When pg_init is running no I/O can be submitted to the underlying
devices, as the path priority etc might change.  When using queue_io for
this, requests will be piling up within multipath as the block I/O
scheduler just sees a _very fast_ device.  All of this queued I/O has to
be resubmitted from within multipathing once pg_init is done.

This approach has the problem that it's virtually impossible to
abort I/O when pg_init is running, and we're adding heavy load
to the devices after pg_init since all of the queued I/O needs to be
resubmitted _before_ any requests can be pulled off of the request queue
and normal operation continues.

This patch will requeue the I/O that triggers the pg_init call, and
return 'busy' when pg_init is in progress.  With these changes the block
I/O scheduler will stop submitting I/O during pg_init, resulting in a
quicker path switch and less I/O pressure (and memory consumption) after
pg_init.
Signed-off-by: NHannes Reinecke <hare@suse.de>
[patch header edited for clarity and typos by Mike Snitzer]
Signed-off-by: NMike Snitzer <snitzer@redhat.com>

Whenever multipath_dtr() is happening we must prevent queueing any
further path activation work.  Implement this by adding a new
'pg_init_disabled' flag to the multipath structure that denotes future
path activation work should be skipped if it is set.  By disabling
pg_init and then re-enabling in flush_multipath_work() we also avoid the
potential for pg_init to be initiated while suspending an mpath device.

Without this patch a race condition exists that may result in a kernel
panic:

1) If after pg_init_done() decrements pg_init_in_progress to 0, a call
   to wait_for_pg_init_completion() assumes there are no more pending path
   management commands.
2) If pg_init_required is set by pg_init_done(), due to retryable
   mode_select errors, then process_queued_ios() will again queue the
   path activation work.
3) If free_multipath() completes before activate_path() work is called a
   NULL pointer dereference like the following can be seen when
   accessing members of the recently destructed multipath:

BUG: unable to handle kernel NULL pointer dereference at 0000000000000090
RIP: 0010:[<ffffffffa003db1b>]  [<ffffffffa003db1b>] activate_path+0x1b/0x30 [dm_multipath]
[<ffffffff81090ac0>] worker_thread+0x170/0x2a0
[<ffffffff81096c80>] ? autoremove_wake_function+0x0/0x40

[switch to disabling pg_init in flush_multipath_work & header edits by Mike Snitzer]
Signed-off-by: NShiva Krishna Merla <shivakrishna.merla@netapp.com>
Reviewed-by: NKrishnasamy Somasundaram <somasundaram.krishnasamy@netapp.com>
Tested-by: NSpeagle Andy <Andy.Speagle@netapp.com>
Acked-by: NJunichi Nomura <j-nomura@ce.jp.nec.com>
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Cc: stable@vger.kernel.org

dm-mpath and dm-thin must process messages even if some device is
suspended, so we allocate argv buffer with GFP_NOIO. These messages have
a small fixed number of arguments.

On the other hand, dm-switch needs to process bulk data using messages
so excessive use of GFP_NOIO could cause trouble.

The patch also lowers the default number of arguments from 64 to 8, so
that there is smaller load on GFP_NOIO allocations.
Signed-off-by: NMikulas Patocka <mpatocka@redhat.com>
Cc: stable@vger.kernel.org
Acked-by: NAlasdair G Kergon <agk@redhat.com>
Signed-off-by: NMike Snitzer <snitzer@redhat.com>

Commit c0f04d88 ("bcache: Fix flushes in writeback mode") was fixing
a reported data corruption bug, but it seems some last minute
refactoring or rebasing introduced a null pointer deref.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>
Cc: linux-stable <stable@vger.kernel.org> # >= v3.10
Reported-by: NGabriel de Perthuis <g2p.code@gmail.com>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

In writeback mode, when we get a cache flush we need to make sure we
issue a flush to the backing device.

The code for sending down an extra flush was wrong - by cloning the bio
we were probably getting flags that didn't make sense for a bare flush,
and also the old code was firing for FUA bios, for which we don't need
to send a flush to the backing device.

This was causing data corruption somehow - the mechanism was never
determined, but this patch fixes it for the users that were seeing it.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>
Cc: linux-stable <stable@vger.kernel.org> # >= v3.10
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

btree_sort_fixup() was overly clever, because it was trying to avoid
pulling a key off the btree iterator in more than one place.

This led to a really obscure bug where we'd break early from the loop in
btree_sort_fixup() if the current key overlapped with keys in more than
one older set, and the next key it overlapped with was zero size.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>
Cc: linux-stable <stable@vger.kernel.org> # >= v3.10
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

GFP_NOIO means we could be getting called recursively - mca_alloc() ->
mca_data_alloc() - definitely can't use mutex_lock(bucket_lock) then.
Whoops.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>
Cc: linux-stable <stable@vger.kernel.org> # >= v3.10
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

schedule_timeout() != schedule_timeout_uninterruptible()
Signed-off-by: NKent Overstreet <kmo@daterainc.com>
Cc: linux-stable <stable@vger.kernel.org> # >= v3.10
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

bch_journal_meta() was missing the flush to make the journal write
actually go down (instead of waiting up to journal_delay_ms)...

Whoops
Signed-off-by: NKent Overstreet <kmo@daterainc.com>
Cc: linux-stable <stable@vger.kernel.org> # >= v3.10
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Background writeback works by scanning the btree for dirty data and
adding those keys into a fixed size buffer, then for each dirty key in
the keybuf writing it to the backing device.

When read_dirty() finishes and it's time to scan for more dirty data, we
need to wait for the outstanding writeback IO to finish - they still
take up slots in the keybuf (so that foreground writes can check for
them to avoid races) - without that wait, we'll continually rescan when
we'll be able to add at most a key or two to the keybuf, and that takes
locks that starves foreground IO.  Doh.
Signed-off-by: NKent Overstreet <kmo@daterainc.com>
Cc: linux-stable <stable@vger.kernel.org> # >= v3.10
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Fix

  drivers/md/bcache/btree.c: In function ‘bch_btree_node_read’:
  drivers/md/bcache/btree.c:259: warning: format ‘%lu’ expects type ‘long unsigned int’, but argument 3 has type ‘size_t’
Signed-off-by: NGeert Uytterhoeven <geert@linux-m68k.org>
Signed-off-by: NKent Overstreet <kmo@daterainc.com>
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

The journal replay code didn't handle this case, causing it to go into
an infinite loop...
Signed-off-by: NKent Overstreet <kmo@daterainc.com>
Cc: linux-stable <stable@vger.kernel.org> # >= v3.10
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

sysfs attributes with unusual characters have crappy failure modes
in Squeeze (udev 164); later versions of udev are unaffected.

This should make these characters more unusual.
Signed-off-by: NGabriel de Perthuis <g2p.code@gmail.com>
Signed-off-by: NKent Overstreet <kmo@daterainc.com>
Cc: linux-stable <stable@vger.kernel.org> # >= v3.10
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

That switch statement was obviously wrong, leading to some sort of weird
spinning on rare occasion with discards enabled...
Signed-off-by: NKent Overstreet <kmo@daterainc.com>
Cc: linux-stable <stable@vger.kernel.org> # >= v3.10
Signed-off-by: NLinus Torvalds <torvalds@linux-foundation.org>

Allow user to change the number of IOs that are reserved by
bio-based DM's mempools by writing to this file:
/sys/module/dm_mod/parameters/reserved_bio_based_ios

The default value is RESERVED_BIO_BASED_IOS (16).  The maximum allowed
value is RESERVED_MAX_IOS (1024).

Export dm_get_reserved_bio_based_ios() for use by DM targets and core
code.  Switch to sizing dm-io's mempool and bioset using DM core's
configurable 'reserved_bio_based_ios'.
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Signed-off-by: NFrank Mayhar <fmayhar@google.com>

Allow user to change the number of IOs that are reserved by
request-based DM's mempools by writing to this file:
/sys/module/dm_mod/parameters/reserved_rq_based_ios

The default value is RESERVED_REQUEST_BASED_IOS (256).  The maximum
allowed value is RESERVED_MAX_IOS (1024).

Export dm_get_reserved_rq_based_ios() for use by DM targets and core
code.  Switch to sizing dm-mpath's mempool using DM core's configurable
'reserved_rq_based_ios'.
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Signed-off-by: NFrank Mayhar <fmayhar@google.com>
Acked-by: NMikulas Patocka <mpatocka@redhat.com>

Bio-based device mapper processing doesn't need larger mempools (like
request-based DM does), so lower the number of reserved entries for
bio-based operation.  16 was already used for bio-based DM's bioset
but mistakenly wasn't used for it's _io_cache.

Formalize difference between bio-based and request-based defaults by
introducing RESERVED_BIO_BASED_IOS and RESERVED_REQUEST_BASED_IOS.

(based on older code from Mikulas Patocka)
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Signed-off-by: NFrank Mayhar <fmayhar@google.com>
Acked-by: NMikulas Patocka <mpatocka@redhat.com>

Fix issue where the block layer would stack the discard limits of the
pool's data device even if the "ignore_discard" pool feature was
specified.

The pool and thin device(s) still had discards disabled because the
QUEUE_FLAG_DISCARD request_queue flag wasn't set.  But to avoid user
confusion when "ignore_discard" is used: both the pool device and the
thin device(s) have zeroes for all discard limits.

Also, always set discard_zeroes_data_unsupported in targets because they
should never advertise the 'discard_zeroes_data' capability (even if the
pool's data device supports it).
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Acked-by: NJoe Thornber <ejt@redhat.com>

Workaround the SCSI layer's problematic WRITE SAME heuristics by
disabling WRITE SAME in the DM multipath device's queue_limits if an
underlying device disabled it.

The WRITE SAME heuristics, with both the original commit 5db44863
("[SCSI] sd: Implement support for WRITE SAME") and the updated commit
66c28f97 ("[SCSI] sd: Update WRITE SAME heuristics"), default to enabling
WRITE SAME(10) even without successfully determining it is supported.
After the first failed WRITE SAME the SCSI layer will disable WRITE SAME
for the device (by setting sdkp->device->no_write_same which results in
'max_write_same_sectors' in device's queue_limits to be set to 0).

When a device is stacked ontop of such a SCSI device any changes to that
SCSI device's queue_limits do not automatically propagate up the stack.
As such, a DM multipath device will not have its WRITE SAME support
disabled.  This causes the block layer to continue to issue WRITE SAME
requests to the mpath device which causes paths to fail and (if mpath IO
isn't configured to queue when no paths are available) it will result in
actual IO errors to the upper layers.

This fix doesn't help configurations that have additional devices
stacked ontop of the mpath device (e.g. LVM created linear DM devices
ontop).  A proper fix that restacks all the queue_limits from the bottom
of the device stack up will need to be explored if SCSI will continue to
use this model of optimistically allowing op codes and then disabling
them after they fail for the first time.

Before this patch:

EXT4-fs (dm-6): mounted filesystem with ordered data mode. Opts: (null)
device-mapper: multipath: XXX snitm debugging: got -EREMOTEIO (-121)
device-mapper: multipath: XXX snitm debugging: failing WRITE SAME IO with error=-121
end_request: critical target error, dev dm-6, sector 528
dm-6: WRITE SAME failed. Manually zeroing.
device-mapper: multipath: Failing path 8:112.
end_request: I/O error, dev dm-6, sector 4616
dm-6: WRITE SAME failed. Manually zeroing.
end_request: I/O error, dev dm-6, sector 4616
end_request: I/O error, dev dm-6, sector 5640
end_request: I/O error, dev dm-6, sector 6664
end_request: I/O error, dev dm-6, sector 7688
end_request: I/O error, dev dm-6, sector 524288
Buffer I/O error on device dm-6, logical block 65536
lost page write due to I/O error on dm-6
JBD2: Error -5 detected when updating journal superblock for dm-6-8.
end_request: I/O error, dev dm-6, sector 524296
Aborting journal on device dm-6-8.
end_request: I/O error, dev dm-6, sector 524288
Buffer I/O error on device dm-6, logical block 65536
lost page write due to I/O error on dm-6
JBD2: Error -5 detected when updating journal superblock for dm-6-8.

# cat /sys/block/sdh/queue/write_same_max_bytes
0
# cat /sys/block/dm-6/queue/write_same_max_bytes
33553920

After this patch:

EXT4-fs (dm-6): mounted filesystem with ordered data mode. Opts: (null)
device-mapper: multipath: XXX snitm debugging: got -EREMOTEIO (-121)
device-mapper: multipath: XXX snitm debugging: WRITE SAME I/O failed with error=-121
end_request: critical target error, dev dm-6, sector 528
dm-6: WRITE SAME failed. Manually zeroing.

# cat /sys/block/sdh/queue/write_same_max_bytes
0
# cat /sys/block/dm-6/queue/write_same_max_bytes
0

It should be noted that WRITE SAME support wasn't enabled in DM
multipath until v3.10.
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Cc: Martin K. Petersen <martin.petersen@oracle.com>
Cc: Hannes Reinecke <hare@suse.de>
Cc: stable@vger.kernel.org # 3.10+

LVM2, since version 2.02.96, creates origin with zero size, then loads
the snapshot driver and then loads the origin.  Consequently, the
snapshot driver sees the origin size zero and sets the hash size to the
lower bound 64.  Such small hash table causes performance degradation.

This patch changes it so that the hash size is determined by the size of
snapshot volume, not minimum of origin and snapshot size.  It doesn't
make sense to set the snapshot size significantly larger than the origin
size, so we do not need to take origin size into account when
calculating the hash size.
Signed-off-by: NMikulas Patocka <mpatocka@redhat.com>
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Cc: stable@vger.kernel.org

The kernel reports a lockdep warning if a snapshot is invalidated because
it runs out of space.

The lockdep warning was triggered by commit 0976dfc1
("workqueue: Catch more locking problems with flush_work()") in v3.5.

The warning is false positive.  The real cause for the warning is that
the lockdep engine treats different instances of md->lock as a single
lock.

This patch is a workaround - we use flush_workqueue instead of flush_work.
This code path is not performance sensitive (it is called only on
initialization or invalidation), thus it doesn't matter that we flush the
whole workqueue.

The real fix for the problem would be to teach the lockdep engine to treat
different instances of md->lock as separate locks.
Signed-off-by: NMikulas Patocka <mpatocka@redhat.com>
Acked-by: NAlasdair G Kergon <agk@redhat.com>
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Cc: stable@vger.kernel.org # 3.5+

There was a deliberate race condition in dm_stat_for_entry() to avoid the
overhead of disabling and enabling interrupts.  The race could result in
some events not being counted on 64-bit architectures.

However, on 32-bit architectures, operations on long long variables are
not atomic, so the race condition could cause the counter to jump by 2^32.
Such jumps could be disruptive, so we need to do proper locking on 32-bit
architectures.
Signed-off-by: NMikulas Patocka <mpatocka@redhat.com>
Cc: Alasdair G. Kergon <agk@redhat.com>
Signed-off-by: NMike Snitzer <snitzer@redhat.com>

Since ENOSPC is a target-side error, dm-mpath should just pass the error
information to upper layer instead of retrying itself with path failover.
Otherwise it will end up failing all paths down while path checkers find
all paths ok.

ENOSPC can now be returned from SCSI device after commit a9d6ceb8
("[SCSI] return ENOSPC on thin provisioning failure").
Signed-off-by: NJun'ichi Nomura <j-nomura@ce.jp.nec.com>
Acked-by: NHannes Reinecke <hare@suse.de>
Signed-off-by: NMike Snitzer <snitzer@redhat.com>

Convert the driver shrinkers to the new API.  Most changes are compile
tested only because I either don't have the hardware or it's staging
stuff.

FWIW, the md and android code is pretty good, but the rest of it makes me
want to claw my eyes out.  The amount of broken code I just encountered is
mind boggling.  I've added comments explaining what is broken, but I fear
that some of the code would be best dealt with by being dragged behind the
bike shed, burying in mud up to it's neck and then run over repeatedly
with a blunt lawn mower.

Special mention goes to the zcache/zcache2 drivers.  They can't co-exist
in the build at the same time, they are under different menu options in
menuconfig, they only show up when you've got the right set of mm
subsystem options configured and so even compile testing is an exercise in
pulling teeth.  And that doesn't even take into account the horrible,
broken code...

[glommer@openvz.org: fixes for i915, android lowmem, zcache, bcache]
Signed-off-by: NDave Chinner <dchinner@redhat.com>
Signed-off-by: NGlauber Costa <glommer@openvz.org>
Acked-by: NMel Gorman <mgorman@suse.de>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: David Rientjes <rientjes@google.com>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Cc: Adrian Hunter <adrian.hunter@intel.com>
Cc: Al Viro <viro@zeniv.linux.org.uk>
Cc: Artem Bityutskiy <artem.bityutskiy@linux.intel.com>
Cc: Arve Hjønnevåg <arve@android.com>
Cc: Carlos Maiolino <cmaiolino@redhat.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Chuck Lever <chuck.lever@oracle.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: David Rientjes <rientjes@google.com>
Cc: Gleb Natapov <gleb@redhat.com>
Cc: Greg Thelen <gthelen@google.com>
Cc: J. Bruce Fields <bfields@redhat.com>
Cc: Jan Kara <jack@suse.cz>
Cc: Jerome Glisse <jglisse@redhat.com>
Cc: John Stultz <john.stultz@linaro.org>
Cc: KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>
Cc: Kent Overstreet <koverstreet@google.com>
Cc: Kirill A. Shutemov <kirill.shutemov@linux.intel.com>
Cc: Marcelo Tosatti <mtosatti@redhat.com>
Cc: Mel Gorman <mgorman@suse.de>
Cc: Steven Whitehouse <swhiteho@redhat.com>
Cc: Thomas Hellstrom <thellstrom@vmware.com>
Cc: Trond Myklebust <Trond.Myklebust@netapp.com>
Signed-off-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NAl Viro <viro@zeniv.linux.org.uk>

Eliminate the following sparse warnings:
drivers/md/dm-stripe.c:443:12: warning: symbol 'dm_stripe_init' was not declared. Should it be static?
drivers/md/dm-stripe.c:456:6: warning: symbol 'dm_stripe_exit' was not declared. Should it be static?
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Signed-off-by: NAlasdair G Kergon <agk@redhat.com>

Support the collection of I/O statistics on user-defined regions of
a DM device.  If no regions are defined no statistics are collected so
there isn't any performance impact.  Only bio-based DM devices are
currently supported.

Each user-defined region specifies a starting sector, length and step.
Individual statistics will be collected for each step-sized area within
the range specified.

The I/O statistics counters for each step-sized area of a region are
in the same format as /sys/block/*/stat or /proc/diskstats but extra
counters (12 and 13) are provided: total time spent reading and
writing in milliseconds.  All these counters may be accessed by sending
the @stats_print message to the appropriate DM device via dmsetup.

The creation of DM statistics will allocate memory via kmalloc or
fallback to using vmalloc space.  At most, 1/4 of the overall system
memory may be allocated by DM statistics.  The admin can see how much
memory is used by reading
/sys/module/dm_mod/parameters/stats_current_allocated_bytes

See Documentation/device-mapper/statistics.txt for more details.
Signed-off-by: NMikulas Patocka <mpatocka@redhat.com>
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Signed-off-by: NAlasdair G Kergon <agk@redhat.com>

If pool has 'no_free_space' set it means a previous allocation already
determined the pool has no free space (and failed that allocation with
-ENOSPC).  By always returning -ENOSPC if 'no_free_space' is set, we do
not allow the pool to oscillate between allocating blocks and then not.

But a side-effect of this determinism is that if a user wants to be able
to allocate new blocks they'll need to reload the pool's table (to clear
the 'no_free_space' flag).  This reload will happen automatically if the
pool's data volume is resized.  But if the user takes action to free a
lot of space by deleting snapshot volumes, etc the pool will no longer
allow data allocations to continue without an intervening table reload.
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Acked-by: NJoe Thornber <ejt@redhat.com>
Signed-off-by: NAlasdair G Kergon <agk@redhat.com>

Make use of common cleanup code.
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Signed-off-by: NAlasdair G Kergon <agk@redhat.com>

Hold the mapped device's type_lock before calling populate_table() since
it is where the table's type is determined based on the specified
targets.  There is no need to allow concurrent table loads to race to
establish the table's targets or type.

This eliminates the need to grab the lock in dm_table_set_type().

Also verify that the type_lock is held in both dm_set_md_type() and
dm_get_md_type().
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Signed-off-by: NAlasdair G Kergon <agk@redhat.com>

A device-mapper device must always have a name consisting of a non-empty
string.  If the device also has a uuid, this similarly must not be an
empty string.

The DM_DEV_CREATE ioctl enforces these rules when the device is created,
but this patch is needed to enforce them when DM_DEV_RENAME is used to
change the name or uuid.
Reported-by: NZdenek Kabelac <zkabelac@redhat.com>
Signed-off-by: NAlasdair G Kergon <agk@redhat.com>
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Acked-by: NMikulas Patocka <mpatocka@redhat.com>

break_sharing() now handles an arbitrary alloc_data_block() error
the same way as provision_block(): marks pool read-only and errors the
cell.
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Acked-by: NJoe Thornber <ejt@redhat.com>
Signed-off-by: NAlasdair G Kergon <agk@redhat.com>

Useful to know which pool is experiencing the error.
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Acked-by: NJoe Thornber <ejt@redhat.com>
Signed-off-by: NAlasdair G Kergon <agk@redhat.com>

It may be useful to switch a request-based table to the "error" target.
Enhance the DM core to allow a hybrid target_type which is capable of
handling either bios (via .map) or requests (via .map_rq).

Add a request-based map function (.map_rq) to the "error" target_type;
making it DM's first hybrid target.  Train dm_table_set_type() to prefer
the mapped device's established type (request-based or bio-based).  If
the mapped device doesn't have an established type default to making the
table with the hybrid target(s) bio-based.

Tested 'dmsetup wipe_table' to work on both bio-based and request-based
devices.
Signed-off-by: NMike Snitzer <snitzer@redhat.com>
Signed-off-by: NJoe Jin <joe.jin@oracle.com>
Signed-off-by: NJun'ichi Nomura <j-nomura@ce.jp.nec.com>
Acked-by: NMikulas Patocka <mpatocka@redhat.com>
Signed-off-by: NAlasdair G Kergon <agk@redhat.com>

If there are not enough stripes to handle, we'd better not always
queue all available work_structs. If one worker can only handle small
or even none stripes, it will impact request merge and create lock
contention.

With this patch, the number of work_struct running will depend on
pending stripes number. Note: some statistics info used in the patch
are accessed without locking protection. This should doesn't matter,
we just try best to avoid queue unnecessary work_struct.
Signed-off-by: NShaohua Li <shli@fusionio.com>
Signed-off-by: NNeilBrown <neilb@suse.de>

Some requests - particularly 'discard' and 'read' are handled
differently depending on whether a reshape is active or not.

It is harmless to assume reshape is active if it isn't but wrong
to act as though reshape is not active when it is.

So when we start reshape - after making clear to all requests that
reshape has started - use mddev_suspend/mddev_resume to flush out all
requests.  This will ensure that no requests will be assuming the
absence of reshape once it really starts.
Signed-off-by: NNeilBrown <neilb@suse.de>

make_request() access various shape parameters (raid_disks, chunk_size
etc) which might be changed by raid5_start_reshape().

If the later is called at and awkward time during the form, the wrong
stripe_head might be used.

So introduce a 'seqcount' and after finding a stripe_head make sure
there is no reason to expect that we got the wrong one.
Signed-off-by: NNeilBrown <neilb@suse.de>

Add a sysfs entry to control running workqueue thread number. If
group_thread_cnt is set to 0, we will disable workqueue offload handling of
stripes.
Signed-off-by: NShaohua Li <shli@fusionio.com>
Signed-off-by: NNeilBrown <neilb@suse.de>

This is another attempt to create multiple threads to handle raid5 stripes.
This time I use workqueue.

raid5 handles request (especially write) in stripe unit. A stripe is page size
aligned/long and acrosses all disks. Writing to any disk sector, raid5 runs a
state machine for the corresponding stripe, which includes reading some disks
of the stripe, calculating parity, and writing some disks of the stripe. The
state machine is running in raid5d thread currently. Since there is only one
thread, it doesn't scale well for high speed storage. An obvious solution is
multi-threading.

To get better performance, we have some requirements:
a. locality. stripe corresponding to request submitted from one cpu is better
handled in thread in local cpu or local node. local cpu is preferred but some
times could be a bottleneck, for example, parity calculation is too heavy.
local node running has wide adaptability.
b. configurablity. Different setup of raid5 array might need diffent
configuration. Especially the thread number. More threads don't always mean
better performance because of lock contentions.

My original implementation is creating some kernel threads. There are
interfaces to control which cpu's stripe each thread should handle. And
userspace can set affinity of the threads. This provides biggest flexibility
and configurability. But it's hard to use and apparently a new thread pool
implementation is disfavor.

Recent workqueue improvement is quite promising. unbound workqueue will be
bound to numa node. If WQ_SYSFS is set in workqueue, there are sysfs option to
do affinity setting. For example, we can only include one HT sibling in
affinity. Since work is non-reentrant by default, and we can control running
thread number by limiting dispatched work_struct number.

In this patch, I created several stripe worker group. A group is a numa node.
stripes from cpus of one node will be added to a group list. Workqueue thread
of one node will only handle stripes of worker group of the node. In this way,
stripe handling has numa node locality. And as I said, we can control thread
number by limiting dispatched work_struct number.

The work_struct callback function handles several stripes in one run. A typical
work queue usage is to run one unit in each work_struct. In raid5 case, the
unit is a stripe. But we can't do that:
a. Though handling a stripe doesn't need lock because of reference accounting
and stripe isn't in any list, queuing a work_struct for each stripe will make
workqueue lock contended very heavily.
b. blk_start_plug()/blk_finish_plug() should surround stripe handle, as we
might dispatch request. If each work_struct only handles one stripe, such block
plug is meaningless.

This implementation can't do very fine grained configuration. But the numa
binding is most popular usage model, should be enough for most workloads.

Note: since we have only one stripe queue, switching to multi-thread might
decrease request size dispatching down to low level layer. The impact depends
on thread number, raid configuration and workload. So multi-thread raid5 might
not be proper for all setups.

Changes V1 -> V2:
1. remove WQ_NON_REENTRANT
2. disabling multi-threading by default
3. Add more descriptions in changelog
Signed-off-by: NShaohua Li <shli@fusionio.com>
Signed-off-by: NNeilBrown <neilb@suse.de>

patch "make release_stripe lockless" changes the order stripes are released.
Originally I thought block layer can take care of request merge, but it appears
there are still some requests not merged. It's easy to fix the order.
Signed-off-by: NShaohua Li <shli@fusionio.com>
Signed-off-by: NNeilBrown <neilb@suse.de>

release_stripe still has big lock contention. We just add the stripe to a llist
without taking device_lock. We let the raid5d thread to do the real stripe
release, which must hold device_lock anyway. In this way, release_stripe
doesn't hold any locks.

The side effect is the released stripes order is changed. But sounds not a big
deal, stripes are never handled in order. And I thought block layer can already
do nice request merge, which means order isn't that important.

I kept the unplug release batch, which is unnecessary with this patch from lock
contention avoid point of view, and actually if we delete it, the stripe_head
release_list and lru can share storage. But the unplug release batch is also
helpful for request merge. We probably can delay wakeup raid5d till unplug, but
I'm still afraid of the case which raid5d is running.
Signed-off-by: NShaohua Li <shli@fusionio.com>
Signed-off-by: NNeilBrown <neilb@suse.de>