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

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

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

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

Cc: Alex Elder <elder@linaro.org>
Cc: stable@vger.kernel.org # 3.10+
Reported-by: Shawn Edwards <lesser.evil@gmail.com>
Reviewed-by: Sage Weil <sage@redhat.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>

After the switch to blk-mq rbd_wq processes requests, not devices.

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>

Set QUEUE_FLAG_NONROT.  Following commit b277da0a8a59 ("block: disable
entropy contributions for nonrot devices") we should also clear
QUEUE_FLAG_ADD_RANDOM, but it's off by default for blk-mq drivers, so
just note it in the comment.

Also remove physical block size assignment - no sense in repeating
defaults that are not going to change.

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>

pr_info what exactly was the culprit: missing pool, image or snap.

Signed-off-by: Ilya Dryomov <idryomov@gmail.com>

This converts the rbd driver to use the blk-mq infrastructure.  Except
for switching to a per-request work item this is almost mechanical.

This was tested by Alexandre DERUMIER in November, and found to give
him 120000 iops, although the only comparism available was an old
3.10 kernel which gave 80000iops.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Reviewed-by: Alex Elder <elder@linaro.org>
[idryomov@gmail.com: context, blk_mq_init_queue() EH]
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>

If the clone is resized down to 0, it becomes standalone.  If such
resize is carried over while an image is mapped we would detect this
and call rbd_dev_parent_put() which means "let go of all parent state,
including the spec(s) of parent images(s)".  This leads to a mismatch
between "rbd info" and sysfs parent fields, so a fix is in order.

    # rbd create --image-format 2 --size 1 foo
    # rbd snap create foo@snap
    # rbd snap protect foo@snap
    # rbd clone foo@snap bar
    # DEV=$(rbd map bar)
    # rbd resize --allow-shrink --size 0 bar
    # rbd resize --size 1 bar
    # rbd info bar | grep parent
            parent: rbd/foo@snap

Before:

    # cat /sys/bus/rbd/devices/0/parent
    (no parent image)

After:

    # cat /sys/bus/rbd/devices/0/parent
    pool_id 0
    pool_name rbd
    image_id 10056b8b4567
    image_name foo
    snap_id 2
    snap_name snap
    overlap 0

Signed-off-by: Ilya Dryomov <idryomov@redhat.com>
Reviewed-by: Josh Durgin <jdurgin@redhat.com>
Reviewed-by: Alex Elder <elder@linaro.org>

header_rwsem should be released on errors.  Also remove useless
rbd_dev->mapping.size != rbd_dev->header.image_size test.

Signed-off-by: Ilya Dryomov <idryomov@redhat.com>

It's been largely superseded by dup_token() and unused for over
2 years, identified by cppcheck.

Signed-off-by: Rickard Strandqvist <rickard_strandqvist@spectrumdigital.se>
[idryomov@redhat.com: changelog]
Signed-off-by: Ilya Dryomov <idryomov@redhat.com>

This effectively reverts the last hunk of 392a9dad7e77 ("rbd: detect
when clone image is flattened").

The problem with parent_overlap != 0 condition is that it's possible
and completely valid to have an image with parent_overlap == 0 whose
parent state needs to be cleaned up on unmap.  The next commit, which
drops the "clone image now standalone" logic, opens up another window
of opportunity to hit this, but even without it

    # cat parent-ref.sh
    #!/bin/bash
    rbd create --image-format 2 --size 1 foo
    rbd snap create foo@snap
    rbd snap protect foo@snap
    rbd clone foo@snap bar
    rbd resize --allow-shrink --size 0 bar
    rbd resize --size 1 bar
    DEV=$(rbd map bar)
    rbd unmap $DEV

leaves rbd_device/rbd_spec/etc and rbd_client along with ceph_client
hanging around.

My thinking behind calling rbd_dev_parent_put() unconditionally is that
there shouldn't be any requests in flight at that point in time as we
are deep into unmap sequence.  Hence, even if rbd_dev_unparent() caused
by flatten is delayed by in-flight requests, it will have finished by
the time we reach rbd_dev_unprobe() caused by unmap, thus turning
unconditional rbd_dev_parent_put() into a no-op.

Fixes: http://tracker.ceph.com/issues/10352

Cc: stable@vger.kernel.org # 3.11+
Signed-off-by: Ilya Dryomov <idryomov@redhat.com>
Reviewed-by: Josh Durgin <jdurgin@redhat.com>
Reviewed-by: Alex Elder <elder@linaro.org>

The comment for rbd_dev_parent_get() said

    * We must get the reference before checking for the overlap to
    * coordinate properly with zeroing the parent overlap in
    * rbd_dev_v2_parent_info() when an image gets flattened.  We
    * drop it again if there is no overlap.

but the "drop it again if there is no overlap" part was missing from
the implementation.  This lead to absurd parent_ref values for images
with parent_overlap == 0, as parent_ref was incremented for each
img_request and virtually never decremented.

Fix this by leveraging the fact that refresh path calls
rbd_dev_v2_parent_info() under header_rwsem and use it for read in
rbd_dev_parent_get(), instead of messing around with atomics.  Get rid
of barriers in rbd_dev_v2_parent_info() while at it - I don't see what
they'd pair with now and I suspect we are in a pretty miserable
situation as far as proper locking goes regardless.

Cc: stable@vger.kernel.org # 3.11+
Signed-off-by: Ilya Dryomov <idryomov@redhat.com>
Reviewed-by: Josh Durgin <jdurgin@redhat.com>
Reviewed-by: Alex Elder <elder@linaro.org>