platform/kernel/linux-rpi3.git
8 years agonvme/pci: Provide SR-IOV support
Keith Busch [Mon, 20 Jun 2016 15:41:06 +0000 (09:41 -0600)]
nvme/pci: Provide SR-IOV support

This registers an sr-iov callback for nvme.

Signed-off-by: Keith Busch <keith.busch@intel.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agonvme: initialize variable before logical OR'ing it
Jay Freyensee [Thu, 21 Jul 2016 03:26:16 +0000 (21:26 -0600)]
nvme: initialize variable before logical OR'ing it

It is typically not good coding or secure coding practice
to logical OR a variable without an initialization value first.
Here on this line:

integrity.flags |= BLK_INTEGRITY_DEVICE_CAPABLE;

BLK_INTEGRITY_DEVICE_CAPABLE is being OR'ed to a member variable
never set to an initial value. This patch fixes that.

Signed-off-by: Jay Freyensee <james.p.freyensee@intel.com>
Reviewed-by: Ming Lin <ming.l@samsung.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agoblock: unexport various bio mapping helpers
Christoph Hellwig [Tue, 19 Jul 2016 09:31:54 +0000 (11:31 +0200)]
block: unexport various bio mapping helpers

They are unused and potential new users really should use the
blk_rq_map* versions.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agoscsi/osd: open code blk_make_request
Christoph Hellwig [Tue, 19 Jul 2016 09:31:53 +0000 (11:31 +0200)]
scsi/osd: open code blk_make_request

I wish the OSD code could simply use blk_rq_map_* helpers like
everyone else, but the complex nature of deciding if we have
DATA IN and/or DATA OUT buffers might make this impossible
(at least for a mere human like me).

But using blk_rq_append_bio at least allows sharing the setup code
between request with or without dat a buffers, and given that this
is the last user of blk_make_request it allows getting rid of that
somewhat awkward interface.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Boaz Harrosh <ooo@electrozaur.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agotarget: stop using blk_make_request
Christoph Hellwig [Tue, 19 Jul 2016 09:31:52 +0000 (11:31 +0200)]
target: stop using blk_make_request

Using blk_rq_append_bio allows to append the bios to the request
directly instead of having to build up a list first, and also
allows to have a single code path for requests with or without
data attached to them.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agoblock: simplify and export blk_rq_append_bio
Christoph Hellwig [Tue, 19 Jul 2016 09:31:51 +0000 (11:31 +0200)]
block: simplify and export blk_rq_append_bio

The target SCSI passthrough backend is much better served with the low-level
blk_rq_append_bio construct then the helpers built on top of it, so export it.

Also use the opportunity to remove the pointless request_queue argument and
make the code flow a little more readable.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agoblock: ensure bios return from blk_get_request are properly initialized
Christoph Hellwig [Tue, 19 Jul 2016 09:31:50 +0000 (11:31 +0200)]
block: ensure bios return from blk_get_request are properly initialized

blk_get_request is used for BLOCK_PC and similar passthrough requests.
Currently we always need to call blk_rq_set_block_pc or an open coded
version of it to allow appending bios using the request mapping helpers
later on, which is a somewhat awkward API.  Instead move the
initialization part of blk_rq_set_block_pc into blk_get_request, so that
we always have a safe to use request.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agovirtio_blk: use blk_rq_map_kern
Christoph Hellwig [Tue, 19 Jul 2016 09:31:49 +0000 (11:31 +0200)]
virtio_blk: use blk_rq_map_kern

Similar to how SCSI and NVMe prepare passthrough requests.  This avoids
poking into request internals too much.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agomemstick: don't allow REQ_TYPE_BLOCK_PC requests
Christoph Hellwig [Tue, 19 Jul 2016 09:31:48 +0000 (11:31 +0200)]
memstick: don't allow REQ_TYPE_BLOCK_PC requests

There is no code to issue or handle REQ_TYPE_BLOCK_PC request in the
memstick drivers, so remove the bogus conditional.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agoblock: shrink bio size again
Christoph Hellwig [Tue, 19 Jul 2016 09:28:43 +0000 (11:28 +0200)]
block: shrink bio size again

The recent ops split grew the bio by adding the new ioprio field.
Shrink it again by using a 16-bit field for the bi_flags value and
filling the holes near the beginning of the structure.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Mike Christie <mchristi@redhat.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agoblock: simplify and cleanup bvec pool handling
Christoph Hellwig [Tue, 19 Jul 2016 09:28:42 +0000 (11:28 +0200)]
block: simplify and cleanup bvec pool handling

Instead of a flag and an index just make sure an index of 0 means
no need to free the bvec array.  Also move the constants related
to the bvec pools together and use a consistent naming scheme for
them.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Mike Christie <mchristi@redhat.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agoblock: get rid of bio_rw and READA
Christoph Hellwig [Tue, 19 Jul 2016 09:28:41 +0000 (11:28 +0200)]
block: get rid of bio_rw and READA

These two are confusing leftover of the old world order, combining
values of the REQ_OP_ and REQ_ namespaces.  For callers that don't
special case we mostly just replace bi_rw with bio_data_dir or
op_is_write, except for the few cases where a switch over the REQ_OP_
values makes more sense.  Any check for READA is replaced with an
explicit check for REQ_RAHEAD.  Also remove the READA alias for
REQ_RAHEAD.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Mike Christie <mchristi@redhat.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agoblock: don't ignore -EOPNOTSUPP blkdev_issue_write_same
Christoph Hellwig [Tue, 19 Jul 2016 09:23:34 +0000 (11:23 +0200)]
block: don't ignore -EOPNOTSUPP blkdev_issue_write_same

WRITE SAME is a data integrity operation and we can't simply ignore
errors.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agoblock: introduce BLKDEV_DISCARD_ZERO to fix zeroout
Christoph Hellwig [Tue, 19 Jul 2016 09:23:33 +0000 (11:23 +0200)]
block: introduce BLKDEV_DISCARD_ZERO to fix zeroout

Currently blkdev_issue_zeroout cascades down from discards (if the driver
guarantees that discards zero data), to WRITE SAME and then to a loop
writing zeroes.  Unfortunately we ignore run-time EOPNOTSUPP errors in the
block layer blkdev_issue_discard helper to work around DM volumes that
may have mixed discard support underneath.

This patch intoroduces a new BLKDEV_DISCARD_ZERO flag to
blkdev_issue_discard that indicates we are called for zeroing operation.
This allows both to ignore the EOPNOTSUPP hack and actually consolidating
the discard_zeroes_data check into the function.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agoNVMe: don't allocate unused nvme_major
NeilBrown [Wed, 13 Jul 2016 18:03:58 +0000 (11:03 -0700)]
NVMe: don't allocate unused nvme_major

When alloc_disk(0) is used, the ->major number is ignored.  All device
numbers are allocated with a major of BLOCK_EXT_MAJOR.

So remove all references to nvme_major.

[akpm@linux-foundation.org: one unregister_blkdev() was missed]
Link: http://lkml.kernel.org/r/20160602064318.4403.93301.stgit@noble
Signed-off-by: NeilBrown <neilb@suse.com>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Cc: Jens Axboe <axboe@kernel.dk>
Cc: Maxim Levitsky <maximlevitsky@gmail.com>
Cc: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agonvme: avoid crashes when node 0 is memoryless node.
Masayoshi Mizuma [Mon, 20 Jun 2016 00:33:17 +0000 (09:33 +0900)]
nvme: avoid crashes when node 0 is memoryless node.

When CONFIG_NUMA is enabled and node 0 is memoryless, the system
crashes because nvme_probe() sets the device->numa_node to 0 by
set_dev_node(&pdev->dev, 0), so it tries to allocate memory from node 0.
To avoid the crash, we should change the 0 to first_memory_node.

Signed-off-by: Masayoshi Mizuma <m.mizuma@jp.fujitsu.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agonvme: Limit command retries
Keith Busch [Tue, 12 Jul 2016 23:20:31 +0000 (16:20 -0700)]
nvme: Limit command retries

Many controller implementations will return errors to commands that will
not succeed, but without the DNR bit set. The driver previously retried
these commands an unlimited number of times until the command timeout
has exceeded, which takes an unnecessarilly long period of time.

This patch limits the number of retries a command can have, defaulting
to 5, but is user tunable at load or runtime.

The struct request's 'retries' field is used to track the number of
retries attempted. This is in contrast with scsi's use of this field,
which indicates how many retries are allowed.

Signed-off-by: Keith Busch <keith.busch@intel.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agoloop: Make user notify for adding loop device failed
Minfei Huang [Tue, 7 Jun 2016 02:05:15 +0000 (10:05 +0800)]
loop: Make user notify for adding loop device failed

There is no error number returned if loop driver fails in function
alloc_disk to add new loop device. Add a correct error number to make
user notify in this case.

Signed-off-by: Minfei Huang <mnghuan@gmail.com>
Reviewed-by: Ming Lei <tom.leiming@gmail.com>
Reviewed-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agonvme-loop: fix nvme-loop Kconfig dependencies
Arnd Bergmann [Tue, 12 Jul 2016 15:36:40 +0000 (08:36 -0700)]
nvme-loop: fix nvme-loop Kconfig dependencies

I ran into the same problem on NVME_TARGET_RDMA now,
which otherwise needs dependencies on both CONFIG_BLOCK and
CONFIGFS_FS:

    warning: (NVME_TARGET_LOOP && NVME_TARGET_RDMA) selects NVME_TARGET which has unmet direct dependencies (BLOCK && CONFIGFS_FS)
    0xA002B368 Mon Jul 11 18:00:45 CEST 2016 failed
    In file included from ../drivers/nvme/target/core.c:16:0:
    drivers/nvme/target/nvmet.h:222:14: error: field 'inline_bio' has incomplete type
      struct bio  inline_bio;
                  ^~~~~~~~~~
    drivers/nvme/target/core.c: In function 'nvmet_async_event_work':
    drivers/nvme/target/core.c:98:3: error: implicit declaration of function 'kfree' [-Werror=implicit-function-declaration]
       kfree(aen);
       ^~~~~
    ../drivers/nvme/target/core.c: In function 'nvmet_ns_enable':
    ../drivers/nvme/target/core.c:269:13: error: implicit declaration of function 'blkdev_get_by_path' [-Werror=implicit-function-declaration]
      ns->bdev = blkdev_get_by_path(ns->device_path, FMODE_READ | FMODE_WRITE,

Folding in my patch below should address that too.

Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Reported-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agonvmet: fix return value check in nvmet_subsys_alloc()
Wei Yongjun [Wed, 6 Jul 2016 12:02:09 +0000 (12:02 +0000)]
nvmet: fix return value check in nvmet_subsys_alloc()

In case of error, the function kstrndup() returns NULL pointer
not ERR_PTR(). The IS_ERR() test in the return value check
should be replaced with NULL test.

Signed-off-by: Wei Yongjun <yongjun_wei@trendmicro.com.cn>
Reviewed-by: Jay Freyensee <james_p_freyensee@linux.intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agonvme-fabrics: add-remove ctrl repeat fix
Ming Lin [Fri, 1 Jul 2016 19:13:32 +0000 (12:13 -0700)]
nvme-fabrics: add-remove ctrl repeat fix

Repeatedly adding then removing the same NVMe-over-Fabrics controller
over and over again (shown below) can cause a kernel crash (also shown
below).  This patch fixes that.

[nvmf]# ./setup_nvme_connections.sh
traddr=192.168.1.100,transport=rdma,trsvcid=4420,nqn=darkside
-nqn,hostnqn=evil-wins-nqn,nr_io_queues=16 > /dev/nvme-fabrics
traddr=192.168.1.100,transport=rdma,trsvcid=4420,nqn=lightside
-nqn,hostnqn=good-wins-nqn > /dev/nvme-fabrics
[nvmf]# ./remove_nvme_connections.sh 2
echo 1 > /sys/class/nvme/nvme0/delete_controller
echo 1 > /sys/class/nvme/nvme1/delete_controller
[nvmf]# ./setup_nvme_connections.sh
traddr=192.168.1.100,transport=rdma,trsvcid=4420,nqn=darkside
-nqn,hostnqn=evil-wins-nqn,nr_io_queues=16 > /dev/nvme-fabrics
Killed

[nvmf]# dmesg
[  313.416908] nvme nvme0: creating 16 I/O queues.
[  313.523908] nvme nvme0: new ctrl: NQN "darkside-nqn", addr
192.168.1.100:4420
[  313.524857] BUG: unable to handle kernel NULL pointer dereference at
0000000000000010
[  313.525262] IP: [<ffffffff8136c60e>] strcmp+0xe/0x30
[  313.525490] PGD 0
[  313.525726] Oops: 0000 [#1] SMP
[  313.525900] Modules linked in: nvme_rdma nvme_fabrics nvme_core
ib_ipoib rdma_ucm ib_ucm ib_uverbs ib_umad rdma_cm ib_cm iw_cm mlx4_en
mlx4_ib ib_core mlx4_core
[  313.527085] CPU: 15 PID: 5856 Comm: setup_nvme_conn Not tainted
4.7.0-rc2+ #2
[  313.527259] Hardware name: Supermicro X9DRT-F/IBQF/IBFF/X9DRT
-F/IBQF/IBFF, BIOS 1.0a 10/09/2012
[  313.527551] task: ffff88027646cd40 ti: ffff88025b980000 task.ti:
ffff88025b980000
[  313.527879] RIP: 0010:[<ffffffff8136c60e>]  [<ffffffff8136c60e>]
strcmp+0xe/0x30
[  313.528232] RSP: 0018:ffff88025b983db0  EFLAGS: 00010206
[  313.528403] RAX: 0000000000000000 RBX: ffff880471879880 RCX:
fffffffffffffff1
[  313.528594] RDX: 0000000000000000 RSI: ffff880474afa860 RDI:
0000000000000011
[  313.528778] RBP: ffff88025b983db0 R08: ffff880474afa860 R09:
ffff880471879058
[  313.528956] R10: 000000000000002c R11: ffff88047f415000 R12:
ffff880471879800
[  313.529129] R13: ffff880471879000 R14: ffff880474afa860 R15:
fffffffffffffff8
[  313.529303] FS:  00007f778f510700(0000) GS:ffff88047fbc0000(0000)
knlGS:0000000000000000
[  313.529629] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[  313.529817] CR2: 0000000000000010 CR3: 0000000274174000 CR4:
00000000000406e0
[  313.529989] Stack:
[  313.530154]  ffff88025b983e48 ffffffffa0171c74 0000000000000001
0000000000000059
[  313.530621]  ffff880476f32400 ffff88047e8add80 0000010074b33aa0
ffff880471879059
[  313.531162]  ffff88047187904b ffff880471879058 0000000000000000
ffff88047736e000
[  313.531629] Call Trace:
[  313.531797]  [<ffffffffa0171c74>] nvmf_dev_write+0x674/0x840
[nvme_fabrics]
[  313.531974]  [<ffffffff81180b53>] __vfs_write+0x23/0x120
[  313.532146]  [<ffffffff8119daff>] ? __fd_install+0x1f/0xc0
[  313.532316]  [<ffffffff8119d97a>] ? __alloc_fd+0x3a/0x170
[  313.532487]  [<ffffffff811811f3>] vfs_write+0xb3/0x1b0
[  313.532658]  [<ffffffff8117e321>] ? filp_close+0x51/0x70
[  313.532845]  [<ffffffff811824e1>] SyS_write+0x41/0xa0
[  313.533016]  [<ffffffff8183055b>]
entry_SYSCALL_64_fastpath+0x13/0x8f
[  313.533188] Code: 80 3a 00 75 f7 48 83 c6 01 0f b6 4e ff 48 83 c2 01
84 c9 88 4a ff 75 ed 5d c3 0f 1f 00 55 48 89 e5 eb 04 84 c0 74 18 48 83
c7 01 <0f> b6 47 ff 48 83 c6 01 3a 46 ff 74 eb 19 c0 83 c8 01 5d c3 31
[  313.536563] RIP  [<ffffffff8136c60e>] strcmp+0xe/0x30
[  313.536815]  RSP <ffff88025b983db0>
[  313.536981] CR2: 0000000000000010
[  313.537151] ---[ end trace 3d952e590e7bc2d5 ]---

Reported-and-tested-by: Jay Freyensee <james.p.freyensee@intel.com>
Signed-off-by: Ming Lin <mlin@kernel.org>
Signed-off-by: Jay Freyensee <james.p.freyensee@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agonvme-fabrics: Remove tl_retry_count
Sagi Grimberg [Wed, 22 Jun 2016 12:06:01 +0000 (15:06 +0300)]
nvme-fabrics: Remove tl_retry_count

The timeout before error recovery logic kicks in is
dictated by the nvme keep-alive, so we don't really need
a transport layer retry count. transports can retry for
as much as they like.

Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agonvme-rdma: Don't use tl_retry_count
Sagi Grimberg [Wed, 22 Jun 2016 12:06:00 +0000 (15:06 +0300)]
nvme-rdma: Don't use tl_retry_count

Always use the maximum qp retry count as the
error recovery timeout is dictated from the nvme
keep-alive.

Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agonvme-rdma: fix the return value of nvme_rdma_reinit_request()
Wei Yongjun [Tue, 12 Jul 2016 11:06:17 +0000 (11:06 +0000)]
nvme-rdma: fix the return value of nvme_rdma_reinit_request()

PTR_ERR should be applied before its argument is reassigned, otherwise the
return value will be set to 0, not error code.

Signed-off-by: Wei Yongjun <yongjun_wei@trendmicro.com.cn>
Reviewed-by: Jay Freyensee <james_p_freyensee@linux.intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agocdrom: support read sub-channel command in LBA format
vchannaiah [Wed, 29 Jun 2016 12:18:25 +0000 (08:18 -0400)]
cdrom: support read sub-channel command in LBA format

userspace application can send READ_SUB_CHANNEL command with time bit
enabled and disabled. The time bit allows selection of address reporting
format. If the time bit is disabled the response is in logical block
address(CDROM_LBA) format, represented as a 32-bit integer with ms-byte
first. If the time bit is enabled the response is in time format i.e.,
minutes, second, frame (CDROM_MSF) format.

Signed-off-by: vchannaiah <vanitha.channaiah@in.bosch.com>
Signed-off-by: Mahendran Kuppusamy <mahendran.kuppusamy@in.bosch.com>
[veeraiyan.chidambaram@in.bosch.com: updated Documentation/ioctl/cdrom.txt]
Signed-off-by: Veeraiyan Chidambaram <veeraiyan.chidambaram@in.bosch.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agonvme/quirk: Add a delay before checking for adapter readiness
Guilherme G. Piccoli [Tue, 14 Jun 2016 21:22:41 +0000 (18:22 -0300)]
nvme/quirk: Add a delay before checking for adapter readiness

When disabling the controller, the specification says the register
NVME_REG_CC should be written and then driver needs to wait the
adapter to be ready, which is checked by reading another register
bit (NVME_CSTS_RDY). There's a timeout validation in this checking,
so in case this timeout is reached the driver gives up and removes
the adapter from the system.

After a firmware activation procedure, the PCI_DEVICE(0x1c58, 0x0003)
(HGST adapter) end up being removed if we issue a reset_controller,
because driver keeps verifying the NVME_REG_CSTS until the timeout is
reached. This patch adds a necessary quirk for this adapter, by
introducing a delay before nvme_wait_ready(), so the reset procedure
is able to be completed. This quirk is needed because just increasing
the timeout is not enough in case of this adapter - the driver must
wait before start reading NVME_REG_CSTS register on this specific
device.

Signed-off-by: Guilherme G. Piccoli <gpiccoli@linux.vnet.ibm.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agoMerge branch 'for-4.8/block' of git://git.kernel.org/pub/scm/linux/kernel/git/nvdimm...
Jens Axboe [Fri, 8 Jul 2016 22:04:11 +0000 (16:04 -0600)]
Merge branch 'for-4.8/block' of git://git./linux/kernel/git/nvdimm/nvdimm into for-4.8/drivers

Dan writes:

"The removal of ->driverfs_dev in favor of just passing the parent
device in as a parameter to add_disk().  See below, it has received a
"Reviewed-by" from Christoph, Bart, and Johannes.

It is also a pre-requisite for Fam Zheng's work to cleanup gendisk
uevents vs attribute visibility [1].  We would extend device_add_disk()
to take an attribute_group list.

This is based off a branch of block.git/for-4.8/drivers and has
received a positive build success notification from the kbuild robot
across several configs.

[1]: "gendisk: Generate uevent after attribute available"
http://marc.info/?l=linux-virtualization&m=146725201522201&w=2"

8 years agonvme-rdma: add a NVMe over Fabrics RDMA host driver
Christoph Hellwig [Wed, 6 Jul 2016 12:55:52 +0000 (21:55 +0900)]
nvme-rdma: add a NVMe over Fabrics RDMA host driver

This patch implements the RDMA host (initiator in SCSI speak) driver.  It
can be used to connect to remote NVMe over Fabrics controllers over
Infiniband, RoCE or iWarp, and uses the existing NVMe core driver as well
a the new fabrics library.

To connect to all NVMe over Fabrics controller reachable on a given taget
port using RDMA/CM use the following command:

nvme connect-all -t rdma -a $IPADDR

This requires the latest version of nvme-cli with Fabrics support.

Signed-off-by: Jay Freyensee <james.p.freyensee@intel.com>
Signed-off-by: Ming Lin <ming.l@ssi.samsung.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agonvmet-rdma: add a NVMe over Fabrics RDMA target driver
Christoph Hellwig [Wed, 6 Jul 2016 12:55:51 +0000 (21:55 +0900)]
nvmet-rdma: add a NVMe over Fabrics RDMA target driver

This patch implements the RDMA transport for the NVMe over Fabrics target,
which allows exporting NVMe over Fabrics functionality over RDMA fabrics
(Infiniband, RoCE, iWARP).

All NVMe logic is in the generic target and this module just provides a
small glue between it and the generic code in the RDMA subsystem.

Signed-off-by: Armen Baloyan <armenx.baloyan@intel.com>,
Signed-off-by: Jay Freyensee <james.p.freyensee@intel.com>
Signed-off-by: Ming Lin <ming.l@ssi.samsung.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agonvme-rdma.h: Add includes for nvme rdma_cm negotiation
Sagi Grimberg [Wed, 6 Jul 2016 12:55:50 +0000 (21:55 +0900)]
nvme-rdma.h: Add includes for nvme rdma_cm negotiation

NVMe over Fabrics RDMA transport defines a connection establishment
protocol over the RDMA connection manager. This header will be used by
both the host and target drivers to negotiate the connection
establishment parameters.

Signed-off-by: Jay Freyensee <james.p.freyensee@intel.com>
Signed-off-by: Ming Lin <ming.l@ssi.samsung.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agonvme: add new reconnecting controller state
Christoph Hellwig [Wed, 6 Jul 2016 12:55:49 +0000 (21:55 +0900)]
nvme: add new reconnecting controller state

The nvme fabric (RDMA, FC, etc...) can introduce port, link or node
failures that may require a reconnect to re-establish the connection.

Add a new reconnecting state that will initially be used by the RDMA
driver.

Reviewed-by: Jay Freyensee <james.p.freyensee@intel.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agoblk-mq: Introduce blk_mq_reinit_tagset
Sagi Grimberg [Wed, 6 Jul 2016 12:55:48 +0000 (21:55 +0900)]
blk-mq: Introduce blk_mq_reinit_tagset

The new nvme-rdma driver will need to reinitialize all the tags as part of
the error recovery procedure (realloc the tag memory region). Add a helper
in blk-mq for it that can iterate over all requests in a tagset to make
this easier.

Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Tested-by: Ming Lin <ming.l@ssi.samsung.com>
Reviewed-by: Stephen Bates <Stephen.Bates@pmcs.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>
Tested-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agolightnvm: make __nvm_submit_ppa static
Matias Bjørling [Thu, 7 Jul 2016 07:54:23 +0000 (09:54 +0200)]
lightnvm: make __nvm_submit_ppa static

The __nvm_submit_ppa() function is not used outside lightnvm core.

Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agolightnvm: make ppa_list const in nvm_set_rqd_list
Matias Bjørling [Thu, 7 Jul 2016 07:54:22 +0000 (09:54 +0200)]
lightnvm: make ppa_list const in nvm_set_rqd_list

The passed by reference ppa list in nvm_set_rqd_list() is updated when
multiple planes are available. In that case, each PPA plane is
incremented when the device side PPA list is created. This prevents the
caller to rely on the PPA list to be unmodified after a call.

Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agolightnvm: fix lun offset calculation for mark blk
Matias Bjørling [Thu, 7 Jul 2016 07:54:21 +0000 (09:54 +0200)]
lightnvm: fix lun offset calculation for mark blk

The gen_mark_blk_bad function marks the wrong block when a block is on
a different channel. Fix the index calculation, so that it updates the
correct block.

Reported-by: Javier Gonzalez <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agolightnvm: make rrpc_map_page call nvm_get_blk outside locks
Matias Bjørling [Thu, 7 Jul 2016 07:54:20 +0000 (09:54 +0200)]
lightnvm: make rrpc_map_page call nvm_get_blk outside locks

The nvm_get_blk() function is called with rlun->lock held. This is ok
when the media manager implementation doesn't go out of its atomic
context. However, if a media manager persists its metadata, and
guarantees that the block is given to the target, this is no longer
a viable approach. Therefore, clean up the flow of rrpc_map_page,
and make sure that nvm_get_blk() is called without any locks acquired.

Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agolightnvm: remove _unlocked variant of [get/put]_blk
Matias Bjørling [Thu, 7 Jul 2016 07:54:19 +0000 (09:54 +0200)]
lightnvm: remove _unlocked variant of [get/put]_blk

The [get/put]_blk API enables targets to get ownership of blocks at
runtime. This information is currently not recorded on disk, and the
information is therefore lost on power failure. To restore the
metadata, the [get/put]_blk must persist its metadata. In that case,
we need to control the outer lock, so that we can disable them while
updating the on-disk metadata. Fortunately, the _unlocked versions can
be removed, which allows us to move the lock into the [get/put]_blk
functions.

Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agolightnvm: remove unused lists from struct rrpc_block
Matias Bjørling [Thu, 7 Jul 2016 07:54:18 +0000 (09:54 +0200)]
lightnvm: remove unused lists from struct rrpc_block

The ->list, ->open_list, and ->closed_list lists were previously used
for statistics. However, their usage have been removed, and thus these
can safely be removed.

Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agolightnvm: remove nested lock conflict with mm
Matias Bjørling [Thu, 7 Jul 2016 07:54:17 +0000 (09:54 +0200)]
lightnvm: remove nested lock conflict with mm

If a media manager tries to initialize it targets upon media manager
initialization, the media manager will need to know which target types
are available in LightNVM. The lists of which managers and target types
are available shares the same lock.

Therefore, on initialization, the nvm_lock is taken by LightNVM core,
which later leads to a deadlock when target types are enumerated by the
media manager.

Add an exclusive lock for target types to resolve this conflict.

Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agolightnvm: move target mgmt into media mgr
Matias Bjørling [Thu, 7 Jul 2016 07:54:16 +0000 (09:54 +0200)]
lightnvm: move target mgmt into media mgr

To enable persistent block management to easily control creation and
removal of targets, we move target management into the media
manager. The LightNVM core continues to maintain which target types are
registered, while the media manager now keeps track of its initialized
targets.

Two new callbacks for the media manager are introduced. create_tgt and
remove_tgt. Note that remove_tgt returns 0 on successfully removing a
target, and returns 1 if the target was not found.

Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agolightnvm: rename gennvm and update description
Matias Bjørling [Thu, 7 Jul 2016 07:54:15 +0000 (09:54 +0200)]
lightnvm: rename gennvm and update description

The generic manager should be called the general media manager, and
instead of using the rather long name of "gennvm" in front of each data
structures, use "gen" instead to shorten it. Update the description of
the media manager as well to make the media manager purpose clearer.

Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agolightnvm: remove open/close statistics for gennvm
Matias Bjørling [Thu, 7 Jul 2016 07:54:14 +0000 (09:54 +0200)]
lightnvm: remove open/close statistics for gennvm

The responsibility of the media manager is not to keep track of
open/closed blocks. This is better maintained within a target,
that already manages this information on writes.

Remove the statistics and merge the states NVM_BLK_ST_OPEN and
NVM_BLK_ST_CLOSED.

Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agolightnvm: fix checkpatch terse errors
Matias Bjørling [Thu, 7 Jul 2016 07:54:13 +0000 (09:54 +0200)]
lightnvm: fix checkpatch terse errors

A couple of small checkpatch fixups to stop it from complaining.

./drivers/lightnvm/core.c:360: WARNING: line over 80 characters
./drivers/lightnvm/core.c:360: ERROR: trailing statements should be on
next line
./drivers/lightnvm/core.c:503: WARNING: Block comments use a trailing */
on a separate line

Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agolightnvm: remove checkpatch warning for unsigned ints
Matias Bjørling [Thu, 7 Jul 2016 07:54:12 +0000 (09:54 +0200)]
lightnvm: remove checkpatch warning for unsigned ints

Checkpatch found two incidents where the type was preferred to be
written out in full.

./drivers/lightnvm/rrpc.h:184: WARNING: Prefer 'unsigned int' to bare
use of 'unsigned'
./drivers/lightnvm/rrpc.h:209: WARNING: Prefer 'unsigned int' to bare
use of 'unsigned'
./drivers/lightnvm/rrpc.c:51: WARNING: Prefer 'unsigned int' to bare use
of 'unsigned'

Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agolightnvm: Make functions not used by ouside static
Johannes Thumshirn [Thu, 7 Jul 2016 07:54:11 +0000 (09:54 +0200)]
lightnvm: Make functions not used by ouside static

Mark functions not used by ouside of thier implementing file as static.

Signed-off-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agonvme: lightnvm: make MLC num_pairs little endian
Johannes Thumshirn [Thu, 7 Jul 2016 07:54:10 +0000 (09:54 +0200)]
nvme: lightnvm: make MLC num_pairs little endian

According to the OpenChannel SSD interface specification the NAND flash
MLC page pairing information's number of page page pairings field is the
first two bytes in the MLC Page Pairing data structure. The hardware's
data structure itself is little endian so annotate it as such, like the
rest of lighnvm's data structures.

Signed-off-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agolightnvm: initialize ppa_addr in dev_to_generic_addr()
Javier González [Thu, 7 Jul 2016 07:54:09 +0000 (09:54 +0200)]
lightnvm: initialize ppa_addr in dev_to_generic_addr()

The ->reserved bit is not initialized when allocated on stack.
This may lead targets to misinterpret the PPA as cached.

Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agolightnvm: add media manager mark_blk helper
Javier González [Thu, 7 Jul 2016 07:54:08 +0000 (09:54 +0200)]
lightnvm: add media manager mark_blk helper

Expose media manager mark_blk() to targets, as done for the rest of the
media manager callback functions.

Signed-off-by: Javier González <javier@cnexlabs.com>
Updated description
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agolightnvm: break the loop when rqd is not null
Wenwei Tao [Thu, 7 Jul 2016 07:54:07 +0000 (09:54 +0200)]
lightnvm: break the loop when rqd is not null

Break the loop when rqd is not null to reduce
an unnecessary schedule.

Signed-off-by: Wenwei Tao <ww.tao0320@gmail.com>
Signed-off-by: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agonvmet: fix an error code
Dan Carpenter [Thu, 7 Jul 2016 08:15:26 +0000 (11:15 +0300)]
nvmet: fix an error code

We accidentally return zero here when ERR_PTR(-ENOMEM) is intended.

Fixes: a07b4970f464 ('nvmet: add a generic NVMe target')
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agonvme-loop: add configfs dependency
Arnd Bergmann [Thu, 7 Jul 2016 09:01:26 +0000 (11:01 +0200)]
nvme-loop: add configfs dependency

CONFIG_NVME_TARGET has a correct CONFIG_CONFIGFS_FS dependency, but the
newly added NVME_TARGET_LOOP is missing this, resulting in a link
failure:

drivers/nvme/built-in.o: In function `nvmet_init_configfs':
loop.c:(.init.text+0x2a0): undefined reference to `config_group_init'
loop.c:(.init.text+0x2c0): undefined reference to `config_group_init_type_name'
loop.c:(.init.text+0x318): undefined reference to `configfs_register_subsystem'
drivers/nvme/built-in.o: In function `nvmet_exit_configfs':
loop.c:(.exit.text+0x9c): undefined reference to `configfs_unregister_subsystem'

This adds the same dependency here.

Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Fixes: 3a85a5de29ea ("nvme-loop: add a NVMe loopback host driver")
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agobcache: Remove redundant block_size assignment
Yijing Wang [Mon, 4 Jul 2016 01:23:41 +0000 (09:23 +0800)]
bcache: Remove redundant block_size assignment

We have assigned sb->block_size before the switch,
so remove the redundant one.

Reviewed-by: Coly Li <colyli@suse.de>
Signed-off-by: Yijing Wang <wangyijing@huawei.com>
Acked-by: Eric Wheeler <bcache@lists.ewheeler.net>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agobcache: update document info
Yijing Wang [Mon, 4 Jul 2016 01:23:34 +0000 (09:23 +0800)]
bcache: update document info

There is no return in continue_at(), update the documentation.

Signed-off-by: Yijing Wang <wangyijing@huawei.com>
Acked-by: Coly Li <colyli@suse.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agobcache: Remove redundant parameter for cache_alloc()
Yijing Wang [Mon, 4 Jul 2016 01:23:25 +0000 (09:23 +0800)]
bcache: Remove redundant parameter for cache_alloc()

Cache_sb is not used in cache_alloc, and we have copied
sb info to cache->sb already, remove it.

Reviewed-by: Coly Li <colyli@suse.de>
Signed-off-by: Yijing Wang <wangyijing@huawei.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agonvme-loop: add a NVMe loopback host driver
Christoph Hellwig [Tue, 21 Jun 2016 16:04:21 +0000 (18:04 +0200)]
nvme-loop: add a NVMe loopback host driver

This patch implements adds nvme-loop which allows to access local devices
exported as NVMe over Fabrics namespaces. This module can be useful for
easy evaluation, testing and also feature experimentation.

To createa nvme-loop device you need to configure the NVMe target to
export a loop port (see the nvmetcli documentaton for that) and then
connect to it using

nvme connect-all -t loop

which requires the very latest nvme-cli version with Fabrics support.

Signed-off-by: Jay Freyensee <james.p.freyensee@intel.com>
Signed-off-by: Ming Lin <ming.l@ssi.samsung.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agonvmet: add a generic NVMe target
Christoph Hellwig [Tue, 21 Jun 2016 16:04:20 +0000 (18:04 +0200)]
nvmet: add a generic NVMe target

This patch introduces a implementation of NVMe subsystems,
controllers and discovery service which allows to export
NVMe namespaces across fabrics such as Ethernet, FC etc.

The implementation conforms to the NVMe 1.2.1 specification
and interoperates with NVMe over fabrics host implementations.

Configuration works using configfs, and is best performed using
the nvmetcli tool from http://git.infradead.org/users/hch/nvmetcli.git,
which also has a detailed explanation of the required steps in the
README file.

Signed-off-by: Armen Baloyan <armenx.baloyan@intel.com>
Signed-off-by: Anthony Knapp <anthony.j.knapp@intel.com>
Signed-off-by: Jay Freyensee <james.p.freyensee@intel.com>
Signed-off-by: Ming Lin <ming.l@ssi.samsung.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agoblock: Export blk_poll
Sagi Grimberg [Tue, 21 Jun 2016 16:04:19 +0000 (18:04 +0200)]
block: Export blk_poll

The new NVMe over fabrics target will make use of this outside from a
module.

Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Steve Wise <swise@opengridcomputing.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agonvme: add keep-alive support
Sagi Grimberg [Mon, 13 Jun 2016 14:45:28 +0000 (16:45 +0200)]
nvme: add keep-alive support

Periodic keep-alive is a mandatory feature in NVMe over Fabrics, and
optional in NVMe 1.2.1 for PCIe.  This patch adds periodic keep-alive
sent from the host to verify that the controller is still responsive
and vice-versa.  The keep-alive timeout is user-defined (with
keep_alive_tmo connection parameter) and defaults to 5 seconds.

In order to avoid a race condition where the host sends a keep-alive
competing with the target side keep-alive timeout expiration, the host
adds a grace period of 10 seconds when publishing the keep-alive timeout
to the target.

In case a keep-alive failed (or timed out), a transport specific error
recovery kicks in.

For now only NVMe over Fabrics is wired up to support keep alive, but
we can add PCIe support easily once controllers actually supporting it
become available.

Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Reviewed-by: Steve Wise <swise@chelsio.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agonvme.h: Add keep-alive opcode and identify controller attribute
Sagi Grimberg [Mon, 13 Jun 2016 14:45:27 +0000 (16:45 +0200)]
nvme.h: Add keep-alive opcode and identify controller attribute

KAS: keep-alive support and granularity of kato in units of 100 ms
nvme_admin_keep_alive opcode: 0x18

Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agonvme-fabrics: add a generic NVMe over Fabrics library
Christoph Hellwig [Mon, 13 Jun 2016 14:45:26 +0000 (16:45 +0200)]
nvme-fabrics: add a generic NVMe over Fabrics library

The NVMe over Fabrics library provides an interface for both transports
and the nvme core to handle fabrics specific commands and attributes
independent of the underlying transport.

In addition, the fabrics library adds a misc device interface that allow
actually creating a fabrics controller, as we can't just autodiscover
it like in the PCI case.  The nvme-cli utility has been enhanced to use
this interface to support fabric connect and discovery.

Signed-off-by: Armen Baloyan <armenx.baloyan@intel.com>,
Signed-off-by: Jay Freyensee <james.p.freyensee@intel.com>,
Signed-off-by: Ming Lin <ming.l@ssi.samsung.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agonvme.h: add NVMe over Fabrics definitions
Christoph Hellwig [Mon, 13 Jun 2016 14:45:25 +0000 (16:45 +0200)]
nvme.h: add NVMe over Fabrics definitions

The NVMe over Fabrics specification defines a protocol interface and
related extensions to NVMe that enable operation over network protocols.
The NVMe over Fabrics specification has an NVMe Transport binding for
each NVMe Transport.

This patch adds the fabrics related definitions:
- fabric specific command set and error codes
- transport addressing and binding definitions
- fabrics sgl extensions
- controller identification fabrics enhancements
- discovery log page definition

Signed-off-by: Armen Baloyan <armenx.baloyan@intel.com>
Signed-off-by: James Smart <james.smart@broadcom.com>
Signed-off-by: Jay Freyensee <james.p.freyensee@intel.com>
Signed-off-by: Ming Lin <ming.l@ssi.samsung.com>
Signed-off-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agonvme: add fabrics sysfs attributes
Ming Lin [Mon, 13 Jun 2016 14:45:24 +0000 (16:45 +0200)]
nvme: add fabrics sysfs attributes

- delete_controller: This attribute allows to delete a controller.
  A driver is not obligated to support it (pci doesn't) so it is
  created only if the driver supports it. The new fabrics drivers
  will support it (essentialy a disconnect operation).

  Usage:
  echo > /sys/class/nvme/nvme0/delete_controller

- subsysnqn: This attribute shows the subsystem nqn of the configured
  device. If a driver does not implement the get_subsysnqn method, the
  file will not appear in sysfs.

- transport: This attribute shows the transport name. Added a "name"
  field to struct nvme_ctrl_ops.

  For loop,
  cat /sys/class/nvme/nvme0/transport
  loop

  For RDMA,
  cat /sys/class/nvme/nvme0/transport
  rdma

  For PCIe,
  cat /sys/class/nvme/nvme0/transport
  pcie

- address: This attributes shows the controller address. The fabrics
  drivers that will implement get_address can show the address of the
  connected controller.

  example:
  cat /sys/class/nvme/nvme0/address
  traddr=192.168.2.2,trsvcid=1023

Signed-off-by: Ming Lin <ming.l@ssi.samsung.com>
Reviewed-by: Jay Freyensee <james.p.freyensee@intel.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agonvme: Modify and export sync command submission for fabrics
Christoph Hellwig [Mon, 13 Jun 2016 14:45:23 +0000 (16:45 +0200)]
nvme: Modify and export sync command submission for fabrics

NVMe over fabrics will use __nvme_submit_sync_cmd in the the
transport and require a few tweaks to it.  For that we export it
and add a few more paramters:

1. allow passing a queue ID to the block layer

   For the NVMe over Fabrics connect command we need to able to specify a
   queue ID that we want to send the command on.  Add a qid parameter to
   the relevant functions to enable this behavior.

2. allow submitting at_head commands

   In cases where we want to (re)connect to a controller
   where we have inflight queued commands we want to first
   connect and only then allow the other queued commands to
   be kicked. This will prevents failures in controller resets
   and reconnects.

3. allow passing flags to blk_mq_allocate_request

   Both for Fabrics connect the the keep-alive feature in NVMe 1.2.1 we
   want to be able to use reserved requests.

Reviewed-by: Jay Freyensee <james.p.freyensee@intel.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Tested-by: Ming Lin <ming.l@ssi.samsung.com>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agonvme: allow transitioning from NEW to LIVE state
Christoph Hellwig [Mon, 13 Jun 2016 14:45:22 +0000 (16:45 +0200)]
nvme: allow transitioning from NEW to LIVE state

For Fabrics we're not going through an intermediate reset state
(at least for now).

Reviewed-by: Jay Freyensee <james.p.freyensee@intel.com>
Reviewed-by: Sagi Grimberg <sagi@grimberg.me>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agoblk-mq: add blk_mq_alloc_request_hctx
Ming Lin [Mon, 13 Jun 2016 14:45:21 +0000 (16:45 +0200)]
blk-mq: add blk_mq_alloc_request_hctx

For some protocols like NVMe over Fabrics we need to be able to send
initialization commands to a specific queue.

Based on an earlier patch from Christoph Hellwig <hch@lst.de>.

Signed-off-by: Ming Lin <ming.l@ssi.samsung.com>
[hch: disallow sleeping allocation, req_op fixes]
Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Keith Busch <keith.busch@intel.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agomg_disk: fix error path in mg_probe()
Bartlomiej Zolnierkiewicz [Tue, 28 Jun 2016 13:42:41 +0000 (15:42 +0200)]
mg_disk: fix error path in mg_probe()

MG_DISK_MAJ is defined as 0 so dynamic block major number
allocation is used by the driver and the assigned major
number is stored in host->major.  This patch fixes error
path in mg_probe() to use host->major instead of using
MG_DISK_MAJ.

Cc: unsik Kim <donari75@gmail.com>
Signed-off-by: Bartlomiej Zolnierkiewicz <b.zolnierkie@samsung.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agoblock: remove ->driverfs_dev
Dan Williams [Thu, 16 Jun 2016 02:43:07 +0000 (19:43 -0700)]
block: remove ->driverfs_dev

Now that all drivers that specify a ->driverfs_dev have been converted
to device_add_disk(), the pointer can be removed from struct gendisk.

Cc: Jens Axboe <axboe@fb.com>
Cc: Ross Zwisler <ross.zwisler@linux.intel.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
8 years agoblock: convert to device_add_disk()
Dan Williams [Thu, 16 Jun 2016 02:44:20 +0000 (19:44 -0700)]
block: convert to device_add_disk()

For block drivers that specify a parent device, convert them to use
device_add_disk().

This conversion was done with the following semantic patch:

    @@
    struct gendisk *disk;
    expression E;
    @@

    - disk->driverfs_dev = E;
    ...
    - add_disk(disk);
    + device_add_disk(E, disk);

    @@
    struct gendisk *disk;
    expression E1, E2;
    @@

    - disk->driverfs_dev = E1;
    ...
    E2 = disk;
    ...
    - add_disk(E2);
    + device_add_disk(E1, E2);

...plus some manual fixups for a few missed conversions.

Cc: Jens Axboe <axboe@fb.com>
Cc: Keith Busch <keith.busch@intel.com>
Cc: Michael S. Tsirkin <mst@redhat.com>
Cc: David Woodhouse <dwmw2@infradead.org>
Cc: David S. Miller <davem@davemloft.net>
Cc: James Bottomley <James.Bottomley@hansenpartnership.com>
Cc: Ross Zwisler <ross.zwisler@linux.intel.com>
Cc: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Cc: Martin K. Petersen <martin.petersen@oracle.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
8 years agoum: track 'parent' device in a local variable
Dan Williams [Mon, 20 Jun 2016 17:44:32 +0000 (10:44 -0700)]
um: track 'parent' device in a local variable

In preparation for the removal of 'driverfs_dev' from 'struct gendisk'
use a local variable to track the parented vs un-parented case in
ubd_disk_register().

Cc: Jeff Dike <jdike@addtoit.com>
Cc: Richard Weinberger <richard@nod.at>
Cc: Bart Van Assche <bart.vanassche@sandisk.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
8 years agommc: move 'parent' tracking to mmc_blk_data
Dan Williams [Mon, 20 Jun 2016 17:40:44 +0000 (10:40 -0700)]
mmc: move 'parent' tracking to mmc_blk_data

In preparation for the removal of 'driverfs_dev' from 'struct gendisk',
carry this data in mmc_blk_data. It is used for registration of parent
disks and partitions.

Cc: Ulf Hansson <ulf.hansson@linaro.org>
Cc: Bart Van Assche <bart.vanassche@sandisk.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
8 years agoblock: introduce device_add_disk()
Dan Williams [Thu, 16 Jun 2016 01:17:27 +0000 (18:17 -0700)]
block: introduce device_add_disk()

In preparation for removing the ->driverfs_dev member of a gendisk, add
an api that takes the parent device as a parameter to add_disk().  For
now this maintains the status quo of WARN()ing on failure, but not
return a error code.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Reviewed-by: Bart Van Assche <bart.vanassche@sandisk.com>
Signed-off-by: Dan Williams <dan.j.williams@intel.com>
8 years agodrbd: correctly handle failed crypto_alloc_hash
Lars Ellenberg [Mon, 13 Jun 2016 22:26:39 +0000 (00:26 +0200)]
drbd: correctly handle failed crypto_alloc_hash

crypto_alloc_hash returns an ERR_PTR(), not NULL.

Also reset peer_integrity_tfm to NULL, to not call crypto_free_hash()
on an errno in the cleanup path.

Reported-by: Insu Yun <wuninsu@gmail.com>
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: al_write_transaction: skip re-scanning of bitmap page pointer array
Lars Ellenberg [Mon, 13 Jun 2016 22:26:38 +0000 (00:26 +0200)]
drbd: al_write_transaction: skip re-scanning of bitmap page pointer array

For larger devices, the array of bitmap page pointers can grow very
large (8000 pointers per TB of storage).

For each activity log transaction, we need to flush the associated
bitmap pages to stable storage. Currently, we just "mark" the respective
pages while setting up the transaction, then tell the bitmap code to
write out all marked pages, but skip unchanged pages.

But one such transaction can affect only a small number of bitmap pages,
there is no need to scan the full array of several (ten-)thousand
page pointers to find the few marked ones.

Instead, remember the index numbers of the few affected pages,
and later only re-check those to skip duplicates and unchanged ones.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: finally report ms, not jiffies, in log message
Lars Ellenberg [Mon, 13 Jun 2016 22:26:37 +0000 (00:26 +0200)]
drbd: finally report ms, not jiffies, in log message

Also skip the message unless bitmap IO took longer than 5 ms.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: get rid of empty statement in is_valid_state
Roland Kammerer [Mon, 13 Jun 2016 22:26:36 +0000 (00:26 +0200)]
drbd: get rid of empty statement in is_valid_state

This should silence a warning about an empty statement. Thanks to Fabian
Frederick <fabf@skynet.be> who sent a patch I modified to be smaller and
avoids an additional indent level.

Signed-off-by: Roland Kammerer <roland.kammerer@linbit.com>
Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: code cleanups without semantic changes
Fabian Frederick [Mon, 13 Jun 2016 22:26:35 +0000 (00:26 +0200)]
drbd: code cleanups without semantic changes

This contains various cosmetic fixes ranging from simple typos to
const-ifying, and using booleans properly.

Original commit messages from Fabian's patch set:
drbd: debugfs: constify drbd_version_fops
drbd: use seq_put instead of seq_print where possible
drbd: include linux/uaccess.h instead of asm/uaccess.h
drbd: use const char * const for drbd strings
drbd: kerneldoc warning fix in w_e_end_data_req()
drbd: use unsigned for one bit fields
drbd: use bool for peer is_ states
drbd: fix typo
drbd: use | for bitmask combination
drbd: use true/false for bool
drbd: fix drbd_bm_init() comments
drbd: introduce peer state union
drbd: fix maybe_pull_ahead() locking comments
drbd: use bool for growing
drbd: remove redundant declarations
drbd: replace if/BUG by BUG_ON

Signed-off-by: Fabian Frederick <fabf@skynet.be>
Signed-off-by: Roland Kammerer <roland.kammerer@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: bump current uuid when resuming IO with diskless peer
Lars Ellenberg [Mon, 13 Jun 2016 22:26:34 +0000 (00:26 +0200)]
drbd: bump current uuid when resuming IO with diskless peer

Scenario, starting with normal operation
 Connected Primary/Secondary UpToDate/UpToDate
 NetworkFailure Primary/Unknown UpToDate/DUnknown (frozen)
 ... more failures happen, secondary loses it's disk,
 but eventually is able to re-establish the replication link ...
 Connected Primary/Secondary UpToDate/Diskless (resumed; needs to bump uuid!)

We used to just resume/resent suspended requests,
without bumping the UUID.

Which will lead to problems later, when we want to re-attach the disk on
the peer, without first disconnecting, or if we experience additional
failures, because we now have diverging data without being able to
recognize it.

Make sure we also bump the current data generation UUID,
if we notice "peer disk unknown" -> "peer disk known bad".

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: disallow promotion during resync handshake, avoid deadlock and hard reset
Lars Ellenberg [Mon, 13 Jun 2016 22:26:33 +0000 (00:26 +0200)]
drbd: disallow promotion during resync handshake, avoid deadlock and hard reset

We already serialize connection state changes,
and other, non-connection state changes (role changes)
while we are establishing a connection.

But if we have an established connection,
then trigger a resync handshake (by primary --force or similar),
until now we just had to be "lucky".

Consider this sequence (e.g. deployment scenario):
create-md; up;
  -> Connected Secondary/Secondary Inconsistent/Inconsistent
then do a racy primary --force on both peers.

 block drbd0: drbd_sync_handshake:
 block drbd0: self 0000000000000004:0000000000000000:0000000000000000:0000000000000000 bits:25590 flags:0
 block drbd0: peer 0000000000000004:0000000000000000:0000000000000000:0000000000000000 bits:25590 flags:0
 block drbd0: peer( Unknown -> Secondary ) conn( WFReportParams -> Connected ) pdsk( DUnknown -> Inconsistent )
 block drbd0: peer( Secondary -> Primary ) pdsk( Inconsistent -> UpToDate )
  *** HERE things go wrong. ***
 block drbd0: role( Secondary -> Primary )
 block drbd0: drbd_sync_handshake:
 block drbd0: self 0000000000000005:0000000000000000:0000000000000000:0000000000000000 bits:25590 flags:0
 block drbd0: peer C90D2FC716D232AB:0000000000000004:0000000000000000:0000000000000000 bits:25590 flags:0
 block drbd0: Becoming sync target due to disk states.
 block drbd0: Writing the whole bitmap, full sync required after drbd_sync_handshake.
 block drbd0: Remote failed to finish a request within 6007ms > ko-count (2) * timeout (30 * 0.1s)
 drbd s0: peer( Primary -> Unknown ) conn( Connected -> Timeout ) pdsk( UpToDate -> DUnknown )

The problem here is that the local promotion happens before the sync handshake
triggered by the remote promotion was completed.  Some assumptions elsewhere
become wrong, and when the expected resync handshake is then received and
processed, we get stuck in a deadlock, which can only be recovered by reboot :-(

Fix: if we know the peer has good data,
and our own disk is present, but NOT good,
and there is no resync going on yet,
we expect a sync handshake to happen "soon".
So reject a racy promotion with SS_IN_TRANSIENT_STATE.

Result:
 ... as above ...
 block drbd0: peer( Secondary -> Primary ) pdsk( Inconsistent -> UpToDate )
  *** local promotion being postponed until ... ***
 block drbd0: drbd_sync_handshake:
 block drbd0: self 0000000000000004:0000000000000000:0000000000000000:0000000000000000 bits:25590 flags:0
 block drbd0: peer 77868BDA836E12A5:0000000000000004:0000000000000000:0000000000000000 bits:25590 flags:0
  ...
 block drbd0: conn( WFBitMapT -> WFSyncUUID )
 block drbd0: updated sync uuid 85D06D0E8887AD44:0000000000000000:0000000000000000:0000000000000000
 block drbd0: conn( WFSyncUUID -> SyncTarget )
  *** ... after the resync handshake ***
 block drbd0: role( Secondary -> Primary )

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: sync_handshake: handle identical uuids with current (frozen) Primary
Lars Ellenberg [Mon, 13 Jun 2016 22:26:32 +0000 (00:26 +0200)]
drbd: sync_handshake: handle identical uuids with current (frozen) Primary

If in a two-primary scenario, we lost our peer, freeze IO,
and are still frozen (no UUID rotation) when the peer comes back
as Secondary after a hard crash, we will see identical UUIDs.

The "rule_nr = 40" chose to use the "CRASHED_PRIMARY" bit as
arbitration, but that would cause the still running (but frozen) Primary
to become SyncTarget (which it typically refuses), and the handshake is
declined.

Fix: check current roles.
If we have *one* current primary, the Primary wins.
(rule_nr = 41)

Since that is a protocol change, use the newly introduced DRBD_FF_WSAME
to determine if rule_nr = 41 can be applied.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: introduce WRITE_SAME support
Lars Ellenberg [Mon, 13 Jun 2016 22:26:31 +0000 (00:26 +0200)]
drbd: introduce WRITE_SAME support

We will support WRITE_SAME, if
 * all peers support WRITE_SAME (both in kernel and DRBD version),
 * all peer devices support WRITE_SAME
 * logical_block_size is identical on all peers.

We may at some point introduce a fallback on the receiving side
for devices/kernels that do not support WRITE_SAME,
by open-coding a submit loop. But not yet.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: report sizes if rejecting too small peer disk
Lars Ellenberg [Mon, 13 Jun 2016 22:26:30 +0000 (00:26 +0200)]
drbd: report sizes if rejecting too small peer disk

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: discard_zeroes_if_aligned allows "thin" resync for discard_zeroes_data=0
Lars Ellenberg [Mon, 13 Jun 2016 22:26:29 +0000 (00:26 +0200)]
drbd: discard_zeroes_if_aligned allows "thin" resync for discard_zeroes_data=0

Even if discard_zeroes_data != 0,
if discard_zeroes_if_aligned is set, we assume we can reliably
zero-out/discard using the drbd_issue_peer_discard() helper.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: only restart frozen disk io when D_UP_TO_DATE
Lars Ellenberg [Mon, 13 Jun 2016 22:26:28 +0000 (00:26 +0200)]
drbd: only restart frozen disk io when D_UP_TO_DATE

When re-attaching the local backend device to a C_STANDALONE D_DISKLESS
R_PRIMARY with OND_SUSPEND_IO, we may only resume IO if we recognize the
backend that is being attached as D_UP_TO_DATE.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: if there is no good data accessible, writes should be IO errors
Lars Ellenberg [Mon, 13 Jun 2016 22:26:27 +0000 (00:26 +0200)]
drbd: if there is no good data accessible, writes should be IO errors

If DRBD lost all path to good data,
and the on-no-data-accessible policy is OND_SUSPEND_IO,
all pending and new IO requests are suspended (will block).

If that setting is OND_IO_ERROR, IO will still be completed.
READ to "clean" areas (e.g. on an D_INCONSISTENT device,
and bitmap indicates a block is already in sync) will succeed.
READ to "unclean" areas (bitmap indicates block is out-of-sync),
will return EIO.

If we are already D_DISKLESS (or D_FAILED), we also return EIO.

Unfortunately, on a former R_PRIMARY C_SYNC_TARGET D_INCONSISTENT,
after replication link loss, new WRITE requests still went through OK.

The would also set the "out-of-sync" bit on their way, so READ after
WRITE would still return EIO. Also, the data generation UUIDs had not
been bumped, we would cause data divergence, without being able to
detect it on the next sync handshake, given the right sequence of events
in a multiple error scenario and "improper" order of recovery actions.

The right thing to do is to return EIO for all new writes,
unless we have access to good, current, D_UP_TO_DATE data.

The "established best practices" way to avoid these situations in the
first place is to set OND_SUSPEND_IO, or even do a hard-reset from
the pri-on-incon-degr policy helper hook.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: don't forget error completion when "unsuspending" IO
Lars Ellenberg [Mon, 13 Jun 2016 22:26:26 +0000 (00:26 +0200)]
drbd: don't forget error completion when "unsuspending" IO

Possibly sequence of events:
SyncTarget is made Primary, then loses replication link
(only path to good data on SyncSource).

Behavior is then controlled by the on-no-data-accessible policy,
which defaults to OND_IO_ERROR (may be set to OND_SUSPEND_IO).

If OND_IO_ERROR is in fact the current policy, we clear the susp_fen
(IO suspended due to fencing policy) flag, do NOT set the susp_nod
(IO suspended due to no data) flag.

But we forgot to call the IO error completion for all pending,
suspended, requests.

While at it, also add a race check for a theoretically possible
race with a new handshake (network hickup), we may be able to
re-send requests, and can avoid passing IO errors up the stack.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: introduce unfence-peer handler
Lars Ellenberg [Mon, 13 Jun 2016 22:26:25 +0000 (00:26 +0200)]
drbd: introduce unfence-peer handler

When resync is finished, we already call the "after-resync-target"
handler (on the former sync target, obviously), once per volume.

Paired with the before-resync-target handler, you can create snapshots,
before the resync causes the volumes to become inconsistent,
and discard those snapshots again, once they are no longer needed.

It was also overloaded to be paired with the "fence-peer" handler,
to "unfence" once the volumes are up-to-date and known good.

This has some disadvantages, though: we call "fence-peer" for the whole
connection (once for the group of volumes), but would call unfence as
side-effect of after-resync-target once for each volume.

Also, we fence on a (current, or about to become) Primary,
which will later become the sync-source.

Calling unfence only as a side effect of the after-resync-target
handler opens a race window, between a new fence on the Primary
(SyncTarget) and the unfence on the SyncTarget, which is difficult to
close without some kind of "cluster wide lock" in those handlers.

We would not need those handlers if we could still communicate.
Which makes trying to aquire a cluster wide lock from those handlers
seem like a very bad idea.

This introduces the "unfence-peer" handler, which will be called
per connection (once for the group of volumes), just like the fence
handler, only once all volumes are back in sync, and on the SyncSource.

Which is expected to be the node that previously called "fence", the
node that is currently allowed to be Primary, and thus the only node
that could trigger a new "fence" that could race with this unfence.

Which makes us not need any cluster wide synchronization here,
serializing two scripts running on the same node is trivial.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: finish resync on sync source only by notification from sync target
Lars Ellenberg [Mon, 13 Jun 2016 22:26:24 +0000 (00:26 +0200)]
drbd: finish resync on sync source only by notification from sync target

If the replication link breaks exactly during "resync finished" detection,
finishing too early on the sync source could again lead to UUIDs rotated
too fast, and potentially a spurious full resync on next handshake.

Always wait for explicit resync finished state change notification from
the sync target.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: allow larger max_discard_sectors
Lars Ellenberg [Mon, 13 Jun 2016 22:26:23 +0000 (00:26 +0200)]
drbd: allow larger max_discard_sectors

Make sure we have at least 67 (> AL_UPDATES_PER_TRANSACTION)
al-extents available, and allow up to half of that to be
discarded in one bio.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: zero-out partial unaligned discards on local backend
Lars Ellenberg [Mon, 13 Jun 2016 22:26:22 +0000 (00:26 +0200)]
drbd: zero-out partial unaligned discards on local backend

For consistency, also zero-out partial unaligned chunks of discard
requests on the local backend.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: possibly disable discard support, if backend has discard_zeroes_data=0
Lars Ellenberg [Mon, 13 Jun 2016 22:26:21 +0000 (00:26 +0200)]
drbd: possibly disable discard support, if backend has discard_zeroes_data=0

Now that we have the discard_zeroes_if_aligned setting, we should also
check it when setting up our queue parameters on the primary,
not only on the receiving side.

We announce discard support,
UNLESS

 * we are connected to a peer that does not support TRIM
   on the DRBD protocol level.  Otherwise, it would either discard, or
   do a fallback to zero-out, depending on its backend and configuration.

 * our local backend does not support discards,
   or (discard_zeroes_data=0 AND discard_zeroes_if_aligned=no).

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: when receiving P_TRIM, zero-out partial unaligned chunks
Lars Ellenberg [Mon, 13 Jun 2016 22:26:20 +0000 (00:26 +0200)]
drbd: when receiving P_TRIM, zero-out partial unaligned chunks

We can avoid spurious data divergence caused by partially-ignored
discards on certain backends with discard_zeroes_data=0, if we
translate partial unaligned discard requests into explicit zero-out.

The relevant use case is LVM/DM thin.

If on different nodes, DRBD is backed by devices with differing
discard characteristics, discards may lead to data divergence
(old data or garbage left over on one backend, zeroes due to
unmapped areas on the other backend). Online verify would now
potentially report tons of spurious differences.

While probably harmless for most use cases (fstrim on a file system),
DRBD cannot have that, it would violate our promise to upper layers
that our data instances on the nodes are identical.

To be correct and play safe (make sure data is identical on both copies),
we would have to disable discard support, if our local backend (on a
Primary) does not support "discard_zeroes_data=true".

We'd also have to translate discards to explicit zero-out on the
receiving (typically: Secondary) side, unless the receiving side
supports "discard_zeroes_data=true".

Which both would allocate those blocks, instead of unmapping them,
in contrast with expectations.

LVM/DM thin does set discard_zeroes_data=0,
because it silently ignores discards to partial chunks.

We can work around this by checking the alignment first.
For unaligned (wrt. alignment and granularity) or too small discards,
we zero-out the initial (and/or) trailing unaligned partial chunks,
but discard all the aligned full chunks.

At least for LVM/DM thin, the result is effectively "discard_zeroes_data=1".

Arguably it should behave this way internally, by default,
and we'll try to make that happen.

But our workaround is still valid for already deployed setups,
and for other devices that may behave this way.

Setting discard-zeroes-if-aligned=yes will allow DRBD to use
discards, and to announce discard_zeroes_data=true, even on
backends that announce discard_zeroes_data=false.

Setting discard-zeroes-if-aligned=no will cause DRBD to always
fall-back to zero-out on the receiving side, and to not even
announce discard capabilities on the Primary, if the respective
backend announces discard_zeroes_data=false.

We used to ignore the discard_zeroes_data setting completely.
To not break established and expected behaviour, and suddenly
cause fstrim on thin-provisioned LVs to run out-of-space,
instead of freeing up space, the default value is "yes".

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: allow parallel flushes for multi-volume resources
Lars Ellenberg [Mon, 13 Jun 2016 22:26:19 +0000 (00:26 +0200)]
drbd: allow parallel flushes for multi-volume resources

To maintain write-order fidelity accros all volumes in a DRBD resource,
the receiver of a P_BARRIER needs to issue flushes to all volumes.
We used to do this by calling blkdev_issue_flush(), synchronously,
one volume at a time.

We now submit all flushes to all volumes in parallel, then wait for all
completions, to reduce worst-case latencies on multi-volume resources.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: fix for truncated minor number in callback command line
Lars Ellenberg [Mon, 13 Jun 2016 22:26:18 +0000 (00:26 +0200)]
drbd: fix for truncated minor number in callback command line

The command line parameter the kernel module uses to communicate the
device minor to userland helper is flawed in a way that the device
indentifier "minor-%d" is being truncated to minors with a maximum
of 5 digits.

But DRBD 8.4 allows 2^20 == 1048576 minors,
thus a minimum of 7 digits must be supported.

Reported by Veit Wahlich on drbd-dev.

Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: fix regression: protocol A sometimes synchronous, C sometimes double-latency
Lars Ellenberg [Mon, 13 Jun 2016 22:26:17 +0000 (00:26 +0200)]
drbd: fix regression: protocol A sometimes synchronous, C sometimes double-latency

Regression introduced with 8.4.5
 drbd: application writes may set-in-sync in protocol != C

Overwriting the same block (LBA) while a former version is still
"in-flight" to the peer (to be exact: we did not receive the
P_BARRIER_ACK for its epoch yet) would wait for the full epoch of that
former version to be acknowledged by the peer.

In synchronous and quasi-synchronous protocols C and B,
this may double the latency on overwrites.

With protocol A, which is supposed to be asynchronous and only wait for
local completion, it is even worse: it would make overwrites
quasi-synchronous, they would be hit by the full RTT, which protocol A
was specifically meant to avoid, and possibly the additional time it
takes to drain the buffers first.

Particularly bad for databases, or anything else that
does frequent updates to the same blocks (various file system meta data).

No impact if >= rtt passes between updates to the same block.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: adjust assert in w_bitmap_io to account for BM_LOCKED_CHANGE_ALLOWED
Lars Ellenberg [Mon, 13 Jun 2016 22:26:16 +0000 (00:26 +0200)]
drbd: adjust assert in w_bitmap_io to account for BM_LOCKED_CHANGE_ALLOWED

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: Create the protocol feature THIN_RESYNC
Philipp Reisner [Mon, 13 Jun 2016 22:26:15 +0000 (00:26 +0200)]
drbd: Create the protocol feature THIN_RESYNC

If thinly provisioned volumes are used, during a resync the sync source
tries to find out if a block is deallocated. If it is deallocated, then
the resync target uses block_dev_issue_zeroout() on the range in
question.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: Introduce new disk config option rs-discard-granularity
Philipp Reisner [Mon, 13 Jun 2016 22:26:14 +0000 (00:26 +0200)]
drbd: Introduce new disk config option rs-discard-granularity

As long as the value is 0 the feature is disabled. With setting
it to a positive value, DRBD limits and aligns its resync requests
to the rs-discard-granularity setting. If the sync source detects
all zeros in such a block, the resync target discards the range
on disk.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: Implement handling of thinly provisioned storage on resync target nodes
Philipp Reisner [Mon, 13 Jun 2016 22:26:13 +0000 (00:26 +0200)]
drbd: Implement handling of thinly provisioned storage on resync target nodes

If during resync we read only zeroes for a range of sectors assume
that these secotors can be discarded on the sync target node.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: Kill code duplication
Philipp Reisner [Mon, 13 Jun 2016 22:26:12 +0000 (00:26 +0200)]
drbd: Kill code duplication

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
8 years agodrbd: change bitmap write-out when leaving resync states
Lars Ellenberg [Mon, 13 Jun 2016 22:26:11 +0000 (00:26 +0200)]
drbd: change bitmap write-out when leaving resync states

When leaving resync states because of disconnect,
do the bitmap write-out synchronously in the drbd_disconnected() path.

When leaving resync states because we go back to AHEAD/BEHIND, or
because resync actually finished, or some disk was lost during resync,
trigger the write-out from after_state_ch().

The bitmap write-out for resync -> ahead/behind was missing completely before.

Note that this is all only an optimization to avoid double-resyncs of
already completed blocks in case this node crashes.

Signed-off-by: Philipp Reisner <philipp.reisner@linbit.com>
Signed-off-by: Lars Ellenberg <lars.ellenberg@linbit.com>
Signed-off-by: Jens Axboe <axboe@fb.com>