The default setting of 256 stripe_heads is probably
much too small for many configurations.  So it is best to make it
auto-configure.

Shrinking the cache under memory pressure is easy.  The only
interesting part here is that we put a fairly high cost
('seeks') on shrinking the cache as the cost is greater than
just having to read more data, it reduces parallelism.

Growing the cache on demand needs to be done carefully.  If we allow
fast growth, that can upset memory balance as lots of dirty memory can
quickly turn into lots of memory queued in the stripe_cache.
It is important for the raid5 block device to appear congested to
allow write-throttling to work.

So we only add stripes slowly. We set a flag when an allocation
fails because all stripes are in use, allocate at a convenient
time when that flag is set, and don't allow it to be set again
until at least one stripe_head has been released for re-use.

This means that a spurt of requests will only cause one stripe_head
to be allocated, but a steady stream of requests will slowly
increase the cache size - until memory pressure puts it back again.

It could take hours to reach a steady state.

The value written to, and displayed in, stripe_cache_size is
used as a minimum.  The cache can grow above this and shrink back
down to it.  The actual size is not directly visible, though it can
be deduced to some extent by watching stripe_cache_active.
Signed-off-by: NNeilBrown <neilb@suse.de>

This allows us to easily add more (atomic) flags.
Signed-off-by: NNeilBrown <neilb@suse.de>

Rather than adjusting max_nr_stripes whenever {grow,drop}_one_stripe()
succeeds, do it inside the functions.

Also choose the correct hash to handle next inside the functions.

This removes duplication and will help with future new uses of
{grow,drop}_one_stripe.

This also fixes a minor bug where the "md/raid:%md: allocate XXkB"
message always said "0kB".
Signed-off-by: NNeilBrown <neilb@suse.de>

This is needed for future improvement to stripe cache management.
Signed-off-by: NNeilBrown <neilb@suse.de>

Depending on the available coding we allow optimized rmw logic for write
operations. To support easier testing this patch allows manual control
of the rmw/rcw descision through the interface /sys/block/mdX/md/rmw_level.

The configuration can handle three levels of control.

rmw_level=0: Disable rmw for all RAID types. Hardware assisted P/Q
calculation has no implementation path yet to factor in/out chunks of
a syndrome. Enforcing this level can be benefical for slow CPUs with
hardware syndrome support and fast SSDs.

rmw_level=1: Estimate rmw IOs and rcw IOs. Execute rmw only if we will
save IOs. This equals the "old" unpatched behaviour and will be the
default.

rmw_level=2: Execute rmw even if calculated IOs for rmw and rcw are
equal. We might have higher CPU consumption because of calculating the
parity twice but it can be benefical otherwise. E.g. RAID4 with fast
dedicated parity disk/SSD. The option is implemented just to be
forward-looking and will ONLY work with this patch!
Signed-off-by: NMarkus Stockhausen <stockhausen@collogia.de>
Signed-off-by: NNeilBrown <neilb@suse.de>

Glue it altogehter. The raid6 rmw path should work the same as the
already existing raid5 logic. So emulate the prexor handling/flags
and split functions as needed.

1) Enable xor_syndrome() in the async layer.

2) Split ops_run_prexor() into RAID4/5 and RAID6 logic. Xor the syndrome
at the start of a rmw run as we did it before for the single parity.

3) Take care of rmw run in ops_run_reconstruct6(). Again process only
the changed pages to get syndrome back into sync.

4) Enhance set_syndrome_sources() to fill NULL pages if we are in a rmw
run. The lower layers will calculate start & end pages from that and
call the xor_syndrome() correspondingly.

5) Adapt the several places where we ignored Q handling up to now.

Performance numbers for a single E5630 system with a mix of 10 7200k
desktop/server disks. 300 seconds random write with 8 threads onto a
3,2TB (10*400GB) RAID6 64K chunk without spare (group_thread_cnt=4)

bsize   rmw_level=1   rmw_level=0   rmw_level=1   rmw_level=0
        skip_copy=1   skip_copy=1   skip_copy=0   skip_copy=0
   4K      115 KB/s      141 KB/s      165 KB/s      140 KB/s
   8K      225 KB/s      275 KB/s      324 KB/s      274 KB/s
  16K      434 KB/s      536 KB/s      640 KB/s      534 KB/s
  32K      751 KB/s    1,051 KB/s    1,234 KB/s    1,045 KB/s
  64K    1,339 KB/s    1,958 KB/s    2,282 KB/s    1,962 KB/s
 128K    2,673 KB/s    3,862 KB/s    4,113 KB/s    3,898 KB/s
 256K    7,685 KB/s    7,539 KB/s    7,557 KB/s    7,638 KB/s
 512K   19,556 KB/s   19,558 KB/s   19,652 KB/s   19,688 Kb/s
Signed-off-by: NMarkus Stockhausen <stockhausen@collogia.de>
Signed-off-by: NNeilBrown <neilb@suse.de>

The second and (last) optimized XOR syndrome calculation. This version
supports right and left side optimization. All CPUs with architecture
older than Haswell will benefit from it.

It should be noted that SSE2 movntdq kills performance for memory areas
that are read and written simultaneously in chunks smaller than cache
line size. So use movdqa instead for P/Q writes in sse21 and sse22 XOR
functions.
Signed-off-by: NMarkus Stockhausen <stockhausen@collogia.de>
Signed-off-by: NNeilBrown <neilb@suse.de>

Start the algorithms with the very basic one. It is left and right
optimized. That means we can avoid all calculations for unneeded pages
above the right stop offset. For pages below the left start offset we
still need the syndrome multiplication but without reading data pages.
Signed-off-by: NMarkus Stockhausen <stockhausen@collogia.de>
Signed-off-by: NNeilBrown <neilb@suse.de>

It is always helpful to have a test tool in place if we implement
new data critical algorithms. So add some test routines to the raid6
checker that can prove if the new xor_syndrome() works as expected.

Run through all permutations of start/stop pages per algorithm and
simulate a xor_syndrome() assisted rmw run. After each rmw check if
the recovery algorithm still confirms that the stripe is fine.
Signed-off-by: NMarkus Stockhausen <stockhausen@collogia.de>
Signed-off-by: NNeilBrown <neilb@suse.de>

v3: s-o-b comment, explanation of performance and descision for
the start/stop implementation

Implementing rmw functionality for RAID6 requires optimized syndrome
calculation. Up to now we can only generate a complete syndrome. The
target P/Q pages are always overwritten. With this patch we provide
a framework for inplace P/Q modification. In the first place simply
fill those functions with NULL values.

xor_syndrome() has two additional parameters: start & stop. These
will indicate the first and last page that are changing during a
rmw run. That makes it possible to avoid several unneccessary loops
and speed up calculation. The caller needs to implement the following
logic to make the functions work.

1) xor_syndrome(disks, start, stop, ...): "Remove" all data of source
blocks inside P/Q between (and including) start and end.

2) modify any block with start <= block <= stop

3) xor_syndrome(disks, start, stop, ...): "Reinsert" all data of
source blocks into P/Q between (and including) start and end.

Pages between start and stop that won't be changed should be filled
with a pointer to the kernel zero page. The reasons for not taking NULL
pages are:

1) Algorithms cross the whole source data line by line. Thus avoid
additional branches.

2) Having a NULL page avoids calculating the XOR P parity but still
need calulation steps for the Q parity. Depending on the algorithm
unrolling that might be only a difference of 2 instructions per loop.

The benchmark numbers of the gen_syndrome() functions are displayed in
the kernel log. Do the same for the xor_syndrome() functions. This
will help to analyze performance problems and give an rough estimate
how well the algorithm works. The choice of the fastest algorithm will
still depend on the gen_syndrome() performance.

With the start/stop page implementation the speed can vary a lot in real
life. E.g. a change of page 0 & page 15 on a stripe will be harder to
compute than the case where page 0 & page 1 are XOR candidates. To be not
to enthusiatic about the expected speeds we will run a worse case test
that simulates a change on the upper half of the stripe. So we do:

1) calculation of P/Q for the upper pages

2) continuation of Q for the lower (empty) pages
Signed-off-by: NMarkus Stockhausen <stockhausen@collogia.de>
Signed-off-by: NNeilBrown <neilb@suse.de>

expansion/resync can grab a stripe when the stripe is in batch list. Since all
stripes in batch list must be in the same state, we can't allow some stripes
run into expansion/resync. So we delay expansion/resync for stripe in batch
list.
Signed-off-by: NShaohua Li <shli@fusionio.com>
Signed-off-by: NNeilBrown <neilb@suse.de>

If io error happens in any stripe of a batch list, the batch list will be
split, then normal process will run for the stripes in the list.
Signed-off-by: NShaohua Li <shli@fusionio.com>
Signed-off-by: NNeilBrown <neilb@suse.de>

stripe cache is 4k size. Even adjacent full stripe writes are handled in 4k
unit. Idealy we should use big size for adjacent full stripe writes. Bigger
stripe cache size means less stripes runing in the state machine so can reduce
cpu overhead. And also bigger size can cause bigger IO size dispatched to under
layer disks.

With below patch, we will automatically batch adjacent full stripe write
together. Such stripes will be added to the batch list. Only the first stripe
of the list will be put to handle_list and so run handle_stripe(). Some steps
of handle_stripe() are extended to cover all stripes of the list, including
ops_run_io, ops_run_biodrain and so on. With this patch, we have less stripes
running in handle_stripe() and we send IO of whole stripe list together to
increase IO size.

Stripes added to a batch list have some limitations. A batch list can only
include full stripe write and can't cross chunk boundary to make sure stripes
have the same parity disks. Stripes in a batch list must be in the same state
(no written, toread and so on). If a stripe is in a batch list, all new
read/write to add_stripe_bio will be blocked to overlap conflict till the batch
list is handled. The limitations will make sure stripes in a batch list be in
exactly the same state in the life circly.

I did test running 160k randwrite in a RAID5 array with 32k chunk size and 6
PCIe SSD. This patch improves around 30% performance and IO size to under layer
disk is exactly 32k. I also run a 4k randwrite test in the same array to make
sure the performance isn't changed with the patch.
Signed-off-by: NShaohua Li <shli@fusionio.com>
Signed-off-by: NNeilBrown <neilb@suse.de>

Track overwrite disk count, so we can know if a stripe is a full stripe write.
Signed-off-by: NShaohua Li <shli@fusionio.com>
Signed-off-by: NNeilBrown <neilb@suse.de>

A freshly new stripe with write request can be batched. Any time the stripe is
handled or new read is queued, the flag will be cleared.
Signed-off-by: NShaohua Li <shli@fusionio.com>
Signed-off-by: NNeilBrown <neilb@suse.de>

Use flex_array for scribble data. Next patch will batch several stripes
together, so scribble data should be able to cover several stripes, so this
patch also allocates scribble data for stripes across a chunk.
Signed-off-by: NShaohua Li <shli@fusionio.com>
Signed-off-by: NNeilBrown <neilb@suse.de>

md raid0: access mddev->queue (request queue member) conditionally because it is not set when accessed from dm-raid

The patch makes 3 references to mddev->queue in the raid0 personality
conditional in order to allow for it to be accessed from dm-raid.
Mandatory, because md instances underneath dm-raid don't manage
a request queue of their own which'd lead to oopses without the patch.
Signed-off-by: NHeinz Mauelshagen <heinzm@redhat.com>
Tested-by: NHeinz Mauelshagen <heinzm@redhat.com>
Signed-off-by: NNeilBrown <neilb@suse.de>

When md notices non-sync IO happening while it is trying
to resync (or reshape or recover) it slows down to the
set minimum.

The default minimum might have made sense many years ago
but the drives have become faster.  Changing the default
to match the times isn't really a long term solution.

This patch changes the code so that instead of waiting until the speed
has dropped to the target, it just waits until pending requests
have completed.
This means that the delay inserted is a function of the speed
of the devices.

Testing shows that:
 - for some loads, the resync speed is unchanged.  For those loads
   increasing the minimum doesn't change the speed either.
   So this is a good result.  To increase resync speed under such
   loads we would probably need to increase the resync window
   size.

 - for other loads, resync speed does increase to a reasonable
   fraction (e.g. 20%) of maximum possible, and throughput of
   the load only drops a little bit (e.g. 10%)

 - for other loads, throughput of the non-sync load drops quite a bit
   more.  These seem to be latency-sensitive loads.

So it isn't a perfect solution, but it is mostly an improvement.
Signed-off-by: NNeilBrown <neilb@suse.de>

This option is not well justified and testing suggests that
it hardly ever makes any difference.

The comment suggests there might be a need to wait for non-resync
activity indicated by ->nr_waiting, however raise_barrier()
already waits for all of that.

So just remove it to simplify reasoning about speed limiting.

This allows us to remove a 'FIXME' comment from raid5.c as that
never used the flag.
Signed-off-by: NNeilBrown <neilb@suse.de>

There is really no need for sync_min to be a multiple of
chunk_size, and values read from here often aren't.
That means you cannot read a value and expect to be able
to write it back later.

So remove the chunk_size check, and round down to a multiple
of 4K, to be sure everything works with 4K-sector devices.
Signed-off-by: NNeilBrown <neilb@suse.de>

When "re-add" is writted to /sys/block/mdXX/md/dev-YYY/state,
the clustered md:

1. Sends RE_ADD message with the desc_nr. Nodes receiving the message
   clear the Faulty bit in their respective rdev->flags.
2. The node initiating re-add, gathers the bitmaps of all nodes
   and copies them into the local bitmap. It does not clear the bitmap
   from which it is copying.
3. Initiating node schedules a md recovery to sync the devices.
Signed-off-by: NGuoqing Jiang <gqjiang@suse.com>
Signed-off-by: NGoldwyn Rodrigues <rgoldwyn@suse.com>
Signed-off-by: NNeilBrown <neilb@suse.de>

This adds the capability of re-adding a failed disk by
writing "re-add" to /sys/block/mdXX/md/dev-YYY/state.

This facilitates adding disks which have encountered a temporary
error such as a network disconnection/hiccup in an iSCSI device,
or a SAN cable disconnection which has been restored. In such
a situation, you do not need to remove and re-add the device.
Writing re-add to the failed device's state would add it again
to the array and perform the recovery of only the blocks which
were written after the device failed.

This works for generic md, and is not related to clustering. However,
this patch is to ease re-add operations listed above in clustering
environments.
Signed-off-by: NGoldwyn Rodrigues <rgoldwyn@suse.com>
Signed-off-by: NNeilBrown <neilb@suse.de>

This adds "remove" capabilities for the clustered environment.
When a user initiates removal of a device from the array, a
REMOVE message with disk number in the array is sent to all
the nodes which kick the respective device in their own array.

This facilitates the removal of failed devices.
Signed-off-by: NGoldwyn Rodrigues <rgoldwyn@suse.com>
Signed-off-by: NNeilBrown <neilb@suse.de>

This is required by the clustering module (patches to follow) to
find the device to remove or re-add.
Signed-off-by: NGoldwyn Rodrigues <rgoldwyn@suse.com>
Signed-off-by: NNeilBrown <neilb@suse.de>

This export is required for clustering module in order to
co-ordinate remove/readd a rdev from all nodes.
Signed-off-by: NGoldwyn Rodrigues <rgoldwyn@suse.com>
Signed-off-by: NNeilBrown <neilb@suse.de>

Since the node num of md-cluster is from zero, and
cinfo->slot_number represents the slot num of dlm,
no need to check for equality.
Signed-off-by: NGuoqing Jiang <gqjiang@suse.com>
Signed-off-by: NGoldwyn Rodrigues <rgoldwyn@suse.com>
Signed-off-by: NNeilBrown <neilb@suse.de>

Since commit 20d0189b
in v3.14-rc1 RAID0 has performed incorrect calculations
when the chunksize is not a power of 2.

This happens because "sector_div()" modifies its first argument, but
this wasn't taken into account in the patch.

So restore that first arg before re-using the variable.
Reported-by: NJoe Landman <joe.landman@gmail.com>
Reported-by: NDave Chinner <david@fromorbit.com>
Fixes: 20d0189b
Cc: stable@vger.kernel.org (3.14 and later).
Signed-off-by: NNeilBrown <neilb@suse.de>

Simon reported the md io stats accounting issue:
"
I'm seeing "iostat -x -k 1" print this after a RAID1 rebuild on 4.0-rc5.
It's not abnormal other than it's 3-disk, with one being SSD (sdc) and
the other two being write-mostly:

Device:         rrqm/s   wrqm/s     r/s     w/s    rkB/s    wkB/s avgrq-sz avgqu-sz   await r_await w_await  svctm  %util
sda               0.00     0.00    0.00    0.00     0.00     0.00     0.00     0.00    0.00    0.00    0.00   0.00   0.00
sdb               0.00     0.00    0.00    0.00     0.00     0.00     0.00     0.00    0.00    0.00    0.00   0.00   0.00
sdc               0.00     0.00    0.00    0.00     0.00     0.00     0.00     0.00    0.00    0.00    0.00   0.00   0.00
md0               0.00     0.00    0.00    0.00     0.00     0.00     0.00   345.00    0.00    0.00    0.00   0.00 100.00
md2               0.00     0.00    0.00    0.00     0.00     0.00     0.00 58779.00    0.00    0.00    0.00   0.00 100.00
md1               0.00     0.00    0.00    0.00     0.00     0.00     0.00    12.00    0.00    0.00    0.00   0.00 100.00
"
The cause is commit "18c0b223" uses the
generic_start_io_acct to account the disk stats rather than the open code,
but it also introduced the increase to .in_flight[rw] which is needless to
md. So we re-use the open code here to fix it.
Reported-by: NSimon Kirby <sim@hostway.ca>
Cc: <stable@vger.kernel.org> 3.19
Signed-off-by: NGu Zheng <guz.fnst@cn.fujitsu.com>
Signed-off-by: NNeilBrown <neilb@suse.de>

Pull networking fixes from David Miller:

 1) In TCP, don't register an FRTO for cumulatively ACK'd data that was
    previously SACK'd, from Neal Cardwell.

 2) Need to hold RNL mutex in ipv4 multicast code namespace cleanup,
    from Cong WANG.

 3) Similarly we have to hold RNL mutex for fib_rules_unregister(), also
    from Cong WANG.

 4) Revert and rework netns nsid allocation fix, from Nicolas Dichtel.

 5) When we encapsulate for a tunnel device, skb->sk still points to the
    user socket.  So this leads to cases where we retraverse the
    ipv4/ipv6 output path with skb->sk being of some other address
    family (f.e. AF_PACKET).  This can cause things to crash since the
    ipv4 output path is dereferencing an AF_PACKET socket as if it were
    an ipv4 one.

    The short term fix for 'net' and -stable is to elide these socket
    checks once we've entered an encapsulation sequence by testing
    xmit_recursion.

    Longer term we have a better solution wherein we pass the tunnel's
    socket down through the output paths, but that is way too invasive
    for 'net' and -stable.

    From Hannes Frederic Sowa.

 6) l2tp_init() failure path forgets to unregister per-net ops, from
    Cong WANG.

* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net:
  net/mlx4_core: Fix error message deprecation for ConnectX-2 cards
  net: dsa: fix filling routing table from OF description
  l2tp: unregister l2tp_net_ops on failure path
  mvneta: dont call mvneta_adjust_link() manually
  ipv6: protect skb->sk accesses from recursive dereference inside the stack
  netns: don't allocate an id for dead netns
  Revert "netns: don't clear nsid too early on removal"
  ip6mr: call del_timer_sync() in ip6mr_free_table()
  net: move fib_rules_unregister() under rtnl lock
  ipv4: take rtnl_lock and mark mrt table as freed on namespace cleanup
  tcp: fix FRTO undo on cumulative ACK of SACKed range
  xen-netfront: transmit fully GSO-sized packets

Commit 1daa4303 ("net/mlx4_core: Deprecate error message at
ConnectX-2 cards startup to debug") did the deprecation only for port 1
of the card. Need to deprecate for port 2 as well.

Fixes: 1daa4303 ("net/mlx4_core: Deprecate error message at ConnectX-2 cards startup to debug")
Signed-off-by: NJack Morgenstein <jackm@dev.mellanox.co.il>
Signed-off-by: NAmir Vadai <amirv@mellanox.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

According to description in 'include/net/dsa.h', in cascade switches
configurations where there are more than one interconnected devices,
'rtable' array in 'dsa_chip_data' structure is used to indicate which
port on this switch should be used to send packets to that are destined
for corresponding switch.

However, dsa_of_setup_routing_table() fills 'rtable' with port numbers
of the _target_ switch, but not current one.

This commit removes redundant devicetree parsing and adds needed port
number as a function argument. So dsa_of_setup_routing_table() now just
looks for target switch number by parsing parent of 'link' device node.

To remove possible misunderstandings with the way of determining target
switch number, a corresponding comment was added to the source code and
to the DSA device tree bindings documentation file.

This was tested on a custom board with two Marvell 88E6095 switches with
following corresponding routing tables: { -1, 10 } and { 8, -1 }.
Signed-off-by: NPavel Nakonechny <pavel.nakonechny@skitlab.ru>
Reviewed-by: NAndrew Lunn <andrew@lunn.ch>
Reviewed-by: NFlorian Fainelli <f.fainelli@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Pull input fixes from Dmitry Torokhov:
 "Updates for the input subsystem - two more tweaks for ALPS driver to
  work out kinks after splitting the touchpad, trackstick, and potential
  external PS/2 mouse into separate input devices.

  Changes to support ALPS SS4 devices (protocol V8) will be coming in
  4.1..."

* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/dtor/input:
  Input: alps - document stick behavior for protocol V2
  Input: alps - report V2 Dualpoint Stick events via the right evdev node
  Input: alps - report interleaved bare PS/2 packets via dev3

Signed-off-by: NCong Wang <xiyou.wangcong@gmail.com>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

mvneta_adjust_link() is a callback for of_phy_connect() and should
not be called directly. The result of calling it directly is as below:
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

We should not consult skb->sk for output decisions in xmit recursion
levels > 0 in the stack. Otherwise local socket settings could influence
the result of e.g. tunnel encapsulation process.

ipv6 does not conform with this in three places:

1) ip6_fragment: we do consult ipv6_npinfo for frag_size

2) sk_mc_loop in ipv6 uses skb->sk and checks if we should
   loop the packet back to the local socket

3) ip6_skb_dst_mtu could query the settings from the user socket and
   force a wrong MTU

Furthermore:
In sk_mc_loop we could potentially land in WARN_ON(1) if we use a
PF_PACKET socket ontop of an IPv6-backed vxlan device.

Reuse xmit_recursion as we are currently only interested in protecting
tunnel devices.

Cc: Jiri Pirko <jiri@resnulli.us>
Signed-off-by: NHannes Frederic Sowa <hannes@stressinduktion.org>
Signed-off-by: NDavid S. Miller <davem@davemloft.net>

Document that protocol V2 uses standard (bare) PS/2 mouse packets for the
DualPoint stick.
Signed-off-by: NHans de Goede <hdegoede@redhat.com>
Acked-By: NPali Rohár <pali.rohar@gmail.com>
Signed-off-by: NDmitry Torokhov <dmitry.torokhov@gmail.com>

On V2 devices the DualPoint Stick reports bare packets, these should be
reported via the "AlpsPS/2 ALPS DualPoint Stick" dev2 evdev node, which also
has the INPUT_PROP_POINTING_STICK propbit set.

Note that since there is no way to distinguish these packets from an external
PS/2 mouse (insofar as these laptops have an external PS/2 port) this means
that we will be reporting PS/2 mouse events via this evdev node too, as we've
been doing in kernel 3.19 and older.

This has been tested on a Dell Latitude D620 and a Dell Latitude E6400,
which both have a V2 touchpad + a DualPoint Stick which reports bare packets.
Signed-off-by: NHans de Goede <hdegoede@redhat.com>
Reviewed-by: NPali Rohár <pali.rohar@gmail.com>
Signed-off-by: NDmitry Torokhov <dmitry.torokhov@gmail.com>

Bare packets should be reported via the same evdev device independent on
whether they are detected on the beginning of a packet or in the middle
of a packet.

This has been tested on a Dell Latitude E6400, where the DualPoint Stick
reports bare packets, which get reported via dev3 when the touchpad is
idle, and via dev2 when the touchpad and stick are used simultaneously.

This commit fixes this inconsistency by always reporting bare packets via
dev3. Note that since the come from a DualPoint Stick they really should be
reported via dev2, this gets fixed in a later commit.
Signed-off-by: NHans de Goede <hdegoede@redhat.com>
Reviewed-by: NPali Rohár <pali.rohar@gmail.com>
Signed-off-by: NDmitry Torokhov <dmitry.torokhov@gmail.com>