]> Pileus Git - ~andy/linux/log
~andy/linux
11 years agorbd: don't destroy ceph_opts in rbd_add()
Alex Elder [Thu, 16 May 2013 20:04:20 +0000 (15:04 -0500)]
rbd: don't destroy ceph_opts in rbd_add()

Whether rbd_client_create() successfully creates a new client or
not, it takes responsibility for getting the ceph_opts structure
it's passed destroyed.  If successful, the structure becomes
associated with the created client; if not, rbd_client_create()
will destroy it.

Previously, rbd_get_client() would call ceph_destroy_options()
if rbd_get_client() failed, and that meant it got called twice.
That led freeing various pointers more than once, which is never a
good idea.

This resolves:
    http://tracker.ceph.com/issues/4559

Cc: stable@vger.kernel.org # 3.8+
Reported-by: Dan van der Ster <dan@vanderster.com>
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agoceph: ceph_pagelist_append might sleep while atomic
Jim Schutt [Wed, 15 May 2013 18:03:35 +0000 (13:03 -0500)]
ceph: ceph_pagelist_append might sleep while atomic

Ceph's encode_caps_cb() worked hard to not call __page_cache_alloc()
while holding a lock, but it's spoiled because ceph_pagelist_addpage()
always calls kmap(), which might sleep.  Here's the result:

[13439.295457] ceph: mds0 reconnect start
[13439.300572] BUG: sleeping function called from invalid context at include/linux/highmem.h:58
[13439.309243] in_atomic(): 1, irqs_disabled(): 0, pid: 12059, name: kworker/1:1
    . . .
[13439.376225] Call Trace:
[13439.378757]  [<ffffffff81076f4c>] __might_sleep+0xfc/0x110
[13439.384353]  [<ffffffffa03f4ce0>] ceph_pagelist_append+0x120/0x1b0 [libceph]
[13439.391491]  [<ffffffffa0448fe9>] ceph_encode_locks+0x89/0x190 [ceph]
[13439.398035]  [<ffffffff814ee849>] ? _raw_spin_lock+0x49/0x50
[13439.403775]  [<ffffffff811cadf5>] ? lock_flocks+0x15/0x20
[13439.409277]  [<ffffffffa045e2af>] encode_caps_cb+0x41f/0x4a0 [ceph]
[13439.415622]  [<ffffffff81196748>] ? igrab+0x28/0x70
[13439.420610]  [<ffffffffa045e9f8>] ? iterate_session_caps+0xe8/0x250 [ceph]
[13439.427584]  [<ffffffffa045ea25>] iterate_session_caps+0x115/0x250 [ceph]
[13439.434499]  [<ffffffffa045de90>] ? set_request_path_attr+0x2d0/0x2d0 [ceph]
[13439.441646]  [<ffffffffa0462888>] send_mds_reconnect+0x238/0x450 [ceph]
[13439.448363]  [<ffffffffa0464542>] ? ceph_mdsmap_decode+0x5e2/0x770 [ceph]
[13439.455250]  [<ffffffffa0462e42>] check_new_map+0x352/0x500 [ceph]
[13439.461534]  [<ffffffffa04631ad>] ceph_mdsc_handle_map+0x1bd/0x260 [ceph]
[13439.468432]  [<ffffffff814ebc7e>] ? mutex_unlock+0xe/0x10
[13439.473934]  [<ffffffffa043c612>] extra_mon_dispatch+0x22/0x30 [ceph]
[13439.480464]  [<ffffffffa03f6c2c>] dispatch+0xbc/0x110 [libceph]
[13439.486492]  [<ffffffffa03eec3d>] process_message+0x1ad/0x1d0 [libceph]
[13439.493190]  [<ffffffffa03f1498>] ? read_partial_message+0x3e8/0x520 [libceph]
    . . .
[13439.587132] ceph: mds0 reconnect success
[13490.720032] ceph: mds0 caps stale
[13501.235257] ceph: mds0 recovery completed
[13501.300419] ceph: mds0 caps renewed

Fix it up by encoding locks into a buffer first, and when the number
of encoded locks is stable, copy that into a ceph_pagelist.

[elder@inktank.com: abbreviated the stack info a bit.]

Cc: stable@vger.kernel.org # 3.4+
Signed-off-by: Jim Schutt <jaschut@sandia.gov>
Reviewed-by: Alex Elder <elder@inktank.com>
11 years agoceph: add cpu_to_le32() calls when encoding a reconnect capability
Jim Schutt [Wed, 15 May 2013 18:03:35 +0000 (13:03 -0500)]
ceph: add cpu_to_le32() calls when encoding a reconnect capability

In his review, Alex Elder mentioned that he hadn't checked that
num_fcntl_locks and num_flock_locks were properly decoded on the
server side, from a le32 over-the-wire type to a cpu type.
I checked, and AFAICS it is done; those interested can consult
    Locker::_do_cap_update()
in src/mds/Locker.cc and src/include/encoding.h in the Ceph server
code (git://github.com/ceph/ceph).

I also checked the server side for flock_len decoding, and I believe
that also happens correctly, by virtue of having been declared
__le32 in struct ceph_mds_cap_reconnect, in src/include/ceph_fs.h.

Cc: stable@vger.kernel.org # 3.4+
Signed-off-by: Jim Schutt <jaschut@sandia.gov>
Reviewed-by: Alex Elder <elder@inktank.com>
11 years agolibceph: must hold mutex for reset_changed_osds()
Alex Elder [Wed, 15 May 2013 21:28:33 +0000 (16:28 -0500)]
libceph: must hold mutex for reset_changed_osds()

An osd client has a red-black tree describing its osds, and
occasionally we would get crashes due to one of these trees tree
becoming corrupt somehow.

The problem turned out to be that reset_changed_osds() was being
called without protection of the osd client request mutex.  That
function would call __reset_osd() for any osd that had changed, and
__reset_osd() would call __remove_osd() for any osd with no
outstanding requests, and finally __remove_osd() would remove the
corresponding entry from the red-black tree.  Thus, the tree was
getting modified without having any lock protection, and was
vulnerable to problems due to concurrent updates.

This appears to be the only osd tree updating path that has this
problem.  It can be fairly easily fixed by moving the call up
a few lines, to just before the request mutex gets dropped
in kick_requests().

This resolves:
    http://tracker.ceph.com/issues/5043

Cc: stable@vger.kernel.org # 3.4+
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
11 years agorbd: re-submit flattened write request (part 2)
Alex Elder [Mon, 6 May 2013 22:40:33 +0000 (17:40 -0500)]
rbd: re-submit flattened write request (part 2)

Add code to rbd_img_obj_exists_callback() to detect when a clone's
parent image has disappeared, and re-submit the original write
request in that case.

Kill off some redundant assertions.

This completes the resolution for:
    http://tracker.ceph.com/issues/3763

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: re-submit write request for flattened clone
Alex Elder [Mon, 6 May 2013 22:40:33 +0000 (17:40 -0500)]
rbd: re-submit write request for flattened clone

Add code to rbd_img_parent_read_full_callback() to detect when a
clone's parent image has disappeared, and re-submit the original
write request in that case.  (See the previous commit for more
reasoning about why this is appropriate.)

Rename some variables in rbd_img_obj_parent_read_full_callback()
to match the convention used in the previous patch.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: re-submit read request for flattened clone
Alex Elder [Mon, 6 May 2013 22:40:33 +0000 (17:40 -0500)]
rbd: re-submit read request for flattened clone

If a clone image gets flattened while a parent read request is
underway, the original rbd object request needs to be resubmitted.

The reason is that by the time we get the response to the parent
read request, the data read from the parent may be out of date.
In other words, we could see this sequence of events:

    rbd client                      parent image/osd
    ----------                      ----------------
    original object ENOENT;
        issue parent read
                                    respond to parent read
                                    child image flattened
    original image header refresh
             <--- original object written independently here
    parent read response received

Add code to rbd_img_parent_read_callback() to detect when a clone's
parent image has disappeared (as evidenced by its parent overlap
becoming 0), and re-submit the original read request in that case.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: detect when clone image is flattened
Alex Elder [Mon, 6 May 2013 22:40:33 +0000 (17:40 -0500)]
rbd: detect when clone image is flattened

A format 2 clone image can be the subject of a "flatten" operation,
during which all of its data gets "copied up" from its parent image,
leaving the image fully populated.  Once this is complete, the
clone's association with the parent is abolished.

Since this can occur when a clone is mapped, we need to detect when
it has occurred and handle it accordingly.  We know an image has
been flattened when we know it at one time had a parent, but we have
learned (via a "get_parent" object class method call) it no longer
has one.

There might be in-flight requests at the point we learn an image has
been flattened, so we can't simply clean up parent data structures
right away.  Instead, we'll drop the initial parent reference when
the parent has disappeared (rather than when the image gets
destroyed), which will allow the last in-flight reference to clean
things up when it's complete.

We leverage the fact that a zero parent overlap renders an image
effectively unlayered.  We set the overlap to 0 at the point we
detect the clone image has flattened, which allows the unlayered
behavior to take effect immediately, while keeping other parent
structures in place until in-flight requests to complete.

This and the next few patches resolve:
    http://tracker.ceph.com/issues/3763

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: reference count parent requests
Alex Elder [Thu, 9 May 2013 03:50:04 +0000 (22:50 -0500)]
rbd: reference count parent requests

Keep a reference count for uses of the parent information for an rbd
device.

An initial reference is set in rbd_img_request_create() if the
target image has a parent (with non-zero overlap).  Each image
request for an image with a non-zero parent overlap gets another
reference when it's created, and that reference is dropped when the
request is destroyed.

The initial reference is dropped when the image gets torn down.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: define parent image request routines
Alex Elder [Thu, 9 May 2013 03:50:04 +0000 (22:50 -0500)]
rbd: define parent image request routines

Define rbd_parent_request_create() and rbd_parent_request_destroy()
to handle the creation of parent image requests submitted for
layered image objects.  For simplicity, let rbd_img_request_put()
handle dropping the reference to any image request (parent or not),
and call whichever destructor is appropriate on the last put.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: define rbd_dev_unparent()
Alex Elder [Thu, 9 May 2013 03:50:04 +0000 (22:50 -0500)]
rbd: define rbd_dev_unparent()

Define rbd_dev_unparent() to encapsulate cleaning up parent data
structures from a layered rbd image.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: don't release write request until necessary
Alex Elder [Thu, 9 May 2013 15:08:49 +0000 (10:08 -0500)]
rbd: don't release write request until necessary

Previously when a layered write was going to involve a copyup
request, the original osd request was released before submitting the
parent full-object read.  The osd request for the copyup would then
be allocated in rbd_img_obj_parent_read_full_callback().

Shortly we will be handling the event of mapped layered images
getting flattened, and when that occurs we need to resubmit the
original request.  We therefore don't want to release the osd
request until we really konw we're going to replace it--in the
callback function.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: get parent info on refresh
Alex Elder [Mon, 6 May 2013 22:40:33 +0000 (17:40 -0500)]
rbd: get parent info on refresh

Get parent info for format 2 images on every refresh (rather than
just during the initial probe).  This will be needed to detect the
disappearance of the parent image in the event a mapped image
becomes unlayered (i.e., flattened).  Avoid leaking the previous
parent spec on the second and subsequent times this information is
requested by dropping the previous one (if any) before updating it.
(Also, extract the pool id into a local variable before assigning
it into the parent spec.)

Switch to using a non-zero parent overlap value rather than the
existence of a parent (a non-null parent_spec pointer) to determine
whether to mark a request layered.  It will soon be possible for
a layered image to become unlayered while a request is in flight.

This means that the layered flag for an image request indicates that
there was a non-zero parent overlap at the time the image request
was created.  The parent overlap can change thereafter, which may
lead to special handling at request submission or completion time.

This and the next several patches are related to:
    http://tracker.ceph.com/issues/3763

NOTE:
If an error occurs while refreshing the parent info (i.e.,
requesting it after initial probe), the old parent info will
persist.  This is not really correct, and is a scenario that needs
to be addressed.  For now we'll assert that the failure mode is
unlikely, but the issue has been documented in tracker issue 5040.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: ignore zero-overlap parent
Alex Elder [Mon, 6 May 2013 22:40:33 +0000 (17:40 -0500)]
rbd: ignore zero-overlap parent

An rbd clone image that has an overlap with its parent of 0 is
effectively not a layered image at all.  Detect this case and treat
such an image as non-layered.  Issue a warning to be sure the user
knows what's going on.

This resolves:
    http://tracker.ceph.com/issues/5028

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: support reading parent page data for writes
Alex Elder [Fri, 10 May 2013 21:29:22 +0000 (16:29 -0500)]
rbd: support reading parent page data for writes

Currently, rbd_img_obj_parent_read_full() assumes the incoming
object request contains bio data.  But if a layered image is part of
a multi-layer stack of images it will result in read requests of
page data to parent images.

This is handling the same kind of issue as was resolved by this
commit:
    5b2ab72d  rbd: support reading parent page data

This resolves:
    http://tracker.ceph.com/issues/5027

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: fix parent request size assumption
Alex Elder [Fri, 10 May 2013 21:29:22 +0000 (16:29 -0500)]
rbd: fix parent request size assumption

The code that reads object data from the parent for a copyup on
write request currently assumes that the size of that request is the
size of a "full" object from the original target image.

That is not necessarily the case.  The parent overlap could reduce
the request size below that.  To fix that assumption we need to
record the number of pages in the copyup_pages array, for both an
image request and an object request.  Rename a local variable in
rbd_img_obj_parent_read_full_callback() to reflect we're recording
the length of the parent read request, not the size of the target
object.

This resolves:
    http://tracker.ceph.com/issues/5038

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agolibceph: init sent and completed when starting
Alex Elder [Thu, 9 May 2013 19:56:32 +0000 (14:56 -0500)]
libceph: init sent and completed when starting

The rbd code has a need to be able to restart an osd request that
has already been started and completed once before.  This currently
wouldn't work right because the osd client code assumes an osd
request will be started exactly once  Certain fields in a request
are never cleared and this leads to trouble if you try to reuse it.

Specifically, the r_sent, r_got_reply, and r_completed fields are
never cleared.  The r_sent field records the osd incarnation at the
time the request was sent to that osd.  If that's non-zero, the
message won't get re-mapped to a target osd properly, and won't be
put on the unsafe requests list the first time it's sent as it
should.  The r_got_reply field is used in handle_reply() to ensure
the reply to a request is processed only once.  And the r_completed
field is used for lingering requests to avoid calling the callback
function every time the osd client re-sends the request on behalf of
its initiator.

Each osd request passes through ceph_osdc_start_request() when
responsibility for the request is handed over to the osd client for
completion.  We can safely zero these three fields there each time a
request gets started.

One last related change--clear the r_linger flag when a request
is no longer registered as a linger request.

This resolves:
    http://tracker.ceph.com/issues/5026

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: kill rbd_img_request_get()
Alex Elder [Wed, 8 May 2013 13:49:52 +0000 (08:49 -0500)]
rbd: kill rbd_img_request_get()

Get rid of rbd_img_request_get(), because it isn't used, and maybe
won't ever be needed.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: only set up watch for mapped images
Alex Elder [Mon, 6 May 2013 22:40:33 +0000 (17:40 -0500)]
rbd: only set up watch for mapped images

Any changes to parent images are immaterial to any mapped clone.
So there is no need to have a watch event registered on header
objects except for the header object of an image that is mapped.
In fact, a watch request is a write operation, and we may only
have read access to a parent image.

We can't set up the watch request until we know the name of the
header object though.  So pass a flag to rbd_dev_image_probe() to
indicate whether this probe is for a mapping or for a parent image.

Change the second parameter to rbd_dev_header_watch_sync() be
Boolean while we're at it.

This resolves:
    http://tracker.ceph.com/issues/4941

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: set mapping read-only flag in rbd_add()
Alex Elder [Mon, 6 May 2013 22:40:33 +0000 (17:40 -0500)]
rbd: set mapping read-only flag in rbd_add()

The rbd_dev->mapping field for a parent image is not meaningful.
Since rbd_image_probe() is used both for images being mapped and
their parents, it doesn't make sense to set that flag in that
function.

So move the setting of the mapping.read_only flag out of
rbd_dev_image_probe() and into rbd_add() instead.

This resolves:
    http://tracker.ceph.com/issues/4940

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: support reading parent page data
Alex Elder [Mon, 6 May 2013 22:40:33 +0000 (17:40 -0500)]
rbd: support reading parent page data

Currently, rbd_img_parent_read() assumes the incoming object request
contains bio data.  But if a layered image is part of a multi-layer
stack of images it will result in read requests of page data to parent
images.

Fortunately, it's not hard to add support for page data.

This resolves:
    http://tracker.ceph.com/issues/4939

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: fix an incorrect assertion condition
Alex Elder [Mon, 6 May 2013 22:40:32 +0000 (17:40 -0500)]
rbd: fix an incorrect assertion condition

In rbd_img_obj_parent_read_full_callback() there is an assertion
intended to verify the size of the image request for a full parent
read was the size of the original request's target object.  But
assertion was looking at the parent image order rather than the
original one, and these values can differ.

Fix that.

This resolves:
    http://tracker.ceph.com/issues/4938

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: define rbd_dev_v2_header_info()
Alex Elder [Mon, 6 May 2013 14:51:30 +0000 (09:51 -0500)]
rbd: define rbd_dev_v2_header_info()

This rearranges rbd_dev_v2_refresh() so it works more like
rbd_dev_v1_header_info().  While format 1 images need to read the
whole header object to get any information, format 2 can collect
almost all information selectively.  So the one-time initialization
will remain in a separate function--based on rbd_dev_v2_probe().

Rename rbd_dev_v2_refresh() to be rbd_dev_v2_header_info(), and have
it call rbd_dev_v2_header_onetime() if it's being called for the
first time for the given rbd device.

Rename rbd_dev_v2_probe() to be rbd_dev_v2_header_onetime() and
remove the image size and snapshot context calls it held in
common with the refresh function.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: get rid of trivial v1 header wrappers
Alex Elder [Mon, 6 May 2013 14:51:30 +0000 (09:51 -0500)]
rbd: get rid of trivial v1 header wrappers

Get rid of the trivial wrapper functions rbd_dev_v1_refresh() and
rbd_dev_v1_probe(), substituting rbd_dev_v1_header_read() calls
in their place.

Rename rbd_dev_v1_header_read() to be rbd_dev_v1_header_info(), to
be more generic (it will better reflect what happens with format 2
images).

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: simplify rbd_dev_v1_probe()
Alex Elder [Mon, 6 May 2013 14:51:30 +0000 (09:51 -0500)]
rbd: simplify rbd_dev_v1_probe()

An rbd_dev structure's fields are all zero-filled for an initial
probe, so there's no need to explicitly zero the parent_spec
and parent_overlap fields in rbd_dev_v1_probe().  Removing these
assignments makes rbd_dev_v1_probe() *almost* trivial.

Move the dout() message that announces discovery of an image into
rbd_dev_image_probe(), generalize to support images in either format
and only show it if an image is fully discovered.

This highlights that are some unnecessary cleanups in the error
path for rbd_dev_v1_probe(), so they can be removed.

Now rbd_dev_v1_probe() *is* a trivial wrapper function.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: update in-core header directly
Alex Elder [Mon, 6 May 2013 14:51:29 +0000 (09:51 -0500)]
rbd: update in-core header directly

Now that rbd_header_from_disk() only fills in one-time fields once,
we can extend it slightly so it releases the other fields before
replacing their values.  This way there's no need to pass a
temporary buffer and then copy all the results in.  Just use the rbd
device header structure in rbd_header_from_disk() so its values get
updated directly.

Note that this means we need to take the header semaphore at the
point we update things.  So pass the rbd_dev rather than the address
of its header as its first argument to rbd_header_from_disk(), and
have it return an error code.

As a result, rbd_dev_v1_header_read() does all the work,
rbd_read_header() becomes unnecessary, and rbd_dev_v1_refresh()
becomes a very simple wrapper.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: refactor rbd_header_from_disk()
Alex Elder [Mon, 6 May 2013 14:51:29 +0000 (09:51 -0500)]
rbd: refactor rbd_header_from_disk()

This rearranges rbd_header_from_disk so that it:
    - allocates the snapshot context right away
    - keeps results in local variables, not changing the passed-in
      header until it's known we'll succeed
    - does initialization of set-once fields in a header only if
      they have not already been set

The last point is moot at the moment, because rbd_read_header()
(the only caller) always supplies a zero-filled header buffer.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: zero format 1 header structure earlier
Alex Elder [Mon, 6 May 2013 14:51:29 +0000 (09:51 -0500)]
rbd: zero format 1 header structure earlier

The passed-in header structure is zeroed in rbd_header_from_disk().
Instead, have the caller do it.  Note that there are two callers,
rbd_dev_v1_refresh() and rbd_dev_v1_probe().  The latter already has
a zeroed header structure so zeroing it isn't necessary there.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: set the mapping size and features later
Alex Elder [Mon, 6 May 2013 14:51:29 +0000 (09:51 -0500)]
rbd: set the mapping size and features later

Defer setting the size and features fields of a mapped image until
after the Linux disk structure is set up.  Set the capacity of the
disk after that.

Rearrange the definition of rbd_image_header, separating the fields
that are set only once from those that can be updated.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: always set read-only flag in rbd_add()
Alex Elder [Mon, 6 May 2013 12:40:30 +0000 (07:40 -0500)]
rbd: always set read-only flag in rbd_add()

Hold off setting the read-only flag in rbd_add() for an image being
mapped until we have successfully probed the image.  At that point
we know whether it's a snapshot mapping or not, so we can set the
read-only flag in that one place rather than doing so (for
snapshots) in rbd_dev_mapping_set().  To do this, pass a flag to the
image probe routine indicating whether we want a read-only mapping.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: kill rbd_dev_clear_mapping()
Alex Elder [Mon, 6 May 2013 12:40:30 +0000 (07:40 -0500)]
rbd: kill rbd_dev_clear_mapping()

This function is a duplicate of rbd_dev_mapping_clear(), and was
added by mistake.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: don't look up snapshot id in rbd_dev_mapping_set()
Alex Elder [Mon, 6 May 2013 12:40:30 +0000 (07:40 -0500)]
rbd: don't look up snapshot id in rbd_dev_mapping_set()

Currently rbd_dev_mapping_set() looks up the snapshot id for the
snapshot whose name is found in the rbd device's spec structure.

That function gets called by rbd_dev_device_setup(), which is
called by rbd_add() *after* rbd_dev_image_probe().  If the
image probe succeeds, the rbd device's spec will already have
been updated to include names and ids for all fields.

Therefore there's no need to look up the snapshot id in
rbd_dev_mapping_set().

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: don't print warning if not mapping a parent
Alex Elder [Mon, 6 May 2013 12:40:30 +0000 (07:40 -0500)]
rbd: don't print warning if not mapping a parent

The presence of the LAYERING bit in an rbd image's feature mask does
not guarantee the image actually has a parent image.  Currently that
bit is set only when a clone (i.e., image with a parent) is created,
but it is (currently) not cleared if that clone gets flattened back
into a "normal" image.  A "parent_id" query will leave the
parent_spec for the image being mapped a null pointer, but will not
return an error.

Currently, whenever an image with the LAYERED feature gets mapped, a
warning about the use of layered images gets printed.  But we don't
want to do this for a flattened image, so print the warning only
if we find there is a parent spec after the probe.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: kill rbd_update_mapping_size()
Alex Elder [Mon, 6 May 2013 12:40:30 +0000 (07:40 -0500)]
rbd: kill rbd_update_mapping_size()

Since rbd_update_mapping_size() is now a trivial wrapper, just open
code it in its two callers.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: update capacity in rbd_dev_refresh()
Alex Elder [Mon, 6 May 2013 12:40:30 +0000 (07:40 -0500)]
rbd: update capacity in rbd_dev_refresh()

When a mapped image changes size, we change the capacity recorded
for the Linux disk associated with it, in rbd_update_mapping_size().
That function is called in two places--the format 1 and format 2
refresh routines.

There is no need to set the capacity while holding the header
semaphore.  Instead, do it in the common rbd_dev_refresh(), using
the logic that's already there to initiate disk revalidation.

Add handling in the request function, just in case a request
that exceeds the capacity of the device comes in (perhaps one
that was started before a refresh shrunk the device).

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: revalidate only for mapping size changes
Alex Elder [Mon, 6 May 2013 12:40:30 +0000 (07:40 -0500)]
rbd: revalidate only for mapping size changes

This commit:
    d98df63e rbd: revalidate_disk upon rbd resize
instituted a call to revalidate_disk() to notify interested parties
that a mapped image has changed size.  This works well, as long as
the the rbd device doesn't map a snapshot.

A snapshot will never change size.  However, the base image the
snapshot is associated with can, and it can do so while the snapshot
is mapped.

The problem is that the test for the size is looking at the size of
the base image, not the size of the mapped snapshot.  This patch
corrects that.

Update the warning message shown in the event of error, and move
it into the callers.

This resolves:
    http://tracker.ceph.com/issues/4911

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: fix leak of format 2 snapshot context
Alex Elder [Mon, 6 May 2013 13:37:00 +0000 (08:37 -0500)]
rbd: fix leak of format 2 snapshot context

When rbd_dev_v2_refresh() is called, the rbd device already has a
snapshot context associated with it.  But that never gets freed,
the pointer just gets overwritten.

Fix this by dropping the rbd device's reference to the snapshot
context before overwriting the pointer.

Because ceph_put_snap_context() already handles for a null pointer
we don't need to check for that (for the probe case, where no
context has yet been assigned).

This resolves:
    http://tracker.ceph.com/issues/4912

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: fix image request leak on parent read
Alex Elder [Thu, 2 May 2013 02:37:07 +0000 (21:37 -0500)]
rbd: fix image request leak on parent read

When a read for a layered image object finds the target object
doesn't exist, a read image request for the parent image is created
and submitted.  When that completes, the callback routine was
not releasing that parent image request.  Fix that.

The slab allocation stuff just added has greatly simplified the
search for the source of this memory leak.

This resolves:
    http://tracker.ceph.com/issues/4803

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agolibceph: use slab cache for osd client requests
Alex Elder [Wed, 1 May 2013 17:43:04 +0000 (12:43 -0500)]
libceph: use slab cache for osd client requests

Create a slab cache to manage allocation of ceph_osdc_request
structures.

This resolves:
    http://tracker.ceph.com/issues/3926

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agolibceph: allocate ceph message data with a slab allocator
Alex Elder [Wed, 1 May 2013 17:43:04 +0000 (12:43 -0500)]
libceph: allocate ceph message data with a slab allocator

Create a slab cache to manage ceph_msg_data structure allocation.

This is part of:
    http://tracker.ceph.com/issues/3926

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agolibceph: allocate ceph messages with a slab allocator
Alex Elder [Wed, 1 May 2013 17:43:04 +0000 (12:43 -0500)]
libceph: allocate ceph messages with a slab allocator

Create a slab cache to manage ceph_msg structure allocation.

This is part of:
    http://tracker.ceph.com/issues/3926

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: allocate image object names with a slab allocator
Alex Elder [Wed, 1 May 2013 17:43:04 +0000 (12:43 -0500)]
rbd: allocate image object names with a slab allocator

The names of objects used for image object requests are always fixed
size.  So create a slab cache to manage them.  Define a new function
rbd_segment_name_free() to match rbd_segment_name() (which is what
supplies the dynamically-allocated name buffer).

This is part of:
    http://tracker.ceph.com/issues/3926

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: allocate object requests with a slab allocator
Alex Elder [Wed, 1 May 2013 17:43:03 +0000 (12:43 -0500)]
rbd: allocate object requests with a slab allocator

Create a slab cache to manage rbd_obj_request allocation.  We aren't
using a constructor, and we'll zero-fill object request structures
when they're allocated.

This is part of:
    http://tracker.ceph.com/issues/3926

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: allocate name separate from obj_request
Alex Elder [Wed, 1 May 2013 17:43:03 +0000 (12:43 -0500)]
rbd: allocate name separate from obj_request

The next patch will define a slab allocator for a object requests.
To use that we'll need to allocate the name of an object separate
from the request structure itself.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: allocate image requests with a slab allocator
Alex Elder [Wed, 1 May 2013 17:43:03 +0000 (12:43 -0500)]
rbd: allocate image requests with a slab allocator

Create a slab cache to manage rbd_img_request allocation.  Nothing
too fancy at this point--we'll still initialize everything at
allocation time (no constructor)

This is part of:
    http://tracker.ceph.com/issues/3926

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: use binary search for snapshot lookup
Alex Elder [Wed, 1 May 2013 17:43:03 +0000 (12:43 -0500)]
rbd: use binary search for snapshot lookup

Use bsearch(3) to make snapshot lookup by id more efficient.  (There
could be thousands of snapshots, and conceivably many more.)

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: clear EXISTS flag if mapped snapshot disappears
Alex Elder [Wed, 1 May 2013 17:43:03 +0000 (12:43 -0500)]
rbd: clear EXISTS flag if mapped snapshot disappears

This functionality inadvertently disappeared in the last patch.

Image snapshots can get removed at just about any time.  In
particular it can disappear even if it is in use by an rbd
client as a mapped image.

The rbd client deals with such a disappearance by responding to new
requests with ENXIO.  This is implemented by each rbd device
maintaining an EXISTS flag, which is normally set but cleared if a
snapshot disappears.

This patch (re-)implements the clearing of that flag.

Whenever mapped image header information is refreshed, if the
mapping is for a snapshot, verify the mapped snapshot is still
present in the updated snapshot context.  If it is not, clear the
flag.

It is not necessary to check this in the initial probe, because the
probe will not succeed if the snapshot doesn't exist.

This resolves:
    http://tracker.ceph.com/issues/4880

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: kill off the snapshot list
Alex Elder [Tue, 30 Apr 2013 05:44:33 +0000 (00:44 -0500)]
rbd: kill off the snapshot list

We no longer use the snapshot list for anything.  When we need to
look up a snapshot name, id, size, or feature mask, we just do it
directly rather than relying on this list being updated with every
refresh.  The main reason it existed was for the benefit of the
device/sysfs entries that previously were associated with snapshots.

So get rid of the snapshot list, and struct rbd_snap, and the
hundreds of lines of code that supported them.

This resolves:
    http://tracker.ceph.com/issues/4868

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: define rbd_snap_size() and rbd_snap_features()
Alex Elder [Tue, 30 Apr 2013 05:44:33 +0000 (00:44 -0500)]
rbd: define rbd_snap_size() and rbd_snap_features()

This patch defines a handful of new functions that will allow
us to get rid of the rbd device structure's list of snapshots.

Define rbd_snap_id_by_name() to look up a snapshot id given its
name.  This is efficient for format 1 images but not for format 2.
Fortunately it only gets called at mapping time so it's not that
critical.

Use rbd_snap_id_by_name() to find out the id for a snapshot getting
mapped, and pass that id to new functions rbd_snap_size() and
rbd_snap_features() to look up information about a given snapshot's
size and feature mask given its snapshot id.  All this gets done
in rbd_dev_mapping_set().

As a result, snap_by_name() is no longer needed, so get rid of it.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: use snap_id not index to look up snap info
Alex Elder [Tue, 30 Apr 2013 05:44:33 +0000 (00:44 -0500)]
rbd: use snap_id not index to look up snap info

In order to align with what was needed for format 1 rbd images,
rbd_dev_v2_snap_info() was set up to take as argument an index into
the array of snapshot ids in a rbd device's snapshot context.

This switches that around, so we pass the snapshot id instead.
In doing this, rbd_snap_name() now returns a dynamically-allocated
string rather than a fixed one, so there's no need to make a
duplicate in its caller, rbd_dev_spec_update().

This means the following functions take a snapshot id where they
previously used an index value:
    rbd_dev_snap_info()
    rbd_dev_v1_snap_info()
    rbd_dev_v2_snap_info()

A new function, rbd_dev_snap_index(), determines the snap index for
format 1 images and uses it to look up the name.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: look up snapshot name in names buffer
Alex Elder [Tue, 30 Apr 2013 05:44:33 +0000 (00:44 -0500)]
rbd: look up snapshot name in names buffer

Rather than scanning the list of snapshot structures for it, scan
the snapshot context buffer containing snapshot names in order to
determine for a format 1 image the name associated with a given
snapshot id.

Pull out the part of rbd_dev_v1_snap_info() that does this scan into
a new function, _rbd_dev_v1_snap_name().  Have that function return
a dynamically-allocated copy of the name, and don't duplicate it in
rbd_dev_v1_snap_info().

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: drop obj_request->version
Alex Elder [Tue, 30 Apr 2013 05:44:33 +0000 (00:44 -0500)]
rbd: drop obj_request->version

Nothing ever uses the version field maintained in the object request
structure any more, so get rid of it.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: drop rbd_obj_method_sync() version parameter
Alex Elder [Tue, 30 Apr 2013 05:44:33 +0000 (00:44 -0500)]
rbd: drop rbd_obj_method_sync() version parameter

Only NULL is passed as the version argument to rbd_obj_method_sync(),
so get rid of it.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: more version parameter removal
Alex Elder [Tue, 30 Apr 2013 05:44:33 +0000 (00:44 -0500)]
rbd: more version parameter removal

Continued from the last patch, more parameters that can go away
because we no longer have a need to track object versions.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: get rid of some version parameters
Alex Elder [Tue, 30 Apr 2013 05:44:33 +0000 (00:44 -0500)]
rbd: get rid of some version parameters

Several functions in rbd have parameters meant to allow the version
of an object to be passed in or out.  The purpose of those was to
allow the version of a header object to be maintained, but we no
longer do that.  As a result, these parameters are never actually
needed or used, so get rid of them.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: stop tracking header object version
Alex Elder [Tue, 30 Apr 2013 05:44:32 +0000 (00:44 -0500)]
rbd: stop tracking header object version

The rbd code takes care to maintain the version of the header
object.  This was done in hopes of using it to detect a change in
the object between reading it and setting up a watch request to
be notified of changes.

The mechanism was never fully implemented, however.  And we now
avoid the original problem by setting up the watch request before
ever reading the content of the header.

The osd doesn't interpret the object version supplied with a WATCH
osd op, nor does it use the version supplied with a NOTIFY_ACK op
(we can just supply 0 for both).  There is therefore no need to
maintain the header's object version any more, so stop doing so.

We'll be able to simplify some more rbd code in the next few patches
as a result of this.

This resolves:
    http://tracker.ceph.com/issues/3952

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: snap names are pointer to constant data
Alex Elder [Tue, 30 Apr 2013 05:44:33 +0000 (00:44 -0500)]
rbd: snap names are pointer to constant data

Make explicit that snapshot names don't change by making functions
return and take parameters that that point to const qualified data.

This resolves:
    http://tracker.ceph.com/issues/4867

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: don't revalidate so much
Alex Elder [Tue, 30 Apr 2013 05:44:32 +0000 (00:44 -0500)]
rbd: don't revalidate so much

Whenever a header object event causes a mapped rbd image to refresh
its header information, revalidate_disk() is being called.  This was
done in rbd_dev_refresh() outside the control mutex in order to
avoid a lock inversion.  Although a an event like this *might*
indicate the image has changed size, most of the time it does not.

Record the image size before and after the refresh, and only
call revalidate_disk() if it changes.

This resolves:
    http://tracker.ceph.com/issues/4867

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: fix up the layering warning message
Alex Elder [Tue, 30 Apr 2013 05:44:32 +0000 (00:44 -0500)]
rbd: fix up the layering warning message

A warning gets spewed for any image being probed, including parent
images.  Set up a condition such that the warning message only gets
printed for the image being mapped, not any of its parents.

Also, I didn't like the way the warning ended up being so long.
Make it a terse warning instead.  People experimenting with layering
will know what the message means.

This is part of:
    http://tracker.ceph.com/issues/4867

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agoceph: use ceph_create_snap_context()
Alex Elder [Tue, 30 Apr 2013 05:44:32 +0000 (00:44 -0500)]
ceph: use ceph_create_snap_context()

Now that we have a library routine to create snap contexts, use it.

This is part of:
    http://tracker.ceph.com/issues/4857

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agolibceph: create source file "net/ceph/snapshot.c"
Alex Elder [Tue, 30 Apr 2013 05:44:32 +0000 (00:44 -0500)]
libceph: create source file "net/ceph/snapshot.c"

This creates a new source file "net/ceph/snapshot.c" to contain
utility routines related to ceph snapshot contexts.  The main
motivation was to define ceph_create_snap_context() as a common way
to create these structures, but I've moved the definitions of
ceph_get_snap_context() and ceph_put_snap_context() there too.
(The benefit of inlining those is very small, and I'd rather
keep this collection of functions together.)

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: set up devices only for mapped images
Alex Elder [Mon, 29 Apr 2013 04:32:34 +0000 (23:32 -0500)]
rbd: set up devices only for mapped images

Stop setting up Linux devices during the image probe operation.
Instead, set up the devices as a separate step after the image
probe, in rbd_add().

A consequence of this is that only mapped images get devices
assigned to them, which is pretty sweet.

This resolves:
    http://tracker.ceph.com/issues/4774

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: don't have device release destroy rbd_dev
Alex Elder [Mon, 29 Apr 2013 04:32:34 +0000 (23:32 -0500)]
rbd: don't have device release destroy rbd_dev

Currently an rbd_device structure gets destroyed from the release
routine for the device embedded within it.  Stop doing that, instead
calling rbd_dev_image_release() right after rbd_bus_del_dev()
wherever the latter is called.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: define rbd_dev_unprobe()
Alex Elder [Mon, 29 Apr 2013 04:32:34 +0000 (23:32 -0500)]
rbd: define rbd_dev_unprobe()

Define a new function rbd_dev_unprobe() which undoes state changes
that occur from calling rbd_dev_v1_probe() or rbd_dev_v2_probe().
Note that this is a superset of rbd_header_free(), which is now
getting removed (it seems to have been used improperly anyway).

Flesh out rbd_dev_image_release() so it undoes exactly what
rbd_dev_image_probe() does.

This means that:
    - rbd_dev_device_release() gets called when the last device
      reference gets dropped;
    - that undoes everything done by the rbd_dev_device_setup() call
      at the end of rbd_dev_image_probe() (and nothing more), ending
      by calling rbd_dev_image_release(); and
    - rbd_dev_image_release() undoes everything else done by
      rbd_dev_image_probe() (and this includes a call to
      rbd_dev_unprobe().

This means the image and device portions of an rbd device are fairly
cleanly separated now, so error paths should be a little easier to
verify than they used to be.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: don't destroy rbd_dev in device release function
Alex Elder [Mon, 29 Apr 2013 04:32:34 +0000 (23:32 -0500)]
rbd: don't destroy rbd_dev in device release function

Rename rbd_dev_probe_finish() to be rbd_dev_device_setup().  Its
purpose is to set up the Linux side of an rbd device mapping.
Rename rbd_dev_release() to be rbd_dev_device_release(), making
it more obvious it serves as the inverse of the setup function
(or it will).

Encapsulate some of what was done in rbd_dev_release() into a new
function rbd_dev_image_release(), which serves as the inverse of
setting up the ceph side of the mapped rbd image.

Define a new helper rbd_dev_clear_mapping() to simply zero out the
fields of a mapping structure--the inverse of rbd_dev_set_mapping().

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: drop module later
Alex Elder [Mon, 29 Apr 2013 04:32:34 +0000 (23:32 -0500)]
rbd: drop module later

Drop the module reference at the end of rbd_remove() for symmetry
with adding a reference at the top of rbd_add().

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: set up watch in rbd_dev_image_probe()
Alex Elder [Sat, 27 Apr 2013 14:59:31 +0000 (09:59 -0500)]
rbd: set up watch in rbd_dev_image_probe()

Move setting up the watch request for an image so it's done in
rbd_dev_image_probe() rather than rbd_dev_probe_finish().  Move
it all the way up to before doing the initial probe.  This avoids
a potential race condition, in which we get (and use) the initial
snapshot context for an image, and it gets changed between that
time and the time we get the watch set up.

This resolves:
    http://tracker.ceph.com/issues/3871

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: don't bother checking whether order changes
Alex Elder [Sat, 27 Apr 2013 14:59:31 +0000 (09:59 -0500)]
rbd: don't bother checking whether order changes

When a format 2 image is refreshed, code is in place to verify that
the object order never changes from what it was originally.  This
relies on the fact that the refresh will occur *after* an initial
load of information about the image.

An upcoming patch makes it possible for the refresh to occur first,
so we can no longer make this order check.  The order really can't
ever change anyway--this was just a sanity check.  So get rid of it.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: don't clean up watch in device release function
Alex Elder [Sat, 27 Apr 2013 14:59:30 +0000 (09:59 -0500)]
rbd: don't clean up watch in device release function

Currently, a watch on an rbd device header object gets torn down
when its final Linux device reference gets dropped.  Instead, tear
it down when removing the device.  If an error occurs cleaning up
the watch event when unmapping, abort the unmap request.

All images (including parents) still get watch requests set up, so
tear these down also, in rbd_dev_remove_parent().  For now, ignore
any errors that occur in this case.

Get rid of local variable "rc" in rbd_remove(); use "ret" instead
(they both somehow ended up defined in the function and only one is
needed).

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: define rbd_header_name()
Alex Elder [Sat, 27 Apr 2013 14:59:30 +0000 (09:59 -0500)]
rbd: define rbd_header_name()

Define a new function rbd_header_name(), which allocates and formats
the name of the header object for the rbd device.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: move more initialization into rbd_dev_image_probe()
Alex Elder [Sat, 27 Apr 2013 14:59:30 +0000 (09:59 -0500)]
rbd: move more initialization into rbd_dev_image_probe()

Move a block of initialization related to the "ceph-side" of an rbd
image out of rbd_dev_probe_finish() and into rbd_dev_image_probe().

Add appropriate error handling to clean things up in the event any
of these new functions return an error.

We know that rbd_dev_snaps_update(), rbd_dev_spec_update(), and
rbd_dev_probe_parent() all clean up after themselves before they
return an error, so no special cleanup is required except when an
earlier call succeeds.  Since rbd_dev_spec_update() only updates the
spec field (whose cleanup will be handled by dropping the last
reference to the spec) there is no cleanup action associatied with
that.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: probe for the parent earlier
Alex Elder [Fri, 26 Apr 2013 20:44:37 +0000 (15:44 -0500)]
rbd: probe for the parent earlier

Probe for a parent device earlier in rbd_dev_probe_finish(), before
starting to set up the Linux side of the rbd device.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: remove parent devices on probe error
Alex Elder [Fri, 26 Apr 2013 20:44:36 +0000 (15:44 -0500)]
rbd: remove parent devices on probe error

When an error occurs while finishing probing a device it is assumed
that parent devices get cleaned up when deleting a device.  They
don't.  Add a call to clean them up.  Note that this means the
parent spec will already be cleaned up so it doesn't have to be
in one of the rbd_add() error paths.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: fix rbd_dev_remove_parent()
Alex Elder [Fri, 26 Apr 2013 20:44:36 +0000 (15:44 -0500)]
rbd: fix rbd_dev_remove_parent()

In certain error paths, it is possible for an rbd device to have a
parent spec but no parent rbd_dev.  In rbd_dev_remove_parent() use
the parent field rather than parent_spec in determining whether to
try to remove any parent devices.  Use assertions to indicate that
any non-null parent pointer has parent_spec associated with it.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: kill __rbd_remove()
Alex Elder [Fri, 26 Apr 2013 20:44:36 +0000 (15:44 -0500)]
rbd: kill __rbd_remove()

The function __rbd_remove() is used in two spots, and it's fairly
simple.  It combines cleanup of part of the ceph-side state as well
as cleaning up the Linux-side state.  Just open code it in the two
callers and eliminate the function.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: set mapping info earlier
Alex Elder [Sat, 27 Apr 2013 14:59:30 +0000 (09:59 -0500)]
rbd: set mapping info earlier

Set the mapping size and features earlier in rbd_dev_probe_finish().

Define rbd_dev_mapping_clear() as an inverse for setting those
fields, and use it both in error handling in rbd_dev_image_probe()
and in the final cleanup in rbd_dev_release().  Change the name
of rbd_dev_set_mapping() to of rbd_dev_mapping_set().

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: encapsulate removing parent devices
Alex Elder [Fri, 26 Apr 2013 20:44:36 +0000 (15:44 -0500)]
rbd: encapsulate removing parent devices

Encapsulate the code that removes an rbd device's parent images into
a new function, rbd_dev_remove_parent().

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: encapsulate probing for parent devices
Alex Elder [Fri, 26 Apr 2013 20:44:36 +0000 (15:44 -0500)]
rbd: encapsulate probing for parent devices

Encapsulate the code that probes for an rbd device's parent images
into a new function, rbd_dev_probe_parent().

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: defer setting disk capacity
Alex Elder [Fri, 26 Apr 2013 20:44:36 +0000 (15:44 -0500)]
rbd: defer setting disk capacity

Don't set the disk capacity until right before we announce the
device as available for use.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: only set device exists flag when ready
Alex Elder [Fri, 26 Apr 2013 20:44:36 +0000 (15:44 -0500)]
rbd: only set device exists flag when ready

Hold off setting the EXISTS rbd device flag until just before we
announce the disk as available for use.  There's no point in doing
so any earlier than that, and at that point the device truly is
fully set up and ready to use.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: fix up some sysfs stuff
Alex Elder [Fri, 26 Apr 2013 20:44:36 +0000 (15:44 -0500)]
rbd: fix up some sysfs stuff

This just tweaks a few things in the routines that implement
rbd sysfs files.

All of the entries for an rbd device in /sys/bus/rbd/devices/<id>/
will represent information whose valid values are known by the time
they are accessible.

Right now we get the size of the mapped image by a call to
get_capacity().  There's no need to do this, because that will
return what we last set the capacity to, which is just the size
recorded for the mapping.  So just show that value instead.

We also get this under protection of the header semaphore, in order
to provide a precisely correct value.  This isn't really necessary;
these files are really informational only and it's not necessary to
be so careful.

Finally, print a special value in case the major device number is
not recorded.  Right now that won't matter much but soon the parent
images won't have devices associated with them.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: fix a bug in resizing a mapping
Alex Elder [Fri, 26 Apr 2013 20:44:35 +0000 (15:44 -0500)]
rbd: fix a bug in resizing a mapping

When a snapshot context update occurs, rbd_update_mapping_size() is
called to set the capacity of the disk to record the updated
size of the image in case it has changed.

There's a bug though.  The mapping size is in units of *bytes*.  The
code that updates the mapping size field is assigning a value that
has been scaled down to *sectors*.

Fix that.  Also, check to see if the size has actually changed, and
don't bother updating things (specifically, calling set_capacity())
if it has not.

This resolves:
    http://tracker.ceph.com/issues/4833

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: refactor rbd_dev_probe_update_spec()
Alex Elder [Fri, 26 Apr 2013 14:43:48 +0000 (09:43 -0500)]
rbd: refactor rbd_dev_probe_update_spec()

Fairly straightforward refactoring of rbd_dev_probe_update_spec().
The name is changed to rbd_dev_spec_update().

Rearrange it so nothing gets assigned to the spec until all of the
names have been successfully acquired.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: rename rbd_dev_probe()
Alex Elder [Fri, 26 Apr 2013 14:43:48 +0000 (09:43 -0500)]
rbd: rename rbd_dev_probe()

Rename rbd_dev_probe() to be rbd_dev_image_probe().  Its purpose
will eventually be to probe for the existence of a valid rbd image
for the rbd device--focusing only on the ceph side and not the Linux
device side of initialization.

For now the two "sides" are not fully separated, and this function
is still the entry point for initializing the full rbd device.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: make rbd_dev_destroy() match rbd_dev_create()
Alex Elder [Fri, 26 Apr 2013 14:43:47 +0000 (09:43 -0500)]
rbd: make rbd_dev_destroy() match rbd_dev_create()

Currently, rbd_dev_destroy() does more than just the inverse of what
rbd_dev_create() does.  Stop doing that, and move the two extra
things it does into the three call sites.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: define rbd snap context routines
Alex Elder [Fri, 26 Apr 2013 14:43:47 +0000 (09:43 -0500)]
rbd: define rbd snap context routines

Encapsulate the creation of a snapshot context for rbd in a new
function rbd_snap_context_create().  Define rbd wrappers for getting
and dropping references to them once they're created.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: use rbd_warn(), not WARN_ON()
Alex Elder [Fri, 26 Apr 2013 14:43:47 +0000 (09:43 -0500)]
rbd: use rbd_warn(), not WARN_ON()

Change some calls to WARN_ON() so they use rbd_warn() instead, so we
get consistent messaging.  A few remain but they can probably just
go away eventually.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: move stripe_unit and stripe_count into header
Alex Elder [Fri, 26 Apr 2013 14:43:47 +0000 (09:43 -0500)]
rbd: move stripe_unit and stripe_count into header

This commit added fetching if fancy striping parameters:
    09186ddb rbd: get and check striping parameters

They are almost unused, but the two fields storing the information
really belonged in the rbd_image_header structure.

This patch moves them there.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: make rbd spec names pointer to const
Alex Elder [Fri, 26 Apr 2013 14:43:47 +0000 (09:43 -0500)]
rbd: make rbd spec names pointer to const

Make the names and image id in an rbd_spec be pointers to constant
data.  This required the use of a local variable to hold the
snapshot name in rbd_add_parse_args() to avoid a warning.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: set snapshot id in rbd_dev_probe_update_spec()
Alex Elder [Fri, 26 Apr 2013 04:15:08 +0000 (23:15 -0500)]
rbd: set snapshot id in rbd_dev_probe_update_spec()

Set the rbd spec's snapshot id for an image getting mapped in
rbd_dev_probe_update_spec() rather than rbd_dev_set_mapping().
This is the more logical place for that to happen (even though
it means we might look up the snapshot by name twice).

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: have snap_by_name() return a snapshot
Alex Elder [Fri, 26 Apr 2013 04:15:08 +0000 (23:15 -0500)]
rbd: have snap_by_name() return a snapshot

A function called snap_by_name() ought to just look up a snapshot by
name.  It does that, but then it assigns some stuff to the rbd
device structure as well.

Change the function to do just the lookup, and have the caller do
the assignments that follow.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: fix image id leak in initial probe
Alex Elder [Fri, 26 Apr 2013 04:15:08 +0000 (23:15 -0500)]
rbd: fix image id leak in initial probe

If a format 2 image id is found for an image being mapped, but the
subsequent probe of the image fails, rbd_dev_probe() quits without
freeing the image id.  Fix that.

Also drop a redundant hunk of code in rbd_dev_image_id().

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: have rbd_dev_image_id() set format 1 image id
Alex Elder [Fri, 26 Apr 2013 04:15:08 +0000 (23:15 -0500)]
rbd: have rbd_dev_image_id() set format 1 image id

Currently, rbd_dev_probe() assumes that any error returned by
rbd_dev_image_id() is most likely -ENOENT, and responds by
calling the format 1 probe routine, rbd_dev_v1_probe().  Then,
at the top of rbd_dev_v1_probe(), an empty string is allocated
for the image id.

This is sort of unbalanced.  Fix this by having rbd_dev_image_id()
look for -ENOENT from its "get_id" method call.  If that is seen,
have it allocate the empty string there rather than depending on
rbd_dev_v1_probe() to do it.

Given that this is effectively defining the format of the image,
set rbd_dev->image_format inside rbd_dev_image_id() rather than in
the format-specific probe routines.

Also drop a redundant hunk of code in rbd_dev_image_id().

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: avoid dropping extra reference in rbd_free_disk()
Alex Elder [Fri, 26 Apr 2013 04:15:08 +0000 (23:15 -0500)]
rbd: avoid dropping extra reference in rbd_free_disk()

I found during some failure injection testing that the call to
rbd_free_disk() in the error path of rbd_dev_probe_finish() was
dropping an extra reference to the disk queue.  The problem
occurred when put_disk tried to drop a reference to the disk's
queue.  A call to blk_cleanup_queue() just prior to that will have
also dropped a reference to the queue.

The problem is that the reference dropped by put_disk() is assumed
to have been taken by add_disk().  Our code has error paths that can
occur after the disk and its queue are initialized, but before the
call to add_disk(), and in those paths we won't have that extra
reference.

The fix is easy though.  In rbd_free_disk() we're already checking
the disk's GENHD_FL_UP flag.  That flag is an indication that
add_disk() has been called, so just call blk_cleanup_queue()
conditional on that flag being set.

This resolves:
    http://tracker.ceph.com/issues/4800

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: use rbd_obj_method_sync() return value
Alex Elder [Thu, 25 Apr 2013 20:09:42 +0000 (15:09 -0500)]
rbd: use rbd_obj_method_sync() return value

Now that rbd_obj_method_sync() returns the number of bytes
returned by the method call, that value should be used by
callers to ensure we don't overrun the valid portion of the
buffer.

Fix the two spots that remained that weren't doing that,
rbd_dev_image_name() and rbd_dev_v2_snap_name().

Rearrange the error path slightly in rbd_dev_v2_snap_name().

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: fix leak of format 2 snapshot names
Alex Elder [Thu, 25 Apr 2013 20:09:42 +0000 (15:09 -0500)]
rbd: fix leak of format 2 snapshot names

When the snapshot context for an rbd device gets updated (or the
initial one is recorded) a a list of snapshot structures is created
to represent them, one entry per snapshot.  Each entry includes a
dynamically-allocated copy of the snapshot name.

Currently the name is allocated in rbd_snap_create(), as a duplicate
of the passed-in name.

For format 1 images, the snapshot name provided is just a pointer to
an existing name.  But for format 2 images, the passed-in name is
already dynamically allocated, and in the the process of duplicating
it here we are leaking the passed-in name.

Fix this by dynamically allocating the name for format 1 snapshots
also, and then stop allocating a duplicate in rbd_snap_create().

Change rbd_dev_v1_snap_info() so none of its parameters is
side-effected unless it's going to return success.

This is part of:
    http://tracker.ceph.com/issues/4803

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: rename __rbd_add_snap_dev()
Alex Elder [Thu, 25 Apr 2013 20:09:41 +0000 (15:09 -0500)]
rbd: rename __rbd_add_snap_dev()

Rename __rbd_add_snap_dev() to be rbd_snap_create().  We no longer
have devices for non-mapped snapshots, and we're not actually
"adding" it to the list in this function, just creating it.

Rename rbd_remove_snap_dev() to be rbd_snap_destroy() for reasons
similar to the above.  Stop having this function delete the snapshot
from its list (to be symmetrical with its create counterpart) and do
that in the caller instead.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: only update values on snap_info success
Alex Elder [Thu, 25 Apr 2013 20:09:41 +0000 (15:09 -0500)]
rbd: only update values on snap_info success

Change rbd_dev_v2_snap_info() so it only ever sets values of the
size and features parameters if looking up the snapshot name was
successful.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: make snap_size order parameter optional
Alex Elder [Thu, 25 Apr 2013 20:09:41 +0000 (15:09 -0500)]
rbd: make snap_size order parameter optional

Only one of the two callers of _rbd_dev_v2_snap_size() needs the
order value returned.  So make that an optional argument--a null
pointer if the caller doesn't need it.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
11 years agorbd: fix leak of snapshots during initial probe
Alex Elder [Thu, 25 Apr 2013 20:09:41 +0000 (15:09 -0500)]
rbd: fix leak of snapshots during initial probe

When an rbd image is initially mapped, its snapshot context is
collected, and then a list of snapshot entries representing the
snapshots in that context is created.  The list is created using
rbd_dev_snaps_update().  (This function also supports updating an
existing snapshot list based on a new snapshot context.)

If an error occurs, updating the list is aborted, and the list is
currently left as-is, in an inconsistent state.  At that point,
there may be a partially-constructed list, but the calling functions
(rbd_dev_probe_finish() from rbd_dev_probe() from rbd_add()) never
clean them up.  So this constitutes a leak.

A snapshot list that is inconsistent with the current snapshot
context is of no use, and might even be actively bad.  So rather
than just having the caller clean it up, have rbd_dev_snaps_update()
just clear out the entire snapshot list in the event an error
occurs.

The other place rbd_dev_snaps_update() is used is when a refresh is
triggered, either because of a watch callback or via a write to the
/sys/bus/rbd/devices/<id>/refresh interface.  An error while
updating the snapshots has no substantive effect in either of those
cases, but one of them issues a warning.  Move that warning to the
common rbd_dev_refresh() function so it gets issued regardless of
how it got initiated.

This is part of:
    http://tracker.ceph.com/issues/4803

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>