GitHub/moto-9609/android_kernel_motorola_exynos9610.git
9 years agoxen/block: add multi-page ring support
Bob Liu [Wed, 3 Jun 2015 05:40:03 +0000 (13:40 +0800)]
xen/block: add multi-page ring support

Extend xen/block to support multi-page ring, so that more requests can be
issued by using more than one pages as the request ring between blkfront
and backend.
As a result, the performance can get improved significantly.

We got some impressive improvements on our highend iscsi storage cluster
backend. If using 64 pages as the ring, the IOPS increased about 15 times
for the throughput testing and above doubled for the latency testing.

The reason was the limit on outstanding requests is 32 if use only one-page
ring, but in our case the iscsi lun was spread across about 100 physical
drives, 32 was really not enough to keep them busy.

Changes in v2:
 - Rebased to 4.0-rc6.
 - Document on how multi-page ring feature working to linux io/blkif.h.

Changes in v3:
 - Remove changes to linux io/blkif.h and follow the protocol defined
   in io/blkif.h of XEN tree.
 - Rebased to 4.1-rc3

Changes in v4:
 - Turn to use 'ring-page-order' and 'max-ring-page-order'.
 - A few comments from Roger.

Changes in v5:
 - Clarify with 4k granularity to comment
 - Address more comments from Roger

Signed-off-by: Bob Liu <bob.liu@oracle.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
9 years agodriver: xen-blkfront: move talk_to_blkback to a more suitable place
Bob Liu [Wed, 3 Jun 2015 05:40:02 +0000 (13:40 +0800)]
driver: xen-blkfront: move talk_to_blkback to a more suitable place

The major responsibility of talk_to_blkback() is allocate and initialize
the request ring and write the ring info to xenstore.
But this work should be done after backend entered 'XenbusStateInitWait' as
defined in the protocol file.
See xen/include/public/io/blkif.h in XEN git tree:
Front                                Back
=================================    =====================================
XenbusStateInitialising              XenbusStateInitialising
 o Query virtual device               o Query backend device identification
   properties.                          data.
 o Setup OS device instance.          o Open and validate backend device.
                                      o Publish backend features and
                                        transport parameters.
                                                     |
                                                     |
                                                     V
                                     XenbusStateInitWait

o Query backend features and
  transport parameters.
o Allocate and initialize the
  request ring.

There is no problem with this yet, but it is an violation of the design and
furthermore it would not allow frontend/backend to negotiate 'multi-page'
and 'multi-queue' features.

Changes in v2:
 - Re-write the commit message to be more clear.

Signed-off-by: Bob Liu <bob.liu@oracle.com>
Acked-by: Roger Pau Monné <roger.pau@citrix.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
9 years agodrivers: xen-blkback: delay pending_req allocation to connect_ring
Bob Liu [Wed, 3 Jun 2015 05:40:01 +0000 (13:40 +0800)]
drivers: xen-blkback: delay pending_req allocation to connect_ring

This is a pre-patch for multi-page ring feature.
In connect_ring, we can know exactly how many pages are used for the shared
ring, delay pending_req allocation here so that we won't waste too much memory.

Signed-off-by: Bob Liu <bob.liu@oracle.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
9 years agoNVMe: Automatic namespace rescan
Keith Busch [Mon, 1 Jun 2015 20:28:14 +0000 (14:28 -0600)]
NVMe: Automatic namespace rescan

Namespaces may be dynamically allocated and deleted or attached and
detached. This has the driver rescan the device for namespace changes
after each device reset or namespace change asynchronous event.

There could potentially be many detached namespaces that we don't want
polluting /dev/ with unusable block handles, so this will delete disks
if the namespace is not active as indicated by the response from identify
namespace. This also skips adding the disk if no capacity is provisioned
to the namespace in the first place.

Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoMerge branch 'for-4.2/core' into for-4.2/drivers
Jens Axboe [Fri, 5 Jun 2015 16:58:28 +0000 (10:58 -0600)]
Merge branch 'for-4.2/core' into for-4.2/drivers

9 years agoblock: add blk_set_queue_dying() to blkdev.h
Jens Axboe [Fri, 5 Jun 2015 16:57:37 +0000 (10:57 -0600)]
block: add blk_set_queue_dying() to blkdev.h

We export this function and NVMe wants to use it, but for some reason
it was never added to the block header. Do that.

Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoNVMe: Memory barrier before queue_count is incremented
Jon Derrick [Wed, 27 May 2015 18:26:23 +0000 (12:26 -0600)]
NVMe: Memory barrier before queue_count is incremented

Protects against reordering and/or preempting which would allow the
kthread to access the queue descriptor before it is set up

Signed-off-by: Jon Derrick <jonathan.derrick@intel.com>
Acked-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoNVMe: add sysfs and ioctl controller reset
Keith Busch [Fri, 5 Jun 2015 16:30:08 +0000 (10:30 -0600)]
NVMe: add sysfs and ioctl controller reset

We need the ability to perform an nvme controller reset as discussed on
the mailing list thread:

  http://lists.infradead.org/pipermail/linux-nvme/2015-March/001585.html

This adds a sysfs entry that when written to will reset perform an NVMe
controller reset if the controller was successfully initialized in the
first place.

This also adds locking around resetting the device in the async probe
method so the driver can't schedule two resets.

Signed-off-by: Keith Busch <keith.busch@intel.com>
Cc: Brandon Schultz <brandon.schulz@hgst.com>
Cc: David Sariel <david.sariel@pmcs.com>
Updated by Jens to:

1) Merge this with the ioctl reset patch from David Sariel. The ioctl
   path now shares the reset code from the sysfs path.

2) Don't flush work if we fail issuing the reset.

Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agonull_blk: restart request processing on completion handler
Akinobu Mita [Mon, 1 Jun 2015 23:35:10 +0000 (08:35 +0900)]
null_blk: restart request processing on completion handler

When irqmode=2 (IRQ completion handler is timer) and queue_mode=1
(Block interface to use is rq), the completion handler should restart
request handling for any pending requests on a queue because request
processing stops when the number of commands are queued more than
hw_queue_depth (null_rq_prep_fn returns BLKPREP_DEFER).

Without this change, the following command cannot finish.

# modprobe null_blk irqmode=2 queue_mode=1 hw_queue_depth=1
# fio --name=t --rw=read --size=1g --direct=1 \
  --ioengine=libaio --iodepth=64 --filename=/dev/nullb0

Signed-off-by: Akinobu Mita <akinobu.mita@gmail.com>
Cc: Jens Axboe <axboe@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agonull_blk: prevent timer handler running on a different CPU where started
Akinobu Mita [Mon, 1 Jun 2015 23:35:09 +0000 (08:35 +0900)]
null_blk: prevent timer handler running on a different CPU where started

When irqmode=2 (IRQ completion handler is timer), timer handler should
be called on the same CPU where the timer has been started.

Since completion_queues are per-cpu and the completion handler only
touches completion_queue for local CPU, we need to prevent the handler
from running on a different CPU where the timer has been started.
Otherwise, the IO cannot be completed until another completion handler
is executed on that CPU.

Signed-off-by: Akinobu Mita <akinobu.mita@gmail.com>
Cc: Jens Axboe <axboe@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoNVMe: Remove hctx reliance for multi-namespace
Keith Busch [Mon, 1 Jun 2015 15:29:54 +0000 (09:29 -0600)]
NVMe: Remove hctx reliance for multi-namespace

The driver needs to track shared tags to support multiple namespaces
that may be dynamically allocated or deleted. Relying on the first
request_queue's hctx's is not appropriate as we cannot clear outstanding
tags for all namespaces using this handle, nor can the driver easily track
all request_queue's hctx as namespaces are attached/detached. Instead,
this patch uses the nvme_dev's tagset to get the shared tag resources
instead of through a request_queue hctx.

Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoMerge branch 'for-4.2/core' into for-4.2/drivers
Jens Axboe [Mon, 1 Jun 2015 20:36:02 +0000 (14:36 -0600)]
Merge branch 'for-4.2/core' into for-4.2/drivers

9 years agoblk-mq: Shared tag enhancements
Keith Busch [Mon, 1 Jun 2015 15:29:53 +0000 (09:29 -0600)]
blk-mq: Shared tag enhancements

Storage controllers may expose multiple block devices that share hardware
resources managed by blk-mq. This patch enhances the shared tags so a
low-level driver can access the shared resources not tied to the unshared
h/w contexts. This way the LLD can dynamically add and delete disks and
request queues without having to track all the request_queue hctx's to
iterate outstanding tags.

Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblock: don't honor chunk sizes for data-less IO
Jens Axboe [Fri, 29 May 2015 19:11:32 +0000 (13:11 -0600)]
block: don't honor chunk sizes for data-less IO

We don't need to honor chunk sizes for IO that doesn't carry any
data.

Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblock: only honor SG gap prevention for merges that contain data
Jens Axboe [Fri, 29 May 2015 19:10:23 +0000 (13:10 -0600)]
block: only honor SG gap prevention for merges that contain data

We can safely merge anything that wont generate an SG list entry,
so if the bio is data-less (discard), don't look at potential
SG gaps.

Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoNVMe: End sync requests immediately on failure
Keith Busch [Thu, 28 May 2015 15:48:55 +0000 (09:48 -0600)]
NVMe: End sync requests immediately on failure

Do not retry failed sync commands so the original status may be seen
without issuing unnecessary retries.

Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoNVMe: Use requested sync command timeout
Keith Busch [Thu, 28 May 2015 15:48:54 +0000 (09:48 -0600)]
NVMe: Use requested sync command timeout

Signed-off-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblock: fix returnvar.cocci warnings
Julia Lawall [Tue, 26 May 2015 19:59:53 +0000 (21:59 +0200)]
block: fix returnvar.cocci warnings

Remove unneeded variable used to store return value.

Generated by: scripts/coccinelle/misc/returnvar.cocci

Signed-off-by: Fengguang Wu <fengguang.wu@intel.com>
Signed-off-by: Julia Lawall <julia.lawall@lip6.fr>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoNVMe: Fix obtaining command result
Keith Busch [Fri, 22 May 2015 18:28:31 +0000 (12:28 -0600)]
NVMe: Fix obtaining command result

Replaces req->sense_len usage, which is not owned by the LLD, to
req->special to contain the command result for driver created commands,
and sets the result unconditionally on completion.

Signed-off-by: Keith Busch <keith.busch@intel.com>
Cc: Christoph Hellwig <hch@lst.de>
Cc: Jens Axboe <axboe@fb.com>
Fixes: d29ec8241c10 ("nvme: submit internal commands through the block layer")
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblock, dm: don't copy bios for request clones
Christoph Hellwig [Fri, 22 May 2015 13:14:04 +0000 (09:14 -0400)]
block, dm: don't copy bios for request clones

Currently dm-multipath has to clone the bios for every request sent
to the lower devices, which wastes cpu cycles and ties down memory.

This patch instead adds a new REQ_CLONE flag that instructs req_bio_endio
to not complete bios attached to a request, which we set on clone
requests similar to bios in a flush sequence.  With this change I/O
errors on a path failure only get propagated to dm-multipath, which
can then either resubmit the I/O or complete the bios on the original
request.

I've done some basic testing of this on a Linux target with ALUA support,
and it survives path failures during I/O nicely.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblock: remove management of bi_remaining when restoring original bi_end_io
Mike Snitzer [Fri, 22 May 2015 13:14:03 +0000 (09:14 -0400)]
block: remove management of bi_remaining when restoring original bi_end_io

Commit c4cf5261 ("bio: skip atomic inc/dec of ->bi_remaining for
non-chains") regressed all existing callers that followed this pattern:
 1) saving a bio's original bi_end_io
 2) wiring up an intermediate bi_end_io
 3) restoring the original bi_end_io from intermediate bi_end_io
 4) calling bio_endio() to execute the restored original bi_end_io

The regression was due to BIO_CHAIN only ever getting set if
bio_inc_remaining() is called.  For the above pattern it isn't set until
step 3 above (step 2 would've needed to establish BIO_CHAIN).  As such
the first bio_endio(), in step 2 above, never decremented __bi_remaining
before calling the intermediate bi_end_io -- leaving __bi_remaining with
the value 1 instead of 0.  When bio_inc_remaining() occurred during step
3 it brought it to a value of 2.  When the second bio_endio() was
called, in step 4 above, it should've called the original bi_end_io but
it didn't because there was an extra reference that wasn't dropped (due
to atomic operations being optimized away since BIO_CHAIN wasn't set
upfront).

Fix this issue by removing the __bi_remaining management complexity for
all callers that use the above pattern -- bio_chain() is the only
interface that _needs_ to be concerned with __bi_remaining.  For the
above pattern callers just expect the bi_end_io they set to get called!
Remove bio_endio_nodec() and also remove all bio_inc_remaining() calls
that aren't associated with the bio_chain() interface.

Also, the bio_inc_remaining() interface has been moved local to bio.c.

Fixes: c4cf5261 ("bio: skip atomic inc/dec of ->bi_remaining for non-chains")
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Mike Snitzer <snitzer@redhat.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agonvme: submit internal commands through the block layer
Christoph Hellwig [Fri, 22 May 2015 09:12:46 +0000 (11:12 +0200)]
nvme: submit internal commands through the block layer

Use block layer queues with an internal cmd_type to submit internally
generated NVMe commands.  This both simplifies the code a lot and allow
for a better structure.  For example now the LighNVM code can construct
commands without knowing the details of the underlying I/O descriptors.
Or a future NVMe over network target could inject commands, as well as
could the SCSI translation and ioctl code be reused for such a beast.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agonvme: fail SCSI read/write command with unsupported protection bit
Christoph Hellwig [Fri, 22 May 2015 09:12:45 +0000 (11:12 +0200)]
nvme: fail SCSI read/write command with unsupported protection bit

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agonvme: report the DPOFUA in MODE_SENSE
Christoph Hellwig [Fri, 22 May 2015 09:12:44 +0000 (11:12 +0200)]
nvme: report the DPOFUA in MODE_SENSE

NVMe device always support the FUA bit, and the SCSI translations
accepts the DPO bit, which doesn't have much of a meaning for us.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agonvme: simplify and cleanup the READ/WRITE SCSI CDB parsing code
Christoph Hellwig [Fri, 22 May 2015 09:12:43 +0000 (11:12 +0200)]
nvme: simplify and cleanup the READ/WRITE SCSI CDB parsing code

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agonvme: first round at deobsfucating the SCSI translation code
Christoph Hellwig [Fri, 22 May 2015 09:12:42 +0000 (11:12 +0200)]
nvme: first round at deobsfucating the SCSI translation code

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agonvme: fix scsi translation error handling
Christoph Hellwig [Fri, 22 May 2015 09:12:41 +0000 (11:12 +0200)]
nvme: fix scsi translation error handling

Erorr handling for the scsi translation was completely broken, as there
were two different positive error number spaces overlapping.  Fix this
up by removing one of them, and centralizing the generation of the other
positive values in a single place.  Also fix up a few places that didn't
handle the NVMe error codes properly.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agonvme: split nvme_trans_send_fw_cmd
Christoph Hellwig [Fri, 22 May 2015 09:12:40 +0000 (11:12 +0200)]
nvme: split nvme_trans_send_fw_cmd

This function handles two totally different opcodes, so split it.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agonvme: store a struct device pointer in struct nvme_dev
Christoph Hellwig [Fri, 22 May 2015 09:12:39 +0000 (11:12 +0200)]
nvme: store a struct device pointer in struct nvme_dev

Most users want the generic device, so store that in struct nvme_dev
instead of the pci_dev.  This also happens to be a nice step towards
making some code reusable for non-PCI transports.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agonvme: consolidate synchronous command submission helpers
Christoph Hellwig [Fri, 22 May 2015 09:12:38 +0000 (11:12 +0200)]
nvme: consolidate synchronous command submission helpers

Note that we keep the unused timeout argument, but allow callers to
pass 0 instead of a timeout if they want the default.  This will allow
adding a timeout to the pass through path later on.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoloop: remove (now) unused 'out' label
Jens Axboe [Wed, 20 May 2015 15:54:35 +0000 (09:54 -0600)]
loop: remove (now) unused 'out' label

gcc, righfully, complains:

drivers/block/loop.c:1369:1: warning: label 'out' defined but not used [-Wunused-label]

Kill it.

Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agos390/block/dasd: remove obsolete while -EBUSY loop
Jarod Wilson [Wed, 6 May 2015 04:26:28 +0000 (12:26 +0800)]
s390/block/dasd: remove obsolete while -EBUSY loop

With the mutex_trylock bit gone from blkdev_reread_part(), the retry logic
in dasd_scan_partitions() shouldn't be necessary.

CC: Christoph Hellwig <hch@infradead.org>
CC: Jens Axboe <axboe@kernel.dk>
CC: Tejun Heo <tj@kernel.org>
CC: Alexander Viro <viro@zeniv.linux.org.uk>
CC: Markus Pargmann <mpa@pengutronix.de>
CC: Stefan Weinhuber <wein@de.ibm.com>
CC: Stefan Haberland <stefan.haberland@de.ibm.com>
CC: Sebastian Ott <sebott@linux.vnet.ibm.com>
CC: Fabian Frederick <fabf@skynet.be>
CC: Ming Lei <ming.lei@canonical.com>
CC: David Herrmann <dh.herrmann@gmail.com>
CC: Andrew Morton <akpm@linux-foundation.org>
CC: Peter Zijlstra <peterz@infradead.org>
CC: nbd-general@lists.sourceforge.net
CC: linux-s390@vger.kernel.org
Reviewed-by: Christoph Hellwig <hch@lst.de>
Acked-by: Sebastian Ott <sebott@linux.vnet.ibm.com>
Signed-off-by: Ming Lei <ming.lei@canonical.com>
Signed-off-by: Jarod Wilson <jarod@redhat.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblock: dasd_genhd: convert to blkdev_reread_part
Ming Lei [Wed, 6 May 2015 04:26:26 +0000 (12:26 +0800)]
block: dasd_genhd: convert to blkdev_reread_part

Also remove the obsolete comment.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Tested-by: Jarod Wilson <jarod@redhat.com>
Acked-by: Jarod Wilson <jarod@redhat.com>
Acked-by: Sebastian Ott <sebott@linux.vnet.ibm.com>
Signed-off-by: Ming Lei <ming.lei@canonical.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblock: nbd: convert to blkdev_reread_part()
Ming Lei [Wed, 6 May 2015 04:26:25 +0000 (12:26 +0800)]
block: nbd: convert to blkdev_reread_part()

Reviewed-by: Christoph Hellwig <hch@lst.de>
Tested-by: Jarod Wilson <jarod@redhat.com>
Acked-by: Jarod Wilson <jarod@redhat.com>
Signed-off-by: Ming Lei <ming.lei@canonical.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblock: loop: fix another reread part failure
Ming Lei [Wed, 6 May 2015 04:26:24 +0000 (12:26 +0800)]
block: loop: fix another reread part failure

loop_clr_fd() can be run piggyback with lo_release(), and
under this situation, reread partition may always fail because
bd_mutex has been held already.

This patch detects the situation by the reference count, and
call __blkdev_reread_part() to avoid acquiring the lock again.

In the meantime, this patch switches to new kernel APIs
of blkdev_reread_part() and __blkdev_reread_part().

Reviewed-by: Christoph Hellwig <hch@lst.de>
Tested-by: Jarod Wilson <jarod@redhat.com>
Acked-by: Jarod Wilson <jarod@redhat.com>
Signed-off-by: Jarod Wilson <jarod@redhat.com>
Signed-off-by: Ming Lei <ming.lei@canonical.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblock: loop: don't hold lo_ctl_mutex in lo_open
Ming Lei [Wed, 6 May 2015 04:26:23 +0000 (12:26 +0800)]
block: loop: don't hold lo_ctl_mutex in lo_open

The lo_ctl_mutex is held for running all ioctl handlers, and
in some ioctl handlers, ioctl_by_bdev(BLKRRPART) is called for
rereading partitions, which requires bd_mutex.

So it is easy to cause failure because trylock(bd_mutex) may
fail inside blkdev_reread_part(), and follows the lock context:

blkid or other application:
->open()
->mutex_lock(bd_mutex)
->lo_open()
->mutex_lock(lo_ctl_mutex)

losetup(set fd ioctl):
->mutex_lock(lo_ctl_mutex)
->ioctl_by_bdev(BLKRRPART)
->trylock(bd_mutex)

This patch trys to eliminate the ABBA lock dependency by removing
lo_ctl_mutext in lo_open() with the following approach:

1) make lo_refcnt as atomic_t and avoid acquiring lo_ctl_mutex in lo_open():
- for open vs. add/del loop, no any problem because of loop_index_mutex
- freeze request queue during clr_fd, so I/O can't come until
  clearing fd is completed, like the effect of holding lo_ctl_mutex
  in lo_open
- both open() and release() have been serialized by bd_mutex already

2) don't hold lo_ctl_mutex for decreasing/checking lo_refcnt in
lo_release(), then lo_ctl_mutex is only required for the last release.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Tested-by: Jarod Wilson <jarod@redhat.com>
Acked-by: Jarod Wilson <jarod@redhat.com>
Signed-off-by: Ming Lei <ming.lei@canonical.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoMerge branch 'for-4.2/core' into for-4.2/drivers
Jens Axboe [Wed, 20 May 2015 15:05:57 +0000 (09:05 -0600)]
Merge branch 'for-4.2/core' into for-4.2/drivers

We need the blkdev_reread_part() changes for drivers to adapt.

9 years agoblock: replace trylock with mutex_lock in blkdev_reread_part()
Ming Lei [Wed, 6 May 2015 04:26:27 +0000 (12:26 +0800)]
block: replace trylock with mutex_lock in blkdev_reread_part()

The only possible problem of using mutex_lock() instead of trylock
is about deadlock.

If there aren't any locks held before calling blkdev_reread_part(),
deadlock can't be caused by this conversion.

If there are locks held before calling blkdev_reread_part(),
and if these locks arn't required in open, close handler and I/O
path, deadlock shouldn't be caused too.

Both user space's ioctl(BLKRRPART) and md_setup_drive() from
init/do_mounts_md.c belongs to the 1st case, so the conversion is safe
for the two cases.

For loop, the previous patches in this pathset has fixed the ABBA lock
dependency, so the conversion is OK.

For nbd, tx_lock is held when calling the function:

- both open and release won't hold the lock
- when blkdev_reread_part() is run, I/O thread has been stopped
already, so tx_lock won't be acquired in I/O path at that time.
- so the conversion won't cause deadlock for nbd

For dasd, both dasd_open(), dasd_release() and request function don't
acquire any mutex/semphone, so the conversion should be safe.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Tested-by: Jarod Wilson <jarod@redhat.com>
Acked-by: Jarod Wilson <jarod@redhat.com>
Signed-off-by: Ming Lei <ming.lei@canonical.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblock: export blkdev_reread_part() and __blkdev_reread_part()
Jarod Wilson [Wed, 6 May 2015 04:26:22 +0000 (12:26 +0800)]
block: export blkdev_reread_part() and __blkdev_reread_part()

This patch exports blkdev_reread_part() for block drivers, also
introduce __blkdev_reread_part().

For some drivers, such as loop, reread of partitions can be run
from the release path, and bd_mutex may already be held prior to
calling ioctl_by_bdev(bdev, BLKRRPART, 0), so introduce
__blkdev_reread_part for use in such cases.

CC: Christoph Hellwig <hch@lst.de>
CC: Jens Axboe <axboe@kernel.dk>
CC: Tejun Heo <tj@kernel.org>
CC: Alexander Viro <viro@zeniv.linux.org.uk>
CC: Markus Pargmann <mpa@pengutronix.de>
CC: Stefan Weinhuber <wein@de.ibm.com>
CC: Stefan Haberland <stefan.haberland@de.ibm.com>
CC: Sebastian Ott <sebott@linux.vnet.ibm.com>
CC: Fabian Frederick <fabf@skynet.be>
CC: Ming Lei <ming.lei@canonical.com>
CC: David Herrmann <dh.herrmann@gmail.com>
CC: Andrew Morton <akpm@linux-foundation.org>
CC: Peter Zijlstra <peterz@infradead.org>
CC: nbd-general@lists.sourceforge.net
CC: linux-s390@vger.kernel.org
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jarod Wilson <jarod@redhat.com>
Signed-off-by: Ming Lei <ming.lei@canonical.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agosuspend: simplify block I/O handling
Christoph Hellwig [Tue, 19 May 2015 07:23:23 +0000 (09:23 +0200)]
suspend: simplify block I/O handling

Stop abusing struct page functionality and the swap end_io handler, and
instead add a modified version of the blk-lib.c bio_batch helpers.

Also move the block I/O code into swap.c as they are directly tied into
each other.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Tested-by: Pavel Machek <pavel@ucw.cz>
Tested-by: Ming Lin <mlin@kernel.org>
Acked-by: Pavel Machek <pavel@ucw.cz>
Acked-by: Rafael J. Wysocki <rjw@rjwysocki.net>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblock: collapse bio bit space
Jens Axboe [Tue, 19 May 2015 15:18:28 +0000 (09:18 -0600)]
block: collapse bio bit space

Various previous patches removed bits and left holes, collapse them
all. Leave the reset start bit where it is, we don't need to change
that.

Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblock: remove unused BIO_RW_BLOCK and BIO_EOF flags
Christoph Hellwig [Fri, 24 Apr 2015 19:41:02 +0000 (21:41 +0200)]
block: remove unused BIO_RW_BLOCK and BIO_EOF flags

Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblock: remove BIO_EOPNOTSUPP
Christoph Hellwig [Fri, 24 Apr 2015 19:41:01 +0000 (21:41 +0200)]
block: remove BIO_EOPNOTSUPP

Since the big barrier rewrite/removal in 2007 we never fail FLUSH or
FUA requests, which means we can remove the magic BIO_EOPNOTSUPP flag
to help propagating those to the buffer_head layer.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Jeff Moyer <jmoyer@redhat.com>
Acked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agonvme: disable irqs in nvme_freeze_queues
Christoph Hellwig [Thu, 7 May 2015 07:38:14 +0000 (09:38 +0200)]
nvme: disable irqs in nvme_freeze_queues

The queue_lock needs to be taken with irqs disabled.  This is mostly
due to the old pre blk-mq usage pattern, but we've also picked it up
in most of the few places where we use the queue_lock with blk-mq.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblock: use an atomic_t for mq_freeze_depth
Christoph Hellwig [Thu, 7 May 2015 07:38:13 +0000 (09:38 +0200)]
block: use an atomic_t for mq_freeze_depth

lockdep gets unhappy about the not disabling irqs when using the queue_lock
around it.  Instead of trying to fix that up just switch to an atomic_t
and get rid of the lock.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agocciss: correct the non-resettable board list
Tomas Henzl [Tue, 17 Feb 2015 16:40:22 +0000 (17:40 +0100)]
cciss: correct the non-resettable board list

The hpsa driver carries a more recent version,
copy the table from there.

Signed-off-by: Tomas Henzl <thenzl@redhat.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agocciss: remove duplicate entries from board_type struct
Tomas Henzl [Tue, 17 Feb 2015 16:40:21 +0000 (17:40 +0100)]
cciss: remove duplicate entries from board_type struct

and devices not supported by this driver from unresettable list

Signed-off-by: Tomas Henzl <thenzl@redhat.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblk-mq: make plug work for mutiple disks and queues
Shaohua Li [Fri, 8 May 2015 17:51:33 +0000 (10:51 -0700)]
blk-mq: make plug work for mutiple disks and queues

Last patch makes plug work for multiple queue case. However it only
works for single disk case, because it assumes only one request in the
plug list. If a task is accessing multiple disks, eg MD/DM, the
assumption is wrong. Let blk_attempt_plug_merge() record request from
the same queue.

V2: use NULL parameter in !mq case. Fix a bug. Add comments in
blk_attempt_plug_merge to make it less (hopefully) confusion.

Cc: Jens Axboe <axboe@fb.com>
Cc: Christoph Hellwig <hch@lst.de>
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblk-mq: do limited block plug for multiple queue case
Shaohua Li [Fri, 8 May 2015 17:51:32 +0000 (10:51 -0700)]
blk-mq: do limited block plug for multiple queue case

plug is still helpful for workload with IO merge, but it can be harmful
otherwise especially with multiple hardware queues, as there is
(supposed) no lock contention in this case and plug can introduce
latency. For multiple queues, we do limited plug, eg plug only if there
is request merge. If a request doesn't have merge with following
request, the requet will be dispatched immediately.

V2: check blk_queue_nomerges() as suggested by Jeff.

Cc: Jens Axboe <axboe@fb.com>
Cc: Christoph Hellwig <hch@lst.de>
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblk-mq: avoid re-initialize request which is failed in direct dispatch
Shaohua Li [Fri, 8 May 2015 17:51:31 +0000 (10:51 -0700)]
blk-mq: avoid re-initialize request which is failed in direct dispatch

If we directly issue a request and it fails, we use
blk_mq_merge_queue_io(). But we already assigned bio to a request in
blk_mq_bio_to_request. blk_mq_merge_queue_io shouldn't run
blk_mq_bio_to_request again.

Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblk-mq: fix plugging in blk_sq_make_request
Jeff Moyer [Fri, 8 May 2015 17:51:30 +0000 (10:51 -0700)]
blk-mq: fix plugging in blk_sq_make_request

The following appears in blk_sq_make_request:

/*
 * If we have multiple hardware queues, just go directly to
 * one of those for sync IO.
 */

We clearly don't have multiple hardware queues, here!  This comment was
introduced with this commit 07068d5b8e (blk-mq: split make request
handler for multi and single queue):

    We want slightly different behavior from them:

    - On single queue devices, we currently use the per-process plug
      for deferred IO and for merging.

    - On multi queue devices, we don't use the per-process plug, but
      we want to go straight to hardware for SYNC IO.

The old code had this:

        use_plug = !is_flush_fua && ((q->nr_hw_queues == 1) || !is_sync);

and that was converted to:

use_plug = !is_flush_fua && !is_sync;

which is not equivalent.  For the single queue case, that second half of
the && expression is always true.  So, what I think was actually inteded
follows (and this more closely matches what is done in blk_queue_bio).

V2: delete the 'likely', which should not be a big deal

Signed-off-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Shaohua Li <shli@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agosched: always use blk_schedule_flush_plug in io_schedule_out
Shaohua Li [Fri, 8 May 2015 17:51:29 +0000 (10:51 -0700)]
sched: always use blk_schedule_flush_plug in io_schedule_out

block plug callback could sleep, so we introduce a parameter
'from_schedule' and corresponding drivers can use it to destinguish a
schedule plug flush or a plug finish. Unfortunately io_schedule_out
still uses blk_flush_plug(). This causes below output (Note, I added a
might_sleep() in raid1_unplug to make it trigger faster, but the whole
thing doesn't matter if I add might_sleep). In raid1/10, this can cause
deadlock.

This patch makes io_schedule_out always uses blk_schedule_flush_plug.
This should only impact drivers (as far as I know, raid 1/10) which are
sensitive to the 'from_schedule' parameter.

[  370.817949] ------------[ cut here ]------------
[  370.817960] WARNING: CPU: 7 PID: 145 at ../kernel/sched/core.c:7306 __might_sleep+0x7f/0x90()
[  370.817969] do not call blocking ops when !TASK_RUNNING; state=2 set at [<ffffffff81092fcf>] prepare_to_wait+0x2f/0x90
[  370.817971] Modules linked in: raid1
[  370.817976] CPU: 7 PID: 145 Comm: kworker/u16:9 Tainted: G        W       4.0.0+ #361
[  370.817977] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996), BIOS 1.7.5-20140709_153802- 04/01/2014
[  370.817983] Workqueue: writeback bdi_writeback_workfn (flush-9:1)
[  370.817985]  ffffffff81cd83be ffff8800ba8cb298 ffffffff819dd7af 0000000000000001
[  370.817988]  ffff8800ba8cb2e8 ffff8800ba8cb2d8 ffffffff81051afc ffff8800ba8cb2c8
[  370.817990]  ffffffffa00061a8 000000000000041e 0000000000000000 ffff8800ba8cba28
[  370.817993] Call Trace:
[  370.817999]  [<ffffffff819dd7af>] dump_stack+0x4f/0x7b
[  370.818002]  [<ffffffff81051afc>] warn_slowpath_common+0x8c/0xd0
[  370.818004]  [<ffffffff81051b86>] warn_slowpath_fmt+0x46/0x50
[  370.818006]  [<ffffffff81092fcf>] ? prepare_to_wait+0x2f/0x90
[  370.818008]  [<ffffffff81092fcf>] ? prepare_to_wait+0x2f/0x90
[  370.818010]  [<ffffffff810776ef>] __might_sleep+0x7f/0x90
[  370.818014]  [<ffffffffa0000c03>] raid1_unplug+0xd3/0x170 [raid1]
[  370.818024]  [<ffffffff81421d9a>] blk_flush_plug_list+0x8a/0x1e0
[  370.818028]  [<ffffffff819e3550>] ? bit_wait+0x50/0x50
[  370.818031]  [<ffffffff819e21b0>] io_schedule_timeout+0x130/0x140
[  370.818033]  [<ffffffff819e3586>] bit_wait_io+0x36/0x50
[  370.818034]  [<ffffffff819e31b5>] __wait_on_bit+0x65/0x90
[  370.818041]  [<ffffffff8125b67c>] ? ext4_read_block_bitmap_nowait+0xbc/0x630
[  370.818043]  [<ffffffff819e3550>] ? bit_wait+0x50/0x50
[  370.818045]  [<ffffffff819e3302>] out_of_line_wait_on_bit+0x72/0x80
[  370.818047]  [<ffffffff810935e0>] ? autoremove_wake_function+0x40/0x40
[  370.818050]  [<ffffffff811de744>] __wait_on_buffer+0x44/0x50
[  370.818053]  [<ffffffff8125ae80>] ext4_wait_block_bitmap+0xe0/0xf0
[  370.818058]  [<ffffffff812975d6>] ext4_mb_init_cache+0x206/0x790
[  370.818062]  [<ffffffff8114bc6c>] ? lru_cache_add+0x1c/0x50
[  370.818064]  [<ffffffff81297c7e>] ext4_mb_init_group+0x11e/0x200
[  370.818066]  [<ffffffff81298231>] ext4_mb_load_buddy+0x341/0x360
[  370.818068]  [<ffffffff8129a1a3>] ext4_mb_find_by_goal+0x93/0x2f0
[  370.818070]  [<ffffffff81295b54>] ? ext4_mb_normalize_request+0x1e4/0x5b0
[  370.818072]  [<ffffffff8129ab67>] ext4_mb_regular_allocator+0x67/0x460
[  370.818074]  [<ffffffff81295b54>] ? ext4_mb_normalize_request+0x1e4/0x5b0
[  370.818076]  [<ffffffff8129ca4b>] ext4_mb_new_blocks+0x4cb/0x620
[  370.818079]  [<ffffffff81290956>] ext4_ext_map_blocks+0x4c6/0x14d0
[  370.818081]  [<ffffffff812a4d4e>] ? ext4_es_lookup_extent+0x4e/0x290
[  370.818085]  [<ffffffff8126399d>] ext4_map_blocks+0x14d/0x4f0
[  370.818088]  [<ffffffff81266fbd>] ext4_writepages+0x76d/0xe50
[  370.818094]  [<ffffffff81149691>] do_writepages+0x21/0x50
[  370.818097]  [<ffffffff811d5c00>] __writeback_single_inode+0x60/0x490
[  370.818099]  [<ffffffff811d630a>] writeback_sb_inodes+0x2da/0x590
[  370.818103]  [<ffffffff811abf4b>] ? trylock_super+0x1b/0x50
[  370.818105]  [<ffffffff811abf4b>] ? trylock_super+0x1b/0x50
[  370.818107]  [<ffffffff811d665f>] __writeback_inodes_wb+0x9f/0xd0
[  370.818109]  [<ffffffff811d69db>] wb_writeback+0x34b/0x3c0
[  370.818111]  [<ffffffff811d70df>] bdi_writeback_workfn+0x23f/0x550
[  370.818116]  [<ffffffff8106bbd8>] process_one_work+0x1c8/0x570
[  370.818117]  [<ffffffff8106bb5b>] ? process_one_work+0x14b/0x570
[  370.818119]  [<ffffffff8106c09b>] worker_thread+0x11b/0x470
[  370.818121]  [<ffffffff8106bf80>] ? process_one_work+0x570/0x570
[  370.818124]  [<ffffffff81071868>] kthread+0xf8/0x110
[  370.818126]  [<ffffffff81071770>] ? kthread_create_on_node+0x210/0x210
[  370.818129]  [<ffffffff819e9322>] ret_from_fork+0x42/0x70
[  370.818131]  [<ffffffff81071770>] ? kthread_create_on_node+0x210/0x210
[  370.818132] ---[ end trace 7b4deb71e68b6605 ]---

V2: don't change ->in_iowait

Cc: NeilBrown <neilb@suse.de>
Signed-off-by: Shaohua Li <shli@fb.com>
Reviewed-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblk: clean up plug
Shaohua Li [Fri, 8 May 2015 17:51:28 +0000 (10:51 -0700)]
blk: clean up plug

Current code looks like inner plug gets flushed with a
blk_finish_plug(). Actually it's a nop. All requests/callbacks are added
to current->plug, while only outmost plug is assigned to current->plug.
So inner plug always has empty request/callback list, which makes
blk_flush_plug_list() a nop. This tries to make the code more clear.

Signed-off-by: Shaohua Li <shli@fb.com>
Reviewed-by: Jeff Moyer <jmoyer@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblock: loop: avoiding too many pending per work I/O
Ming Lei [Tue, 5 May 2015 11:49:55 +0000 (19:49 +0800)]
block: loop: avoiding too many pending per work I/O

If there are too many pending per work I/O, too many
high priority work thread can be generated so that
system performance can be effected.

This patch limits the max_active parameter of workqueue as 16.

This patch fixes Fedora 22 live booting performance
regression when it is booted from squashfs over dm
based on loop, and looks the following reasons are
related with the problem:

- not like other filesyststems(such as ext4), squashfs
is a bit special, and I observed that increasing I/O jobs
to access file in squashfs only improve I/O performance a
little, but it can make big difference for ext4

- nested loop: both squashfs.img and ext3fs.img are mounted
as loop block, and ext3fs.img is inside the squashfs

- during booting, lots of tasks may run concurrently

Fixes: b5dd2f6047ca108001328aac0e8588edd15f1778
Cc: stable@vger.kernel.org (v4.0)
Cc: Justin M. Forbes <jforbes@fedoraproject.org>
Signed-off-by: Ming Lei <ming.lei@canonical.com>
Acked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblock: loop: convert to per-device workqueue
Ming Lei [Tue, 5 May 2015 11:49:54 +0000 (19:49 +0800)]
block: loop: convert to per-device workqueue

Documentation/workqueue.txt:
If there is dependency among multiple work items used
during memory reclaim, they should be queued to separate
wq each with WQ_MEM_RECLAIM.

Loop devices can be stacked, so we have to convert to per-device
workqueue. One example is Fedora live CD.

Fixes: b5dd2f6047ca108001328aac0e8588edd15f1778
Cc: stable@vger.kernel.org (v4.0)
Cc: Justin M. Forbes <jforbes@fedoraproject.org>
Signed-off-by: Ming Lei <ming.lei@canonical.com>
Acked-by: Tejun Heo <tj@kernel.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agonbd: stop using req->cmd
Christoph Hellwig [Fri, 17 Apr 2015 20:37:21 +0000 (22:37 +0200)]
nbd: stop using req->cmd

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblock: move PM request support to IDE
Christoph Hellwig [Fri, 17 Apr 2015 20:37:20 +0000 (22:37 +0200)]
block: move PM request support to IDE

This removes the request types and hacks from the block code and into the
old IDE driver.  There is a small amunt of code duplication due to this,
but it's not too bad.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblock: remove REQ_TYPE_PM_SHUTDOWN
Christoph Hellwig [Fri, 17 Apr 2015 20:37:19 +0000 (22:37 +0200)]
block: remove REQ_TYPE_PM_SHUTDOWN

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblock: move REQ_TYPE_SENSE to the ide driver
Christoph Hellwig [Fri, 17 Apr 2015 20:37:18 +0000 (22:37 +0200)]
block: move REQ_TYPE_SENSE to the ide driver

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblock: move REQ_TYPE_ATA_TASKFILE and REQ_TYPE_ATA_PC to ide.h
Christoph Hellwig [Fri, 17 Apr 2015 20:37:17 +0000 (22:37 +0200)]
block: move REQ_TYPE_ATA_TASKFILE and REQ_TYPE_ATA_PC to ide.h

These values are only used by the IDE driver, so move them into it
by allowing drivers to take cmd_type values after the first private
one.  Note that we have to turn cmd_type into a plain unsigned integer
so that gcc doesn't complain about mismatching enum types.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoblock: rename REQ_TYPE_SPECIAL to REQ_TYPE_DRV_PRIV
Christoph Hellwig [Fri, 17 Apr 2015 20:37:16 +0000 (22:37 +0200)]
block: rename REQ_TYPE_SPECIAL to REQ_TYPE_DRV_PRIV

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agobio: skip atomic inc/dec of ->bi_cnt for most use cases
Jens Axboe [Fri, 17 Apr 2015 22:23:59 +0000 (16:23 -0600)]
bio: skip atomic inc/dec of ->bi_cnt for most use cases

Struct bio has a reference count that controls when it can be freed.
Most uses cases is allocating the bio, which then returns with a
single reference to it, doing IO, and then dropping that single
reference. We can remove this atomic_dec_and_test() in the completion
path, if nobody else is holding a reference to the bio.

If someone does call bio_get() on the bio, then we flag the bio as
now having valid count and that we must properly honor the reference
count when it's being put.

Tested-by: Robert Elliott <elliott@hp.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agobio: skip atomic inc/dec of ->bi_remaining for non-chains
Jens Axboe [Fri, 17 Apr 2015 22:15:18 +0000 (16:15 -0600)]
bio: skip atomic inc/dec of ->bi_remaining for non-chains

Struct bio has an atomic ref count for chained bio's, and we use this
to know when to end IO on the bio. However, most bio's are not chained,
so we don't need to always introduce this atomic operation as part of
ending IO.

Add a helper to elevate the bi_remaining count, and flag the bio as
now actually needing the decrement at end_io time. Rename the field
to __bi_remaining to catch any current users of this doing the
incrementing manually.

For high IOPS workloads, this reduces the overhead of bio_endio()
substantially.

Tested-by: Robert Elliott <elliott@hp.com>
Acked-by: Kent Overstreet <kent.overstreet@gmail.com>
Reviewed-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
9 years agoMerge git://git.kernel.org/pub/scm/linux/kernel/git/herbert/crypto-2.6
Linus Torvalds [Tue, 5 May 2015 16:03:52 +0000 (09:03 -0700)]
Merge git://git./linux/kernel/git/herbert/crypto-2.6

Pull crypto fixes from Herbert Xu:
 "This fixes a build problem with bcm63xx and yet another fix to the
  memzero_explicit function to ensure that the memset is not elided"

* git://git.kernel.org/pub/scm/linux/kernel/git/herbert/crypto-2.6:
  hwrng: bcm63xx - Fix driver compilation
  lib: make memzero_explicit more robust against dead store elimination

9 years agoMerge tag 'media/v4.1-3' of git://git.kernel.org/pub/scm/linux/kernel/git/mchehab...
Linus Torvalds [Tue, 5 May 2015 15:42:06 +0000 (08:42 -0700)]
Merge tag 'media/v4.1-3' of git://git./linux/kernel/git/mchehab/linux-media

Pull media fixes from Mauro Carvalho Chehab:
 "Three driver fixes:

   - fix for omap4, fixing a regression due to a subsystem API that got
     removed for 4.1 (commit efde234674d9);

   - fix for one of the formats supported by Marvel ccic driver;

   - fix rcar_vin driver that, when stopping abnormally, the driver
     can't return from wait_for_completion"

* tag 'media/v4.1-3' of git://git.kernel.org/pub/scm/linux/kernel/git/mchehab/linux-media:
  [media] v4l: omap4iss: Replace outdated OMAP4 control pad API with syscon
  [media] media: soc_camera: rcar_vin: Fix wait_for_completion
  [media] marvell-ccic: fix Y'CbCr ordering

9 years agohwrng: bcm63xx - Fix driver compilation
Álvaro Fernández Rojas [Sat, 2 May 2015 10:08:42 +0000 (12:08 +0200)]
hwrng: bcm63xx - Fix driver compilation

- s/clk_didsable_unprepare/clk_disable_unprepare
- s/prov/priv
- s/error/ret (bcm63xx_rng_probe)

Fixes: 6229c16060fe ("hwrng: bcm63xx - make use of devm_hwrng_register")
Signed-off-by: Álvaro Fernández Rojas <noltari@gmail.com>
Acked-by: Florian Fainelli <f.fainelli@gmail.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
9 years agolib: make memzero_explicit more robust against dead store elimination
Daniel Borkmann [Thu, 30 Apr 2015 02:13:52 +0000 (04:13 +0200)]
lib: make memzero_explicit more robust against dead store elimination

In commit 0b053c951829 ("lib: memzero_explicit: use barrier instead
of OPTIMIZER_HIDE_VAR"), we made memzero_explicit() more robust in
case LTO would decide to inline memzero_explicit() and eventually
find out it could be elimiated as dead store.

While using barrier() works well for the case of gcc, recent efforts
from LLVMLinux people suggest to use llvm as an alternative to gcc,
and there, Stephan found in a simple stand-alone user space example
that llvm could nevertheless optimize and thus elimitate the memset().
A similar issue has been observed in the referenced llvm bug report,
which is regarded as not-a-bug.

Based on some experiments, icc is a bit special on its own, while it
doesn't seem to eliminate the memset(), it could do so with an own
implementation, and then result in similar findings as with llvm.

The fix in this patch now works for all three compilers (also tested
with more aggressive optimization levels). Arguably, in the current
kernel tree it's more of a theoretical issue, but imho, it's better
to be pedantic about it.

It's clearly visible with gcc/llvm though, with the below code: if we
would have used barrier() only here, llvm would have omitted clearing,
not so with barrier_data() variant:

  static inline void memzero_explicit(void *s, size_t count)
  {
    memset(s, 0, count);
    barrier_data(s);
  }

  int main(void)
  {
    char buff[20];
    memzero_explicit(buff, sizeof(buff));
    return 0;
  }

  $ gcc -O2 test.c
  $ gdb a.out
  (gdb) disassemble main
  Dump of assembler code for function main:
   0x0000000000400400  <+0>: lea   -0x28(%rsp),%rax
   0x0000000000400405  <+5>: movq  $0x0,-0x28(%rsp)
   0x000000000040040e <+14>: movq  $0x0,-0x20(%rsp)
   0x0000000000400417 <+23>: movl  $0x0,-0x18(%rsp)
   0x000000000040041f <+31>: xor   %eax,%eax
   0x0000000000400421 <+33>: retq
  End of assembler dump.

  $ clang -O2 test.c
  $ gdb a.out
  (gdb) disassemble main
  Dump of assembler code for function main:
   0x00000000004004f0  <+0>: xorps  %xmm0,%xmm0
   0x00000000004004f3  <+3>: movaps %xmm0,-0x18(%rsp)
   0x00000000004004f8  <+8>: movl   $0x0,-0x8(%rsp)
   0x0000000000400500 <+16>: lea    -0x18(%rsp),%rax
   0x0000000000400505 <+21>: xor    %eax,%eax
   0x0000000000400507 <+23>: retq
  End of assembler dump.

As gcc, clang, but also icc defines __GNUC__, it's sufficient to define
this in compiler-gcc.h only to be picked up. For a fallback or otherwise
unsupported compiler, we define it as a barrier. Similarly, for ecc which
does not support gcc inline asm.

Reference: https://llvm.org/bugs/show_bug.cgi?id=15495
Reported-by: Stephan Mueller <smueller@chronox.de>
Tested-by: Stephan Mueller <smueller@chronox.de>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Stephan Mueller <smueller@chronox.de>
Cc: Hannes Frederic Sowa <hannes@stressinduktion.org>
Cc: mancha security <mancha1@zoho.com>
Cc: Mark Charlebois <charlebm@gmail.com>
Cc: Behan Webster <behanw@converseincode.com>
Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
9 years agoLinux 4.1-rc2
Linus Torvalds [Mon, 4 May 2015 02:22:23 +0000 (19:22 -0700)]
Linux 4.1-rc2

9 years agoMerge tag 'for_linus_stable' of git://git.kernel.org/pub/scm/linux/kernel/git/tytso...
Linus Torvalds [Mon, 4 May 2015 01:23:53 +0000 (18:23 -0700)]
Merge tag 'for_linus_stable' of git://git./linux/kernel/git/tytso/ext4

Pull ext4 fixes from Ted Ts'o:
 "Some miscellaneous bug fixes and some final on-disk and ABI changes
  for ext4 encryption which provide better security and performance"

* tag 'for_linus_stable' of git://git.kernel.org/pub/scm/linux/kernel/git/tytso/ext4:
  ext4: fix growing of tiny filesystems
  ext4: move check under lock scope to close a race.
  ext4: fix data corruption caused by unwritten and delayed extents
  ext4 crypto: remove duplicated encryption mode definitions
  ext4 crypto: do not select from EXT4_FS_ENCRYPTION
  ext4 crypto: add padding to filenames before encrypting
  ext4 crypto: simplify and speed up filename encryption

9 years agoMerge branch 'drm-fixes' of git://people.freedesktop.org/~airlied/linux
Linus Torvalds [Mon, 4 May 2015 01:15:48 +0000 (18:15 -0700)]
Merge branch 'drm-fixes' of git://people.freedesktop.org/~airlied/linux

Pull drm fixes from Dave Airlie:
 "One intel fix, one rockchip fix, and a bunch of radeon fixes for some
  regressions from audio rework and vm stability"

* 'drm-fixes' of git://people.freedesktop.org/~airlied/linux:
  drm/i915/chv: Implement WaDisableShadowRegForCpd
  drm/radeon: fix userptr return value checking (v2)
  drm/radeon: check new address before removing old one
  drm/radeon: reset BOs address after clearing it.
  drm/radeon: fix lockup when BOs aren't part of the VM on release
  drm/radeon: add SI DPM quirk for Sapphire R9 270 Dual-X 2G GDDR5
  drm/radeon: adjust pll when audio is not enabled
  drm/radeon: only enable audio streams if the monitor supports it
  drm/radeon: only mark audio as connected if the monitor supports it (v3)
  drm/radeon/audio: don't enable packets until the end
  drm/radeon: drop dce6_dp_enable
  drm/radeon: fix ordering of AVI packet setup
  drm/radeon: Use drm_calloc_ab for CS relocs
  drm/rockchip: fix error check when getting irq
  MAINTAINERS: add entry for Rockchip drm drivers

9 years agoMerge tag 'drm-intel-fixes-2015-04-30' of git://anongit.freedesktop.org/drm-intel...
Dave Airlie [Sun, 3 May 2015 22:56:47 +0000 (08:56 +1000)]
Merge tag 'drm-intel-fixes-2015-04-30' of git://anongit.freedesktop.org/drm-intel into drm-fixes

Just a single intel fix
* tag 'drm-intel-fixes-2015-04-30' of git://anongit.freedesktop.org/drm-intel:
  drm/i915/chv: Implement WaDisableShadowRegForCpd

9 years agoMerge branch 'drm-next0420' of https://github.com/markyzq/kernel-drm-rockchip into...
Dave Airlie [Sun, 3 May 2015 22:56:27 +0000 (08:56 +1000)]
Merge branch 'drm-next0420' of https://github.com/markyzq/kernel-drm-rockchip into drm-fixes

one fix and maintainers update
* 'drm-next0420' of https://github.com/markyzq/kernel-drm-rockchip:
  drm/rockchip: fix error check when getting irq
  MAINTAINERS: add entry for Rockchip drm drivers

9 years agoMerge tag 'scsi-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi
Linus Torvalds [Sun, 3 May 2015 20:22:32 +0000 (13:22 -0700)]
Merge tag 'scsi-fixes' of git://git./linux/kernel/git/jejb/scsi

Pull SCSI fixes from James Bottomley:
 "This is three logical fixes (as 5 patches).

  The 3ware class of drivers were causing an oops with multiqueue by
  tearing down the command mappings after completing the command (where
  the variables in the command used to tear down the mapping were
  no-longer valid).  There's also a fix for the qnap iscsi target which
  was choking on us sending it commands that were too long and a fix for
  the reworked aha1542 allocating GFP_KERNEL under a lock"

* tag 'scsi-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi:
  3w-9xxx: fix command completion race
  3w-xxxx: fix command completion race
  3w-sas: fix command completion race
  aha1542: Allocate memory before taking a lock
  SCSI: add 1024 max sectors black list flag

9 years agoMerge branch 'next' of git://git.infradead.org/users/vkoul/slave-dma
Linus Torvalds [Sun, 3 May 2015 17:49:04 +0000 (10:49 -0700)]
Merge branch 'next' of git://git.infradead.org/users/vkoul/slave-dma

Pull slave dmaengine fixes from Vinod Koul:
 "Here are the fixes in dmaengine subsystem for rc2:

   - privatecnt fix for slave dma request API by Christopher

   - warn fix for PM ifdef in usb-dmac by Geert

   - fix hardware dependency for xgene by Jean"

* 'next' of git://git.infradead.org/users/vkoul/slave-dma:
  dmaengine: increment privatecnt when using dma_get_any_slave_channel
  dmaengine: xgene: Set hardware dependency
  dmaengine: usb-dmac: Protect PM-only functions to kill warning

9 years agoMerge tag 'powerpc-4.1-3' of git://git.kernel.org/pub/scm/linux/kernel/git/mpe/linux
Linus Torvalds [Sun, 3 May 2015 17:28:36 +0000 (10:28 -0700)]
Merge tag 'powerpc-4.1-3' of git://git./linux/kernel/git/mpe/linux

Pull powerpc fixes from Michael Ellerman:
 - build fix for SMP=n in book3s_xics.c
 - fix for Daniel's pci_controller_ops on powernv.
 - revert the TM syscall abort patch for now.
 - CPU affinity fix from Nathan.
 - two EEH fixes from Gavin.
 - fix for CR corruption from Sam.
 - selftest build fix.

* tag 'powerpc-4.1-3' of git://git.kernel.org/pub/scm/linux/kernel/git/mpe/linux:
  powerpc/powernv: Restore non-volatile CRs after nap
  powerpc/eeh: Delay probing EEH device during hotplug
  powerpc/eeh: Fix race condition in pcibios_set_pcie_reset_state()
  powerpc/pseries: Correct cpu affinity for dlpar added cpus
  selftests/powerpc: Fix the pmu install rule
  Revert "powerpc/tm: Abort syscalls in active transactions"
  powerpc/powernv: Fix early pci_controller_ops loading.
  powerpc/kvm: Fix SMP=n build error in book3s_xics.c

9 years agoext4: fix growing of tiny filesystems
Jan Kara [Sun, 3 May 2015 03:58:32 +0000 (23:58 -0400)]
ext4: fix growing of tiny filesystems

The estimate of necessary transaction credits in ext4_flex_group_add()
is too pessimistic. It reserves credit for sb, resize inode, and resize
inode dindirect block for each group added in a flex group although they
are always the same block and thus it is enough to account them only
once. Also the number of modified GDT block is overestimated since we
fit EXT4_DESC_PER_BLOCK(sb) descriptors in one block.

Make the estimation more precise. That reduces number of requested
credits enough that we can grow 20 MB filesystem (which has 1 MB
journal, 79 reserved GDT blocks, and flex group size 16 by default).

Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Reviewed-by: Eric Sandeen <sandeen@redhat.com>
9 years agoext4: move check under lock scope to close a race.
Davide Italiano [Sun, 3 May 2015 03:21:15 +0000 (23:21 -0400)]
ext4: move check under lock scope to close a race.

fallocate() checks that the file is extent-based and returns
EOPNOTSUPP in case is not. Other tasks can convert from and to
indirect and extent so it's safe to check only after grabbing
the inode mutex.

Signed-off-by: Davide Italiano <dccitaliano@gmail.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Cc: stable@vger.kernel.org
9 years agoext4: fix data corruption caused by unwritten and delayed extents
Lukas Czerner [Sun, 3 May 2015 01:36:55 +0000 (21:36 -0400)]
ext4: fix data corruption caused by unwritten and delayed extents

Currently it is possible to lose whole file system block worth of data
when we hit the specific interaction with unwritten and delayed extents
in status extent tree.

The problem is that when we insert delayed extent into extent status
tree the only way to get rid of it is when we write out delayed buffer.
However there is a limitation in the extent status tree implementation
so that when inserting unwritten extent should there be even a single
delayed block the whole unwritten extent would be marked as delayed.

At this point, there is no way to get rid of the delayed extents,
because there are no delayed buffers to write out. So when a we write
into said unwritten extent we will convert it to written, but it still
remains delayed.

When we try to write into that block later ext4_da_map_blocks() will set
the buffer new and delayed and map it to invalid block which causes
the rest of the block to be zeroed loosing already written data.

For now we can fix this by simply not allowing to set delayed status on
written extent in the extent status tree. Also add WARN_ON() to make
sure that we notice if this happens in the future.

This problem can be easily reproduced by running the following xfs_io.

xfs_io -f -c "pwrite -S 0xaa 4096 2048" \
          -c "falloc 0 131072" \
          -c "pwrite -S 0xbb 65536 2048" \
          -c "fsync" /mnt/test/fff

echo 3 > /proc/sys/vm/drop_caches
xfs_io -c "pwrite -S 0xdd 67584 2048" /mnt/test/fff

This can be theoretically also reproduced by at random by running fsx,
but it's not very reliable, though on machines with bigger page size
(like ppc) this can be seen more often (especially xfstest generic/127)

Signed-off-by: Lukas Czerner <lczerner@redhat.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
Cc: stable@vger.kernel.org
9 years agoext4 crypto: remove duplicated encryption mode definitions
Chanho Park [Sat, 2 May 2015 14:29:22 +0000 (10:29 -0400)]
ext4 crypto: remove duplicated encryption mode definitions

This patch removes duplicated encryption modes which were already in
ext4.h. They were duplicated from commit 3edc18d and commit f542fb.

Cc: Theodore Ts'o <tytso@mit.edu>
Cc: Michael Halcrow <mhalcrow@google.com>
Cc: Andreas Dilger <adilger.kernel@dilger.ca>
Signed-off-by: Chanho Park <chanho61.park@samsung.com>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
9 years agoext4 crypto: do not select from EXT4_FS_ENCRYPTION
Herbert Xu [Sat, 2 May 2015 14:29:19 +0000 (10:29 -0400)]
ext4 crypto: do not select from EXT4_FS_ENCRYPTION

This patch adds a tristate EXT4_ENCRYPTION to do the selections
for EXT4_FS_ENCRYPTION because selecting from a bool causes all
the selected options to be built-in, even if EXT4 itself is a
module.

Signed-off-by: Herbert Xu <herbert@gondor.apana.org.au>
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
9 years agoMerge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net
Linus Torvalds [Sat, 2 May 2015 03:51:04 +0000 (20:51 -0700)]
Merge git://git./linux/kernel/git/davem/net

Pull networking fixes from David Miller:

 1) Receive packet length needs to be adjust by 2 on RX to accomodate
    the two padding bytes in altera_tse driver.  From Vlastimil Setka.

 2) If rx frame is dropped due to out of memory in macb driver, we leave
    the receive ring descriptors in an undefined state.  From Punnaiah
    Choudary Kalluri

 3) Some netlink subsystems erroneously signal NLM_F_MULTI.  That is
    only for dumps.  Fix from Nicolas Dichtel.

 4) Fix mis-use of raw rt->rt_pmtu value in ipv4, one must always go via
    the ipv4_mtu() helper.  From Herbert Xu.

 5) Fix null deref in bridge netfilter, and miscalculated lengths in
    jump/goto nf_tables verdicts.  From Florian Westphal.

 6) Unhash ping sockets properly.

 7) Software implementation of BPF divide did 64/32 rather than 64/64
    bit divide.  The JITs got it right.  Fix from Alexei Starovoitov.

* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net: (30 commits)
  ipv4: Missing sk_nulls_node_init() in ping_unhash().
  net: fec: Fix RGMII-ID mode
  net/mlx4_en: Schedule napi when RX buffers allocation fails
  netxen_nic: use spin_[un]lock_bh around tx_clean_lock
  net/mlx4_core: Fix unaligned accesses
  mlx4_en: Use correct loop cursor in error path.
  cxgb4: Fix MC1 memory offset calculation
  bnx2x: Delay during kdump load
  net: Fix Kernel Panic in bonding driver debugfs file: rlb_hash_table
  net: dsa: Fix scope of eeprom-length property
  net: macb: Fix race condition in driver when Rx frame is dropped
  hv_netvsc: Fix a bug in netvsc_start_xmit()
  altera_tse: Correct rx packet length
  mlx4: Fix tx ring affinity_mask creation
  tipc: fix problem with parallel link synchronization mechanism
  tipc: remove wrong use of NLM_F_MULTI
  bridge/nl: remove wrong use of NLM_F_MULTI
  bridge/mdb: remove wrong use of NLM_F_MULTI
  net: sched: act_connmark: don't zap skb->nfct
  trivial: net: systemport: bcmsysport.h: fix 0x0x prefix
  ...

9 years agovirtio: fix typo in vring_need_event() doc comment
Stefan Hajnoczi [Fri, 1 May 2015 23:12:29 +0000 (08:42 +0930)]
virtio: fix typo in vring_need_event() doc comment

Here the "other side" refers to the guest or host.

Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
9 years agovirtio: pass baton to Michael Tsirkin
Rusty Russell [Fri, 1 May 2015 23:12:38 +0000 (08:42 +0930)]
virtio: pass baton to Michael Tsirkin

With my job change kernel work will be "own time"; I'm keeping lguest
and modules (and the virtio standards work), but virtio kernel has to
go.

This makes it clear that Michael is in charge.  He's good, but having
me watch over his shoulder won't help.

Good luck Michael!

Signed-off-by: Rusty Russell <rusty@rustcorp.com.au>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
9 years agoMerge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/sage/ceph...
Linus Torvalds [Sat, 2 May 2015 03:35:39 +0000 (20:35 -0700)]
Merge branch 'for-linus' of git://git./linux/kernel/git/sage/ceph-client

Pull Ceph RBD fix from Sage Weil.

* 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/sage/ceph-client:
  rbd: end I/O the entire obj_request on error

9 years agoipv4: Missing sk_nulls_node_init() in ping_unhash().
David S. Miller [Sat, 2 May 2015 02:02:47 +0000 (22:02 -0400)]
ipv4: Missing sk_nulls_node_init() in ping_unhash().

If we don't do that, then the poison value is left in the ->pprev
backlink.

This can cause crashes if we do a disconnect, followed by a connect().

Tested-by: Linus Torvalds <torvalds@linux-foundation.org>
Reported-by: Wen Xu <hotdog3645@gmail.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
9 years agorbd: end I/O the entire obj_request on error
Ilya Dryomov [Sat, 25 Apr 2015 12:56:15 +0000 (15:56 +0300)]
rbd: end I/O the entire obj_request on error

When we end I/O struct request with error, we need to pass
obj_request->length as @nr_bytes so that the entire obj_request worth
of bytes is completed.  Otherwise block layer ends up confused and we
trip on

    rbd_assert(more ^ (which == img_request->obj_request_count));

in rbd_img_obj_callback() due to more being true no matter what.  We
already do it in most cases but we are missing some, in particular
those where we don't even get a chance to submit any obj_requests, due
to an early -ENOMEM for example.

A number of obj_request->xferred assignments seem to be redundant but
I haven't touched any of obj_request->xferred stuff to keep this small
and isolated.

Cc: Alex Elder <elder@linaro.org>
Cc: stable@vger.kernel.org # 3.10+
Reported-by: Shawn Edwards <lesser.evil@gmail.com>
Reviewed-by: Sage Weil <sage@redhat.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
9 years agoext4 crypto: add padding to filenames before encrypting
Theodore Ts'o [Fri, 1 May 2015 20:56:50 +0000 (16:56 -0400)]
ext4 crypto: add padding to filenames before encrypting

This obscures the length of the filenames, to decrease the amount of
information leakage.  By default, we pad the filenames to the next 4
byte boundaries.  This costs nothing, since the directory entries are
aligned to 4 byte boundaries anyway.  Filenames can also be padded to
8, 16, or 32 bytes, which will consume more directory space.

Change-Id: Ibb7a0fb76d2c48e2061240a709358ff40b14f322
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
9 years agoext4 crypto: simplify and speed up filename encryption
Theodore Ts'o [Fri, 1 May 2015 20:56:45 +0000 (16:56 -0400)]
ext4 crypto: simplify and speed up filename encryption

Avoid using SHA-1 when calculating the user-visible filename when the
encryption key is available, and avoid decrypting lots of filenames
when searching for a directory entry in a directory block.

Change-Id: If4655f144784978ba0305b597bfa1c8d7bb69e63
Signed-off-by: Theodore Ts'o <tytso@mit.edu>
9 years agoMerge branch 'for-linus-4.1' of git://git.kernel.org/pub/scm/linux/kernel/git/mason...
Linus Torvalds [Fri, 1 May 2015 14:46:21 +0000 (07:46 -0700)]
Merge branch 'for-linus-4.1' of git://git./linux/kernel/git/mason/linux-btrfs

Pull btrfs fixes from Chris Mason:
 "A few more btrfs fixes.

  These range from corners Filipe found in the new free space cache
  writeback to a grab bag of fixes from the list"

* 'for-linus-4.1' of git://git.kernel.org/pub/scm/linux/kernel/git/mason/linux-btrfs:
  Btrfs: btrfs_release_extent_buffer_page didn't free pages of dummy extent
  Btrfs: fill ->last_trans for delayed inode in btrfs_fill_inode.
  btrfs: unlock i_mutex after attempting to delete subvolume during send
  btrfs: check io_ctl_prepare_pages return in __btrfs_write_out_cache
  btrfs: fix race on ENOMEM in alloc_extent_buffer
  btrfs: handle ENOMEM in btrfs_alloc_tree_block
  Btrfs: fix find_free_dev_extent() malfunction in case device tree has hole
  Btrfs: don't check for delalloc_bytes in cache_save_setup
  Btrfs: fix deadlock when starting writeback of bg caches
  Btrfs: fix race between start dirty bg cache writeout and bg deletion

9 years agoMerge tag 'arm64-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/arm64/linux
Linus Torvalds [Fri, 1 May 2015 14:44:32 +0000 (07:44 -0700)]
Merge tag 'arm64-fixes' of git://git./linux/kernel/git/arm64/linux

Pull arm64 fixes from Will Deacon:
 "Not too much here, but we've addressed a couple of nasty issues in the
  dma-mapping code as well as adding the halfword and byte variants of
  load_acquire/store_release following on from the CSD locking bug that
  you fixed in the core.

   - fix perf devicetree warnings at probe time

   - fix memory leak in __dma_free()

   - ensure DMA buffers are always zeroed

   - show IRQ trigger in /proc/interrupts (for parity with ARM)

   - implement byte and halfword access for smp_{load_acquire,store_release}"

* tag 'arm64-fixes' of git://git.kernel.org/pub/scm/linux/kernel/git/arm64/linux:
  arm64: perf: Fix the pmu node name in warning message
  arm64: perf: don't warn about missing interrupt-affinity property for PPIs
  arm64: add missing PAGE_ALIGN() to __dma_free()
  arm64: dma-mapping: always clear allocated buffers
  ARM64: Enable CONFIG_GENERIC_IRQ_SHOW_LEVEL
  arm64: add missing data types in smp_load_acquire/smp_store_release

9 years agopowerpc/powernv: Restore non-volatile CRs after nap
Sam Bobroff [Fri, 1 May 2015 06:50:34 +0000 (16:50 +1000)]
powerpc/powernv: Restore non-volatile CRs after nap

Patches 7cba160ad "powernv/cpuidle: Redesign idle states management"
and 77b54e9f2 "powernv/powerpc: Add winkle support for offline cpus"
use non-volatile condition registers (cr2, cr3 and cr4) early in the system
reset interrupt handler (system_reset_pSeries()) before it has been determined
if state loss has occurred. If state loss has not occurred, control returns via
the power7_wakeup_noloss() path which does not restore those condition
registers, leaving them corrupted.

Fix this by restoring the condition registers in the power7_wakeup_noloss()
case.

This is apparent when running a KVM guest on hardware that does not
support winkle or sleep and the guest makes use of secondary threads. In
practice this means Power7 machines, though some early unreleased Power8
machines may also be susceptible.

The secondary CPUs are taken off line before the guest is started and
they call pnv_smp_cpu_kill_self(). This checks support for sleep
states (in this case there is no support) and power7_nap() is called.

When the CPU is woken, power7_nap() returns and because the CPU is
still off line, the main while loop executes again. The sleep states
support test is executed again, but because the tested values cannot
have changed, the compiler has optimized the test away and instead we
rely on the result of the first test, which has been left in cr3
and/or cr4. With the result overwritten, the wrong branch is taken and
power7_winkle() is called on a CPU that does not support it, leading
to it stalling.

Fixes: 7cba160ad789 ("powernv/cpuidle: Redesign idle states management")
Fixes: 77b54e9f213f ("powernv/powerpc: Add winkle support for offline cpus")
[mpe: Massage change log a bit more]
Signed-off-by: Sam Bobroff <sam.bobroff@au1.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
9 years agopowerpc/eeh: Delay probing EEH device during hotplug
Gavin Shan [Thu, 30 Apr 2015 23:22:15 +0000 (09:22 +1000)]
powerpc/eeh: Delay probing EEH device during hotplug

Commit 1c509148b ("powerpc/eeh: Do probe on pci_dn") probes EEH
devices in early stage, which is reasonable to pSeries platform.
However, it's wrong for PowerNV platform because the PE# isn't
determined until the resources (IO and MMIO) are assigned to
PE in hotplug case. So we have to delay probing EEH devices
for PowerNV platform until the PE# is assigned.

Fixes: ff57b454ddb9 ("powerpc/eeh: Do probe on pci_dn")
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
9 years agopowerpc/eeh: Fix race condition in pcibios_set_pcie_reset_state()
Gavin Shan [Thu, 30 Apr 2015 23:14:11 +0000 (09:14 +1000)]
powerpc/eeh: Fix race condition in pcibios_set_pcie_reset_state()

When asserting reset in pcibios_set_pcie_reset_state(), the PE
is enforced to (hardware) frozen state in order to drop unexpected
PCI transactions (except PCI config read/write) automatically by
hardware during reset, which would cause recursive EEH error.
However, the (software) frozen state EEH_PE_ISOLATED is missed.
When users get 0xFF from PCI config or MMIO read, EEH_PE_ISOLATED
is set in PE state retrival backend. Unfortunately, nobody (the
reset handler or the EEH recovery functinality in host) will clear
EEH_PE_ISOLATED when the PE has been passed through to guest.

The patch sets and clears EEH_PE_ISOLATED properly during reset
in function pcibios_set_pcie_reset_state() to fix the issue.

Fixes: 28158cd ("Enhance pcibios_set_pcie_reset_state()")
Reported-by: Carol L. Soto <clsoto@us.ibm.com>
Signed-off-by: Gavin Shan <gwshan@linux.vnet.ibm.com>
Tested-by: Carol L. Soto <clsoto@us.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
9 years agopowerpc/pseries: Correct cpu affinity for dlpar added cpus
Nathan Fontenot [Thu, 30 Apr 2015 01:42:06 +0000 (20:42 -0500)]
powerpc/pseries: Correct cpu affinity for dlpar added cpus

The incorrect ordering of operations during cpu dlpar add results in invalid
affinity for the cpu being added. The ibm,associativity property in the
device tree is populated with all zeroes for the added cpu which results in
invalid affinity mappings and all cpus appear to belong to node 0.

This occurs because rtas configure-connector is called prior to making the
rtas set-indicator calls. Phyp does not assign affinity information
for a cpu until the rtas set-indicator calls are made to set the isolation
and allocation state.

Correct the order of operations to make the rtas set-indicator
calls (done in dlpar_acquire_drc) before calling rtas configure-connector.

Fixes: 1a8061c46c46 ("powerpc/pseries: Add kernel based CPU DLPAR handling")

Signed-off-by: Nathan Fontenot <nfont@linux.vnet.ibm.com>
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
9 years agoselftests/powerpc: Fix the pmu install rule
Michael Ellerman [Fri, 1 May 2015 01:10:09 +0000 (11:10 +1000)]
selftests/powerpc: Fix the pmu install rule

My patch to add install support for the powerpc selftests had a typo,
leading to the three tests in the pmu directory itself not being
installed.

Fixes: 6faeeea44b84 ("selftests: Add install support for the powerpc tests")
Signed-off-by: Michael Ellerman <mpe@ellerman.id.au>
9 years agoMerge tag 'pm+acpi-4.1-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael...
Linus Torvalds [Thu, 30 Apr 2015 21:23:31 +0000 (14:23 -0700)]
Merge tag 'pm+acpi-4.1-rc2' of git://git./linux/kernel/git/rafael/linux-pm

Pull power management and ACPI fixes from Rafael Wysocki:
 "Three regression fixes this time, one for a recent regression in the
  cpuidle core affecting multiple systems, one for an inadvertently
  added duplicate typedef in ACPICA that breaks compilation with GCC 4.5
  and one for an ACPI Smart Battery Subsystem driver regression
  introduced during the 3.18 cycle (stable-candidate).

  Specifics:

   - Fix for a regression in the cpuidle core introduced by one of the
     recent commits in the clockevents_notify() removal series that put
     a call to a function which had to be executed with disabled
     interrupts into a code path running with enabled interrupts (Rafael
     J Wysocki)

   - Fix for a build problem in ACPICA (with GCC 4.5) introduced by one
     of the recent ACPICA tools commits that added a duplicate typedef
     to one of the ACPICA's header files by mistake (Olaf Hering)

   - Fix for a regression in the ACPI SBS (Smart Battery Subsystem)
     driver introduced during the 3.18 development cycle causing the
     smart battery manager to be marked as not present when it should be
     marked as present (Chris Bainbridge)"

* tag 'pm+acpi-4.1-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/rafael/linux-pm:
  cpuidle: Run tick_broadcast_exit() with disabled interrupts
  ACPI / SBS: Enable battery manager when present
  ACPICA: remove duplicate u8 typedef

9 years agoMerge tag 'sound-4.1-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/tiwai...
Linus Torvalds [Thu, 30 Apr 2015 21:00:18 +0000 (14:00 -0700)]
Merge tag 'sound-4.1-rc2' of git://git./linux/kernel/git/tiwai/sound

Pull sound fixes from Takashi Iwai:
 "One nice fix is Peter's patch to make the old good SB Audigy PCI to
  work with 32bit DMA instead of 31bit.  This allows the MIDI synth
  running on modern machines again.  Along with it, a few fixes for
  emu10k1 have merged.

  In ASoC side, there is one fix in the common code, but it's just
  trivial additions of static inline functions for CONFIG_PM=n.  The
  rest are various device-specific small fixes.

  Last but not least, a few HD-audio fixes are included, as usual, too"

* tag 'sound-4.1-rc2' of git://git.kernel.org/pub/scm/linux/kernel/git/tiwai/sound: (23 commits)
  ASoC: rt5677: fixed wrong DMIC ref clock
  ALSA: emu10k1: Emu10k2 32 bit DMA mode
  ALSA: emux: Fix mutex deadlock in OSS emulation
  ASoC: Update email-id of Rajeev Kumar
  ASoC: rt5645: Fix mask for setting RT5645_DMIC_2_DP_GPIO12 bit
  ALSA: hda - Fix missing va_end() call in snd_hda_codec_pcm_new()
  ALSA: emux: Fix mutex deadlock at unloading
  ALSA: emu10k1: Fix card shortname string buffer overflow
  ALSA: hda - Add mute-LED mode control to Thinkpad
  ALSA: hda - Fix mute-LED fixed mode
  ALSA: hda - Fix click noise at start on Dell XPS13
  ASoC: rt5645: Add ACPI match ID
  ASoC: rt5677: add register patch for PLL
  ASoC: Intel: fix the makefile for atom code
  ASoC: dapm: Enable autodisable on SOC_DAPM_SINGLE_TLV_AUTODISABLE
  ASoC: add static inline funcs to fix a compiling issue
  ASoC: Intel: sst_byt: remove kfree for memory allocated with devm_kzalloc
  ASoC: samsung: s3c24xx-i2s: Fix return value check in s3c24xx_iis_dev_probe()
  ASoC: tfa9879: Fix return value check in tfa9879_i2c_probe()
  ASoC: fsl_ssi: Fix platform_get_irq() error handling
  ...

9 years agonet: fec: Fix RGMII-ID mode
Markus Pargmann [Thu, 30 Apr 2015 15:07:50 +0000 (17:07 +0200)]
net: fec: Fix RGMII-ID mode

RGMII-ID uses an internal delay within the transmitter or receiver. This
feature is phy specific. The rest of the communication is normal RGMII.

So the fec driver has to check for all RGMII modes, not only
'PHY_INTERFACE_MODE_RGMII'.

Signed-off-by: Markus Pargmann <mpa@pengutronix.de>
Signed-off-by: David S. Miller <davem@davemloft.net>
9 years agonet/mlx4_en: Schedule napi when RX buffers allocation fails
Ido Shamay [Thu, 30 Apr 2015 14:32:46 +0000 (17:32 +0300)]
net/mlx4_en: Schedule napi when RX buffers allocation fails

When system is out of memory, refilling of RX buffers fails while
the driver continue to pass the received packets to the kernel stack.
At some point, when all RX buffers deplete, driver may fall into a
sleep, and not recover when memory for new RX buffers is once again
availible. This is because hardware does not have valid descriptors,
so no interrupt will be generated for the driver to return to work
in napi context. Fix it by schedule the napi poll function from
stats_task delayed workqueue, as long as the allocations fail.

Signed-off-by: Ido Shamay <idos@mellanox.com>
Signed-off-by: Amir Vadai <amirv@mellanox.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
9 years agonetxen_nic: use spin_[un]lock_bh around tx_clean_lock
Tony Camuso [Thu, 30 Apr 2015 11:51:27 +0000 (07:51 -0400)]
netxen_nic: use spin_[un]lock_bh around tx_clean_lock

While testing this driver with DEBUG_LOCKDEP and DEBUG_SPINLOCK
enabled did not produce any traces, it would be more prudent in the
case of tx_clean_lock to use spin_[un]lock_bh, since this lock is
manipulated in both the process and softirq contexts.

This patch was tested for functionality and regressions with netperf
and DEBUG_LOCKDEP and DEBUG_SPINLOCK enabled.

Signed-off-by: Tony Camuso <tcamuso@redhat.com>
Acked-by: Neil Horman <nhorman@tuxdriver.com>
Signed-off-by: David S. Miller <davem@davemloft.net>