02 Nov, 2011
1 commit
-
Replace remaining direct i_nlink updates with a new set_nlink()
updater function.Signed-off-by: Miklos Szeredi
Tested-by: Toshiyuki Okajima
Signed-off-by: Christoph Hellwig
21 Oct, 2011
5 commits
-
Unfortunately, it is not enough to just ignore locked buffers during
the AIL flush from fsync. We need to be able to ignore all buffers
which are locked, dirty or pinned at this stage as they might have
been added subsequent to the log flush earlier in the fsync function.In addition, this means that we no longer need to rely on i_mutex to
keep out writes during fsync, so we can, as a side-effect, remove
that protection too.Signed-off-by: Steven Whitehouse
Tested-By: Abhijith Das -
Since we have ruled out supporting online filesystem shrink,
it is possible to make the resource group list append only
during the life of a super block. This gives several benefits:Firstly, we only need to read new rindex elements as they are added
rather than needing to reread the whole rindex file each time one
element is added.Secondly, the rindex glock can be held for much shorter periods of
time, and is completely removed from the fast path for allocations.
The lock is taken in shared mode only when updating the resource
groups when the first allocation occurs, and after a grow has
taken place.Thirdly, this results in a reduction in code size, and everything
gets a lot simpler to understand in this area.Signed-off-by: Steven Whitehouse
-
Here is an update of Bob's original rbtree patch which, in addition, also
resolves the rather strange ref counting that was being done relating to
the bitmap blocks.Originally we had a dual system for journaling resource groups. The metadata
blocks were journaled and also the rgrp itself was added to a list. The reason
for adding the rgrp to the list in the journal was so that the "repolish
clones" code could be run to update the free space, and potentially send any
discard requests when the log was flushed. This was done by comparing the
"cloned" bitmap with what had been written back on disk during the transaction
commit.Due to this, there was a requirement to hang on to the rgrps' bitmap buffers
until the journal had been flushed. For that reason, there was a rather
complicated set up in the ->go_lock ->go_unlock functions for rgrps involving
both a mutex and a spinlock (the ->sd_rindex_spin) to maintain a reference
count on the buffers.However, the journal maintains a reference count on the buffers anyway, since
they are being journaled as metadata buffers. So by moving the code which deals
with the post-journal accounting for bitmap blocks to the metadata journaling
code, we can entirely dispense with the rather strange buffer ref counting
scheme and also the requirement to journal the rgrps.The net result of all this is that the ->sd_rindex_spin is left to do exactly
one job, and that is to look after the rbtree or rgrps.This patch is designed to be a stepping stone towards using RCU for the rbtree
of resource groups, however the reduction in the number of uses of the
->sd_rindex_spin is likely to have benefits for multi-threaded workloads,
anyway.The patch retains ->go_lock and ->go_unlock for rgrps, however these maybe also
be removed in future in favour of calling the functions directly where required
in the code. That will allow locking of resource groups without needing to
actually read them in - something that could be useful in speeding up statfs.In the mean time though it is valid to dereference ->bi_bh only when the rgrp
is locked. This is basically the same rule as before, modulo the references not
being valid until the following journal flush.Signed-off-by: Steven Whitehouse
Signed-off-by: Bob Peterson
Cc: Benjamin Marzinski -
Journaled data requires that a complete flush of all dirty data for
the file is done, in order that the ail flush which comes after
will succeed.Also the recently enhanced bug trap can trigger falsely in case
an ail flush from fsync races with a page read. This updates the
bug trap such that it will ignore buffers which are locked and
only trigger on dirty and/or pinned buffers when the ail flush
is run from fsync. The original bug trap is retained when ail
flush is run from ->go_sync()Signed-off-by: Steven Whitehouse
-
The assert was being tested under the wrong lock, a
legacy of the original code. Also, if it does trigger,
the resulting information was not always a lot of help.This moves the patch under the correct lock and also
prints out more useful information in tacking down the
source of the problem.Signed-off-by: Steven Whitehouse
15 Jul, 2011
3 commits
-
This adds S_NOSEC support to GFS2. We set/reset the flag either when
a user calls setattr or when we have just regained the glock
from another node. The flag is only set if there are no xattrs
on the inode and there is no suid bit set.Signed-off-by: Steven Whitehouse
Reviewed-by: Andi Kleen
Cc: Al Viro -
This patch is a performance improvement for GFS2 in a clustered
environment. It makes the glock hold time self-adjusting.Signed-off-by: Bob Peterson
Signed-off-by: Steven Whitehouse -
This patch adds a cache for the hash table to the directory code
in order to help simplify the way in which the hash table is
accessed. This is intended to be a first step towards introducing
some performance improvements in the directory code.There are two follow ups that I'm hoping to see fairly shortly. One
is to simplify the hash table reading code now that we always read the
complete hash table, whether we want one entry or all of them. The
other is to introduce readahead on the heads of the hash chains
which are referred to from the table.The hash table is a maximum of 128k in size, so it is not worth trying
to read it in small chunks.Signed-off-by: Steven Whitehouse
14 Jul, 2011
1 commit
-
This patch contains a few misc fixes which resolve a recently
reported issue. This patch has been a real team effort and has
received a lot of testing.The first issue is that the ail lock needs to be held over a few
more operations. The lock thats added into gfs2_releasepage() may
possibly be a candidate for replacing with RCU at some future
point, but at this stage we've gone for the obvious fix.The second issue is that gfs2_write_inode() can end up calling
a glock recursively when called from gfs2_evict_inode() via the
syncing code, so it needs a guard added.The third issue is that we either need to not truncate the metadata
pages of inodes which have zero link count, but which we cannot
deallocate due to them still being in use by other nodes, or we need
to ensure that those pages have all made it through the journal and
ail lists first. This patch takes the former approach, but the
latter has also been tested and there is nothing to choose between
them performance-wise. So again, we could revise that decision
in the future.Also, the inode eviction process is now better documented.
Signed-off-by: Steven Whitehouse
Tested-by: Bob Peterson
Tested-by: Abhijith Das
Reported-by: Barry J. Marson
Reported-by: David Teigland
12 Jul, 2011
1 commit
-
Right now, there is nothing that forces the log to get flushed when a node
drops its rindex glock so that another node can grow the filesystem. If the
log doesn't get flushed, GFS2 can corrupt the sd_log_le_rg list in the
following way.A node puts an rgd on the list in rg_lo_add(), and then the rindex glock is
dropped so the other node can grow the filesystem. When the node reacquires the
rindex glock, that rgd gets deleted in clear_rgrpdi() before ever being
removed from the list by gfs2_log_flush().This code simply forces a log flush when the rindex glock is invalidated,
solving the problem.Signed-off-by: Benjamin Marzinski
Signed-off-by: Steven Whitehouse
09 May, 2011
1 commit
-
Eventually there will only be a single caller of this code, so lets
move it where it can be made static at some future date.Signed-off-by: Steven Whitehouse
20 Apr, 2011
1 commit
-
This patch is designed to clean up GFS2's fsync
implementation and ensure that it really does get everything on
disk. Since ->write_inode() has been updated, we can call that
via the vfs library function sync_inode_metadata() and the only
remaining thing that has to be done is to ensure that we get
any revoke records in the log after the inode has been written back.Signed-off-by: Steven Whitehouse
18 Apr, 2011
1 commit
-
This adds a couple of missing tests to avoid read-only nodes
from attempting to deallocate unlinked inodes.Signed-off-by: Steven Whitehouse
Reported-by: Michel Andre de la Porte
11 Mar, 2011
1 commit
-
The log lock is currently used to protect the AIL lists and
the movements of buffers into and out of them. The lists
are self contained and no log specific items outside the
lists are accessed when starting or emptying the AIL lists.Hence the operation of the AIL does not require the protection
of the log lock so split them out into a new AIL specific lock
to reduce the amount of traffic on the log lock. This will
also reduce the amount of serialisation that occurs when
the gfs2_logd pushes on the AIL to move it forward.This reduces the impact of log pushing on sequential write
throughput.Signed-off-by: Dave Chinner
Signed-off-by: Steven Whitehouse
21 Jan, 2011
1 commit
-
This has a number of advantages:
- Reduces contention on the hash table lock
- Makes the code smaller and simpler
- Should speed up glock dumps when under load
- Removes ref count changing in examine_bucket
- No longer need hash chain lock in glock_put() in common caseThere are some further changes which this enables and which
we may do in the future. One is to look at using SLAB_RCU,
and another is to look at using a per-cpu counter for the
per-sb glock counter, since that is touched twice in the
lifetime of each glock (but only used at umount time).Signed-off-by: Steven Whitehouse
Cc: Paul E. McKenney
16 Dec, 2010
1 commit
-
There is no requirement to flush the delete workqueue before a
gfs2 filesystem is suspended. The workqueue's work will just
be suspended along with the rest of the tasks on the filesystem.The resolves a deadlock situation where the transaction lock's
demotion code was trying to flush the delete workqueue while at
the same time, the workqueue was waiting for the transaction
lock.The delete workqueue is flushed by gfs2_make_fs_ro() already, so
that umount/remount are correctly protected anyway.Signed-off-by: Steven Whitehouse
06 Oct, 2010
1 commit
-
Mostly the glock operations follow the type of the glock. The
one exception is the transaction glock, so we need to check for
that directly.Reported-by: Dr. David Alan Gilbert
Signed-off-by: Steven Whitehouse
20 Sep, 2010
1 commit
-
With the update of the truncate code, ip->i_disksize and
inode->i_size are merely copies of each other. This means
we can remove ip->i_disksize and use inode->i_size exclusively
reducing the size of a GFS2 inode by 8 bytes.Signed-off-by: Steven Whitehouse
30 Mar, 2010
1 commit
-
…it slab.h inclusion from percpu.h
percpu.h is included by sched.h and module.h and thus ends up being
included when building most .c files. percpu.h includes slab.h which
in turn includes gfp.h making everything defined by the two files
universally available and complicating inclusion dependencies.percpu.h -> slab.h dependency is about to be removed. Prepare for
this change by updating users of gfp and slab facilities include those
headers directly instead of assuming availability. As this conversion
needs to touch large number of source files, the following script is
used as the basis of conversion.http://userweb.kernel.org/~tj/misc/slabh-sweep.py
The script does the followings.
* Scan files for gfp and slab usages and update includes such that
only the necessary includes are there. ie. if only gfp is used,
gfp.h, if slab is used, slab.h.* When the script inserts a new include, it looks at the include
blocks and try to put the new include such that its order conforms
to its surrounding. It's put in the include block which contains
core kernel includes, in the same order that the rest are ordered -
alphabetical, Christmas tree, rev-Xmas-tree or at the end if there
doesn't seem to be any matching order.* If the script can't find a place to put a new include (mostly
because the file doesn't have fitting include block), it prints out
an error message indicating which .h file needs to be added to the
file.The conversion was done in the following steps.
1. The initial automatic conversion of all .c files updated slightly
over 4000 files, deleting around 700 includes and adding ~480 gfp.h
and ~3000 slab.h inclusions. The script emitted errors for ~400
files.2. Each error was manually checked. Some didn't need the inclusion,
some needed manual addition while adding it to implementation .h or
embedding .c file was more appropriate for others. This step added
inclusions to around 150 files.3. The script was run again and the output was compared to the edits
from #2 to make sure no file was left behind.4. Several build tests were done and a couple of problems were fixed.
e.g. lib/decompress_*.c used malloc/free() wrappers around slab
APIs requiring slab.h to be added manually.5. The script was run on all .h files but without automatically
editing them as sprinkling gfp.h and slab.h inclusions around .h
files could easily lead to inclusion dependency hell. Most gfp.h
inclusion directives were ignored as stuff from gfp.h was usually
wildly available and often used in preprocessor macros. Each
slab.h inclusion directive was examined and added manually as
necessary.6. percpu.h was updated not to include slab.h.
7. Build test were done on the following configurations and failures
were fixed. CONFIG_GCOV_KERNEL was turned off for all tests (as my
distributed build env didn't work with gcov compiles) and a few
more options had to be turned off depending on archs to make things
build (like ipr on powerpc/64 which failed due to missing writeq).* x86 and x86_64 UP and SMP allmodconfig and a custom test config.
* powerpc and powerpc64 SMP allmodconfig
* sparc and sparc64 SMP allmodconfig
* ia64 SMP allmodconfig
* s390 SMP allmodconfig
* alpha SMP allmodconfig
* um on x86_64 SMP allmodconfig8. percpu.h modifications were reverted so that it could be applied as
a separate patch and serve as bisection point.Given the fact that I had only a couple of failures from tests on step
6, I'm fairly confident about the coverage of this conversion patch.
If there is a breakage, it's likely to be something in one of the arch
headers which should be easily discoverable easily on most builds of
the specific arch.Signed-off-by: Tejun Heo <tj@kernel.org>
Guess-its-ok-by: Christoph Lameter <cl@linux-foundation.org>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Lee Schermerhorn <Lee.Schermerhorn@hp.com>
01 Mar, 2010
1 commit
-
Since the start of GFS2, an "extra" inode has been used to store
the metadata belonging to each inode. The only reason for using
this inode was to have an extra address space, the other fields
were unused. This means that the memory usage was rather inefficient.The reason for keeping each inode's metadata in a separate address
space is that when glocks are requested on remote nodes, we need to
be able to efficiently locate the data and metadata which relating
to that glock (inode) in order to sync or sync and invalidate it
(depending on the remotely requested lock mode).This patch adds a new type of glock, which has in addition to
its normal fields, has an address space. This applies to all
inode and rgrp glocks (but to no other glock types which remain
as before). As a result, we no longer need to have the second
inode.This results in three major improvements:
1. A saving of approx 25% of memory used in caching inodes
2. A removal of the circular dependency between inodes and glocks
3. No confusion between "normal" and "metadata" inodes in super.cAlthough the first of these is the more immediately apparent, the
second is just as important as it now enables a number of clean
ups at umount time. Those will be the subject of future patches.Signed-off-by: Steven Whitehouse
03 Dec, 2009
1 commit
-
Invalidate all the cached ACLs when we drop the glock.
Signed-off-by: Steven Whitehouse
30 Jul, 2009
1 commit
-
When a file is deleted from a gfs2 filesystem on one node, a dcache
entry for it may still exist on other nodes in the cluster. If this
happens, gfs2 will be unable to free this file on disk. Because of this,
it's possible to have a gfs2 filesystem with no files on it and no free
space. With this patch, when a node receives a callback notifying it
that the file is being deleted on another node, it schedules a new
workqueue thread to remove the file's dcache entry.Signed-off-by: Benjamin Marzinski
Signed-off-by: Steven Whitehouse
20 May, 2009
1 commit
-
This patch improves the error handling in the case where we
discover that the summary information in the resource group
doesn't match the bitmap information while in the process of
allocating blocks. Originally this resulted in a kernel bug,
but this patch changes that so that we return -EIO and print
some messages explaining what went wrong, and how to fix it.We also remember locally not to try and allocate from the
same rgrp again, so that a subsequent allocation in a
different rgrp should succeed.Signed-off-by: Steven Whitehouse
20 Apr, 2009
1 commit
-
The dirty bit can get set during the inode glock sync. Its too
complicated to change that at the moment, so this is the quick
fix - to clear the bit again at the end of the function.Signed-off-by: Steven Whitehouse
24 Mar, 2009
4 commits
-
This cleans up a number of bits of code mostly based in glops.c.
A couple of simple functions have been merged into the callers
to make it more obvious what is going on, the mysterious raising
of i_writecount around the truncate_inode_pages() call has been
removed. The meta_go_* operations have been renamed rgrp_go_*
since that is the only lock type that they are used with.The unused argument of gfs2_read_sb has been removed. Also
a bug has been fixed where a check for the rindex inode was
in the wrong callback. More comments are added, and the
debugging code is improved too.Signed-off-by: Steven Whitehouse
-
This adds a sysfs file called demote_rq to GFS2's
per filesystem directory. Its possible to use this
file to demote arbitrary glocks in exactly the same
way as if a request had come in from a remote node.This is intended for testing issues relating to caching
of data under glocks. Despite that, the interface is
generic enough to send requests to any type of glock,
but be careful as its not always safe to send an
arbitrary message to an arbitrary glock. For that reason
and to prevent DoS, this interface is restricted to root
only.The messages look like this:
:
Example:
echo -n "2:13324 EX" >/sys/fs/gfs2/unity:myfs/demote_rq
Which means "please demote inode glock (type 2) number 13324 so that
I can get an EX (exclusive) lock". The lock modes are those which
would normally be sent by a remote node in its callback so if you
want to unlock a glock, you use EX, to demote to shared, use SH or PR
(depending on whether you like GFS2 or DLM lock modes better!).If the glock doesn't exist, you'll get -ENOENT returned. If the
arguments don't make sense, you'll get -EINVAL returned.The plan is that this interface will be used in combination with
the blktrace patch which I recently posted for comments although
it is, of course, still useful in its own right.Signed-off-by: Steven Whitehouse
-
This patch fixes a deadlock when the journal is flushed and there
are dirty inodes other than the one which caused the journal flush.
Originally the journal flushing code was trying to obtain the
transaction glock while running the flush code for an inode glock.
We no longer require the transaction glock at this point in time
since we know that any attempt to get the transaction glock from
another node will result in a journal flush. So if we are flushing
the journal, we can be sure that the transaction lock is still
cached from when the transaction was started.By inlining a version of gfs2_trans_begin() (minus the bit which
gets the transaction glock) we can avoid the deadlock problems
caused if there is a demote request queued up on the transaction
glock.In addition I've also moved the umount rwsem so that it covers
the glock workqueue, since it all demotions are done by this
workqueue now. That fixes a bug on umount which I came across
while fixing the original problem.Reported-by: David Teigland
Signed-off-by: Steven Whitehouse -
This is the big patch that I've been working on for some time
now. There are many reasons for wanting to make this change
such as:
o Reducing overhead by eliminating duplicated fields between structures
o Simplifcation of the code (reduces the code size by a fair bit)
o The locking interface is now the DLM interface itself as proposed
some time ago.
o Fewer lookups of glocks when processing replies from the DLM
o Fewer memory allocations/deallocations for each glock
o Scope to do further optimisations in the future (but this patch is
more than big enough for now!)Please note that (a) this patch relates to the lock_dlm module and
not the DLM itself, that is still a separate module; and (b) that
we retain the ability to build GFS2 as a standalone single node
filesystem with out requiring the DLM.This patch needs a lot of testing, hence my keeping it I restarted
my -git tree after the last merge window. That way, this has the maximum
exposure before its merged. This is (modulo a few minor bug fixes) the
same patch that I've been posting on and off the the last three months
and its passed a number of different tests so far.Signed-off-by: Steven Whitehouse
05 Jan, 2009
4 commits
-
This patch removes the two daemons, gfs2_scand and gfs2_glockd
and replaces them with a shrinker which is called from the VM.The net result is that GFS2 responds better when there is memory
pressure, since it shrinks the glock cache at the same rate
as the VFS shrinks the dcache and icache. There are no longer
any time based criteria for shrinking glocks, they are kept
until such time as the VM asks for more memory and then we
demote just as many glocks as required.There are potential future changes to this code, including the
possibility of sorting the glocks which are to be written back
into inode number order, to get a better I/O ordering. It would
be very useful to have an elevator based workqueue implementation
for this, as that would automatically deal with the read I/O cases
at the same time.This patch is my answer to Andrew Morton's remark, made during
the initial review of GFS2, asking why GFS2 needs so many kernel
threads, the answer being that it doesn't :-) This patch is a
net loss of about 200 lines of code.Signed-off-by: Steven Whitehouse
-
Following on from the recent clean up of gfs2_quotad, this patch moves
the processing of "truncate in progress" inodes from the glock workqueue
into gfs2_quotad. This fixes a hang due to the "truncate in progress"
processing requiring glocks in order to complete.It might seem odd to use gfs2_quotad for this particular item, but
we have to use a pre-existing thread since creating a thread implies
a GFP_KERNEL memory allocation which is not allowed from the glock
workqueue context. Of the existing threads, gfs2_logd and gfs2_recoverd
may deadlock if used for this operation. gfs2_scand and gfs2_glockd are
both scheduled for removal at some (hopefully not too distant) future
point. That leaves only gfs2_quotad whose workload is generally fairly
light and is easily adapted for this extra task.Also, as a result of this change, it opens the way for a future patch to
make the reading of the inode's information asynchronous with respect to
the glock workqueue, which is another improvement that has been on the list
for some time now.Signed-off-by: Steven Whitehouse
-
Although the glock dumps print quite a lot of information about
the glocks themselves, there are more things which can be
usefully added to the dump realting to the objects themselves.This patch adds a few more fields to the inode and resource
group lines, which should be useful for debugging.Signed-off-by: Steven Whitehouse
-
The final field in gfs2_dinode_host was the i_flags field. Thats
renamed to i_diskflags in order to avoid confusion with the existing
inode flags, and moved into the inode proper at a suitable location
to avoid creating a "hole".At that point struct gfs2_dinode_host is no longer needed and as
promised (quite some time ago!) it can now be removed completely.Signed-off-by: Steven Whitehouse
27 Jun, 2008
1 commit
-
This patch implements a number of cleanups to the core of the
GFS2 glock code. As a result a lot of code is removed. It looks
like a really big change, but actually a large part of this patch
is either removing or moving existing code.There are some new bits too though, such as the new run_queue()
function which is considerably streamlined. Highlights of this
patch include:o Fixes a cluster coherency bug during SH -> EX lock conversions
o Removes the "glmutex" code in favour of a single bit lock
o Removes the ->go_xmote_bh() for inodes since it was duplicating
->go_lock()
o We now only use the ->lm_lock() function for both locks and
unlocks (i.e. unlock is a lock with target mode LM_ST_UNLOCKED)
o The fast path is considerably shortly, giving performance gains
especially with lock_nolock
o The glock_workqueue is now used for all the callbacks from the DLM
which allows us to simplify the lock_dlm module (see following patch)
o The way is now open to make further changes such as eliminating the two
threads (gfs2_glockd and gfs2_scand) in favour of a more efficient
scheme.This patch has undergone extensive testing with various test suites
so it should be pretty stable by now.Signed-off-by: Steven Whitehouse
Cc: Bob Peterson
12 May, 2008
1 commit
-
This patch fixes a GFS2 filesystem consistency error reported from
function do_strip. The problem was caused by a timing window
that allowed two vfs inodes to be created in memory that point
to the same file. The problem is fixed by making the vfs's
iget_test, iget_set mechanism check and set a new bit in the
in-core gfs2_inode structure while the vfs inode spin_lock is held.Signed-off-by: Bob Peterson
Signed-off-by: Steven Whitehouse
31 Mar, 2008
1 commit
-
This patch further reduces GFS2's memory requirements by
eliminating the 64-bit version number fields in lieu of
a couple bits.Signed-off-by: Bob Peterson
Signed-off-by: Steven Whitehouse
25 Jan, 2008
3 commits
-
Previously we were doing (write data, wait for data, write metadata, wait
for metadata). After this patch we so (write metadata, write data, wait for
data, wait for metadata) which should be more efficient.Also I noticed that the drop_bh and xmote_bh functions were almost
identical. In fact the only difference was a single test, and that
test is such that in the drop_bh case, it would always evaluate to
the correct result. As such we can use the xmote_bh functions in
all the places where we were using the drop_bh function and remove
the drop_bh functions.Signed-off-by: Steven Whitehouse
-
The i_cache was designed to keep references to the indirect blocks
used during block mapping so that they didn't have to be looked
up continually. The idea failed because there are too many places
where the i_cache needs to be freed, and this has in the past been
the cause of many bugs.In addition there was no performance benefit being gained since the
disk blocks in question were cached anyway. So this patch removes
it in order to simplify the code to prepare for other changes which
would otherwise have had to add further support for this feature.Signed-off-by: Steven Whitehouse
-
This cleans up the mmap() code path for GFS2 by implementing the
page_mkwrite function for GFS2. We are thus able to use the
generic filemap_fault function for our ->fault() implementation.This now means that shared writable mappings will be much more
efficiently shared across the cluster if there is a reasonable
proportion of read activity (the greater proportion, the better).As a side effect, it also reduces the size of the code, removes
special cases from readpage and readpages, and makes the code
path easier to follow.Signed-off-by: Steven Whitehouse
10 Oct, 2007
1 commit
-
The following alters gfs2_trans_add_revoke() to take a struct
gfs2_bufdata as an argument. This eliminates the memory allocation which
was previously required by making use of the already existing struct
gfs2_bufdata. It makes some sanity checks to ensure that the
gfs2_bufdata has been removed from all the lists before its recycled as
a revoke structure. This saves one memory allocation and one free per
revoke structure.Also as a result, and to simplify the locking, since there is no longer
any blocking code in gfs2_trans_add_revoke() we must hold the log lock
whenever this function is called. This reduces the amount of times we
take and unlock the log lock.Signed-off-by: Steven Whitehouse