A whole-object layered discard is implemented as a truncate rather
than a delete: a dummy object is needed to prevent the CoW machinery
from kicking in. However, a truncate on a non-existent object is
a no-op. If the object doesn't exist in HEAD, a discard request is
effectively ignored, which violates our "discard zeroes data" promise
and breaks REQ_OP_WRITE_ZEROES implementation.
A non-exclusive create on an existing object is also a no-op, so the
fix is to do a compound create+truncate instead.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
In preparation for rbd "fancy" striping, replace obj_req->img_offset
with obj_req->img_extents. A single starting offset isn't sufficient
because we want only one OSD request per object and will merge adjacent
object extents in ceph_file_to_extents(). The final object extent may
map into multiple different byte ranges in the image.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
obj_req->object_no -> obj_req->ex.oe_objno
obj_req->offset -> obj_req->ex.oe_off
obj_req->length -> obj_req->ex.oe_len
... and use ex for linking object requests to image requests.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
All object requests are associated with an image request now -- avoid
duplicating the same info in each object request.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Do away with partial request completions and all the associated
complexity. Individual object requests no longer need to be completed
in order -- when the last one becomes ready, we complete the entire
higher level request all at once.
This also wraps up the conversion to a state machine model and
eliminates the recursion described in commit 6d69bb536b ("rbd:
prevent kernel stack blow up on rbd map").
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
It should be void now. Also, object requests are unlinked only in
image request destructor, which can't run before rbd_img_request_put(),
so no need for _safe.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
No need to pass rbd_dev and op_type to rbd_osd_req_create(): there are
no standalone (!IMG_DATA) object requests anymore and osd_req->r_flags
can be set in rbd_osd_req_format_{read,write}().
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
The notable changes are:
- instead of explicitly stat'ing the object to see if it exists before
issuing the write, send the write optimistically along with the stat
in a single OSD request
- zero copyup optimization
- all object requests are associated with an image request and have
a valid ->img_request pointer; there are no standalone (!IMG_DATA)
object requests anymore
- code is structured as a state machine (vs a bunch of callbacks with
implicit state)
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
In preparation for rbd "fancy" striping which requires bio_vec arrays,
wire up BVECS data type and kill off PAGES data type. There is nothing
wrong with using page vectors for copyup requests -- it's just less
iterator boilerplate code to write for the new striping framework.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Alex Elder <elder@linaro.org>
The initiating object request is the proper owner -- save a bit of
space.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Alex Elder <elder@linaro.org>
obj_req->pages is for provided data buffers. stat requests are
internal and should be NODATA.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Alex Elder <elder@linaro.org>
The reason we clone bios is to be able to give each object request
(and consequently each ceph_osd_data/ceph_msg_data item) its own
pointer to a (list of) bio(s). The messenger then initializes its
cursor with cloned bio's ->bi_iter, so it knows where to start reading
from/writing to. That's all the cloned bios are used for: to determine
each object request's starting position in the provided data buffer.
Introduce ceph_bio_iter to do exactly that -- store position within bio
list (i.e. pointer to bio) + position within that bio (i.e. bvec_iter).
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Commit 21acdf45f4 ("rbd: set max_segments to USHRT_MAX") removed the
limit on max_segments. Remove the limit on max_segment_size as well.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Alex Elder <elder@linaro.org>
This feature bit restricts older clients from performing certain
maintenance operations against an image (e.g. clone, snap create).
krbd does not perform maintenance operations.
Cc: stable@vger.kernel.org
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Jason Dillaman <dillaman@redhat.com>
If rbd_img_request_submit() fails, parent_request->obj_request is
NULLed out, triggering an assert in rbd_obj_request_put():
rbd_img_request_put(parent_request)
rbd_parent_request_destroy
rbd_obj_request_put(NULL)
Just remove it -- parent_request->obj_request will be put in
rbd_parent_request_destroy().
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Commit d3834fefcf ("rbd: bump queue_max_segments") bumped
max_segments (unsigned short) to max_hw_sectors (unsigned int).
max_hw_sectors is set to the number of 512-byte sectors in an object
and overflows unsigned short for 32M (largest possible) objects, making
the block layer resort to handing us single segment (i.e. single page
or even smaller) bios in that case.
Cc: stable@vger.kernel.org
Fixes: d3834fefcf ("rbd: bump queue_max_segments")
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Alex Elder <elder@linaro.org>
Otherwise, future operations on this RBD using exclusive-lock are
going to require the lock from a non-existent client id.
Cc: stable@vger.kernel.org
Fixes: 14bb211d32 ("rbd: support updating the lock cookie without releasing the lock")
Link: http://tracker.ceph.com/issues/19929
Signed-off-by: Florian Margaine <florian@platform.sh>
[idryomov@gmail.com: rbd_set_owner_cid() call, __rbd_lock() helper]
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
It's been 3.5 years, let's turn it on by default. Support in rbd(8)
utility goes back to pre-firefly, "rbd map" has been loading the module
with single_major=Y ever since. However, if the module is already
loaded (whether by hand or at boot time), we end up with single_major=N.
Also, some people don't install rbd(8) and use the sysfs interface
directly.
(With single-major=N, a major number is consumed for every mapping,
imposing a limit of ~240 rbd images per host. single-major=Y allows
mapping thousands of rbd images on a single machine.)
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Jason Dillaman <dillaman@redhat.com>
RBD devices are currently incorrectly initialised with the block queue
discard_alignment set to the underlying RADOS object size.
As per Documentation/ABI/testing/sysfs-block:
The discard_alignment parameter indicates how many bytes the beginning
of the device is offset from the internal allocation unit's natural
alignment.
Correcting the discard_alignment parameter from the RADOS object size to
zero (the blk_set_default_limits() default) has no effect on how discard
requests are propagated through the block layer - @alignment in
__blkdev_issue_discard() remains zero. However, it does fix the UNMAP
granularity alignment value advertised to SCSI initiators via the Block
Limits VPD.
Signed-off-by: David Disseldorp <ddiss@suse.de>
Reviewed-by: Ilya Dryomov <idryomov@gmail.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
->open_count/-EBUSY check is bogus and wrong: when an open device is
set read-only, blkdev_write_iter() refuses further writes with -EPERM.
This is standard behaviour and all other block devices allow this.
set_disk_ro() call is also problematic: we affect the entire device
when called on a single partition.
All rbd_ioctl_set_ro() needs to do is refuse ro -> rw transition for
mapped snapshots. Everything else can be handled by generic code.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
rbd_img_obj_exists_submit() and rbd_img_obj_parent_read_full() are on
the writeback path for cloned images -- we attempt a stat on the parent
object to see if it exists and potentially read it in to call copyup.
GFP_NOIO should be used instead of GFP_KERNEL here.
Cc: stable@vger.kernel.org
Link: http://tracker.ceph.com/issues/22014
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: David Disseldorp <ddiss@suse.de>
drivers/block/rbd.c: In function 'rbd_acquire_lock':
drivers/block/rbd.c:3602:44: error: 'ret' may be used uninitialized in this function [-Werror=maybe-uninitialized]
Silence the warning, found it when built old kernel(3.10) with
OBS(opensuse build service).
Signed-off-by: Kefeng Wang <wangkefeng.wang@huawei.com>
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
bio_clone() makes a copy of the bi_io_vec, but rbd never changes that,
so there is no need for a copy.
bio_clone_fast() can be used instead, which avoids making the copy.
This requires that we provide a bio_set. bio_clone() uses fs_bio_set,
but it isn't, in general, safe to use the same bio_set at different
levels of the stack, as that can lead to deadlocks. As filesystems
use fs_bio_set, block devices shouldn't.
As rbd never stacks, it is safe to have a single global bio_set for
all rbd devices to use. So allocate that when the module is
initialised, and use it with bio_clone_fast().
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: NeilBrown <neilb@suse.com>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
-----BEGIN PGP SIGNATURE-----
iQEcBAABAgAGBQJZPdbLAAoJEHm+PkMAQRiGx4wH/1nCjfnl6fE8oJ24/1gEAOUh
biFdqJkYZmlLYHVtYfLm4Ueg4adJdg0wx6qM/4RaAzmQVvLfDV34bc1qBf1+P95G
kVF+osWyXrZo5cTwkwapHW/KNu4VJwAx2D1wrlxKDVG5AOrULH1pYOYGOpApEkZU
4N+q5+M0ce0GJpqtUZX+UnI33ygjdDbBxXoFKsr24B7eA0ouGbAJ7dC88WcaETL+
2/7tT01SvDMo0jBSV0WIqlgXwZ5gp3yPGnklC3F4159Yze6VFrzHMKS/UpPF8o8E
W9EbuzwxsKyXUifX2GY348L1f+47glen/1sedbuKnFhP6E9aqUQQJXvEO7ueQl4=
=m2Gx
-----END PGP SIGNATURE-----
Merge tag 'v4.12-rc5' into for-4.13/block
We've already got a few conflicts and upcoming work depends on some of the
changes that have gone into mainline as regression fixes for this series.
Pull in 4.12-rc5 to resolve these conflicts and make it easier on down stream
trees to continue working on 4.13 changes.
Signed-off-by: Jens Axboe <axboe@kernel.dk>
Use the same values for use for request completion errors as the return
value from ->queue_rq. BLK_STS_RESOURCE is special cased to cause
a requeue, and all the others are completed as-is.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Currently we use nornal Linux errno values in the block layer, and while
we accept any error a few have overloaded magic meanings. This patch
instead introduces a new blk_status_t value that holds block layer specific
status codes and explicitly explains their meaning. Helpers to convert from
and to the previous special meanings are provided for now, but I suspect
we want to get rid of them in the long run - those drivers that have a
errno input (e.g. networking) usually get errnos that don't know about
the special block layer overloads, and similarly returning them to userspace
will usually return somethings that strictly speaking isn't correct
for file system operations, but that's left as an exercise for later.
For now the set of errors is a very limited set that closely corresponds
to the previous overloaded errno values, but there is some low hanging
fruite to improve it.
blk_status_t (ab)uses the sparse __bitwise annotations to allow for sparse
typechecking, so that we can easily catch places passing the wrong values.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@fb.com>
Commit 93c1defedc ("rbd: remove the discard_zeroes_data flag")
explicitly didn't implement REQ_OP_WRITE_ZEROES for rbd, while the
following commit 48920ff2a5 ("block: remove the discard_zeroes_data
flag") dropped ->discard_zeroes_data in favor of REQ_OP_WRITE_ZEROES.
rbd does support efficient zeroing via CEPH_OSD_OP_ZERO opcode and will
release either some or all blocks depending on whether the zeroing
request is rbd_obj_bytes() aligned. This is how we currently implement
discards, so REQ_OP_WRITE_ZEROES can be identical to REQ_OP_DISCARD for
now. Caveats:
- REQ_NOUNMAP is ignored, but AFAICT that's true of at least two other
current implementations - nvme and loop
- there is no ->write_zeroes_alignment and blk_bio_write_zeroes_split()
is hence less helpful than blk_bio_discard_split(), but this can (and
should) be fixed on the rbd side
In the future we will split these into two code paths to respect
REQ_NOUNMAP on zeroout and save on zeroing blocks that couldn't be
released on discard.
Fixes: 93c1defedc ("rbd: remove the discard_zeroes_data flag")
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Jason Dillaman <dillaman@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
lock transfers from myself and the long awaited -ENOSPC handling series
from Jeff. The former will allow rbd users to take advantage of
exclusive lock's built-in blacklist/break-lock functionality while
staying in control of who owns the lock. With the latter in place, we
will abort filesystem writes on -ENOSPC instead of having them block
indefinitely.
Beyond that we've got the usual pile of filesystem fixes from Zheng,
some refcount_t conversion patches from Elena and a patch for an
ancient open() flags handling bug from Alexander.
-----BEGIN PGP SIGNATURE-----
Version: GnuPG v2
iQEcBAABCAAGBQJZEt/kAAoJEEp/3jgCEfOLpzAIAIld0N06DuHKG2F9mHEnLeGl
Y60BZ3Ajo32i9qPT/u9ntI99ZMlkuHcNWg6WpCCh8umbwk2eiAKRP/KcfGcWmmp9
EHj9COCmBR9TRM1pNS1lSMzljDnxf9sQmbIO9cwMQBUya5g19O0OpApzxF1YQhCR
V9B/FYV5IXELC3b/NH45oeDAD9oy/WgwbhQ2feTBQJmzIVJx+Je9hdhR1PH1rI06
ysyg3VujnUi/hoDhvPTBznNOxnHx/HQEecHH8b01MkbaCgxPH88jsUK/h7PYF3Gh
DE/sCN69HXeu1D/al3zKoZdahsJ5GWkj9Q+vvBoQJm+ZPsndC+qpgSj761n9v38=
=vamy
-----END PGP SIGNATURE-----
Merge tag 'ceph-for-4.12-rc1' of git://github.com/ceph/ceph-client
Pull ceph updates from Ilya Dryomov:
"The two main items are support for disabling automatic rbd exclusive
lock transfers from myself and the long awaited -ENOSPC handling
series from Jeff.
The former will allow rbd users to take advantage of exclusive lock's
built-in blacklist/break-lock functionality while staying in control
of who owns the lock. With the latter in place, we will abort
filesystem writes on -ENOSPC instead of having them block
indefinitely.
Beyond that we've got the usual pile of filesystem fixes from Zheng,
some refcount_t conversion patches from Elena and a patch for an
ancient open() flags handling bug from Alexander"
* tag 'ceph-for-4.12-rc1' of git://github.com/ceph/ceph-client: (31 commits)
ceph: fix memory leak in __ceph_setxattr()
ceph: fix file open flags on ppc64
ceph: choose readdir frag based on previous readdir reply
rbd: exclusive map option
rbd: return ResponseMessage result from rbd_handle_request_lock()
rbd: kill rbd_is_lock_supported()
rbd: support updating the lock cookie without releasing the lock
rbd: store lock cookie
rbd: ignore unlock errors
rbd: fix error handling around rbd_init_disk()
rbd: move rbd_unregister_watch() call into rbd_dev_image_release()
rbd: move rbd_dev_destroy() call out of rbd_dev_image_release()
ceph: when seeing write errors on an inode, switch to sync writes
Revert "ceph: SetPageError() for writeback pages if writepages fails"
ceph: handle epoch barriers in cap messages
libceph: add an epoch_barrier field to struct ceph_osd_client
libceph: abort already submitted but abortable requests when map or pool goes full
libceph: allow requests to return immediately on full conditions if caller wishes
libceph: remove req->r_replay_version
ceph: make seeky readdir more efficient
...
CURRENT_TIME is not y2038 safe. The macro will be deleted and all the
references to it will be replaced by ktime_get_* apis.
struct timespec is also not y2038 safe. Retain timespec for timestamp
representation here as ceph uses it internally everywhere. These
references will be changed to use struct timespec64 in a separate patch.
The current_fs_time() api is being changed to use vfs struct inode* as
an argument instead of struct super_block*.
Set the new mds client request r_stamp field using ktime_get_real_ts()
instead of using current_fs_time().
Also, since r_stamp is used as mtime on the server, use timespec_trunc()
to truncate the timestamp, using the right granularity from the
superblock.
This api will be transitioned to be y2038 safe along with vfs.
Link: http://lkml.kernel.org/r/1491613030-11599-5-git-send-email-deepa.kernel@gmail.com
Signed-off-by: Deepa Dinamani <deepa.kernel@gmail.com>
Reviewed-by: Arnd Bergmann <arnd@arndb.de>
M: Ilya Dryomov <idryomov@gmail.com>
M: "Yan, Zheng" <zyan@redhat.com>
M: Sage Weil <sage@redhat.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Support disabling automatic exclusive lock transfers to allow users
to be in charge of which node should own the lock while being able to
reuse exclusive lock's built-in blacklist/break-lock functionality.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Jason Dillaman <dillaman@redhat.com>
Right now it's just 0, but "no automatic exclusive lock transfers" mode
code will need -EROFS.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Jason Dillaman <dillaman@redhat.com>
Currently the exclusive lock is acquired only if the mapping is
writable, i.e. an image HEAD mapped in rw mode. This means that we
don't acquire the lock for executing a read from a snapshot or an image
HEAD mapped in ro mode, even if lock_on_read is set. This is somewhat
weird and inconsistent with "no automatic exclusive lock transfers"
mode, where the lock is acquired unconditionally.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Jason Dillaman <dillaman@redhat.com>
As we no longer release the lock before potentially raising BLACKLISTED
in rbd_reregister_watch(), the "either locked or blacklisted" assert in
rbd_queue_workfn() needs to go: we can be both locked and blacklisted
at that point now.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Jason Dillaman <dillaman@redhat.com>
In preparation for supporting set_cookie method (or rather set_cookie
fallback for older OSDs), store the lock cookie on lock and use it on
unlock instead of recalculating from rbd_dev->watch_cookie.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Jason Dillaman <dillaman@redhat.com>
Currently the lock_state is set to UNLOCKED (preventing further I/O),
but RELEASED_LOCK notification isn't sent. Be consistent with userspace
and treat ceph_cls_unlock() errors as the image is unlocked.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Jason Dillaman <dillaman@redhat.com>
add_disk() takes an extra reference on disk->queue, which is put in
put_disk() -> disk_release(). Avoiding blk_cleanup_queue() (which also
puts the queue) until add_disk() sets GENHD_FL_UP works for the queue
itself, but leaks various queue internals. Conditioning tag_set freeing
on GENHD_FL_UP is wrong too: all error paths after rbd_init_disk() leak
the tag_set.
Move the final "announce" steps out of rbd_dev_device_setup() so that
it can be unwound like any other function. Leave "announce" steps to
do_rbd_add/remove().
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Jason Dillaman <dillaman@redhat.com>
rbd_dev->disk tear down vs rbd_watch_cb() race shouldn't be a problem
anymore thanks to EXISTS and REMOVING checks in rbd_dev_update_size().
A similar race could occur on "rbd map", see commit 811c668877
("rbd: fix rbd map vs notify races").
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
Reviewed-by: Jason Dillaman <dillaman@redhat.com>
No reason to hide CephFS-specific features in the rbd case. Recent
feature bits mix RADOS and CephFS-specific stuff together anyway.
Signed-off-by: Ilya Dryomov <idryomov@gmail.com>
