* 'for-linus' of git://neil.brown.name/md: (75 commits)
md/raid10: handle further errors during fix_read_error better.
md/raid10: Handle read errors during recovery better.
md/raid10: simplify read error handling during recovery.
md/raid10: record bad blocks due to write errors during resync/recovery.
md/raid10: attempt to fix read errors during resync/check
md/raid10: Handle write errors by updating badblock log.
md/raid10: clear bad-block record when write succeeds.
md/raid10: avoid writing to known bad blocks on known bad drives.
md/raid10 record bad blocks as needed during recovery.
md/raid10: avoid reading known bad blocks during resync/recovery.
md/raid10 - avoid reading from known bad blocks - part 3
md/raid10: avoid reading from known bad blocks - part 2
md/raid10: avoid reading from known bad blocks - part 1
md/raid10: Split handle_read_error out from raid10d.
md/raid10: simplify/reindent some loops.
md/raid5: Clear bad blocks on successful write.
md/raid5. Don't write to known bad block on doubtful devices.
md/raid5: write errors should be recorded as bad blocks if possible.
md/raid5: use bad-block log to improve handling of uncorrectable read errors.
md/raid5: avoid reading from known bad blocks.
...
When recovering one or more devices, if all the good devices have
bad blocks we should record a bad block on the device being rebuilt.
If this fails, we need to abort the recovery.
To ensure we don't think that we aborted later than we actually did,
we need to move the check for MD_RECOVERY_INTR earlier in md_do_sync,
in particular before mddev->curr_resync is updated.
Signed-off-by: NeilBrown <neilb@suse.de>
It is only safe to choose not to write to a bad block if that bad
block is safely recorded in metadata - i.e. if it has been
'acknowledged'.
If it hasn't we need to wait for the acknowledgement.
We support that using rdev->blocked wait and
md_wait_for_blocked_rdev by introducing a new device flag
'BlockedBadBlock'.
This flag is only advisory.
It is cleared whenever we acknowledge a bad block, so that a waiter
can re-check the particular bad blocks that it is interested it.
It should be set by a caller when they find they need to wait.
This (set after test) is inherently racy, but as
md_wait_for_blocked_rdev already has a timeout, losing the race will
have minimal impact.
When we clear "Blocked" was also clear "BlockedBadBlocks" incase it
was set incorrectly (see above race).
We also modify the way we manage 'Blocked' to fit better with the new
handling of 'BlockedBadBlocks' and to make it consistent between
externally managed and internally managed metadata. This requires
that each raidXd loop checks if the metadata needs to be written and
triggers a write (md_check_recovery) if needed. Otherwise a queued
write request might cause raidXd to wait for the metadata to write,
and only that thread can write it.
Before writing metadata, we set FaultRecorded for all devices that
are Faulty, then after writing the metadata we clear Blocked for any
device for which the Fault was certainly Recorded.
The 'faulty' device flag now appears in sysfs if the device is faulty
*or* it has unacknowledged bad blocks. So user-space which does not
understand bad blocks can continue to function correctly.
User space which does, should not assume a device is faulty until it
sees the 'faulty' flag, and then sees the list of unacknowledged bad
blocks is empty.
Signed-off-by: NeilBrown <neilb@suse.de>
If a device has ever seen a write error, we will want to handle
known-bad-blocks differently.
So create an appropriate state flag and export it via sysfs.
Signed-off-by: NeilBrown <neilb@suse.de>
Reviewed-by: Namhyung Kim <namhyung@gmail.com>
Now that we have a bad block list, we should not read from those
blocks.
There are several main parts to this:
1/ read_balance needs to check for bad blocks, and return not only
the chosen device, but also how many good blocks are available
there.
2/ fix_read_error needs to avoid trying to read from bad blocks.
3/ read submission must be ready to issue multiple reads to
different devices as different bad blocks on different devices
could mean that a single large read cannot be served by any one
device, but can still be served by the array.
This requires keeping count of the number of outstanding requests
per bio. This count is stored in 'bi_phys_segments'
4/ retrying a read needs to also be ready to submit a smaller read
and queue another request for the rest.
This does not yet handle bad blocks when reading to perform resync,
recovery, or check.
'md_trim_bio' will also be used for RAID10, so put it in md.c and
export it.
Signed-off-by: NeilBrown <neilb@suse.de>
Space must have been allocated when array was created.
A feature flag is set when the badblock list is non-empty, to
ensure old kernels don't load and trust the whole device.
We only update the on-disk badblocklist when it has changed.
If the badblocklist (or other metadata) is stored on a bad block, we
don't cope very well.
If metadata has no room for bad block, flag bad-blocks as disabled,
and do the same for 0.90 metadata.
Signed-off-by: NeilBrown <neilb@suse.de>
This can show the log (providing it fits in one page) and
allows bad blocks to be 'acknowledged' meaning that they
have safely been recorded in metadata.
Clearing bad blocks is not allowed via sysfs (except for
code testing). A bad block can only be cleared when
a write to the block succeeds.
Signed-off-by: NeilBrown <neilb@suse.de>
Reviewed-by: Namhyung Kim <namhyung@gmail.com>
This the first step in allowing md to track bad-blocks per-device so
that we can fail individual blocks rather than the whole device.
This patch just adds a data structure for recording bad blocks, with
routines to add, remove, search the list.
Signed-off-by: NeilBrown <neilb@suse.de>
Reviewed-by: Namhyung Kim <namhyung@gmail.com>
When calling bioset_create we pass the size of the front_pad as
sizeof(mddev)
which looks suspicious as mddev is a pointer and so it looks like a
common mistake where
sizeof(*mddev)
was intended.
The size is actually correct as we want to store a pointer in the
front padding of the bios created by the bioset, so make the intent
more explicit by using
sizeof(mddev_t *)
Reported-by: Zdenek Kabelac <zdenek.kabelac@gmail.com>
Signed-off-by: NeilBrown <neilb@suse.de>
This patch causes MD to generate an event (for device-mapper) when the
synchronization thread is reaped. This is expected behavior for device-mapper.
Signed-off-by: Jonathan Brassow <jbrassow@redhat.com>
Signed-off-by: NeilBrown <neilb@suse.de>
page_address() returns void pointer, so the casts can be removed.
Signed-off-by: Namhyung Kim <namhyung@gmail.com>
Signed-off-by: NeilBrown <neilb@suse.de>
If we hit a read error while recovering a mirror, we want to abort the
recovery without necessarily failing the disk - as having a disk this
a read error is better than not having an array at all.
Currently this is managed with a per-array flag "recovery_disabled"
and is only implemented for RAID1. For RAID10 we will need finer
grained control as we might want to disable recovery for individual
devices separately.
So push more of the decision making into the personality.
'recovery_disabled' is now a 'cookie' which is copied when the
personality want to disable recovery and is changed when a device is
added to the array as this is used as a trigger to 'try recovery
again'.
This will allow RAID10 to get the control that it needs.
Signed-off-by: NeilBrown <neilb@suse.de>
Commit c89a8eee61 ("Allow faulty devices to be removed from a
readonly array.") added some work on ro array in the function,
but it couldn't be done since we didn't allow the ro array to be
handled from the beginning. Fix it.
Signed-off-by: Namhyung Kim <namhyung@gmail.com>
Signed-off-by: NeilBrown <neilb@suse.de>
There are places where sysfs links to rdev are handled
in a same way. Add the helper functions to consolidate
them.
Signed-off-by: Namhyung Kim <namhyung@gmail.com>
Signed-off-by: NeilBrown <neilb@suse.de>
Moving the event counter into the dynamically allocated 'struc seq_file'
allows poll() support without the need to allocate its own tracking
structure.
All current users are switched over to use the new counter.
Requested-by: Andrew Morton akpm@linux-foundation.org
Acked-by: NeilBrown <neilb@suse.de>
Tested-by: Lucas De Marchi lucas.demarchi@profusion.mobi
Signed-off-by: Kay Sievers <kay.sievers@vrfy.org>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
If a device fails in a way that causes pending request to take a while
to complete, md will not be able to immediately remove it from the
array in remove_and_add_spares.
It will then incorrectly look like a spare device and md will try to
recover it even though it is failed.
This leads to a recovery process starting and instantly aborting over
and over again.
We should check if the device is faulty before considering it to be a
spare. This will avoid trying to start a recovery that cannot
proceed.
This bug was introduced in 2.6.26 so that patch is suitable for any
kernel since then.
Cc: stable@kernel.org
Reported-by: Jim Paradis <james.paradis@stratus.com>
Signed-off-by: NeilBrown <neilb@suse.de>
Add the 'sync_super' function pointer to MD array structure (struct mddev_s)
If device-mapper (dm-raid.c) is to define its own on-disk superblock and be
able to load it, there must still be a way for MD to initiate superblock
updates. The simplest way to make this happen is to provide a pointer in
the MD array structure that can be set by device-mapper (or other module)
with a function to do this. If the function has been set, it will be used;
otherwise, the method with be looked up via 'super_types' as usual.
Signed-off-by: Jonathan Brassow <jbrassow@redhat.com>
Signed-off-by: NeilBrown <neilb@suse.de>
Move personality and sync/recovery thread starting outside md_run.
Moving the wakeup's of the personality and sync/recovery threads out of
md_run and into do_md_run and mddev_resume solves two issues:
1) It allows bitmap_load to be called before the sync_thread is run and
2) when MD personalities are used by device-mapper (dm-raid.c), the start-up
of the array is better alligned with device-mapper primatives
(CTR/resume/suspend/DTR). I/O - in this case, recovery operations - should
not happen until after a resume has taken place.
Signed-off-by: Jonathan Brassow <jbrassow@redhat.com>
Signed-off-by: NeilBrown <neilb@suse.de>
Make message a bit clearer by s/blocks/k/
I chose 'k' vs 'kiB' or 'kB' because it is what is used earlier in the
message. 'k' may be a bit ambigous, but I think it's better than "blocks"
which normally means 512, but means 1024 in MD.
Signed-off-by: Jonathan Brassow <jbrassow@redhat.com>
Signed-off-by: NeilBrown <neilb@suse.de>
Disallow resync I/O while the RAID array is suspended.
Recovery, resync, and metadata I/O should not be allowed while a device is
suspended.
Signed-off-by: Jonathan Brassow <jbrassow@redhat.com>
Signed-off-by: NeilBrown <neilb@suse.de>
Don't attempt md_integrity_register if there is no gendisk struct available.
When MD arrays are built via device-mapper, the gendisk structure is not
available via mddev.
Signed-off-by: Jonathan Brassow <jbrassow@redhat.com>
Signed-off-by: NeilBrown <neilb@suse.de>
The sysfs attribute 'resync_start' (known internally as recovery_cp),
records where a resync is up to. A value of 0 means the array is
not known to be in-sync at all. A value of MaxSector means the array
is believed to be fully in-sync.
When the size of member devices of an array (RAID1,RAID4/5/6) is
increased, the array can be increased to match. This process sets
resync_start to the old end-of-device offset so that the new part of
the array gets resynced.
However with RAID1 (and RAID6) a resync is not technically necessary
and may be undesirable. So it would be good if the implied resync
after the array is resized could be avoided.
So: change 'resync_start' so the value can be changed while the array
is active, and as a precaution only allow it to be changed while
resync/recovery is 'frozen'. Changing it once resync has started is
not going to be useful anyway.
This allows the array to be resized without a resync by:
write 'frozen' to 'sync_action'
write new size to 'component_size' (this will set resync_start)
write 'none' to 'resync_start'
write 'idle' to 'sync_action'.
Also slightly improve some tests on recovery_cp when resizing
raid1/raid5. Now that an arbitrary value could be set we should be
more careful in our tests.
Signed-off-by: NeilBrown <neilb@suse.de>
The 'add_new_disk' ioctl can be used to add a device either as a
spare, or as an active disk that just needs to be resynced based on
write-intent-bitmap information (re-add)
Currently if a re-add is requested but fails we add as a spare
instead. This makes it impossible for user-space to check for
failure.
So change to require that a re-add attempt will either succeed or
completely fail. User-space can then decide what to do next.
Signed-off-by: NeilBrown <neilb@suse.de>
There is a race when creating an md device by opening /dev/mdXX.
If two processes do this at much the same time they will follow the
call path
__blkdev_get -> get_gendisk -> kobj_lookup
The first will call
-> md_probe -> md_alloc -> add_disk -> blk_register_region
and the race happens when the second gets to kobj_lookup after
add_disk has called blk_register_region but before it returns to
md_alloc.
In the case the second will not call md_probe (as the probe is already
done) but will get a handle on the gendisk, return to __blkdev_get
which will then call md_open (via the ->open) pointer.
As mddev->gendisk hasn't been set yet, md_open will think something is
wrong an return with ERESTARTSYS.
This can loop endlessly while the first thread makes no progress
through add_disk. Nothing is blocking it, but due to scheduler
behaviour it doesn't get a turn.
So this is essentially a live-lock.
We fix this by simply moving the assignment to mddev->gendisk before
the call the add_disk() so md_open doesn't get confused.
Also move blk_queue_flush earlier because add_disk should be as late
as possible.
To make sure that md_open doesn't complete until md_alloc has done all
that is needed, we take mddev->open_mutex during the last part of
md_alloc. md_open will wait for this.
This can cause a lock-up on boot so Cc:ing for stable.
For 2.6.36 and earlier a different patch will be needed as the
'blk_queue_flush' call isn't there.
Signed-off-by: NeilBrown <neilb@suse.de>
Reported-by: Thomas Jarosch <thomas.jarosch@intra2net.com>
Tested-by: Thomas Jarosch <thomas.jarosch@intra2net.com>
Cc: stable@kernel.org
Problem:
After raid4->raid0 takeover operation, another takeover operation
(e.g raid0->raid10) results "kernel oops".
Root cause:
Variables 'degraded' in mddev structure is not cleared
on raid45->raid0 takeover.
This patch reset this variable.
Signed-off-by: Krzysztof Wojcik <krzysztof.wojcik@intel.com>
Signed-off-by: NeilBrown <neilb@suse.de>
When an md device adds a request to a queue, it can call
mddev_check_plugged.
If this succeeds then we know that the md thread will be woken up
shortly, and ->plug_cnt will be non-zero until then, so some
processing can be delayed.
If it fails, then no unplug callback is expected and the make_request
function needs to do whatever is required to make the request happen.
Signed-off-by: NeilBrown <neilb@suse.de>
md has some plugging infrastructure for RAID5 to use because the
normal plugging infrastructure required a 'request_queue', and when
called from dm, RAID5 doesn't have one of those available.
This relied on the ->unplug_fn callback which doesn't exist any more.
So remove all of that code, both in md and raid5. Subsequent patches
with restore the plugging functionality.
Signed-off-by: NeilBrown <neilb@suse.de>
We incorrectly returned -EINVAL when none of the devices in the array
had an integrity profile. This in turn prevented mdadm from starting
the metadevice. Fix this so we only return errors on mismatched
profiles and memory allocation failures.
Reported-by: Giacomo Catenazzi <cate@cateee.net>
Reported-by: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
* 'for-2.6.39/core' of git://git.kernel.dk/linux-2.6-block: (65 commits)
Documentation/iostats.txt: bit-size reference etc.
cfq-iosched: removing unnecessary think time checking
cfq-iosched: Don't clear queue stats when preempt.
blk-throttle: Reset group slice when limits are changed
blk-cgroup: Only give unaccounted_time under debug
cfq-iosched: Don't set active queue in preempt
block: fix non-atomic access to genhd inflight structures
block: attempt to merge with existing requests on plug flush
block: NULL dereference on error path in __blkdev_get()
cfq-iosched: Don't update group weights when on service tree
fs: assign sb->s_bdi to default_backing_dev_info if the bdi is going away
block: Require subsystems to explicitly allocate bio_set integrity mempool
jbd2: finish conversion from WRITE_SYNC_PLUG to WRITE_SYNC and explicit plugging
jbd: finish conversion from WRITE_SYNC_PLUG to WRITE_SYNC and explicit plugging
fs: make fsync_buffers_list() plug
mm: make generic_writepages() use plugging
blk-cgroup: Add unaccounted time to timeslice_used.
block: fixup plugging stubs for !CONFIG_BLOCK
block: remove obsolete comments for blkdev_issue_zeroout.
blktrace: Use rq->cmd_flags directly in blk_add_trace_rq.
...
Fix up conflicts in fs/{aio.c,super.c}
MD and DM create a new bio_set for every metadevice. Each bio_set has an
integrity mempool attached regardless of whether the metadevice is
capable of passing integrity metadata. This is a waste of memory.
Instead we defer the allocation decision to MD and DM since we know at
metadevice creation time whether integrity passthrough is needed or not.
Automatic integrity mempool allocation can then be removed from
bioset_create() and we make an explicit integrity allocation for the
fs_bio_set.
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Reported-by: Zdenek Kabelac <zkabelac@redhat.com>
Acked-by: Mike Snitzer <snizer@redhat.com>
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
* 'for-2.6.39' of git://git.kernel.org/pub/scm/linux/kernel/git/tj/wq:
workqueue: fix build failure introduced by s/freezeable/freezable/
workqueue: add system_freezeable_wq
rds/ib: use system_wq instead of rds_ib_fmr_wq
net/9p: replace p9_poll_task with a work
net/9p: use system_wq instead of p9_mux_wq
xfs: convert to alloc_workqueue()
reiserfs: make commit_wq use the default concurrency level
ocfs2: use system_wq instead of ocfs2_quota_wq
ext4: convert to alloc_workqueue()
scsi/scsi_tgt_lib: scsi_tgtd isn't used in memory reclaim path
scsi/be2iscsi,qla2xxx: convert to alloc_workqueue()
misc/iwmc3200top: use system_wq instead of dedicated workqueues
i2o: use alloc_workqueue() instead of create_workqueue()
acpi: kacpi*_wq don't need WQ_MEM_RECLAIM
fs/aio: aio_wq isn't used in memory reclaim path
input/tps6507x-ts: use system_wq instead of dedicated workqueue
cpufreq: use system_wq instead of dedicated workqueues
wireless/ipw2x00: use system_wq instead of dedicated workqueues
arm/omap: use system_wq in mailbox
workqueue: use WQ_MEM_RECLAIM instead of WQ_RESCUER
With the plugging now being explicitly controlled by the
submitter, callers need not pass down unplugging hints
to the block layer. If they want to unplug, it's because they
manually plugged on their own - in which case, they should just
unplug at will.
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
Code has been converted over to the new explicit on-stack plugging,
and delay users have been converted to use the new API for that.
So lets kill off the old plugging along with aops->sync_page().
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
Revert
b821eaa572
and
f3b99be19d
When I wrote the first of these I had a wrong idea about the
lifetime of 'struct block_device'. It can disappear at any time that
the block device is not open if it falls out of the inode cache.
So relying on the 'size' recorded with it to detect when the
device size has changed and so we need to revalidate, is wrong.
Rather, we really do need the 'changed' attribute stored directly in
the mddev and set/tested as appropriate.
Without this patch, a sequence of:
mknod / open / close / unlink
(which can cause a block_device to be created and then destroyed)
will result in a rescan of the partition table and consequence removal
and addition of partitions.
Several of these in a row can get udev racing to create and unlink and
other code can get confused.
With the patch, the rescan is only performed when needed and so there
are no races.
This is suitable for any stable kernel from 2.6.35.
Reported-by: "Wojcik, Krzysztof" <krzysztof.wojcik@intel.com>
Signed-off-by: NeilBrown <neilb@suse.de>
Cc: stable@kernel.org
'mdp' devices are md devices with preallocated device numbers
for partitions. As such it is possible to mknod and open a partition
before opening the whole device.
this causes md_probe() to be called with a device number of a
partition, which in-turn calls mddev_find with such a number.
However mddev_find expects the number of a 'whole device' and
does the wrong thing with partition numbers.
So add code to mddev_find to remove the 'partition' part of
a device number and just work with the 'whole device'.
This patch addresses https://bugzilla.kernel.org/show_bug.cgi?id=28652
Reported-by: hkmaly@bigfoot.com
Signed-off-by: NeilBrown <neilb@suse.de>
Cc: <stable@kernel.org>
If the desired size of an array is set (via sysfs) before the array is
active (which is the normal sequence), we currrently call set_capacity
immediately.
This means that a subsequent 'open' (as can be caused by some
udev-triggers program) will notice the new size and try to probe for
partitions. However as the array isn't quite ready yet the read will
fail. Then when the array is read, as the size doesn't change again
we don't try to re-probe.
So when setting array size via sysfs, only call set_capacity if the
array is already active.
Signed-off-by: NeilBrown <neilb@suse.de>
md_make_request was calling bio_sectors() for part_stat_add
after it was calling the make_request function. This is
bad because the make_request function can free the bio and
because the bi_size field can change around.
The fix here was suggested by Jens Axboe. It saves the
sector count before the make_request call. I hit this
with CONFIG_DEBUG_PAGEALLOC turned on while trying to break
his pretty fusionio card.
Cc: <stable@kernel.org>
Signed-off-by: Chris Mason <chris.mason@oracle.com>
Signed-off-by: NeilBrown <neilb@suse.de>
Activating a spare in an array while resync/recovery is already
happening can lead the that spare being marked in-sync when it isn't
really.
So don't allow the 'slot' to be set (this activating the device)
while resync/recovery is happening.
Signed-off-by: NeilBrown <neilb@suse.de>
There is no need to set this to zero at this point. It will be
set to zero by remove_and_add_spares or at the start of
md_do_sync at the latest.
And setting it to zero before MD_RECOVERY_RUNNING is cleared can
make a 'zero' appear briefly in the 'sync_completed' sysfs attribute
just as resync is finishing.
So simply remove this setting to zero.
Signed-off-by: NeilBrown <neilb@suse.de>
remove_and_add_spares is called in two places where the needs really
are very different.
remove_and_add_spares should not be called on an array which is about
to be reshaped as some extra devices might have been manually added
and that would remove them. However if the array is 'read-auto',
that will currently happen, which is bad.
So in the 'ro != 0' case don't call remove_and_add_spares but simply
remove the failed devices as the comment suggests is needed.
Signed-off-by: NeilBrown <neilb@suse.de>
This flag is not needed and is used badly.
Devices that are included in a native-metadata array are reserved
exclusively for that array - and currently have AllReserved set.
They all are bd_claimed for the rdev and so cannot be shared.
Devices that are included in external-metadata arrays can be shared
among multiple arrays - providing there is no overlap.
These are bd_claimed for md in general - not for a particular rdev.
When changing the amount of a device that is used in an array we need
to check for overlap. This currently includes a check on AllReserved
So even without overlap, sharing with an AllReserved device is not
allowed.
However the bd_claim usage already precludes sharing with these
devices, so the test on AllReserved is not needed. And in fact it is
wrong.
As this is the only use of AllReserved, simply remove all usage and
definition of AllReserved.
Signed-off-by: NeilBrown <neilb@suse.de>
If we try to update_raid_disks and it fails, we should put
'delta_disks' back to zero. This is important because some code,
such as slot_store, assumes that delta_disks has been validated.
Signed-off-by: NeilBrown <neilb@suse.de>
WQ_RESCUER is now an internal flag and should only be used in the
workqueue implementation proper. Use WQ_MEM_RECLAIM instead.
This doesn't introduce any functional difference.
Signed-off-by: Tejun Heo <tj@kernel.org>
Cc: dm-devel@redhat.com
Cc: Neil Brown <neilb@suse.de>
Commit e09b457b (block: simplify holder symlink handling) incorrectly
assumed that there is only one link at maximum. dm may use multiple
links and expects block layer to track reference count for each link,
which is different from and unrelated to the exclusive device holder
identified by @holder when the device is opened.
Remove the single holder assumption and automatic removal of the link
and revive the per-link reference count tracking. The code
essentially behaves the same as before commit e09b457b sans the
unnecessary kobject reference count dancing.
While at it, note that this facility should not be used by anyone else
than the current ones. Sysfs symlinks shouldn't be abused like this
and the whole thing doesn't belong in the block layer at all.
Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: Milan Broz <mbroz@redhat.com>
Cc: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Cc: Neil Brown <neilb@suse.de>
Cc: linux-raid@vger.kernel.org
Cc: Kay Sievers <kay.sievers@vrfy.org>
Signed-off-by: Jens Axboe <jaxboe@fusionio.com>
