GitHub/moto-9609/android_kernel_motorola_exynos9610.git
7 years agoblock: remove the errors field from struct request
Christoph Hellwig [Thu, 20 Apr 2017 14:03:16 +0000 (16:03 +0200)]
block: remove the errors field from struct request

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Bart Van Assche <Bart.VanAssche@sandisk.com>
Acked-by: Roger Pau Monné <roger.pau@citrix.com>
Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblktrace: remove the unused block_rq_abort tracepoint
Christoph Hellwig [Thu, 20 Apr 2017 14:03:15 +0000 (16:03 +0200)]
blktrace: remove the unused block_rq_abort tracepoint

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoswim3: remove (commented out) printing of req->errors
Christoph Hellwig [Thu, 20 Apr 2017 14:03:14 +0000 (16:03 +0200)]
swim3: remove (commented out) printing of req->errors

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoataflop: switch from req->errors to req->error_count
Christoph Hellwig [Thu, 20 Apr 2017 14:03:13 +0000 (16:03 +0200)]
ataflop: switch from req->errors to req->error_count

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agofloppy: switch from req->errors to req->error_count
Christoph Hellwig [Thu, 20 Apr 2017 14:03:12 +0000 (16:03 +0200)]
floppy: switch from req->errors to req->error_count

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock: add a error_count field to struct request
Christoph Hellwig [Thu, 20 Apr 2017 14:03:11 +0000 (16:03 +0200)]
block: add a error_count field to struct request

This is for the legacy floppy and ataflop drivers that currently abuse
->errors for this purpose.  It's stashed away in a union to not grow
the struct size, the other fields are either used by modern drivers
for different purposes or the I/O scheduler before queing the I/O
to drivers.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblk-mq: simplify __blk_mq_complete_request
Christoph Hellwig [Thu, 20 Apr 2017 14:03:10 +0000 (16:03 +0200)]
blk-mq: simplify __blk_mq_complete_request

Merge blk_mq_ipi_complete_request and blk_mq_stat_add into their only
caller.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblk-mq: remove the error argument to blk_mq_complete_request
Christoph Hellwig [Thu, 20 Apr 2017 14:03:09 +0000 (16:03 +0200)]
blk-mq: remove the error argument to blk_mq_complete_request

Now that all drivers that call blk_mq_complete_requests have a
->complete callback we can remove the direct call to blk_mq_end_request,
as well as the error argument to blk_mq_complete_request.

Signed-off-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: Jens Axboe <axboe@fb.com>
7 years agoxen-blkfront: don't use req->errors
Christoph Hellwig [Thu, 20 Apr 2017 14:03:08 +0000 (16:03 +0200)]
xen-blkfront: don't use req->errors

xen-blkfron is the last users using rq->errros for passing back error to
blk-mq, and I'd like to get rid of that.  In the longer run the driver
should be moving more of the completion processing into .complete, but
this is the minimal change to move forward for now.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Roger Pau Monné <roger.pau@citrix.com>
Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agomtip32xx: add a status field to struct mtip_cmd
Christoph Hellwig [Thu, 20 Apr 2017 14:03:07 +0000 (16:03 +0200)]
mtip32xx: add a status field to struct mtip_cmd

Instead of using req->errors, which will go away.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agonbd: don't use req->errors
Christoph Hellwig [Thu, 20 Apr 2017 14:03:06 +0000 (16:03 +0200)]
nbd: don't use req->errors

Add a nbd-specific field instead.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agodm mpath: don't check for req->errors
Christoph Hellwig [Thu, 20 Apr 2017 14:03:05 +0000 (16:03 +0200)]
dm mpath: don't check for req->errors

We'll get all proper errors reported through ->end_io and ->errors will
go away soon.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agodm rq: don't pass irrelevant error code to blk_mq_complete_request
Christoph Hellwig [Thu, 20 Apr 2017 14:03:04 +0000 (16:03 +0200)]
dm rq: don't pass irrelevant error code to blk_mq_complete_request

dm never uses rq->errors, so there is no need to pass an error argument
to blk_mq_complete_request.

Signed-off-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: Jens Axboe <axboe@fb.com>
7 years agonull_blk: don't pass always-0 req->errors to blk_mq_complete_request
Christoph Hellwig [Thu, 20 Apr 2017 14:03:03 +0000 (16:03 +0200)]
null_blk: don't pass always-0 req->errors to blk_mq_complete_request

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoloop: zero-fill bio on the submitting cpu
Christoph Hellwig [Thu, 20 Apr 2017 14:03:02 +0000 (16:03 +0200)]
loop: zero-fill bio on the submitting cpu

In thruth I've just audited which blk-mq drivers don't currently have a
complete callback, but I think this change is at least borderline useful.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoscsi: introduce a result field in struct scsi_request
Christoph Hellwig [Thu, 20 Apr 2017 14:03:01 +0000 (16:03 +0200)]
scsi: introduce a result field in struct scsi_request

This passes on the scsi_cmnd result field to users of passthrough
requests.  Currently we abuse req->errors for this purpose, but that
field will go away in its current form.

Note that the old IDE code abuses the errors field in very creative
ways and stores all kinds of different values in it.  I didn't dare
to touch this magic, so the abuses are brought forward 1:1.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Reviewed-by: Bart Van Assche <Bart.VanAssche@sandisk.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agovirtio_blk: don't use req->errors
Christoph Hellwig [Thu, 20 Apr 2017 14:03:00 +0000 (16:03 +0200)]
virtio_blk: don't use req->errors

Remove passing req->errors (which at that point is always 0) to
blk_mq_complete_request, and rely on the virtio status code for the
serial number passthrough request.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agovirtio: fix spelling of virtblk_scsi_request_done
Christoph Hellwig [Thu, 20 Apr 2017 14:02:59 +0000 (16:02 +0200)]
virtio: fix spelling of virtblk_scsi_request_done

Signed-off-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: Jens Axboe <axboe@fb.com>
7 years agonvme: make nvme_error_status private
Christoph Hellwig [Thu, 20 Apr 2017 14:02:58 +0000 (16:02 +0200)]
nvme: make nvme_error_status private

Currently it's used by the lighnvm passthrough ioctl, but we'd like to make
it private in preparation of block layer specific error code.  Lighnvm already
returns the real NVMe status anyway, so I think we can just limit it to
returning -EIO for any status set.

This will need a careful audit from the lightnvm folks, though.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agonvme: split nvme status from block req->errors
Christoph Hellwig [Thu, 20 Apr 2017 14:02:57 +0000 (16:02 +0200)]
nvme: split nvme status from block req->errors

We want our own clearly defined error field for NVMe passthrough commands,
and the request errors field is going away in its current form.

Just store the status and result field in the nvme_request field from
hardirq completion context (using a new helper) and then generate a
Linux errno for the block layer only when we actually need it.

Because we can't overload the status value with a negative error code
for cancelled command we now have a flags filed in struct nvme_request
that contains a bit for this condition.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agonvme-fc: fix status code handling in nvme_fc_fcpio_done
Christoph Hellwig [Thu, 20 Apr 2017 14:02:56 +0000 (16:02 +0200)]
nvme-fc: fix status code handling in nvme_fc_fcpio_done

nvme_complete_async_event expects the little endian status code
including the phase bit, and a new completion handler I plan to
introduce will do so as well.

Change the status variable into the little endian format with the
phase bit used in the NVMe CQE to fix / enable this.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock: remove the blk_execute_rq return value
Christoph Hellwig [Thu, 20 Apr 2017 14:02:55 +0000 (16:02 +0200)]
block: remove the blk_execute_rq return value

The function only returns -EIO if rq->errors is non-zero, which is not
very useful and lets a large number of callers ignore the return value.

Just let the callers figure out their error themselves.

Signed-off-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: Jens Axboe <axboe@fb.com>
7 years agopd: don't check blk_execute_rq return value.
Christoph Hellwig [Thu, 20 Apr 2017 14:02:54 +0000 (16:02 +0200)]
pd: don't check blk_execute_rq return value.

The driver never sets req->errors, so blk_execute_rq will always return 0.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Bart Van Assche <Bart.VanAssche@sandisk.com>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agobdi: Drop 'parent' argument from bdi_register[_va]()
Jan Kara [Wed, 12 Apr 2017 10:24:49 +0000 (12:24 +0200)]
bdi: Drop 'parent' argument from bdi_register[_va]()

Drop 'parent' argument of bdi_register() and bdi_register_va().  It is
always NULL.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock: Remove unused functions
Jan Kara [Wed, 12 Apr 2017 10:24:48 +0000 (12:24 +0200)]
block: Remove unused functions

Now that all backing_dev_info structure are allocated separately, we can
drop some unused functions.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agofs: Remove SB_I_DYNBDI flag
Jan Kara [Wed, 12 Apr 2017 10:24:47 +0000 (12:24 +0200)]
fs: Remove SB_I_DYNBDI flag

Now that all bdi structures filesystems use are properly refcounted, we
can remove the SB_I_DYNBDI flag.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoubifs: Convert to separately allocated bdi
Jan Kara [Wed, 12 Apr 2017 10:24:46 +0000 (12:24 +0200)]
ubifs: Convert to separately allocated bdi

Allocate struct backing_dev_info separately instead of embedding it
inside the superblock. This unifies handling of bdi among users.

CC: Richard Weinberger <richard@nod.at>
CC: Artem Bityutskiy <dedekind1@gmail.com>
CC: Adrian Hunter <adrian.hunter@intel.com>
CC: linux-mtd@lists.infradead.org
Acked-by: Richard Weinberger <richard@nod.at>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agonfs: Convert to separately allocated bdi
Jan Kara [Wed, 12 Apr 2017 10:24:45 +0000 (12:24 +0200)]
nfs: Convert to separately allocated bdi

Allocate struct backing_dev_info separately instead of embedding it
inside the superblock. This unifies handling of bdi among users.

CC: Anna Schumaker <anna.schumaker@netapp.com>
CC: linux-nfs@vger.kernel.org
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Acked-by: Trond Myklebust <trond.myklebust@primarydata.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoncpfs: Convert to separately allocated bdi
Jan Kara [Wed, 12 Apr 2017 10:24:44 +0000 (12:24 +0200)]
ncpfs: Convert to separately allocated bdi

Allocate struct backing_dev_info separately instead of embedding it
inside the superblock. This unifies handling of bdi among users.

CC: Petr Vandrovec <petr@vandrovec.name>
Acked-by: Petr Vandrovec <petr@vandrovec.name>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agonilfs2: Convert to properly refcounting bdi
Jan Kara [Wed, 12 Apr 2017 10:24:43 +0000 (12:24 +0200)]
nilfs2: Convert to properly refcounting bdi

Similarly to set_bdev_super() NILFS2 just used block device reference to
bdi. Convert it to properly getting bdi reference. The reference will
get automatically dropped on superblock destruction.

CC: linux-nilfs@vger.kernel.org
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Acked-by: Ryusuke Konishi <konishi.ryusuke@lab.ntt.co.jp>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agogfs2: Convert to properly refcounting bdi
Jan Kara [Wed, 12 Apr 2017 10:24:42 +0000 (12:24 +0200)]
gfs2: Convert to properly refcounting bdi

Similarly to set_bdev_super() GFS2 just used block device reference to
bdi. Convert it to properly getting bdi reference. The reference will
get automatically dropped on superblock destruction.

CC: Steven Whitehouse <swhiteho@redhat.com>
CC: Bob Peterson <rpeterso@redhat.com>
CC: cluster-devel@redhat.com
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agofuse: Get rid of bdi_initialized
Jan Kara [Wed, 12 Apr 2017 10:24:41 +0000 (12:24 +0200)]
fuse: Get rid of bdi_initialized

It is not needed anymore since bdi is initialized whenever superblock
exists.

CC: Miklos Szeredi <miklos@szeredi.hu>
CC: linux-fsdevel@vger.kernel.org
Suggested-by: Miklos Szeredi <mszeredi@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agofuse: Convert to separately allocated bdi
Jan Kara [Wed, 12 Apr 2017 10:24:40 +0000 (12:24 +0200)]
fuse: Convert to separately allocated bdi

Allocate struct backing_dev_info separately instead of embedding it
inside the superblock. This unifies handling of bdi among users.

CC: Miklos Szeredi <miklos@szeredi.hu>
CC: linux-fsdevel@vger.kernel.org
Acked-by: Miklos Szeredi <mszeredi@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoexofs: Convert to separately allocated bdi
Jan Kara [Wed, 12 Apr 2017 10:24:39 +0000 (12:24 +0200)]
exofs: Convert to separately allocated bdi

Allocate struct backing_dev_info separately instead of embedding it
inside the superblock. This unifies handling of bdi among users.

CC: Boaz Harrosh <ooo@electrozaur.com>
CC: Benny Halevy <bhalevy@primarydata.com>
Acked-by: Boaz Harrosh <ooo@electrozaur.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agocoda: Convert to separately allocated bdi
Jan Kara [Wed, 12 Apr 2017 10:24:38 +0000 (12:24 +0200)]
coda: Convert to separately allocated bdi

Allocate struct backing_dev_info separately instead of embedding it
inside the superblock. This unifies handling of bdi among users.

CC: Jan Harkes <jaharkes@cs.cmu.edu>
CC: coda@cs.cmu.edu
CC: codalist@coda.cs.cmu.edu
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agomtd: Convert to dynamically allocated bdi infrastructure
Jan Kara [Wed, 12 Apr 2017 10:24:37 +0000 (12:24 +0200)]
mtd: Convert to dynamically allocated bdi infrastructure

MTD already allocates backing_dev_info dynamically. Convert it to use
generic infrastructure for this including proper refcounting. We drop
mtd->backing_dev_info as its only use was to pass mtd_bdi pointer from
one file into another and if we wanted to keep that in a clean way, we'd
have to make mtd hold and drop bdi reference as needed which seems
pointless for passing one global pointer...

CC: David Woodhouse <dwmw2@infradead.org>
CC: Brian Norris <computersforpeace@gmail.com>
CC: linux-mtd@lists.infradead.org
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoafs: Convert to separately allocated bdi
Jan Kara [Wed, 12 Apr 2017 10:24:36 +0000 (12:24 +0200)]
afs: Convert to separately allocated bdi

Allocate struct backing_dev_info separately instead of embedding it
inside the superblock. This unifies handling of bdi among users.

CC: David Howells <dhowells@redhat.com>
CC: linux-afs@lists.infradead.org
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoecryptfs: Convert to separately allocated bdi
Jan Kara [Wed, 12 Apr 2017 10:24:35 +0000 (12:24 +0200)]
ecryptfs: Convert to separately allocated bdi

Allocate struct backing_dev_info separately instead of embedding it
inside the superblock. This unifies handling of bdi among users.

CC: Tyler Hicks <tyhicks@canonical.com>
CC: ecryptfs@vger.kernel.org
Acked-by: Tyler Hicks <tyhicks@canonical.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agocifs: Convert to separately allocated bdi
Jan Kara [Wed, 12 Apr 2017 10:24:34 +0000 (12:24 +0200)]
cifs: Convert to separately allocated bdi

Allocate struct backing_dev_info separately instead of embedding it
inside superblock. This unifies handling of bdi among users.

CC: Steve French <sfrench@samba.org>
CC: linux-cifs@vger.kernel.org
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoceph: Convert to separately allocated bdi
Jan Kara [Wed, 12 Apr 2017 10:24:33 +0000 (12:24 +0200)]
ceph: Convert to separately allocated bdi

Allocate struct backing_dev_info separately instead of embedding it
inside client structure. This unifies handling of bdi among users.

CC: Ilya Dryomov <idryomov@gmail.com>
CC: "Yan, Zheng" <zyan@redhat.com>
CC: Sage Weil <sage@redhat.com>
CC: ceph-devel@vger.kernel.org
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agobtrfs: Convert to separately allocated bdi
Jan Kara [Wed, 12 Apr 2017 10:24:32 +0000 (12:24 +0200)]
btrfs: Convert to separately allocated bdi

Allocate struct backing_dev_info separately instead of embedding it
inside superblock. This unifies handling of bdi among users.

CC: Chris Mason <clm@fb.com>
CC: Josef Bacik <jbacik@fb.com>
CC: David Sterba <dsterba@suse.com>
CC: linux-btrfs@vger.kernel.org
Reviewed-by: Liu Bo <bo.li.liu@oracle.com>
Reviewed-by: David Sterba <dsterba@suse.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years ago9p: Convert to separately allocated bdi
Jan Kara [Wed, 12 Apr 2017 10:24:31 +0000 (12:24 +0200)]
9p: Convert to separately allocated bdi

Allocate struct backing_dev_info separately instead of embedding it
inside session. This unifies handling of bdi among users.

CC: Eric Van Hensbergen <ericvh@gmail.com>
CC: Ron Minnich <rminnich@sandia.gov>
CC: Latchesar Ionkov <lucho@ionkov.net>
CC: v9fs-developer@lists.sourceforge.net
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agolustre: Convert to separately allocated bdi
Jan Kara [Wed, 12 Apr 2017 10:24:30 +0000 (12:24 +0200)]
lustre: Convert to separately allocated bdi

Allocate struct backing_dev_info separately instead of embedding it
inside superblock. This unifies handling of bdi among users.

CC: Oleg Drokin <oleg.drokin@intel.com>
CC: Andreas Dilger <andreas.dilger@intel.com>
CC: James Simmons <jsimmons@infradead.org>
CC: lustre-devel@lists.lustre.org
Reviewed-by: Andreas Dilger <andreas.dilger@intel.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agofs: Get proper reference for s_bdi
Jan Kara [Wed, 12 Apr 2017 10:24:29 +0000 (12:24 +0200)]
fs: Get proper reference for s_bdi

So far we just relied on block device to hold a bdi reference for us
while the filesystem is mounted. While that works perfectly fine, it is
a bit awkward that we have a pointer to a refcounted structure in the
superblock without proper reference. So make s_bdi hold a proper
reference to block device's BDI. No filesystem using mount_bdev()
actually changes s_bdi so this is safe and will make bdev filesystems
work the same way as filesystems needing to set up their private bdi.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agofs: Provide infrastructure for dynamic BDIs in filesystems
Jan Kara [Wed, 12 Apr 2017 10:24:28 +0000 (12:24 +0200)]
fs: Provide infrastructure for dynamic BDIs in filesystems

Provide helper functions for setting up dynamically allocated
backing_dev_info structures for filesystems and cleaning them up on
superblock destruction.

CC: linux-mtd@lists.infradead.org
CC: linux-nfs@vger.kernel.org
CC: Petr Vandrovec <petr@vandrovec.name>
CC: linux-nilfs@vger.kernel.org
CC: cluster-devel@redhat.com
CC: osd-dev@open-osd.org
CC: codalist@coda.cs.cmu.edu
CC: linux-afs@lists.infradead.org
CC: ecryptfs@vger.kernel.org
CC: linux-cifs@vger.kernel.org
CC: ceph-devel@vger.kernel.org
CC: linux-btrfs@vger.kernel.org
CC: v9fs-developer@lists.sourceforge.net
CC: lustre-devel@lists.lustre.org
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agobdi: Export bdi_alloc_node() and bdi_put()
Jan Kara [Wed, 12 Apr 2017 10:24:27 +0000 (12:24 +0200)]
bdi: Export bdi_alloc_node() and bdi_put()

MTD will want to call bdi_alloc_node() and bdi_put() directly. Export
these functions.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock: Unregister bdi on last reference drop
Jan Kara [Wed, 12 Apr 2017 10:24:26 +0000 (12:24 +0200)]
block: Unregister bdi on last reference drop

Most users will want to unregister bdi when dropping last reference to a
bdi. Only a few users (like block devices) want to play more complex
tricks with bdi registration and unregistration. So unregister bdi when
the last reference to bdi is dropped and just make sure we don't
unregister the bdi the second time if it is already unregistered.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agobdi: Provide bdi_register_va() and bdi_alloc()
Jan Kara [Wed, 12 Apr 2017 10:24:25 +0000 (12:24 +0200)]
bdi: Provide bdi_register_va() and bdi_alloc()

Add function that registers bdi and takes va_list instead of variable
number of arguments.

Add bdi_alloc() as simple wrapper for NUMA-unaware users allocating BDI.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblk-throttle: fix unused variable warning with BLK_DEV_THROTTLING_LOW=n
Jens Axboe [Thu, 20 Apr 2017 15:41:36 +0000 (09:41 -0600)]
blk-throttle: fix unused variable warning with BLK_DEV_THROTTLING_LOW=n

We trigger this warning:

block/blk-throttle.c: In function ‘blk_throtl_bio’:
block/blk-throttle.c:2042:6: warning: variable ‘ret’ set but not used [-Wunused-but-set-variable]
  int ret;
      ^~~

since we only assign 'ret' if BLK_DEV_THROTTLING_LOW is off, we never
check it.

Reported-by: Bart Van Assche <bart.vanassche@sandisk.com>
Reviewed-by: Bart Van Assche <bart.vanassche@sandisk.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agobfq: fix compile error if CONFIG_CGROUPS=n
Jens Axboe [Thu, 20 Apr 2017 15:37:05 +0000 (09:37 -0600)]
bfq: fix compile error if CONFIG_CGROUPS=n

If we don't have CGROUPS enabled, the compile ends in the
following misery:

In file included from ../block/bfq-iosched.c:105:0:
../block/bfq-iosched.h:819:22: error: array type has incomplete element type
 extern struct cftype bfq_blkcg_legacy_files[];
                      ^
../block/bfq-iosched.h:820:22: error: array type has incomplete element type
 extern struct cftype bfq_blkg_files[];
                      ^

Move the declarations under the right ifdef.

Reported-by: Randy Dunlap <rdunlap@infradead.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock, bfq: don't dereference bic before null checking it
Colin Ian King [Thu, 20 Apr 2017 14:07:18 +0000 (15:07 +0100)]
block, bfq: don't dereference bic before null checking it

The call to bfq_check_ioprio_change will dereference bic, however,
the null check for bic is after this call.  Move the the null
check on bic to before the call to avoid any potential null
pointer dereference issues.

Detected by CoverityScan, CID#1430138 ("Dereference before null check")

Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoligtnvm: fix double blk_put_queue on same queue
Rakesh Pandit [Thu, 20 Apr 2017 14:17:47 +0000 (08:17 -0600)]
ligtnvm: fix double blk_put_queue on same queue

On an error path in NVM_DEV_CREATE ioctl blk_put_queue is being called
twice: one via blk_cleanup_queue and another via put_disk.  Straight fix
seems to remove queue pointer so that disk_release never ends up caling
blk_put_queue again.

  [  391.808827] WARNING: CPU: 1 PID: 1250 at lib/refcount.c:128 refcount_sub_and_test+0x70/0x80
  [  391.808830] refcount_t: underflow; use-after-free.
  [ 391.808832] Modules linked in: nf_conntrack_netbios_ns............
  [  391.809052] CPU: 1 PID: 1250 Comm: nvme Not tainted.........
  [  391.809057] Hardware name: QEMU Standard PC (i440FX + PIIX, 1996),
             BIOS rel-1.7.5-0-ge51488c-20140602_164612-nilsson.home.kraxel.org 04/01/2014
  [  391.809060] Call Trace:
  [  391.809079]  dump_stack+0x63/0x86
  [  391.809094]  __warn+0xcb/0xf0
  [  391.809103]  warn_slowpath_fmt+0x5f/0x80
  [  391.809118]  refcount_sub_and_test+0x70/0x80
  [  391.809125]  refcount_dec_and_test+0x11/0x20
  [  391.809136]  kobject_put+0x1f/0x60
  [  391.809149]  blk_put_queue+0x15/0x20
  [  391.809159]  disk_release+0xae/0xf0
  [  391.809172]  device_release+0x32/0x90
  [  391.809184]  kobject_release+0x6a/0x170
  [  391.809196]  kobject_put+0x2f/0x60
  [  391.809206]  put_disk+0x17/0x20
  [  391.809219]  nvm_ioctl_dev_create.isra.16+0x897/0xa30
  [  391.809236]  nvm_ctl_ioctl+0x23c/0x4c0
  [  391.809248]  do_vfs_ioctl+0xa3/0x5f0
  [  391.809258]  SyS_ioctl+0x79/0x90
  [  391.809271]  entry_SYSCALL_64_fastpath+0x1a/0xa9
  [  391.809280] RIP: 0033:0x7f5d3ef363c7
  [  391.809286] RSP: 002b:00007ffc72ed8d78 EFLAGS: 00000206 ORIG_RAX: 0000000000000010
  [  391.809296] RAX: ffffffffffffffda RBX: 00007ffc72edb552 RCX: 00007f5d3ef363c7
  [  391.809301] RDX: 00007ffc72ed8d90 RSI: 0000000040804c22 RDI: 0000000000000003
  [  391.809306] RBP: 0000000000000001 R08: 0000000000000020 R09: 0000000000000001
  [  391.809311] R10: 000000000000053f R11: 0000000000000206 R12: 0000000000000000
  [  391.809316] R13: 0000000000000000 R14: 00007ffc72edb58d R15: 00007ffc72edb581

Signed-off-by: Rakesh Pandit <rakesh@tuxera.com>
Reviewed-by: Matias Bjørling <matias@cnexlabs.com>
Fixes: 7d1ef2f408ab "lightnvm: fix cleanup order of disk on init error"
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock: Optimize ioprio_best()
Bart Van Assche [Wed, 19 Apr 2017 21:01:28 +0000 (14:01 -0700)]
block: Optimize ioprio_best()

Since ioprio_best() translates IOPRIO_CLASS_NONE into IOPRIO_CLASS_BE
and since lower numerical priority values represent a higher priority
a simple numerical comparison is sufficient.

Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Reviewed-by: Adam Manzanares <adam.manzanares@wdc.com>
Tested-by: Adam Manzanares <adam.manzanares@wdc.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Cc: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock: Inline blk_rq_set_prio()
Bart Van Assche [Wed, 19 Apr 2017 21:01:27 +0000 (14:01 -0700)]
block: Inline blk_rq_set_prio()

Since only a single caller remains, inline blk_rq_set_prio(). Initialize
req->ioprio even if no I/O priority has been set in the bio nor in the
I/O context.

Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Reviewed-by: Adam Manzanares <adam.manzanares@wdc.com>
Tested-by: Adam Manzanares <adam.manzanares@wdc.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Cc: Matias Bjørling <m@bjorling.me>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agolightnvm: Use blk_init_request_from_bio() instead of open-coding it
Bart Van Assche [Wed, 19 Apr 2017 21:01:26 +0000 (14:01 -0700)]
lightnvm: Use blk_init_request_from_bio() instead of open-coding it

This patch changes the behavior of the lightnvm driver as follows:
* REQ_FAILFAST_MASK is set for read-ahead requests.
* If no I/O priority has been set in the bio, the I/O priority is
  copied from the I/O context.
* The rq_disk member is initialized if bio->bi_bdev != NULL.
* The bio sector offset is copied into req->__sector instead of
  retaining the value -1 set by blk_mq_alloc_request().
* req->errors is initialized to zero.

Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Cc: Matias Bjørling <m@bjorling.me>
Cc: Adam Manzanares <adam.manzanares@wdc.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agonull_blk: Use blk_init_request_from_bio() instead of open-coding it
Bart Van Assche [Wed, 19 Apr 2017 21:01:25 +0000 (14:01 -0700)]
null_blk: Use blk_init_request_from_bio() instead of open-coding it

This patch changes the behavior of the null_blk driver for the
LightNVM mode as follows:
* REQ_FAILFAST_MASK is set for read-ahead requests.
* If no I/O priority has been set in the bio, the I/O priority is
  copied from the I/O context.
* The rq_disk member is initialized if bio->bi_bdev != NULL.
* req->errors is initialized to zero.

Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Cc: Matias Bjørling <m@bjorling.me>
Cc: Adam Manzanares <adam.manzanares@wdc.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock: Export blk_init_request_from_bio()
Bart Van Assche [Wed, 19 Apr 2017 21:01:24 +0000 (14:01 -0700)]
block: Export blk_init_request_from_bio()

Export this function such that it becomes available to block
drivers.

Signed-off-by: Bart Van Assche <bart.vanassche@sandisk.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
Cc: Matias Bjørling <m@bjorling.me>
Cc: Adam Manzanares <adam.manzanares@wdc.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agolightnvm: assume 64-bit lba numbers
Arnd Bergmann [Wed, 19 Apr 2017 17:39:13 +0000 (19:39 +0200)]
lightnvm: assume 64-bit lba numbers

The driver uses both u64 and sector_t to refer to offsets, and assigns between the
two. This causes one harmless warning when sector_t is 32-bit:

drivers/lightnvm/pblk-rb.c: In function 'pblk_rb_write_entry_gc':
include/linux/lightnvm.h:215:20: error: large integer implicitly truncated to unsigned type [-Werror=overflow]
drivers/lightnvm/pblk-rb.c:324:22: note: in expansion of macro 'ADDR_EMPTY'

As the driver is already doing this inconsistently, changing the type
won't make it worse and is an easy way to avoid the warning.

Fixes: a4bd217b4326 ("lightnvm: physical block device (pblk) target")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock: make __blk_end_bidi_request private
Christoph Hellwig [Wed, 12 Apr 2017 10:13:59 +0000 (12:13 +0200)]
block: make __blk_end_bidi_request private

blk_insert_flush should be using __blk_end_request to start with.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock: remove blk_end_request_cur
Christoph Hellwig [Wed, 12 Apr 2017 10:13:58 +0000 (12:13 +0200)]
block: remove blk_end_request_cur

This function is not used anywhere in the kernel.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock: remove blk_end_request_err and __blk_end_request_err
Christoph Hellwig [Wed, 12 Apr 2017 10:13:57 +0000 (12:13 +0200)]
block: remove blk_end_request_err and __blk_end_request_err

Both functions are entirely unused.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock: remove the osdblk driver
Christoph Hellwig [Wed, 12 Apr 2017 16:01:06 +0000 (18:01 +0200)]
block: remove the osdblk driver

This was just a proof of concept user for the SCSI OSD library, and
never had any real users.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Acked-by: Boaz Harrosh <ooo@electrozaur.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock: Make writeback throttling defaults consistent for SQ devices
Jan Kara [Wed, 19 Apr 2017 09:33:27 +0000 (11:33 +0200)]
block: Make writeback throttling defaults consistent for SQ devices

When CFQ is used as an elevator, it disables writeback throttling
because they don't play well together. Later when a different elevator
is chosen for the device, writeback throttling doesn't get enabled
again as it should. Make sure CFQ enables writeback throttling (if it
should be enabled by default) when we switch from it to another IO
scheduler.

Signed-off-by: Jan Kara <jack@suse.cz>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock, bfq: split bfq-iosched.c into multiple source files
Paolo Valente [Wed, 19 Apr 2017 14:48:24 +0000 (08:48 -0600)]
block, bfq: split bfq-iosched.c into multiple source files

The BFQ I/O scheduler features an optimal fair-queuing
(proportional-share) scheduling algorithm, enriched with several
mechanisms to boost throughput and reduce latency for interactive and
real-time applications. This makes BFQ a large and complex piece of
code. This commit addresses this issue by splitting BFQ into three
main, independent components, and by moving each component into a
separate source file:
1. Main algorithm: handles the interaction with the kernel, and
decides which requests to dispatch; it uses the following two further
components to achieve its goals.
2. Scheduling engine (Hierarchical B-WF2Q+ scheduling algorithm):
computes the schedule, using weights and budgets provided by the above
component.
3. cgroups support: handles group operations (creation, destruction,
move, ...).

Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock, bfq: remove all get and put of I/O contexts
Paolo Valente [Wed, 12 Apr 2017 16:23:21 +0000 (18:23 +0200)]
block, bfq: remove all get and put of I/O contexts

When a bfq queue is set in service and when it is merged, a reference
to the I/O context associated with the queue is taken. This reference
is then released when the queue is deselected from service or
split. More precisely, the release of the reference is postponed to
when the scheduler lock is released, to avoid nesting between the
scheduler and the I/O-context lock. In fact, such nesting would lead
to deadlocks, because of other code paths that take the same locks in
the opposite order. This postponing of I/O-context releases does
complicate code.

This commit addresses these issue by modifying involved operations in
such a way to not need to get the above I/O-context references any
more. Then it also removes any get and release of these references.

Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock, bfq: handle bursts of queue activations
Arianna Avanzini [Wed, 12 Apr 2017 16:23:20 +0000 (18:23 +0200)]
block, bfq: handle bursts of queue activations

Many popular I/O-intensive services or applications spawn or
reactivate many parallel threads/processes during short time
intervals. Examples are systemd during boot or git grep.  These
services or applications benefit mostly from a high throughput: the
quicker the I/O generated by their processes is cumulatively served,
the sooner the target job of these services or applications gets
completed. As a consequence, it is almost always counterproductive to
weight-raise any of the queues associated to the processes of these
services or applications: in most cases it would just lower the
throughput, mainly because weight-raising also implies device idling.

To address this issue, an I/O scheduler needs, first, to detect which
queues are associated with these services or applications. In this
respect, we have that, from the I/O-scheduler standpoint, these
services or applications cause bursts of activations, i.e.,
activations of different queues occurring shortly after each
other. However, a shorter burst of activations may be caused also by
the start of an application that does not consist in a lot of parallel
I/O-bound threads (see the comments on the function bfq_handle_burst
for details).

In view of these facts, this commit introduces:
1) an heuristic to detect (only) bursts of queue activations caused by
   services or applications consisting in many parallel I/O-bound
   threads;
2) the prevention of device idling and weight-raising for the queues
   belonging to these bursts.

Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock, bfq: boost the throughput with random I/O on NCQ-capable HDDs
Paolo Valente [Wed, 12 Apr 2017 16:23:19 +0000 (18:23 +0200)]
block, bfq: boost the throughput with random I/O on NCQ-capable HDDs

This patch is basically the counterpart, for NCQ-capable rotational
devices, of the previous patch. Exactly as the previous patch does on
flash-based devices and for any workload, this patch disables device
idling on rotational devices, but only for random I/O. In fact, only
with these queues disabling idling boosts the throughput on
NCQ-capable rotational devices. To not break service guarantees,
idling is disabled for NCQ-enabled rotational devices only when the
same symmetry conditions considered in the previous patches hold.

Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock, bfq: boost the throughput on NCQ-capable flash-based devices
Paolo Valente [Wed, 12 Apr 2017 16:23:18 +0000 (18:23 +0200)]
block, bfq: boost the throughput on NCQ-capable flash-based devices

This patch boosts the throughput on NCQ-capable flash-based devices,
while still preserving latency guarantees for interactive and soft
real-time applications. The throughput is boosted by just not idling
the device when the in-service queue remains empty, even if the queue
is sync and has a non-null idle window. This helps to keep the drive's
internal queue full, which is necessary to achieve maximum
performance. This solution to boost the throughput is a port of
commits a68bbdd and f7d7b7a for CFQ.

As already highlighted in a previous patch, allowing the device to
prefetch and internally reorder requests trivially causes loss of
control on the request service order, and hence on service guarantees.
Fortunately, as discussed in detail in the comments on the function
bfq_bfqq_may_idle(), if every process has to receive the same
fraction of the throughput, then the service order enforced by the
internal scheduler of a flash-based device is relatively close to that
enforced by BFQ. In particular, it is close enough to let service
guarantees be substantially preserved.

Things change in an asymmetric scenario, i.e., if not every process
has to receive the same fraction of the throughput. In this case, to
guarantee the desired throughput distribution, the device must be
prevented from prefetching requests. This is exactly what this patch
does in asymmetric scenarios.

Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock, bfq: reduce idling only in symmetric scenarios
Arianna Avanzini [Wed, 12 Apr 2017 16:23:17 +0000 (18:23 +0200)]
block, bfq: reduce idling only in symmetric scenarios

A seeky queue (i..e, a queue containing random requests) is assigned a
very small device-idling slice, for throughput issues. Unfortunately,
given the process associated with a seeky queue, this behavior causes
the following problem: if the process, say P, performs sync I/O and
has a higher weight than some other processes doing I/O and associated
with non-seeky queues, then BFQ may fail to guarantee to P its
reserved share of the throughput. The reason is that idling is key
for providing service guarantees to processes doing sync I/O [1].

This commit addresses this issue by allowing the device-idling slice
to be reduced for a seeky queue only if the scenario happens to be
symmetric, i.e., if all the queues are to receive the same share of
the throughput.

[1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
    Scheduler", Proceedings of the First Workshop on Mobile System
    Technologies (MST-2015), May 2015.
    http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf

Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Riccardo Pizzetti <riccardo.pizzetti@gmail.com>
Signed-off-by: Samuele Zecchini <samuele.zecchini92@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock, bfq: add Early Queue Merge (EQM)
Arianna Avanzini [Wed, 12 Apr 2017 16:23:16 +0000 (18:23 +0200)]
block, bfq: add Early Queue Merge (EQM)

A set of processes may happen to perform interleaved reads, i.e.,
read requests whose union would give rise to a sequential read pattern.
There are two typical cases: first, processes reading fixed-size chunks
of data at a fixed distance from each other; second, processes reading
variable-size chunks at variable distances. The latter case occurs for
example with QEMU, which splits the I/O generated by a guest into
multiple chunks, and lets these chunks be served by a pool of I/O
threads, iteratively assigning the next chunk of I/O to the first
available thread. CFQ denotes as 'cooperating' a set of processes that
are doing interleaved I/O, and when it detects cooperating processes,
it merges their queues to obtain a sequential I/O pattern from the union
of their I/O requests, and hence boost the throughput.

Unfortunately, in the following frequent case, the mechanism
implemented in CFQ for detecting cooperating processes and merging
their queues is not responsive enough to handle also the fluctuating
I/O pattern of the second type of processes. Suppose that one process
of the second type issues a request close to the next request to serve
of another process of the same type. At that time the two processes
would be considered as cooperating. But, if the request issued by the
first process is to be merged with some other already-queued request,
then, from the moment at which this request arrives, to the moment
when CFQ controls whether the two processes are cooperating, the two
processes are likely to be already doing I/O in distant zones of the
disk surface or device memory.

CFQ uses however preemption to get a sequential read pattern out of
the read requests performed by the second type of processes too.  As a
consequence, CFQ uses two different mechanisms to achieve the same
goal: boosting the throughput with interleaved I/O.

This patch introduces Early Queue Merge (EQM), a unified mechanism to
get a sequential read pattern with both types of processes. The main
idea is to immediately check whether a newly-arrived request lets some
pair of processes become cooperating, both in the case of actual
request insertion and, to be responsive with the second type of
processes, in the case of request merge. Both types of processes are
then handled by just merging their queues.

Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Mauro Andreolini <mauro.andreolini@unimore.it>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock, bfq: reduce latency during request-pool saturation
Paolo Valente [Wed, 12 Apr 2017 16:23:15 +0000 (18:23 +0200)]
block, bfq: reduce latency during request-pool saturation

This patch introduces an heuristic that reduces latency when the
I/O-request pool is saturated. This goal is achieved by disabling
device idling, for non-weight-raised queues, when there are weight-
raised queues with pending or in-flight requests. In fact, as
explained in more detail in the comment on the function
bfq_bfqq_may_idle(), this reduces the rate at which processes
associated with non-weight-raised queues grab requests from the pool,
thereby increasing the probability that processes associated with
weight-raised queues get a request immediately (or at least soon) when
they need one. Along the same line, if there are weight-raised queues,
then this patch halves the service rate of async (write) requests for
non-weight-raised queues.

Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock, bfq: preserve a low latency also with NCQ-capable drives
Paolo Valente [Wed, 12 Apr 2017 16:23:14 +0000 (18:23 +0200)]
block, bfq: preserve a low latency also with NCQ-capable drives

I/O schedulers typically allow NCQ-capable drives to prefetch I/O
requests, as NCQ boosts the throughput exactly by prefetching and
internally reordering requests.

Unfortunately, as discussed in detail and shown experimentally in [1],
this may cause fairness and latency guarantees to be violated. The
main problem is that the internal scheduler of an NCQ-capable drive
may postpone the service of some unlucky (prefetched) requests as long
as it deems serving other requests more appropriate to boost the
throughput.

This patch addresses this issue by not disabling device idling for
weight-raised queues, even if the device supports NCQ. This allows BFQ
to start serving a new queue, and therefore allows the drive to
prefetch new requests, only after the idling timeout expires. At that
time, all the outstanding requests of the expired queue have been most
certainly served.

[1] P. Valente and M. Andreolini, "Improving Application
    Responsiveness with the BFQ Disk I/O Scheduler", Proceedings of
    the 5th Annual International Systems and Storage Conference
    (SYSTOR '12), June 2012.
    Slightly extended version:
    http://algogroup.unimore.it/people/paolo/disk_sched/bfq-v1-suite-
results.pdf

Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock, bfq: reduce I/O latency for soft real-time applications
Paolo Valente [Wed, 12 Apr 2017 16:23:13 +0000 (18:23 +0200)]
block, bfq: reduce I/O latency for soft real-time applications

To guarantee a low latency also to the I/O requests issued by soft
real-time applications, this patch introduces a further heuristic,
which weight-raises (in the sense explained in the previous patch)
also the queues associated to applications deemed as soft real-time.

To be deemed as soft real-time, an application must meet two
requirements.  First, the application must not require an average
bandwidth higher than the approximate bandwidth required to playback
or record a compressed high-definition video. Second, the request
pattern of the application must be isochronous, i.e., after issuing a
request or a batch of requests, the application must stop issuing new
requests until all its pending requests have been completed. After
that, the application may issue a new batch, and so on.

As for the second requirement, it is critical to require also that,
after all the pending requests of the application have been completed,
an adequate minimum amount of time elapses before the application
starts issuing new requests. This prevents also greedy (i.e.,
I/O-bound) applications from being incorrectly deemed, occasionally,
as soft real-time. In fact, if *any amount of time* is fine, then even
a greedy application may, paradoxically, meet both the above
requirements, if: (1) the application performs random I/O and/or the
device is slow, and (2) the CPU load is high. The reason is the
following.  First, if condition (1) is true, then, during the service
of the application, the throughput may be low enough to let the
application meet the bandwidth requirement.  Second, if condition (2)
is true as well, then the application may occasionally behave in an
apparently isochronous way, because it may simply stop issuing
requests while the CPUs are busy serving other processes.

To address this issue, the heuristic leverages the simple fact that
greedy applications issue *all* their requests as quickly as they can,
whereas soft real-time applications spend some time processing data
after each batch of requests is completed. In particular, the
heuristic works as follows. First, according to the above isochrony
requirement, the heuristic checks whether an application may be soft
real-time, thereby giving to the application the opportunity to be
deemed as such, only when both the following two conditions happen to
hold: 1) the queue associated with the application has expired and is
empty, 2) there is no outstanding request of the application.

Suppose that both conditions hold at time, say, t_c and that the
application issues its next request at time, say, t_i. At time t_c the
heuristic computes the next time instant, called soft_rt_next_start in
the code, such that, only if t_i >= soft_rt_next_start, then both the
next conditions will hold when the application issues its next
request: 1) the application will meet the above bandwidth requirement,
2) a given minimum time interval, say Delta, will have elapsed from
time t_c (so as to filter out greedy application).

The current value of Delta is a little bit higher than the value that
we have found, experimentally, to be adequate on a real,
general-purpose machine. In particular we had to increase Delta to
make the filter quite precise also in slower, embedded systems, and in
KVM/QEMU virtual machines (details in the comments on the code).

If the application actually issues its next request after time
soft_rt_next_start, then its associated queue will be weight-raised
for a relatively short time interval. If, during this time interval,
the application proves again to meet the bandwidth and isochrony
requirements, then the end of the weight-raising period for the queue
is moved forward, and so on. Note that an application whose associated
queue never happens to be empty when it expires will never have the
opportunity to be deemed as soft real-time.

Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock, bfq: improve responsiveness
Paolo Valente [Wed, 12 Apr 2017 16:23:12 +0000 (18:23 +0200)]
block, bfq: improve responsiveness

This patch introduces a simple heuristic to load applications quickly,
and to perform the I/O requested by interactive applications just as
quickly. To this purpose, both a newly-created queue and a queue
associated with an interactive application (we explain in a moment how
BFQ decides whether the associated application is interactive),
receive the following two special treatments:

1) The weight of the queue is raised.

2) The queue unconditionally enjoys device idling when it empties; in
fact, if the requests of a queue are sync, then performing device
idling for the queue is a necessary condition to guarantee that the
queue receives a fraction of the throughput proportional to its weight
(see [1] for details).

For brevity, we call just weight-raising the combination of these
two preferential treatments. For a newly-created queue,
weight-raising starts immediately and lasts for a time interval that:
1) depends on the device speed and type (rotational or
non-rotational), and 2) is equal to the time needed to load (start up)
a large-size application on that device, with cold caches and with no
additional workload.

Finally, as for guaranteeing a fast execution to interactive,
I/O-related tasks (such as opening a file), consider that any
interactive application blocks and waits for user input both after
starting up and after executing some task. After a while, the user may
trigger new operations, after which the application stops again, and
so on. Accordingly, the low-latency heuristic weight-raises again a
queue in case it becomes backlogged after being idle for a
sufficiently long (configurable) time. The weight-raising then lasts
for the same time as for a just-created queue.

According to our experiments, the combination of this low-latency
heuristic and of the improvements described in the previous patch
allows BFQ to guarantee a high application responsiveness.

[1] P. Valente, A. Avanzini, "Evolution of the BFQ Storage I/O
    Scheduler", Proceedings of the First Workshop on Mobile System
    Technologies (MST-2015), May 2015.
    http://algogroup.unimore.it/people/paolo/disk_sched/mst-2015.pdf

Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock, bfq: add more fairness with writes and slow processes
Paolo Valente [Wed, 12 Apr 2017 16:23:11 +0000 (18:23 +0200)]
block, bfq: add more fairness with writes and slow processes

This patch deals with two sources of unfairness, which can also cause
high latencies and throughput loss. The first source is related to
write requests. Write requests tend to starve read requests, basically
because, on one side, writes are slower than reads, whereas, on the
other side, storage devices confuse schedulers by deceptively
signaling the completion of write requests immediately after receiving
them. This patch addresses this issue by just throttling writes. In
particular, after a write request is dispatched for a queue, the
budget of the queue is decremented by the number of sectors to write,
multiplied by an (over)charge coefficient. The value of the
coefficient is the result of our tuning with different devices.

The second source of unfairness has to do with slowness detection:
when the in-service queue is expired, BFQ also controls whether the
queue has been "too slow", i.e., has consumed its last-assigned budget
at such a low rate that it would have been impossible to consume all
of this budget within the maximum time slice T_max (Subsec. 3.5 in
[1]). In this case, the queue is always (over)charged the whole
budget, to reduce its utilization of the device. Both this overcharge
and the slowness-detection criterion may cause unfairness.

First, always charging a full budget to a slow queue is too coarse. It
is much more accurate, and this patch lets BFQ do so, to charge an
amount of service 'equivalent' to the amount of time during which the
queue has been in service. As explained in more detail in the comments
on the code, this enables BFQ to provide time fairness among slow
queues.

Secondly, because of ZBR, a queue may be deemed as slow when its
associated process is performing I/O on the slowest zones of a
disk. However, unless the process is truly too slow, not reducing the
disk utilization of the queue is more profitable in terms of disk
throughput than the opposite. A similar problem is caused by logical
block mapping on non-rotational devices. For this reason, this patch
lets a queue be charged time, and not budget, only if the queue has
consumed less than 2/3 of its assigned budget. As an additional,
important benefit, this tolerance allows BFQ to preserve enough
elasticity to still perform bandwidth, and not time, distribution with
little unlucky or quasi-sequential processes.

Finally, for the same reasons as above, this patch makes slowness
detection itself much less harsh: a queue is deemed slow only if it
has consumed its budget at less than half of the peak rate.

[1] P. Valente and M. Andreolini, "Improving Application
    Responsiveness with the BFQ Disk I/O Scheduler", Proceedings of
    the 5th Annual International Systems and Storage Conference
    (SYSTOR '12), June 2012.
    Slightly extended version:
    http://algogroup.unimore.it/people/paolo/disk_sched/bfq-v1-suite-
results.pdf

Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock, bfq: modify the peak-rate estimator
Paolo Valente [Wed, 12 Apr 2017 16:23:10 +0000 (18:23 +0200)]
block, bfq: modify the peak-rate estimator

Unless the maximum budget B_max that BFQ can assign to a queue is set
explicitly by the user, BFQ automatically updates B_max. In
particular, BFQ dynamically sets B_max to the number of sectors that
can be read, at the current estimated peak rate, during the maximum
time, T_max, allowed before a budget timeout occurs. In formulas, if
we denote as R_est the estimated peak rate, then B_max = T_max ∗
R_est. Hence, the higher R_est is with respect to the actual device
peak rate, the higher the probability that processes incur budget
timeouts unjustly is. Besides, a too high value of B_max unnecessarily
increases the deviation from an ideal, smooth service.

Unfortunately, it is not trivial to estimate the peak rate correctly:
because of the presence of sw and hw queues between the scheduler and
the device components that finally serve I/O requests, it is hard to
say exactly when a given dispatched request is served inside the
device, and for how long. As a consequence, it is hard to know
precisely at what rate a given set of requests is actually served by
the device.

On the opposite end, the dispatch time of any request is trivially
available, and, from this piece of information, the "dispatch rate"
of requests can be immediately computed. So, the idea in the next
function is to use what is known, namely request dispatch times
(plus, when useful, request completion times), to estimate what is
unknown, namely in-device request service rate.

The main issue is that, because of the above facts, the rate at
which a certain set of requests is dispatched over a certain time
interval can vary greatly with respect to the rate at which the
same requests are then served. But, since the size of any
intermediate queue is limited, and the service scheme is lossless
(no request is silently dropped), the following obvious convergence
property holds: the number of requests dispatched MUST become
closer and closer to the number of requests completed as the
observation interval grows. This is the key property used in
this new version of the peak-rate estimator.

Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock, bfq: improve throughput boosting
Paolo Valente [Wed, 12 Apr 2017 16:23:09 +0000 (18:23 +0200)]
block, bfq: improve throughput boosting

The feedback-loop algorithm used by BFQ to compute queue (process)
budgets is basically a set of three update rules, one for each of the
main reasons why a queue may be expired. If many processes suddenly
switch from sporadic I/O to greedy and sequential I/O, then these
rules are quite slow to assign large budgets to these processes, and
hence to achieve a high throughput. On the opposite side, BFQ assigns
the maximum possible budget B_max to a just-created queue. This allows
a high throughput to be achieved immediately if the associated process
is I/O-bound and performs sequential I/O from the beginning. But it
also increases the worst-case latency experienced by the first
requests issued by the process, because the larger the budget of a
queue waiting for service is, the later the queue will be served by
B-WF2Q+ (Subsec 3.3 in [1]). This is detrimental for an interactive or
soft real-time application.

To tackle these throughput and latency problems, on one hand this
patch changes the initial budget value to B_max/2. On the other hand,
it re-tunes the three rules, adopting a more aggressive,
multiplicative increase/linear decrease scheme. This scheme trades
latency for throughput more than before, and tends to assign large
budgets quickly to processes that are or become I/O-bound. For two of
the expiration reasons, the new version of the rules also contains
some more little improvements, briefly described below.

*No more backlog.* In this case, the budget was larger than the number
of sectors actually read/written by the process before it stopped
doing I/O. Hence, to reduce latency for the possible future I/O
requests of the process, the old rule simply set the next budget to
the number of sectors actually consumed by the process. However, if
there are still outstanding requests, then the process may have not
yet issued its next request just because it is still waiting for the
completion of some of the still outstanding ones. If this sub-case
holds true, then the new rule, instead of decreasing the budget,
doubles it, proactively, in the hope that: 1) a larger budget will fit
the actual needs of the process, and 2) the process is sequential and
hence a higher throughput will be achieved by serving the process
longer after granting it access to the device.

*Budget timeout*. The original rule set the new budget to the maximum
value B_max, to maximize throughput and let all processes experiencing
budget timeouts receive the same share of the device time. In our
experiments we verified that this sudden jump to B_max did not provide
sensible benefits; rather it increased the latency of processes
performing sporadic and short I/O. The new rule only doubles the
budget.

[1] P. Valente and M. Andreolini, "Improving Application
    Responsiveness with the BFQ Disk I/O Scheduler", Proceedings of
    the 5th Annual International Systems and Storage Conference
    (SYSTOR '12), June 2012.
    Slightly extended version:
    http://algogroup.unimore.it/people/paolo/disk_sched/bfq-v1-suite-
results.pdf

Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock, bfq: add full hierarchical scheduling and cgroups support
Arianna Avanzini [Wed, 12 Apr 2017 16:23:08 +0000 (18:23 +0200)]
block, bfq: add full hierarchical scheduling and cgroups support

Add complete support for full hierarchical scheduling, with a cgroups
interface. Full hierarchical scheduling is implemented through the
'entity' abstraction: both bfq_queues, i.e., the internal BFQ queues
associated with processes, and groups are represented in general by
entities. Given the bfq_queues associated with the processes belonging
to a given group, the entities representing these queues are sons of
the entity representing the group. At higher levels, if a group, say
G, contains other groups, then the entity representing G is the parent
entity of the entities representing the groups in G.

Hierarchical scheduling is performed as follows: if the timestamps of
a leaf entity (i.e., of a bfq_queue) change, and such a change lets
the entity become the next-to-serve entity for its parent entity, then
the timestamps of the parent entity are recomputed as a function of
the budget of its new next-to-serve leaf entity. If the parent entity
belongs, in its turn, to a group, and its new timestamps let it become
the next-to-serve for its parent entity, then the timestamps of the
latter parent entity are recomputed as well, and so on. When a new
bfq_queue must be set in service, the reverse path is followed: the
next-to-serve highest-level entity is chosen, then its next-to-serve
child entity, and so on, until the next-to-serve leaf entity is
reached, and the bfq_queue that this entity represents is set in
service.

Writeback is accounted for on a per-group basis, i.e., for each group,
the async I/O requests of the processes of the group are enqueued in a
distinct bfq_queue, and the entity associated with this queue is a
child of the entity associated with the group.

Weights can be assigned explicitly to groups and processes through the
cgroups interface, differently from what happens, for single
processes, if the cgroups interface is not used (as explained in the
description of the previous patch). In particular, since each node has
a full scheduler, each group can be assigned its own weight.

Signed-off-by: Fabio Checconi <fchecconi@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoblock, bfq: introduce the BFQ-v0 I/O scheduler as an extra scheduler
Paolo Valente [Wed, 19 Apr 2017 14:29:02 +0000 (08:29 -0600)]
block, bfq: introduce the BFQ-v0 I/O scheduler as an extra scheduler

We tag as v0 the version of BFQ containing only BFQ's engine plus
hierarchical support. BFQ's engine is introduced by this commit, while
hierarchical support is added by next commit. We use the v0 tag to
distinguish this minimal version of BFQ from the versions containing
also the features and the improvements added by next commits. BFQ-v0
coincides with the version of BFQ submitted a few years ago [1], apart
from the introduction of preemption, described below.

BFQ is a proportional-share I/O scheduler, whose general structure,
plus a lot of code, are borrowed from CFQ.

- Each process doing I/O on a device is associated with a weight and a
  (bfq_)queue.

- BFQ grants exclusive access to the device, for a while, to one queue
  (process) at a time, and implements this service model by
  associating every queue with a budget, measured in number of
  sectors.

  - After a queue is granted access to the device, the budget of the
    queue is decremented, on each request dispatch, by the size of the
    request.

  - The in-service queue is expired, i.e., its service is suspended,
    only if one of the following events occurs: 1) the queue finishes
    its budget, 2) the queue empties, 3) a "budget timeout" fires.

    - The budget timeout prevents processes doing random I/O from
      holding the device for too long and dramatically reducing
      throughput.

    - Actually, as in CFQ, a queue associated with a process issuing
      sync requests may not be expired immediately when it empties. In
      contrast, BFQ may idle the device for a short time interval,
      giving the process the chance to go on being served if it issues
      a new request in time. Device idling typically boosts the
      throughput on rotational devices, if processes do synchronous
      and sequential I/O. In addition, under BFQ, device idling is
      also instrumental in guaranteeing the desired throughput
      fraction to processes issuing sync requests (see [2] for
      details).

      - With respect to idling for service guarantees, if several
        processes are competing for the device at the same time, but
        all processes (and groups, after the following commit) have
        the same weight, then BFQ guarantees the expected throughput
        distribution without ever idling the device. Throughput is
        thus as high as possible in this common scenario.

  - Queues are scheduled according to a variant of WF2Q+, named
    B-WF2Q+, and implemented using an augmented rb-tree to preserve an
    O(log N) overall complexity.  See [2] for more details. B-WF2Q+ is
    also ready for hierarchical scheduling. However, for a cleaner
    logical breakdown, the code that enables and completes
    hierarchical support is provided in the next commit, which focuses
    exactly on this feature.

  - B-WF2Q+ guarantees a tight deviation with respect to an ideal,
    perfectly fair, and smooth service. In particular, B-WF2Q+
    guarantees that each queue receives a fraction of the device
    throughput proportional to its weight, even if the throughput
    fluctuates, and regardless of: the device parameters, the current
    workload and the budgets assigned to the queue.

  - The last, budget-independence, property (although probably
    counterintuitive in the first place) is definitely beneficial, for
    the following reasons:

    - First, with any proportional-share scheduler, the maximum
      deviation with respect to an ideal service is proportional to
      the maximum budget (slice) assigned to queues. As a consequence,
      BFQ can keep this deviation tight not only because of the
      accurate service of B-WF2Q+, but also because BFQ *does not*
      need to assign a larger budget to a queue to let the queue
      receive a higher fraction of the device throughput.

    - Second, BFQ is free to choose, for every process (queue), the
      budget that best fits the needs of the process, or best
      leverages the I/O pattern of the process. In particular, BFQ
      updates queue budgets with a simple feedback-loop algorithm that
      allows a high throughput to be achieved, while still providing
      tight latency guarantees to time-sensitive applications. When
      the in-service queue expires, this algorithm computes the next
      budget of the queue so as to:

      - Let large budgets be eventually assigned to the queues
        associated with I/O-bound applications performing sequential
        I/O: in fact, the longer these applications are served once
        got access to the device, the higher the throughput is.

      - Let small budgets be eventually assigned to the queues
        associated with time-sensitive applications (which typically
        perform sporadic and short I/O), because, the smaller the
        budget assigned to a queue waiting for service is, the sooner
        B-WF2Q+ will serve that queue (Subsec 3.3 in [2]).

- Weights can be assigned to processes only indirectly, through I/O
  priorities, and according to the relation:
  weight = 10 * (IOPRIO_BE_NR - ioprio).
  The next patch provides, instead, a cgroups interface through which
  weights can be assigned explicitly.

- If several processes are competing for the device at the same time,
  but all processes and groups have the same weight, then BFQ
  guarantees the expected throughput distribution without ever idling
  the device. It uses preemption instead. Throughput is then much
  higher in this common scenario.

- ioprio classes are served in strict priority order, i.e.,
  lower-priority queues are not served as long as there are
  higher-priority queues.  Among queues in the same class, the
  bandwidth is distributed in proportion to the weight of each
  queue. A very thin extra bandwidth is however guaranteed to the Idle
  class, to prevent it from starving.

- If the strict_guarantees parameter is set (default: unset), then BFQ
     - always performs idling when the in-service queue becomes empty;
     - forces the device to serve one I/O request at a time, by
       dispatching a new request only if there is no outstanding
       request.
  In the presence of differentiated weights or I/O-request sizes,
  both the above conditions are needed to guarantee that every
  queue receives its allotted share of the bandwidth (see
  Documentation/block/bfq-iosched.txt for more details). Setting
  strict_guarantees may evidently affect throughput.

[1] https://lkml.org/lkml/2008/4/1/234
    https://lkml.org/lkml/2008/11/11/148

[2] P. Valente and M. Andreolini, "Improving Application
    Responsiveness with the BFQ Disk I/O Scheduler", Proceedings of
    the 5th Annual International Systems and Storage Conference
    (SYSTOR '12), June 2012.
    Slightly extended version:
    http://algogroup.unimore.it/people/paolo/disk_sched/bfq-v1-suite-
results.pdf

Signed-off-by: Fabio Checconi <fchecconi@gmail.com>
Signed-off-by: Paolo Valente <paolo.valente@linaro.org>
Signed-off-by: Arianna Avanzini <avanzini.arianna@gmail.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agonbd: set the max segment size to UINT_MAX
Josef Bacik [Tue, 18 Apr 2017 20:22:51 +0000 (16:22 -0400)]
nbd: set the max segment size to UINT_MAX

NBD doesn't care about limiting the segment size, let the user push the
largest bio's they want.  This allows us to control the request size
solely through max_sectors_kb.

Signed-off-by: Josef Bacik <jbacik@fb.com>
Reviewed-by: Ming Lei <ming.lei@redhat.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agoMerge branch 'stable/for-jens-4.12' of git://git.kernel.org/pub/scm/linux/kernel...
Jens Axboe [Wed, 19 Apr 2017 14:14:34 +0000 (08:14 -0600)]
Merge branch 'stable/for-jens-4.12' of git://git./linux/kernel/git/konrad/xen into for-4.12/block

Konrad writes:

It has one fix - to emit an uevent whenever the size of the guest disk image
changes.

7 years agoblkfront: add uevent for size change
Marc Olson [Tue, 11 Apr 2017 19:24:09 +0000 (12:24 -0700)]
blkfront: add uevent for size change

When a blkfront device is resized from dom0, emit a KOBJ_CHANGE uevent to
notify the guest about the change. This allows for custom udev rules, such
as automatically resizing a filesystem, when an event occurs.

With this patch you get these udev

KERNEL[577.206230] change   /devices/vbd-51728/block/xvdb (block)
UDEV  [577.226218] change   /devices/vbd-51728/block/xvdb (block)

Signed-off-by: Marc Olson <marcolso@amazon.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
7 years agonbd: add a flag to destroy an nbd device on disconnect
Josef Bacik [Thu, 6 Apr 2017 21:02:07 +0000 (17:02 -0400)]
nbd: add a flag to destroy an nbd device on disconnect

For ease of management it would be nice for users to specify that the
device node for a nbd device is destroyed once it is disconnected and
there are no more users.  Add a client flag and enable this operation to
happen.

Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agonbd: add device refcounting
Josef Bacik [Thu, 6 Apr 2017 21:02:06 +0000 (17:02 -0400)]
nbd: add device refcounting

In order to support deleting the device on disconnect we need to
refcount the actual nbd_device struct.  So add the refcounting framework
and change how we free the normal devices at rmmod time so we can catch
reference leaks.

Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agonbd: add a status netlink command
Josef Bacik [Thu, 6 Apr 2017 21:02:05 +0000 (17:02 -0400)]
nbd: add a status netlink command

Allow users to query the status of existing nbd devices.  Right now this
only returns whether or not the device is connected, but could be
extended in the future to include more information.

Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agonbd: handle dead connections
Josef Bacik [Thu, 6 Apr 2017 21:02:04 +0000 (17:02 -0400)]
nbd: handle dead connections

Sometimes we like to upgrade our server without making all of our
clients freak out and reconnect.  This patch provides a way to specify a
dead connection timeout to allow us to pause all requests and wait for
new connections to be opened.  With this in place I can take down the
nbd server for less than the dead connection timeout time and bring it
back up and everything resumes gracefully.

Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agonbd: only clear the queue on device teardown
Josef Bacik [Thu, 6 Apr 2017 21:02:03 +0000 (17:02 -0400)]
nbd: only clear the queue on device teardown

When running a disconnect torture test I noticed that sometimes we would
crash with a negative ref count on our queue.  This was because we were
ending the same request twice.  Turns out we were racing with
NBD_CLEAR_SOCK clearing the requests as well as the teardown of the
device clearing the requests.  So instead make the ioctl only shutdown
the sockets and make it so that we only ever run nbd_clear_que from the
device teardown.

Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agonbd: multicast dead link notifications
Josef Bacik [Thu, 6 Apr 2017 21:02:02 +0000 (17:02 -0400)]
nbd: multicast dead link notifications

Provide a mechanism to notify userspace that there's been a link problem
on a NBD device.  This will allow userspace to re-establish a connection
and provide the new socket to the device without disrupting the device.

Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agonbd: add a reconfigure netlink command
Josef Bacik [Thu, 6 Apr 2017 21:02:01 +0000 (17:02 -0400)]
nbd: add a reconfigure netlink command

We want to be able to reconnect dead connections to existing block
devices, so add a reconfigure netlink command.  We will also allow users
to change their timeout on the fly, but everything else will require a
disconnect and reconnect.  You won't be able to add more connections
either, simply replace dead connections with new more lively
connections.

Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agonbd: add a basic netlink interface
Josef Bacik [Thu, 6 Apr 2017 21:02:00 +0000 (17:02 -0400)]
nbd: add a basic netlink interface

The existing ioctl interface for configuring NBD devices is a bit
cumbersome and hard to extend.  The other problem is we leave a
userspace app sitting in it's syscall until the device disconnects,
which is less than ideal.

This patch introduces a netlink interface for adding and disconnecting
nbd devices.  This has the benefits of being easily extendable without
breaking older userspace applications, and allows us to configure a nbd
device without leaving a userspace app sitting waiting for the device to
disconnect.

With this interface we also gain the ability to configure more devices
than are preallocated at insmod time.  We also have gained the ability
to not specify a particular device and be provided one for us so that
userspace doesn't need to find a free device to configure.

Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agonbd: stop using the bdev everywhere
Josef Bacik [Thu, 6 Apr 2017 21:01:59 +0000 (17:01 -0400)]
nbd: stop using the bdev everywhere

In preparation for the upcoming netlink interface we need to not rely on
already having the bdev for the NBD device we are doing operations on.
Instead of passing the bdev around, just use it in places where we know
we already have the bdev.

Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agonbd: separate out the config information
Josef Bacik [Thu, 6 Apr 2017 21:01:58 +0000 (17:01 -0400)]
nbd: separate out the config information

In order to properly refcount the various aspects of a NBD device we
need to separate out the configuration elements of the nbd device.  The
configuration of a NBD device has a different lifetime from the actual
device, so it doesn't make sense to bundle these two concepts.  Add a
config_refs to keep track of the configuration structure, that way we
can be sure that we never access it when we've torn down the device.
Add a new nbd_config structure to hold all of the transient
configuration information.  Finally create this when we open the device
so that it is in place when we start to configure the device.  This has
a nice side-effect of fixing a long standing problem where you could end
up with a half-configured nbd device that needed to be "disconnected" in
order to be usable again.  Now once we close our device the
configuration will be discarded.

Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agonbd: handle single path failures gracefully
Josef Bacik [Thu, 6 Apr 2017 21:01:57 +0000 (17:01 -0400)]
nbd: handle single path failures gracefully

Currently if we have multiple connections and one of them goes down we will tear
down the whole device.  However there's no reason we need to do this as we
could have other connections that are working fine.  Deal with this by keeping
track of the state of the different connections, and if we lose one we mark it
as dead and send all IO destined for that socket to one of the other healthy
sockets.  Any outstanding requests that were on the dead socket will timeout and
be re-submitted properly.

Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agonbd: put socket in error cases
Josef Bacik [Thu, 6 Apr 2017 21:01:56 +0000 (17:01 -0400)]
nbd: put socket in error cases

When adding a new socket we look it up and then try to add it to our
configuration.  If any of those steps fail we need to make sure we put
the socket so we don't leak them.

Signed-off-by: Josef Bacik <jbacik@fb.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agolightnvm: fix some error code in pblk-init.c
Dan Carpenter [Sat, 15 Apr 2017 18:55:53 +0000 (20:55 +0200)]
lightnvm: fix some error code in pblk-init.c

There were a bunch of places in pblk_lines_init() where we didn't set an
error code.  And in pblk_writer_init() we accidentally return 1 instead
of a correct error code, which would result in a Oops later.

Fixes: 11a5d6fdf919 ("lightnvm: physical block device (pblk) target")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agolightnvm: fix some WARN() messages
Dan Carpenter [Sat, 15 Apr 2017 18:55:52 +0000 (20:55 +0200)]
lightnvm: fix some WARN() messages

WARN_ON() takes a condition, not an error message.  I slightly tweaked
some conditions so hopefully it's more clear.

Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agolightnvm: pblk-gc: fix an error pointer dereference in init
Dan Carpenter [Sat, 15 Apr 2017 18:55:51 +0000 (20:55 +0200)]
lightnvm: pblk-gc: fix an error pointer dereference in init

These labels are reversed so we could end up dereferencing an error
pointer or leaking.

Fixes: 7f347ba6bb3a ("lightnvm: physical block device (pblk) target")
Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agolightnvm: physical block device (pblk) target
Javier González [Sat, 15 Apr 2017 18:55:50 +0000 (20:55 +0200)]
lightnvm: physical block device (pblk) target

This patch introduces pblk, a host-side translation layer for
Open-Channel SSDs to expose them like block devices. The translation
layer allows data placement decisions, and I/O scheduling to be
managed by the host, enabling users to optimize the SSD for their
specific workloads.

An open-channel SSD has a set of LUNs (parallel units) and a
collection of blocks. Each block can be read in any order, but
writes must be sequential. Writes may also fail, and if a block
requires it, must also be reset before new writes can be
applied.

To manage the constraints, pblk maintains a logical to
physical address (L2P) table,  write cache, garbage
collection logic, recovery scheme, and logic to rate-limit
user I/Os versus garbage collection I/Os.

The L2P table is fully-associative and manages sectors at a
4KB granularity. Pblk stores the L2P table in two places, in
the out-of-band area of the media and on the last page of a
line. In the cause of a power failure, pblk will perform a
scan to recover the L2P table.

The user data is organized into lines. A line is data
striped across blocks and LUNs. The lines enable the host to
reduce the amount of metadata to maintain besides the user
data and makes it easier to implement RAID or erasure coding
in the future.

pblk implements multi-tenant support and can be instantiated
multiple times on the same drive. Each instance owns a
portion of the SSD - both regarding I/O bandwidth and
capacity - providing I/O isolation for each case.

Finally, pblk also exposes a sysfs interface that allows
user-space to peek into the internals of pblk. The interface
is available at /dev/block/*/pblk/ where * is the block
device name exposed.

This work also contains contributions from:
  Matias Bjørling <matias@cnexlabs.com>
  Simon A. F. Lund <slund@cnexlabs.com>
  Young Tack Jin <youngtack.jin@gmail.com>
  Huaicheng Li <huaicheng@cs.uchicago.edu>

Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agolightnvm: convert sprintf into strlcpy
Javier González [Sat, 15 Apr 2017 18:55:49 +0000 (20:55 +0200)]
lightnvm: convert sprintf into strlcpy

Convert sprintf calls to strlcpy in order to make possible buffer
overflow more obvious.

Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
7 years agolightnvm: fix type checks on rrpc
Javier González [Sat, 15 Apr 2017 18:55:48 +0000 (20:55 +0200)]
lightnvm: fix type checks on rrpc

sector_t is always unsigned, therefore avoid < 0 checks on it.

Signed-off-by: Javier González <javier@cnexlabs.com>
Signed-off-by: Matias Bjørling <matias@cnexlabs.com>
Signed-off-by: Jens Axboe <axboe@fb.com>