Add __percpu sparse annotations to places which didn't make it in one
of the previous patches.  All converions are trivial.

These annotations are to make sparse consider percpu variables to be
in a different address space and warn if accessed without going
through percpu accessors.  This patch doesn't affect normal builds.
Signed-off-by: NTejun Heo <tj@kernel.org>
Acked-by: NBorislav Petkov <borislav.petkov@amd.com>
Cc: Dan Williams <dan.j.williams@intel.com>
Cc: Huang Ying <ying.huang@intel.com>
Cc: Len Brown <lenb@kernel.org>
Cc: Neil Brown <neilb@suse.de>

Suggested by  Oren Held <orenhe@il.ibm.com>
Signed-off-by: NNeilBrown <neilb@suse.de>

The post-barrier-flush is sent by md as soon as make_request on the
barrier write completes.  For raid5, the data might not be in the
per-device queues yet.  So for barrier requests, wait for any
pre-reading to be done so that the request will be in the per-device
queues.

We use the 'preread_active' count to check that nothing is still in
the preread phase, and delay the decrement of this count until after
write requests have been submitted to the underlying devices.
Signed-off-by: NNeilBrown <neilb@suse.de>

Previously barriers were only supported on RAID1.  This is because
other levels requires synchronisation across all devices and so needed
a different approach.
Here is that approach.

When a barrier arrives, we send a zero-length barrier to every active
device.  When that completes - and if the original request was not
empty -  we submit the barrier request itself (with the barrier flag
cleared) and then submit a fresh load of zero length barriers.

The barrier request itself is asynchronous, but any subsequent
request will block until the barrier completes.

The reason for clearing the barrier flag is that a barrier request is
allowed to fail.  If we pass a non-empty barrier through a striping
raid level it is conceivable that part of it could succeed and part
could fail.  That would be way too hard to deal with.
So if the first run of zero length barriers succeed, we assume all is
sufficiently well that we send the request and ignore errors in the
second run of barriers.

RAID5 needs extra care as write requests may not have been submitted
to the underlying devices yet.  So we flush the stripe cache before
proceeding with the barrier.

Note that the second set of zero-length barriers are submitted
immediately after the original request is submitted.  Thus when
a personality finds mddev->barrier to be set during make_request,
it should not return from make_request until the corresponding
per-device request(s) have been queued.

That will be done in later patches.
Signed-off-by: NNeilBrown <neilb@suse.de>
Reviewed-by: NAndre Noll <maan@systemlinux.org>

qd_idx is previously declared and given exactly the same value!
Signed-off-by: NNeilBrown <neilb@suse.de>

Normally is it not safe to allow a raid5 that is both dirty and
degraded to be assembled without explicit request from that admin, as
it can cause hidden data corruption.
This is because 'dirty' means that the parity cannot be trusted, and
'degraded' means that the parity needs to be used.

However, if the device that is missing contains only parity, then
there is no issue and assembly can continue.
This particularly applies when a RAID5 is being converted to a RAID6
and there is an unclean shutdown while the conversion is happening.

So check for whether the degraded space only contains parity, and
in that case, allow the assembly.
Signed-off-by: NNeilBrown <neilb@suse.de>

When a reshape finds that it can add spare devices into the array,
those devices might already be 'in_sync' if they are beyond the old
size of the array, or they might not if they are within the array.

The first case happens when we change an N-drive RAID5 to an
N+1-drive RAID5.
The second happens when we convert an N-drive RAID5 to an
N+1-drive RAID6.

So set the flag more carefully.
Also, ->recovery_offset is only meaningful when the flag is clear,
so only set it in that case.

This change needs the preceding two to ensure that the non-in_sync
device doesn't get evicted from the array when it is stopped, in the
case where v0.90 metadata is used.
Signed-off-by: NNeilBrown <neilb@suse.de>

This value is visible through sysfs and is used by mdadm
when it manages a reshape (backing up data that is about to be
rearranged).  So it is important that it is always correct.
Current it does not get updated properly when a reshape
starts which can cause problems when assembling an array
that is in the middle of being reshaped.

This is suitable for 2.6.31.y stable kernels.

Cc: stable@kernel.org
Signed-off-by: NNeilBrown <neilb@suse.de>

Signed-off-by: NDan Williams <dan.j.williams@intel.com>

md/raid6 passes a list of 'struct page *' to the async_tx routines,
which then either DMA map them for offload, or take the page_address
for CPU based calculations.

For RAID6 we sometime leave 'blanks' in the list of pages.
For CPU based calcs, we want to treat theses as a page of zeros.
For offloaded calculations, we simply don't pass a page to the
hardware.

Currently the 'blanks' are encoded as a pointer to
raid6_empty_zero_page.  This is a 4096 byte memory region, not a
'struct page'.  This is mostly handled correctly but is rather ugly.

So change the code to pass and expect a NULL pointer for the blanks.
When taking page_address of a page, we need to check for a NULL and
in that case use raid6_empty_zero_page.
Signed-off-by: NNeilBrown <neilb@suse.de>

When a raid5 (or raid6) array is being reshaped to have fewer devices,
conf->raid_disks is the latter and hence smaller number of devices.
However sometimes we want to use a number which is the total number of
currently required devices - the larger of the 'old' and 'new' sizes.
Before we implemented reducing the number of devices, this was always
'new' i.e. ->raid_disks.
Now we need max(raid_disks, previous_raid_disks) in those places.

This particularly affects assembling an array that was shutdown while
in the middle of a reshape to fewer devices.

md.c needs a similar fix when interpreting the md metadata.
Signed-off-by: NNeilBrown <neilb@suse.de>

Signed-off-by: NNeilBrown <neilb@suse.de>

The percpu conversion allowed a straightforward handoff of stripe
processing to the async subsytem that initially showed some modest gains
(+4%).  However, this model is too simplistic and leads to stripes
bouncing between raid5d and the async thread pool for every invocation
of handle_stripe().  As reported by Holger this can fall into a
pathological situation severely impacting throughput (6x performance
loss).

By downleveling the parallelism to raid_run_ops the pathological
stripe_head bouncing is eliminated.  This version still exhibits an
average 11% throughput loss for:

	mdadm --create /dev/md0 /dev/sd[b-q] -n 16 -l 6
	echo 1024 > /sys/block/md0/md/stripe_cache_size
	dd if=/dev/zero of=/dev/md0 bs=1024k count=2048

...but the results are at least stable and can be used as a base for
further multicore experimentation.
Reported-by: NHolger Kiehl <Holger.Kiehl@dwd.de>
Signed-off-by: NDan Williams <dan.j.williams@intel.com>
Signed-off-by: NNeilBrown <neilb@suse.de>

Deallocating a raid5_conf_t structure requires taking 'device_lock'.
Ensure it is initialized before it is used, i.e. initialize the lock
before attempting any further initializations that might fail.
Signed-off-by: NDan Williams <dan.j.williams@intel.com>
Signed-off-by: NNeilBrown <neilb@suse.de>

This reverts commit df10cfbc.

This patch was based on a misunderstanding and risks introducing a busy-wait loop.
So revert it.
Acked-by: NDan Williams <dan.j.williams@intel.com>
Signed-off-by: NNeilBrown <neilb@suse.de>

This should writeback from coming when the device is temporarily
suspended.
Signed-off-by: NNeilBrown <neilb@suse.de>

The management thread for raid4,5,6 arrays are all called
mdX_raid5, independent of the actual raid level, which is wrong and
can be confusion.

So change md_register_thread to use the name from the personality
unless no alternate name (like 'resync' or 'reshape') is given.

This is simpler and more correct.

Cc: Jinzc <zhenchengjin@gmail.com>
Signed-off-by: NNeilBrown <neilb@suse.de>

Rename some variable and remove some duplicate definitions
to avoid there warnings.  None of them are actual errors.
Signed-off-by: NNeilBrown <neilb@suse.de>

Neil says:
	"It is correct as it stands, but the fact that every branch in
	 the 'if' part ends with a 'return' isn't immediately obvious,
	 so it is clearer if we are explicit about the if / then / else
	 structure."
Signed-off-by: NDan Williams <dan.j.williams@intel.com>

As pointed out by Neil it should be possible to build a driver with all
BUG_ON statements deleted.  It's bad form to have a BUG_ON with a side
effect.
Signed-off-by: NDan Williams <dan.j.williams@intel.com>

Get rid of any functions that test for these bits and make callers
use bio_rw_flagged() directly. Then it is at least directly apparent
what variable and flag they check.
Signed-off-by: NJens Axboe <jens.axboe@oracle.com>

Some engines optimize operation by reading ahead in the descriptor chain
such that descriptor2 may start execution before descriptor1 completes.
If descriptor2 depends on the result from descriptor1 then a fence is
required (on descriptor2) to disable this optimization.  The async_tx
api could implicitly identify dependencies via the 'depend_tx'
parameter, but that would constrain cases where the dependency chain
only specifies a completion order rather than a data dependency.  So,
provide an ASYNC_TX_FENCE to explicitly identify data dependencies.
Signed-off-by: NDan Williams <dan.j.williams@intel.com>

Now that the resources to handle stripe_head operations are allocated
percpu it is possible for raid5d to distribute stripe handling over
multiple cores.  This conversion also adds a call to cond_resched() in
the non-multicore case to prevent one core from getting monopolized for
raid operations.

Cc: Arjan van de Ven <arjan@linux.intel.com>
Signed-off-by: NDan Williams <dan.j.williams@intel.com>

These routines have been replaced by there asynchronous counterparts.
Signed-off-by: NYuri Tikhonov <yur@emcraft.com>
Signed-off-by: NIlya Yanok <yanok@emcraft.com>
Signed-off-by: NDan Williams <dan.j.williams@intel.com>

1/ Use STRIPE_OP_BIOFILL to offload completion of read requests to
   raid_run_ops
2/ Implement a handler for sh->reconstruct_state similar to the raid5 case
   (adds handling of Q parity)
3/ Prevent handle_parity_checks6 from running concurrently with 'compute'
   operations
4/ Hook up raid_run_ops
Signed-off-by: NYuri Tikhonov <yur@emcraft.com>
Signed-off-by: NIlya Yanok <yanok@emcraft.com>
Signed-off-by: NDan Williams <dan.j.williams@intel.com>

[ Based on an original patch by Yuri Tikhonov ]

Implement the state machine for handling the RAID-6 parities check and
repair functionality.  Note that the raid6 case does not need to check
for new failures, like raid5, as it will always writeback the correct
disks.  The raid5 case can be updated to check zero_sum_result to avoid
getting confused by new failures rather than retrying the entire check
operation.
Signed-off-by: NYuri Tikhonov <yur@emcraft.com>
Signed-off-by: NIlya Yanok <yanok@emcraft.com>
Signed-off-by: NDan Williams <dan.j.williams@intel.com>

In the synchronous implementation of stripe dirtying we processed a
degraded stripe with one call to handle_stripe_dirtying6().  I.e.
compute the missing blocks from the other drives, then copy in the new
data and reconstruct the parities.

In the asynchronous case we do not perform stripe operations directly.
Instead, operations are scheduled with flags to be later serviced by
raid_run_ops.  So, for the degraded case the final reconstruction step
can only be carried out after all blocks have been brought up to date by
being read, or computed.  Like the raid5 case schedule_reconstruction()
sets STRIPE_OP_RECONSTRUCT to request a parity generation pass and
through operation chaining can handle compute and reconstruct in a
single raid_run_ops pass.

[dan.j.williams@intel.com: fixup handle_stripe_dirtying6 gating]
Signed-off-by: NYuri Tikhonov <yur@emcraft.com>
Signed-off-by: NIlya Yanok <yanok@emcraft.com>
Signed-off-by: NDan Williams <dan.j.williams@intel.com>

Modify handle_stripe_fill6 to work asynchronously by introducing
fetch_block6 as the raid6 analog of fetch_block5 (schedule compute
operations for missing/out-of-sync disks).

[dan.j.williams@intel.com: compute D+Q in one pass]
Signed-off-by: NYuri Tikhonov <yur@emcraft.com>
Signed-off-by: NIlya Yanok <yanok@emcraft.com>
Signed-off-by: NDan Williams <dan.j.williams@intel.com>

Extend schedule_reconstruction5 for reuse by the raid6 path.  Add
support for generating Q and BUG() if a request is made to perform
'prexor'.
Signed-off-by: NYuri Tikhonov <yur@emcraft.com>
Signed-off-by: NIlya Yanok <yanok@emcraft.com>
Signed-off-by: NDan Williams <dan.j.williams@intel.com>

[ Based on an original patch by Yuri Tikhonov ]

The raid_run_ops routine uses the asynchronous offload api and
the stripe_operations member of a stripe_head to carry out xor+pq+copy
operations asynchronously, outside the lock.

The operations performed by RAID-6 are the same as in the RAID-5 case
except for no support of STRIPE_OP_PREXOR operations. All the others
are supported:
STRIPE_OP_BIOFILL
 - copy data into request buffers to satisfy a read request
STRIPE_OP_COMPUTE_BLK
 - generate missing blocks (1 or 2) in the cache from the other blocks
STRIPE_OP_BIODRAIN
 - copy data out of request buffers to satisfy a write request
STRIPE_OP_RECONSTRUCT
 - recalculate parity for new data that has entered the cache
STRIPE_OP_CHECK
 - verify that the parity is correct

The flow is the same as in the RAID-5 case, and reuses some routines, namely:
1/ ops_complete_postxor (renamed to ops_complete_reconstruct)
2/ ops_complete_compute (updated to set up to 2 targets uptodate)
3/ ops_run_check (renamed to ops_run_check_p for xor parity checks)

[neilb@suse.de: fixes to get it to pass mdadm regression suite]
Reviewed-by: NAndre Noll <maan@systemlinux.org>
Signed-off-by: NYuri Tikhonov <yur@emcraft.com>
Signed-off-by: NIlya Yanok <yanok@emcraft.com>
Signed-off-by: NDan Williams <dan.j.williams@intel.com>

ops_complete_compute5 can be reused in the raid6 path if it is updated to
generically handle a second target.
Signed-off-by: NDan Williams <dan.j.williams@intel.com>

Replace the flat zero_sum_result with a collection of flags to contain
the P (xor) zero-sum result, and the soon to be utilized Q (raid6 reed
solomon syndrome) zero-sum result.  Use the SUM_CHECK_ namespace instead
of DMA_ since these flags will be used on non-dma-zero-sum enabled
platforms.
Reviewed-by: NAndre Noll <maan@systemlinux.org>
Acked-by: NMaciej Sosnowski <maciej.sosnowski@intel.com>
Signed-off-by: NDan Williams <dan.j.williams@intel.com>

Use percpu memory rather than stack for storing the buffer lists used in
parity calculations.  Include space for dma address conversions and pass
that to async_tx via the async_submit_ctl.scribble pointer.

[ Impact: move memory pressure from stack to heap ]
Signed-off-by: NDan Williams <dan.j.williams@intel.com>

In preparation for asynchronous handling of raid6 operations move the
spare page to a percpu allocation to allow multiple simultaneous
synchronous raid6 recovery operations.

Make this allocation cpu hotplug aware to maximize allocation
efficiency.
Signed-off-by: NDan Williams <dan.j.williams@intel.com>

We were removing the drives, from the array, but not
removing symlinks from /sys/.... and not marking the device
as having been removed.
Signed-off-by: NNeilBrown <neilb@suse.de>

This "if" don't allow for the possibility that the number of devices
doesn't change, and so sector_nr isn't set correctly in that case.
So change '>' to '>='.
Signed-off-by: NNeilBrown <neilb@suse.de>

md/raid5 doesn't allow a reshape to restart if it involves writing
over the same part of disk that it would be reading from.
This happens at the beginning of a reshape that increases the number
of devices, at the end of a reshape that decreases the number of
devices, and continuously for a reshape that does not change the
number of devices.

The current code is correct for the "increase number of devices"
case as the critical section at the start is handled by userspace
performing a backup.

It does not work for reducing the number of devices, or the
no-change case.
For 'reducing', we need to invert the test.  For no-change we cannot
really be sure things will be safe, so simply require the array
to be read-only, which is how the user-space code which carefully
starts such arrays works.
Signed-off-by: NNeilBrown <neilb@suse.de>

As revalidate_disk calls check_disk_size_change, it will cause
any capacity change of a gendisk to be propagated to the blockdev
inode.  So use that instead of mucking about with locks and
i_size_write.

Also add a call to revalidate_disk in do_md_run and a few other places
where the gendisk capacity is changed.
Signed-off-by: NNeilBrown <neilb@suse.de>

The ->quiesce method is not supposed to stop resync/recovery/reshape,
just normal IO.
But in raid5 we don't have a way to know which stripes are being
used for normal IO and which for resync etc, so we need to wait for
all stripes to be idle to be sure that all writes have completed.

However reshape keeps at least some stripe busy for an extended period
of time, so a call to raid5_quiesce can block for several seconds
needlessly.
So arrange for reshape etc to pause briefly while raid5_quiesce is
trying to quiesce the array so that the active_stripes count can
drop to zero.
Signed-off-by: NNeilBrown <neilb@suse.de>

As the internal reshape_progress counter is the main driver
for reshape, the fact that reshape_position sometimes starts with the
wrong value has minimal effect.  It is visible in sysfs and that
is all.
Signed-off-by: NNeilBrown <neilb@suse.de>