs/xfs/linux-2.6/xfs_ioctl32.c |    1
 fs/xfs/linux-2.6/xfs_iops.c    |    1
 fs/xfs/linux-2.6/xfs_linux.h   |    1
 fs/xfs/linux-2.6/xfs_lrw.c     |   87 --
 fs/xfs/linux-2.6/xfs_lrw.h     |   45 -
 fs/xfs/linux-2.6/xfs_super.c   |  104 ---
 fs/xfs/linux-2.6/xfs_super.h   |    7
 fs/xfs/linux-2.6/xfs_sync.c    |    1
 fs/xfs/linux-2.6/xfs_trace.c   |   75 ++
 fs/xfs/linux-2.6/xfs_trace.h   | 1369 +++++++++++++++++++++++++++++++++++++++++
 fs/xfs/linux-2.6/xfs_vnode.h   |    4
 fs/xfs/quota/xfs_dquot.c       |  110 ---
 fs/xfs/quota/xfs_dquot.h       |   21
 fs/xfs/quota/xfs_qm.c          |   40 -
 fs/xfs/quota/xfs_qm_syscalls.c |    4
 fs/xfs/support/ktrace.c        |  323 ---------
 fs/xfs/support/ktrace.h        |   85 --
 fs/xfs/xfs.h                   |   16
 fs/xfs/xfs_ag.h                |   14
 fs/xfs/xfs_alloc.c             |  230 +-----
 fs/xfs/xfs_alloc.h             |   27
 fs/xfs/xfs_alloc_btree.c       |    1
 fs/xfs/xfs_attr.c              |  107 ---
 fs/xfs/xfs_attr.h              |   10
 fs/xfs/xfs_attr_leaf.c         |   14
 fs/xfs/xfs_attr_sf.h           |   40 -
 fs/xfs/xfs_bmap.c              |  507 +++------------
 fs/xfs/xfs_bmap.h              |   49 -
 fs/xfs/xfs_bmap_btree.c        |    6
 fs/xfs/xfs_btree.c             |    5
 fs/xfs/xfs_btree_trace.h       |   17
 fs/xfs/xfs_buf_item.c          |   87 --
 fs/xfs/xfs_buf_item.h          |   20
 fs/xfs/xfs_da_btree.c          |    3
 fs/xfs/xfs_da_btree.h          |    7
 fs/xfs/xfs_dfrag.c             |    2
 fs/xfs/xfs_dir2.c              |    8
 fs/xfs/xfs_dir2_block.c        |   20
 fs/xfs/xfs_dir2_leaf.c         |   21
 fs/xfs/xfs_dir2_node.c         |   27
 fs/xfs/xfs_dir2_sf.c           |   26
 fs/xfs/xfs_dir2_trace.c        |  216 ------
 fs/xfs/xfs_dir2_trace.h        |   72 --
 fs/xfs/xfs_filestream.c        |    8
 fs/xfs/xfs_fsops.c             |    2
 fs/xfs/xfs_iget.c              |  111 ---
 fs/xfs/xfs_inode.c             |   67 --
 fs/xfs/xfs_inode.h             |   76 --
 fs/xfs/xfs_inode_item.c        |    5
 fs/xfs/xfs_iomap.c             |   85 --
 fs/xfs/xfs_iomap.h             |    8
 fs/xfs/xfs_log.c               |  181 +----
 fs/xfs/xfs_log_priv.h          |   20
 fs/xfs/xfs_log_recover.c       |    1
 fs/xfs/xfs_mount.c             |    2
 fs/xfs/xfs_quota.h             |    8
 fs/xfs/xfs_rename.c            |    1
 fs/xfs/xfs_rtalloc.c           |    1
 fs/xfs/xfs_rw.c                |    3
 fs/xfs/xfs_trans.h             |   47 +
 fs/xfs/xfs_trans_buf.c         |   62 -
 fs/xfs/xfs_vnodeops.c          |    8
 70 files changed, 2151 insertions(+), 2592 deletions(-)
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NAlex Elder <aelder@sgi.com>

Remove our own STATIC_INLINE macro.  For small function inside
implementation files just use STATIC and let gcc inline it, and for
those in headers do the normal static inline - they are all small
enough to be inlined for debug builds, too.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <david@fromorbit.com>
Signed-off-by: NAlex Elder <aelder@sgi.com>

Currently the low-level buffer cache interfaces are highly confusing
as we have a _flags variant of each that does actually respect the
flags, and one without _flags which has a flags argument that gets
ignored and overriden with a default set.  Given that very few places
use the default arguments get rid of the duplication and convert all
callers to pass the flags explicitly.  Also remove the now confusing
_flags postfix.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <david@fromorbit.com>
Signed-off-by: NAlex Elder <aelder@sgi.com>

When completing I/O requests we must not allow the memory allocator to
recurse into the filesystem, as we might deadlock on waiting for the
I/O completion otherwise.  The only thing currently allocating normal
GFP_KERNEL memory is the allocation of the transaction structure for
the unwritten extent conversion.  Add a memflags argument to
_xfs_trans_alloc to allow controlling the allocator behaviour.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reported-by: NThomas Neumann <tneumann@users.sourceforge.net>
Tested-by: NThomas Neumann <tneumann@users.sourceforge.net>
Reviewed-by: NAlex Elder <aelder@sgi.com>
Signed-off-by: NAlex Elder <aelder@sgi.com>

The xfs_icsb_modify_counters() function no longer needs the cpu variable
if we use this_cpu_ptr() and we can get rid of get/put_cpu().
Acked-by: NTejun Heo <tj@kernel.org>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Acked-by: NOlaf Weber <olaf@sgi.com>
Signed-off-by: NChristoph Lameter <cl@linux-foundation.org>
Signed-off-by: NTejun Heo <tj@kernel.org>

A lot more functions could be made static, but they need
forward declarations; this does some easy ones, and also
found a few unused functions in the process.
Signed-off-by: NEric Sandeen <sandeen@sandeen.net>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NFelix Blyakher <felixb@sgi.com>

A lot more functions could be made static, but they need
forward declarations; this does some easy ones, and also
found a few unused functions in the process.
Signed-off-by: NEric Sandeen <sandeen@sandeen.net>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NFelix Blyakher <felixb@sgi.com>

The noblock parameter of xfs_reclaim_inodes is only ever set to zero. Remove
it and all the conditional code that is never executed.
Signed-off-by: NDave Chinner <david@fromorbit.com>
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NEric Sandeen <sandeen@sandeen.net>

Kill the quota ops function vector and replace it with direct calls or
stubs in the CONFIG_XFS_QUOTA=n case.

Make sure we check XFS_IS_QUOTA_RUNNING in the right spots.  We can remove
the number of those checks because the XFS_TRANS_DQ_DIRTY flag can't be set
otherwise.

This brings us back closer to the way this code worked in IRIX and earlier
Linux versions, but we keep a lot of the more useful factoring of common
code.

Eventually we should also kill xfs_qm_bhv.c, but that's left for a later
patch.

Reduces the size of the source code by about 250 lines and the size of
XFS module by about 1.5 kilobytes with quotas enabled:

   text	   data	    bss	    dec	    hex	filename
 615957	   2960	   3848	 622765	  980ad	fs/xfs/xfs.o
 617231	   3152	   3848	 624231	  98667	fs/xfs/xfs.o.old

Fallout:

 - xfs_qm_dqattach is split into xfs_qm_dqattach_locked which expects
   the inode locked and xfs_qm_dqattach which does the locking around it,
   thus removing XFS_QMOPT_ILOCKED.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NEric Sandeen <sandeen@sandeen.net>

There had been reports where xfs filesystem was randomly
corrupted with fsfuzzer, and xfs failed to handle it
gracefully. This patch fixes couple of reported problem
by providing additional checks in the superblock
validation routine.
Signed-off-by: NOlaf Weber <olaf@sgi.com>
Reviewed-by: NJosef 'Jeff' Sipek <jeffpc@josefsipek.net>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NFelix Blyakher <felixb@sgi.com>

There had been reports where xfs filesystem was randomly
corrupted with fsfuzzer, and xfs failed to handle it
gracefully. This patch fixes couple of reported problem
by providing additional checks in the superblock
validation routine.
Signed-off-by: NOlaf Weber <olaf@sgi.com>
Reviewed-by: NJosef 'Jeff' Sipek <jeffpc@josefsipek.net>
Reviewed-by: NChristoph Hellwig <hch@lst.de>
Signed-off-by: NFelix Blyakher <felixb@sgi.com>

The uuid table handling should not be part of a semi-generic uuid library
but in the XFS code using it, so move those bits to xfs_mount.c and
refactor the whole glob to make it a proper abstraction.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NFelix Blyakher <felixb@sgi.com>

With the upcoming v3 inodes the default attroffset needs to be calculated
for each specific inode, so we can't cache it in the superblock anymore.

Also replace the assert for wrong inode sizes with a proper error check
also included in non-debug builds.  Note that the ENOSYS return for
that might seem odd, but that error is returned by xfs_mount_validate_sb
for all theoretically valid but not supported filesystem geometries.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NJosef 'Jeff' Sipek <jeffpc@josefsipek.net>

Signed-off-by: NMalcolm Parsons <malcolm.parsons@gmail.com>
Reviewed-by: NChristoph Hellwig <hch@lst.de>

With the upcoming v3 inodes the inode data/attr area size needs to be
calculated for each specific inode, so we can't cache it in the superblock
anymore.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NEric Sandeen <sandeen@sandeen.net>
Reviewed-by: NFelix Blyakher <felixb@sgi.com>

Kill the current xfs_log_unmount wrapper and opencode the two function
calls in the only caller.  Rename the current xfs_log_unmount_dealloc to
xfs_log_unmount as it undoes xfs_log_mount and the new name makes that
more clear.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NDave Chinner <david@fromorbit.com>

When mount fails after allocating the real-time inodes we currently leak
them.  Add a new helper to free the real-time inodes which can be used by
both the mount and unmount path.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NFelix Blyakher <felixb@sgi.com>

Clean up the error handling in xfs_mountfs.  Use readable goto label names,
simplify the uuid handling and other error conditions.
Signed-off-by: NChristoph Hellwig <hch@lst.de>
Reviewed-by: NFelix Blyakher <felixb@sgi.com>