When we read the write-intent-bitmap off the device, we currently
read a whole number of pages.
When PAGE_SIZE is 4K, this works due to the alignment we enforce
on the superblock and bitmap.
When PAGE_SIZE is 64K, this case read past the end-of-device
which causes an error.
When we write the superblock, we ensure to clip the last page
to just be the required size. Copy that code into the read path
to just read the required number of sectors.
Signed-off-by: Neil Brown <neilb@suse.de>
Cc: stable@kernel.org
Fix setting of max_segment_size and seg_boundary mask for stacked md/dm
devices.
When stacking devices (LVM over MD over SCSI) some of the request queue
parameters are not set up correctly in some cases by default, namely
max_segment_size and and seg_boundary mask.
If you create MD device over SCSI, these attributes are zeroed.
Problem become when there is over this mapping next device-mapper mapping
- queue attributes are set in DM this way:
request_queue max_segment_size seg_boundary_mask
SCSI 65536 0xffffffff
MD RAID1 0 0
LVM 65536 -1 (64bit)
Unfortunately bio_add_page (resp. bio_phys_segments) calculates number of
physical segments according to these parameters.
During the generic_make_request() is segment cout recalculated and can
increase bio->bi_phys_segments count over the allowed limit. (After
bio_clone() in stack operation.)
Thi is specially problem in CCISS driver, where it produce OOPS here
BUG_ON(creq->nr_phys_segments > MAXSGENTRIES);
(MAXSEGENTRIES is 31 by default.)
Sometimes even this command is enough to cause oops:
dd iflag=direct if=/dev/<vg>/<lv> of=/dev/null bs=128000 count=10
This command generates bios with 250 sectors, allocated in 32 4k-pages
(last page uses only 1024 bytes).
For LVM layer, it allocates bio with 31 segments (still OK for CCISS),
unfortunatelly on lower layer it is recalculated to 32 segments and this
violates CCISS restriction and triggers BUG_ON().
The patch tries to fix it by:
* initializing attributes above in queue request constructor
blk_queue_make_request()
* make sure that blk_queue_stack_limits() inherits setting
(DM uses its own function to set the limits because it
blk_queue_stack_limits() was introduced later. It should probably switch
to use generic stack limit function too.)
* sets the default seg_boundary value in one place (blkdev.h)
* use this mask as default in DM (instead of -1, which differs in 64bit)
Bugs related to this:
https://bugzilla.redhat.com/show_bug.cgi?id=471639http://bugzilla.kernel.org/show_bug.cgi?id=8672
Signed-off-by: Milan Broz <mbroz@redhat.com>
Reviewed-by: Alasdair G Kergon <agk@redhat.com>
Cc: Neil Brown <neilb@suse.de>
Cc: FUJITA Tomonori <fujita.tomonori@lab.ntt.co.jp>
Cc: Tejun Heo <htejun@gmail.com>
Cc: Mike Miller <mike.miller@hp.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>
dm_any_congested() just checks for the DMF_BLOCK_IO and has no
code to make sure that suspend waits for dm_any_congested() to
complete. This patch adds such a check.
Without it, a race can occur with dm_table_put() attempting to
destroying the table in the wrong thread, the one running
dm_any_congested() which is meant to be quick and return
immediately.
Two examples of problems:
1. Sleeping functions called from congested code, the caller
of which holds a spin lock.
2. An ABBA deadlock between pdflush and multipathd. The two locks
in contention are inode lock and kernel lock.
Signed-off-by: Chandra Seetharaman <sekharan@us.ibm.com>
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
This doesn't fix any bug, just moves wake_up immediately after decrementing
md->pending, for better code readability.
It must be clear to anyone manipulating md->pending to wake up
the queue if md->pending reaches zero, so move the wakeup as close to
the decrementing as possible.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Currently dm ignores the parameters provided to hardware handlers
without providing any notifications to the user.
This patch just prints a warning message so that the user knows that
the arguments are ignored.
Signed-off-by: Chandra Seetharaman <sekharan@us.ibm.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Path activation code is called even when the pgpath is NULL. This could
lead to a panic in activate_path(). Such a panic is seen in -rt kernel.
This problem has been there before the pg_init() was moved to a
workqueue.
Signed-off-by: Chandra Seetharaman <sekharan@us.ibm.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Don't proceed if dm_stripe_init() fails to register itself as a dm target.
Signed-off-by: Heinz Mauelshagen <heinzm@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
We queue work on keventd queue --- so this queue must be flushed in the
destructor. Otherwise, keventd could access mirror_set after it was freed.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Cc: stable@kernel.org
We currently oops with a divide error on starting a linear software
raid array consisting of at least two very small (< 500K) devices.
The bug is caused by the calculation of the hash table size which
tries to compute sector_div(sz, base) with "base" being zero due to
the small size of the component devices of the array.
Fix this by requiring the hash spacing to be at least one which
implies that also "base" is non-zero.
This bug has existed since about 2.6.14.
Cc: stable@kernel.org
Signed-off-by: Andre Noll <maan@systemlinux.org>
Signed-off-by: NeilBrown <neilb@suse.de>
Adding a spare to a raid10 doesn't cause recovery to start.
This is due to an silly type in
commit 6c2fce2ef6
and so is a bug in 2.6.27 and .28-rc.
Thanks to Thomas Backlund for bisecting to find this.
Cc: Thomas Backlund <tmb@mandriva.org>
Cc: stable@kernel.org
Signed-off-by: NeilBrown <neilb@suse.de>
It turns out that it is only safe to call blkdev_ioctl when the device
is actually open (as ->bd_disk is set to NULL on last close). And it
is quite possible for do_md_stop to be called when the device is not
open. So discard the call to blkdev_ioctl(BLKRRPART) which was
added in
commit 934d9c23b4
It is just as easy to call this ioctl from userspace when needed (on
mdadm -S) so leave it out of the kernel
Signed-off-by: NeilBrown <neilb@suse.de>
If there are several snapshots sharing an origin and one is removed
while the origin is being written to, the snapshot's mempool may get
deleted while elements are still referenced.
Prior to dm-snapshot-use-per-device-mempools.patch the pending
exceptions may still have been referenced after the snapshot was
destroyed, but this was not a problem because the shared mempool
was still there.
This patch fixes the problem by tracking the number of mempool elements
in use.
The scenario:
- You have an origin and two snapshots 1 and 2.
- Someone writes to the origin.
- It creates two exceptions in the snapshots, snapshot 1 will be primary
exception, snapshot 2's pending_exception->primary_pe will point to the
exception in snapshot 1.
- The exceptions are being relocated, relocation of exception 1 finishes
(but it's pending_exception is still allocated, because it is referenced
by an exception from snapshot 2)
- The user lvremoves snapshot 1 --- it calls just suspend (does nothing)
and destructor. md->pending is zero (there is no I/O submitted to the
snapshot by md layer), so it won't help us.
- The destructor waits for kcopyd jobs to finish on snapshot 1 --- but
there are none.
- The destructor on snapshot 1 cleans up everything.
- The relocation of exception on snapshot 2 finishes, it drops reference
on primary_pe. This frees its primary_pe pointer. Primary_pe points to
pending exception created for snapshot 1. So it frees memory into
non-existing mempool.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
register_snapshot() performs a GFP_KERNEL allocation while holding
_origins_lock for write, but that could write out dirty pages onto a
device that attempts to acquire _origins_lock for read, resulting in
deadlock.
So move the allocation up before taking the lock.
This path is not performance-critical, so it doesn't matter that we
allocate memory and free it if we find that we won't need it.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
md arrays are not currently destroyed when they are stopped - they
remain in /sys/block. Last time I tried this I tripped over locking
too much.
A consequence of this is that udev doesn't remove anything from /dev.
This is rather ugly.
As an interim measure until proper device removal can be achieved,
make sure all partitions are removed using the BLKRRPART ioctl, and
send a KOBJ_CHANGE when an md array is stopped.
Signed-off-by: NeilBrown <neilb@suse.de>
* 'for-linus' of git://neil.brown.name/md:
md: allow extended partitions on md devices.
md: use sysfs_notify_dirent to notify changes to md/dev-xxx/state
md: use sysfs_notify_dirent to notify changes to md/array_state
* git://git.kernel.org/pub/scm/linux/kernel/git/viro/bdev: (66 commits)
[PATCH] kill the rest of struct file propagation in block ioctls
[PATCH] get rid of struct file use in blkdev_ioctl() BLKBSZSET
[PATCH] get rid of blkdev_locked_ioctl()
[PATCH] get rid of blkdev_driver_ioctl()
[PATCH] sanitize blkdev_get() and friends
[PATCH] remember mode of reiserfs journal
[PATCH] propagate mode through swsusp_close()
[PATCH] propagate mode through open_bdev_excl/close_bdev_excl
[PATCH] pass fmode_t to blkdev_put()
[PATCH] kill the unused bsize on the send side of /dev/loop
[PATCH] trim file propagation in block/compat_ioctl.c
[PATCH] end of methods switch: remove the old ones
[PATCH] switch sr
[PATCH] switch sd
[PATCH] switch ide-scsi
[PATCH] switch tape_block
[PATCH] switch dcssblk
[PATCH] switch dasd
[PATCH] switch mtd_blkdevs
[PATCH] switch mmc
...
Now that lookup_bdev is exported and used by dm just use it directly
instead of through a trivial wrapper.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
This patch tidies local_init() in preparation for request-based dm.
No functional change.
Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
This patch removes the DM_WQ_FLUSH_ALL state that is unnecessary.
The dm_queue_flush(md, DM_WQ_FLUSH_ALL, NULL) in dm_suspend()
is never invoked because:
- 'goto flush_and_out' is the same as 'goto out' because
the 'goto flush_and_out' is called only when '!noflush'
- If r is non-zero, then the code above will invoke 'goto out'
and skip this code.
No functional change.
Signed-off-by: Kiyoshi Ueda <k-ueda@ct.jp.nec.com>
Signed-off-by: Jun'ichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: Milan Broz <mbroz@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Separate the region hash code from raid1 so it can be shared by forthcoming
targets. Use BUG_ON() for failed async dm_io() calls.
Signed-off-by: Heinz Mauelshagen <hjm@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
When a bio gets split, mark its fragments with the BIO_CLONED flag.
Signed-off-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Remove waitqueue no longer needed with the async crypto interface.
Signed-off-by: Milan Broz <mbroz@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
When writing io, dm-crypt has to allocate a new cloned bio
and encrypt the data into newly-allocated pages attached to this bio.
In rare cases, because of hw restrictions (e.g. physical segment limit)
or memory pressure, sometimes more than one cloned bio has to be used,
each processing a different fragment of the original.
Currently there is one waitqueue which waits for one fragment to finish
and continues processing the next fragment.
But when using asynchronous crypto this doesn't work, because several
fragments may be processed asynchronously or in parallel and there is
only one crypt context that cannot be shared between the bio fragments.
The result may be corruption of the data contained in the encrypted bio.
The patch fixes this by allocating new dm_crypt_io structs (with new
crypto contexts) and running them independently.
The fragments contains a pointer to the base dm_crypt_io struct to
handle reference counting, so the base one is properly deallocated
after all the fragments are finished.
In a low memory situation, this only uses one additional object from the
mempool. If the mempool is empty, the next allocation simple waits for
previous fragments to complete.
Signed-off-by: Milan Broz <mbroz@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Prepare local sector variable (offset) for later patch.
Do not update io->sector for still-running I/O.
No functional change.
Signed-off-by: Milan Broz <mbroz@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Change #include "dm.h" to #include <linux/device-mapper.h> in all targets.
Targets should not need direct access to internal DM structures.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Move array_too_big to include/linux/device-mapper.h because it is
used by targets.
Remove the test from dm-raid1 as the number of mirror legs is limited
such that it can never fail. (Even for stripes it seems rather
unlikely.)
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
We must zero the next chunk on disk *before* writing out the current chunk, not
after. Otherwise if the machine crashes at the wrong time, the "end of metadata"
marker may be missing.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Cc: stable@kernel.org
Use a separate buffer for writing zeroes to the on-disk snapshot
exception store, make the updating of ps->current_area explicit and
refactor the code in preparation for the fix in the next patch.
No functional change.
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Cc: stable@kernel.org
The last_percent field is unused - remove it.
(It dates from when events were triggered as each X% filled up.)
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Fix a race condition with primary_pe ref_count handling.
put_pending_exception runs under dm_snapshot->lock, it does atomic_dec_and_test
on primary_pe->ref_count, and later does atomic_read primary_pe->ref_count.
__origin_write does atomic_dec_and_test on primary_pe->ref_count without holding
dm_snapshot->lock.
This opens the following race condition:
Assume two CPUs, CPU1 is executing put_pending_exception (and holding
dm_snapshot->lock). CPU2 is executing __origin_write in parallel.
primary_pe->ref_count == 2.
CPU1:
if (primary_pe && atomic_dec_and_test(&primary_pe->ref_count))
origin_bios = bio_list_get(&primary_pe->origin_bios);
... decrements primary_pe->ref_count to 1. Doesn't load origin_bios
CPU2:
if (first && atomic_dec_and_test(&primary_pe->ref_count)) {
flush_bios(bio_list_get(&primary_pe->origin_bios));
free_pending_exception(primary_pe);
/* If we got here, pe_queue is necessarily empty. */
return r;
}
... decrements primary_pe->ref_count to 0, submits pending bios, frees
primary_pe.
CPU1:
if (!primary_pe || primary_pe != pe)
free_pending_exception(pe);
... this has no effect.
if (primary_pe && !atomic_read(&primary_pe->ref_count))
free_pending_exception(primary_pe);
... sees ref_count == 0 (written by CPU 2), does double free !!
This bug can happen only if someone is simultaneously writing to both the
origin and the snapshot.
If someone is writing only to the origin, __origin_write will submit kcopyd
request after it decrements primary_pe->ref_count (so it can't happen that the
finished copy races with primary_pe->ref_count decrementation).
If someone is writing only to the snapshot, __origin_write isn't invoked at all
and the race can't happen.
The race happens when someone writes to the snapshot --- this creates
pending_exception with primary_pe == NULL and starts copying. Then, someone
writes to the same chunk in the snapshot, and __origin_write races with
termination of already submitted request in pending_complete (that calls
put_pending_exception).
This race may be reason for bugs:
http://bugzilla.kernel.org/show_bug.cgi?id=11636https://bugzilla.redhat.com/show_bug.cgi?id=465825
The patch fixes the code to make sure that:
1. If atomic_dec_and_test(&primary_pe->ref_count) returns false, the process
must no longer dereference primary_pe (because someone else may free it under
us).
2. If atomic_dec_and_test(&primary_pe->ref_count) returns true, the process
is responsible for freeing primary_pe.
Signed-off-by: Mikulas Patocka <mpatocka@redhat.com>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Cc: stable@kernel.org
Write throughput to LVM snapshot origin volume is an order
of magnitude slower than those to LV without snapshots or
snapshot target volumes, especially in the case of sequential
writes with O_SYNC on.
The following patch originally written by Kevin Jamieson and
Jan Blunck and slightly modified for the current RCs by myself
tries to improve the performance by modifying the behaviour
of kcopyd, so that it pushes back an I/O job to the head of
the job queue instead of the tail as process_jobs() currently
does when it has to wait for free pages. This way, write
requests aren't shuffled to cause extra seeks.
I tested the patch against 2.6.27-rc5 and got the following results.
The test is a dd command writing to snapshot origin followed by fsync
to the file just created/updated. A couple of filesystem benchmarks
gave me similar results in case of sequential writes, while random
writes didn't suffer much.
dd if=/dev/zero of=<somewhere on snapshot origin> bs=4096 count=...
[conv=notrunc when updating]
1) linux 2.6.27-rc5 without the patch, write to snapshot origin,
average throughput (MB/s)
10M 100M 1000M
create,dd 511.46 610.72 11.81
create,dd+fsync 7.10 6.77 8.13
update,dd 431.63 917.41 12.75
update,dd+fsync 7.79 7.43 8.12
compared with write throughput to LV without any snapshots,
all dd+fsync and 1000 MiB writes perform very poorly.
10M 100M 1000M
create,dd 555.03 608.98 123.29
create,dd+fsync 114.27 72.78 76.65
update,dd 152.34 1267.27 124.04
update,dd+fsync 130.56 77.81 77.84
2) linux 2.6.27-rc5 with the patch, write to snapshot origin,
average throughput (MB/s)
10M 100M 1000M
create,dd 537.06 589.44 46.21
create,dd+fsync 31.63 29.19 29.23
update,dd 487.59 897.65 37.76
update,dd+fsync 34.12 30.07 26.85
Although still not on par with plain LV performance -
cannot be avoided because it's copy on write anyway -
this simple patch successfully improves throughtput
of dd+fsync while not affecting the rest.
Signed-off-by: Jan Blunck <jblunck@suse.de>
Signed-off-by: Kazuo Ito <ito.kazuo@oss.ntt.co.jp>
Signed-off-by: Alasdair G Kergon <agk@redhat.com>
Cc: stable@kernel.org
To keep the size of changesets sane we split the switch by drivers;
to keep the damn thing bisectable we do the following:
1) rename the affected methods, add ones with correct
prototypes, make (few) callers handle both. That's this changeset.
2) for each driver convert to new methods. *ALL* drivers
are converted in this series.
3) kill the old (renamed) methods.
Note that it _is_ a flagday; all in-tree drivers are converted and by the
end of this series no trace of old methods remain. The only reason why
we do that this way is to keep the damn thing bisectable and allow per-driver
debugging if anything goes wrong.
New methods:
open(bdev, mode)
release(disk, mode)
ioctl(bdev, mode, cmd, arg) /* Called without BKL */
compat_ioctl(bdev, mode, cmd, arg)
locked_ioctl(bdev, mode, cmd, arg) /* Called with BKL, legacy */
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Analog of blkdev_driver_ioctl() with sane arguments. For
now uses fake struct file, by the end of the series it won't
and blkdev_driver_ioctl() will become a wrapper around it.
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
The new extended partition support provides a much nicer was
to have partitions on md devices that the 'mdp' alternate major.
We cannot really get rid of 'mdp' at this time, but we can
enable extended partitions as that will probably make life
easier for sysadmins.
Signed-off-by: NeilBrown <neilb@suse.de>
The 'state' file for a device reports, for example, when the device
has failed. Changes should be reported to userspace ASAP without
the possibility of blocking on low-memory. sysfs_notify does
have that possibility (as it takes a mutex which can be held
across a kmalloc) so use sysfs_notify_dirent instead.
Signed-off-by: NeilBrown <neilb@suse.de>
Now that we have sysfs_notify_dirent, use it to notify changes
to md/array_state.
As sysfs_notify_dirent can be called in atomic context, we can
remove the delayed notify and the MD_NOTIFY_ARRAY_STATE flag.
Signed-off-by: NeilBrown <neilb@suse.de>
* 'for-linus' of git://neil.brown.name/md:
md: fix input truncation in safe_delay_store()
md: check for memory allocation failure in faulty personality
md: build failure due to missing delay.h
md: Relax minimum size restrictions on chunk_size.
md: remove space after function name in declaration and call.
md: Remove unnecessary #includes, #defines, and function declarations.
md: Convert remaining 1k representations in linear.c to sectors.
md: linear.c: Make two local variables sector-based.
md: linear: Represent dev_info->size and dev_info->offset in sectors.
md: linear.c: Remove broken debug code.
md: linear.c: Remove pointless initialization of curr_offset.
md: linear.c: Fix typo in comment.
md: Don't try to set an array to 'read-auto' if it is already in that state.
md: Allow metadata_version to be updated for externally managed metadata.
md: Fix rdev_size_store with size == 0
safe_delay_store() currently truncates the last character of input since
it tells strlcpy that the buffer can only hold 'len' characters, off by
one. sysfs already null terminates the buffer, so just increase the
last argument to strlcpy.
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
Signed-off-by: NeilBrown <neilb@suse.de>
It's a fault injection module, but I don't think we should oops here.
Signed-off-by: Sven Wegener <sven.wegener@stealer.net>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Neil Brown <neilb@suse.de>
Today's linux-next build (powerpc ppc64_defconfig) failed like this:
drivers/md/raid1.c: In function 'sync_request':
drivers/md/raid1.c:1759: error: implicit declaration of function 'msleep_interruptible'
make[3]: *** [drivers/md/raid1.o] Error 1
make[3]: *** Waiting for unfinished jobs....
drivers/md/raid10.c: In function 'sync_request':
drivers/md/raid10.c:1749: error: implicit declaration of function 'msleep_interruptible'
make[3]: *** [drivers/md/raid10.o] Error 1
drivers/md/md.c: In function 'md_do_sync':
drivers/md/md.c:5915: error: implicit declaration of function 'msleep'
Caused by commit 6caa3b0bbdb474647f6bdd8a958ffc46f78d8d58 ("md: Remove
unnecessary #includes, #defines, and function declarations"). I added
the following patch.
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Signed-off-by: NeilBrown <neilb@suse.de>
Multipath is best at handling transport errors. If it gets a device
error then there is not much the multipath layer can do. It will just
access the same device but from a different path.
This patch breaks up failfast into device, transport and driver errors.
The multipath layers (md and dm mutlipath) only ask the lower levels to
fast fail transport errors. The user of failfast, read ahead, will ask
to fast fail on all errors.
Note that blk_noretry_request will return true if any failfast bit
is set. This allows drivers that do not support the multipath failfast
bits to continue to fail on any failfast error like before. Drivers
like scsi that are able to fail fast specific errors can check
for the specific fail fast type. In the next patch I will convert
scsi.
Signed-off-by: Mike Christie <michaelc@cs.wisc.edu>
Cc: Jens Axboe <jens.axboe@oracle.com>
Signed-off-by: James Bottomley <James.Bottomley@HansenPartnership.com>
