openwrt/staging/blogic.git
12 years agolibceph: clear all flags on con_close
Sage Weil [Sat, 21 Jul 2012 00:30:40 +0000 (17:30 -0700)]
libceph: clear all flags on con_close

Signed-off-by: Sage Weil <sage@inktank.com>
12 years agolibceph: clean up con flags
Sage Weil [Sat, 21 Jul 2012 00:29:55 +0000 (17:29 -0700)]
libceph: clean up con flags

Rename flags with CON_FLAG prefix, move the definitions into the c file,
and (better) document their meaning.

Signed-off-by: Sage Weil <sage@inktank.com>
12 years agolibceph: replace connection state bits with states
Sage Weil [Sat, 21 Jul 2012 00:24:40 +0000 (17:24 -0700)]
libceph: replace connection state bits with states

Use a simple set of 6 enumerated values for the socket states (CON_STATE_*)
and use those instead of the state bits.  All of the con->state checks are
now under the protection of the con mutex, so this is safe.  It also
simplifies many of the state checks because we can check for anything other
than the expected state instead of various bits for races we can think of.

This appears to hold up well to stress testing both with and without socket
failure injection on the server side.

Signed-off-by: Sage Weil <sage@inktank.com>
12 years agolibceph: drop unnecessary CLOSED check in socket state change callback
Sage Weil [Sat, 21 Jul 2012 00:19:43 +0000 (17:19 -0700)]
libceph: drop unnecessary CLOSED check in socket state change callback

If we are CLOSED, the socket is closed and we won't get these.

Signed-off-by: Sage Weil <sage@inktank.com>
12 years agolibceph: close socket directly from ceph_con_close()
Sage Weil [Fri, 20 Jul 2012 23:45:49 +0000 (16:45 -0700)]
libceph: close socket directly from ceph_con_close()

It is simpler to do this immediately, since we already hold the con mutex.
It also avoids the need to deal with a not-quite-CLOSED socket in con_work.

Signed-off-by: Sage Weil <sage@inktank.com>
12 years agolibceph: drop gratuitous socket close calls in con_work
Sage Weil [Fri, 20 Jul 2012 22:40:04 +0000 (15:40 -0700)]
libceph: drop gratuitous socket close calls in con_work

If the state is CLOSED or OPENING, we shouldn't have a socket.

Signed-off-by: Sage Weil <sage@inktank.com>
12 years agolibceph: move ceph_con_send() closed check under the con mutex
Sage Weil [Fri, 20 Jul 2012 22:34:04 +0000 (15:34 -0700)]
libceph: move ceph_con_send() closed check under the con mutex

Take the con mutex before checking whether the connection is closed to
avoid racing with someone else closing it.

Signed-off-by: Sage Weil <sage@inktank.com>
12 years agolibceph: move msgr clear_standby under con mutex protection
Sage Weil [Fri, 20 Jul 2012 22:33:04 +0000 (15:33 -0700)]
libceph: move msgr clear_standby under con mutex protection

Avoid dropping and retaking con->mutex in the ceph_con_send() case by
leaving locking up to the caller.

Signed-off-by: Sage Weil <sage@inktank.com>
12 years agolibceph: fix fault locking; close socket on lossy fault
Sage Weil [Fri, 20 Jul 2012 22:22:53 +0000 (15:22 -0700)]
libceph: fix fault locking; close socket on lossy fault

If we fault on a lossy connection, we should still close the socket
immediately, and do so under the con mutex.

We should also take the con mutex before printing out the state bits in
the debug output.

Signed-off-by: Sage Weil <sage@inktank.com>
12 years agorbd: drop "object_name" from rbd_req_sync_unwatch()
Alex Elder [Wed, 25 Jul 2012 14:32:41 +0000 (09:32 -0500)]
rbd: drop "object_name" from rbd_req_sync_unwatch()

rbd_req_sync_unwatch() only ever uses rbd_dev->header_name as the
value of its "object_name" parameter, and that value is available
within the function already.  So get rid of the parameter.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: drop "object_name" from rbd_req_sync_notify_ack()
Alex Elder [Wed, 25 Jul 2012 14:32:40 +0000 (09:32 -0500)]
rbd: drop "object_name" from rbd_req_sync_notify_ack()

rbd_req_sync_notify_ack() only ever uses rbd_dev->header_name as the
value of its "object_name" parameter, and that value is available
within the function already.  So get rid of the parameter.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: drop "object_name" from rbd_req_sync_notify()
Alex Elder [Wed, 25 Jul 2012 14:32:40 +0000 (09:32 -0500)]
rbd: drop "object_name" from rbd_req_sync_notify()

rbd_req_sync_notify() only ever uses rbd_dev->header_name as the
value of its "object_name" parameter, and that value is available
within the function already.  So get rid of the parameter.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: drop "object_name" from rbd_req_sync_watch()
Alex Elder [Wed, 25 Jul 2012 14:32:40 +0000 (09:32 -0500)]
rbd: drop "object_name" from rbd_req_sync_watch()

rbd_req_sync_watch() is only called in one place, and in that place
it passes rbd_dev->header_name as the value of the "object_name"
parameter.  This value is available within the function already.

Having the extra parameter leaves the impression the object name
could take on different values, but it does not.

So get rid of the parameter.  We can always add it back again if
we find we want to watch some other object in the future.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: drop rbd_dev parameter in snap functions
Alex Elder [Thu, 19 Jul 2012 14:09:27 +0000 (09:09 -0500)]
rbd: drop rbd_dev parameter in snap functions

Both rbd_register_snap_dev() and __rbd_remove_snap_dev() have
rbd_dev parameters that are unused.  Remove them.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: drop rbd_header_from_disk() gfp_flags parameter
Alex Elder [Thu, 19 Jul 2012 14:09:27 +0000 (09:09 -0500)]
rbd: drop rbd_header_from_disk() gfp_flags parameter

The function rbd_header_from_disk() is only called in one spot, and
it passes GFP_KERNEL as its value for the gfp_flags parameter.

Just drop that parameter and substitute GFP_KERNEL everywhere within
that function it had been used.  (If we find we need the parameter
again in the future it's easy enough to add back again.)

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: snapc is unused in rbd_req_sync_read()
Alex Elder [Thu, 19 Jul 2012 14:09:27 +0000 (09:09 -0500)]
rbd: snapc is unused in rbd_req_sync_read()

The "snapc" parameter to in rbd_req_sync_read() is not used, so
get rid of it.

Reported-by: Josh Durgin <josh.durgin@inktank.com>
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: rename rbd_device->id
Alex Elder [Tue, 3 Jul 2012 21:01:19 +0000 (16:01 -0500)]
rbd: rename rbd_device->id

The "id" field of an rbd device structure represents the unique
client-local device id mapped to the underlying rbd image.  Each rbd
image will have another id--the image id--and each snapshot has its
own id as well.  The simple name "id" no longer conveys the
information one might like to have.

Rename the device "id" field in struct rbd_dev to be "dev_id" to
make it a little more obvious what we're dealing with without having
to think more about context.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: encapsulate header validity test
Alex Elder [Wed, 25 Jul 2012 14:32:40 +0000 (09:32 -0500)]
rbd: encapsulate header validity test

If an rbd image header is read and it doesn't begin with the
expected magic information, a warning is displayed.  This is
a fairly simple test, but it could be extended at some point.
Fix the comparison so it actually looks at the "text" field
rather than the front of the structure.

In any case, encapsulate the validity test in its own function.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agoceph: define snap counts as u32 everywhere
Alex Elder [Sat, 14 Jul 2012 01:35:11 +0000 (20:35 -0500)]
ceph: define snap counts as u32 everywhere

There are two structures in which a count of snapshots are
maintained:

    struct ceph_snap_context {
...
        u32 num_snaps;
...
    }
and
    struct ceph_snap_realm {
...
        u32 num_prior_parent_snaps;   /*  had prior to parent_since */
...
        u32 num_snaps;
...
    }

These fields never take on negative values (e.g., to hold special
meaning), and so are really inherently unsigned.  Furthermore they
take their value from over-the-wire or on-disk formatted 32-bit
values.

So change their definition to have type u32, and change some spots
elsewhere in the code to account for this change.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: clean up a few dout() calls
Alex Elder [Sat, 14 Jul 2012 01:35:11 +0000 (20:35 -0500)]
rbd: clean up a few dout() calls

There was a dout() call in rbd_do_request() that was reporting
the reporting the offset as the length and vice versa.  While
fixing that I did a quick scan of other dout() calls and fixed
a couple of other minor things.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: simplify __rbd_remove_all_snaps()
Alex Elder [Thu, 19 Jul 2012 14:09:27 +0000 (09:09 -0500)]
rbd: simplify __rbd_remove_all_snaps()

This just replaces a while loop with list_for_each_entry_safe()
in __rbd_remove_all_snaps().

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: drop extra header_rwsem init
Alex Elder [Thu, 19 Jul 2012 14:09:27 +0000 (09:09 -0500)]
rbd: drop extra header_rwsem init

In commit c666601a there was inadvertently added an extra
initialization of rbd_dev->header_rwsem.  This gets rid of the
duplicate.

Reported-by: Guangliang Zhao <gzhao@suse.com>
Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: kill rbd_image_header->snap_seq
Alex Elder [Thu, 19 Jul 2012 13:49:18 +0000 (08:49 -0500)]
rbd: kill rbd_image_header->snap_seq

The snap_seq field in an rbd_image_header structure held the value
from the rbd image header when it was last refreshed.  We now
maintain this value in the snapc->seq field.  So get rid of the
other one.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: set snapc->seq only when refreshing header
Alex Elder [Thu, 19 Jul 2012 13:49:18 +0000 (08:49 -0500)]
rbd: set snapc->seq only when refreshing header

In rbd_header_add_snap() there is code to set snapc->seq to the
just-added snapshot id.  This is the only remnant left of the
use of that field for recording which snapshot an rbd_dev was
associated with.  That functionality is no longer supported,
so get rid of that final bit of code.

Doing so means we never actually set snapc->seq any more.  On the
server, the snapshot context's sequence value represents the highest
snapshot id ever issued for a particular rbd image.  So we'll make
it have that meaning here as well.  To do so, set this value
whenever the rbd header is (re-)read.  That way it will always be
consistent with the rest of the snapshot context we maintain.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: preserve snapc->seq in rbd_header_set_snap()
Alex Elder [Thu, 19 Jul 2012 13:49:18 +0000 (08:49 -0500)]
rbd: preserve snapc->seq in rbd_header_set_snap()

In rbd_header_set_snap(), there is logic to make the snap context's
seq field get set to a particular snapshot id, or 0 if there is no
snapshot for the rbd image.

This seems to be an artifact of how the current snapshot id for an
rbd_dev was recorded before the rbd_dev->snap_id field began to be
used for that purpose.

There's no need to update the value of snapc->seq here any more, so
stop doing it.  Tidy up a few local variables in that function
while we're at it.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: don't use snapc->seq that way
Alex Elder [Thu, 19 Jul 2012 13:49:18 +0000 (08:49 -0500)]
rbd: don't use snapc->seq that way

In what appears to be an artifact of a different way of encoding
whether an rbd image maps a snapshot, __rbd_refresh_header() has
code that arranges to update the seq value in an rbd image's
snapshot context to point to the first entry in its snapshot
array if that's where it was pointing initially.

We now use rbd_dev->snap_id to record the snapshot id--using the
special value CEPH_NOSNAP to indicate the rbd_dev is not mapping a
snapshot at all.

There is therefore no need to check for this case, nor to update the
seq value, in __rbd_refresh_header().  Just preserve the seq value
that rbd_read_header() provides (which, at the moment, is nothing).

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: send header version when notifying
Josh Durgin [Tue, 6 Dec 2011 02:10:44 +0000 (18:10 -0800)]
rbd: send header version when notifying

Previously the original header version was sent. Now, we update it
when the header changes.

Signed-off-by: Josh Durgin <josh.durgin@dreamhost.com>
Reviewed-by: Alex Elder <elder@inktank.com>
12 years agorbd: use reference counting for the snap context
Josh Durgin [Mon, 5 Dec 2011 22:03:05 +0000 (14:03 -0800)]
rbd: use reference counting for the snap context

This prevents a race between requests with a given snap context and
header updates that free it. The osd client was already expecting the
snap context to be reference counted, since it get()s it in
ceph_osdc_build_request and put()s it when the request completes.

Also remove the second down_read()/up_read() on header_rwsem in
rbd_do_request, which wasn't actually preventing this race or
protecting any other data.

Signed-off-by: Josh Durgin <josh.durgin@dreamhost.com>
Reviewed-by: Alex Elder <elder@inktank.com>
12 years agorbd: set image size when header is updated
Josh Durgin [Mon, 5 Dec 2011 18:41:28 +0000 (10:41 -0800)]
rbd: set image size when header is updated

The image may have been resized.

Signed-off-by: Josh Durgin <josh.durgin@dreamhost.com>
Reviewed-by: Alex Elder <elder@inktank.com>
12 years agorbd: expose the correct size of the device in sysfs
Josh Durgin [Mon, 5 Dec 2011 18:35:04 +0000 (10:35 -0800)]
rbd: expose the correct size of the device in sysfs

If an image was mapped to a snapshot, the size of the head version
would be shown. Protect capacity with header_rwsem, since it may
change.

Signed-off-by: Josh Durgin <josh.durgin@dreamhost.com>
Reviewed-by: Alex Elder <elder@inktank.com>
12 years agorbd: only reset capacity when pointing to head
Josh Durgin [Tue, 22 Nov 2011 01:13:54 +0000 (17:13 -0800)]
rbd: only reset capacity when pointing to head

Snapshots cannot be resized, and the new capacity of head should not
be reflected by the snapshot.

Signed-off-by: Josh Durgin <josh.durgin@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
12 years agorbd: return errors for mapped but deleted snapshot
Josh Durgin [Tue, 22 Nov 2011 02:14:25 +0000 (18:14 -0800)]
rbd: return errors for mapped but deleted snapshot

When a snapshot is deleted, the OSD will return ENOENT when reading
from it. This is normally interpreted as a hole by rbd, which will
return zeroes. To minimize the time in which this can happen, stop
requests early when we are notified that our snapshot no longer
exists.

[elder@inktank.com: updated __rbd_init_snaps_header() logic]

Signed-off-by: Josh Durgin <josh.durgin@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
12 years agolibceph: trivial fix for the incorrect debug output
Jiaju Zhang [Fri, 20 Jul 2012 13:18:36 +0000 (08:18 -0500)]
libceph: trivial fix for the incorrect debug output

This is a trivial fix for the debug output, as it is inconsistent
with the function name so may confuse people when debugging.

[elder@inktank.com: switched to use __func__]

Signed-off-by: Jiaju Zhang <jjzhang@suse.de>
Reviewed-by: Alex Elder <elder@inktank.com>
12 years agoceph: fix potential double free
Alan Cox [Fri, 20 Jul 2012 13:18:36 +0000 (08:18 -0500)]
ceph: fix potential double free

We re-run the loop but we don't re-set the attrs pointer back to NULL.

Signed-off-by: Alan Cox <alan@linux.intel.com>
Reviewed-by: Alex Elder <elder@inktank.com>
12 years agolibceph: reset connection retry on successfully negotiation
Sage Weil [Mon, 30 Jul 2012 23:22:05 +0000 (16:22 -0700)]
libceph: reset connection retry on successfully negotiation

We exponentially back off when we encounter connection errors.  If several
errors accumulate, we will eventually wait ages before even trying to
reconnect.

Fix this by resetting the backoff counter after a successful negotiation/
connection with the remote node.  Fixes ceph issue #2802.

Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Yehuda Sadeh <yehuda@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
12 years agolibceph: protect ceph_con_open() with mutex
Sage Weil [Mon, 30 Jul 2012 23:21:40 +0000 (16:21 -0700)]
libceph: protect ceph_con_open() with mutex

Take the con mutex while we are initiating a ceph open.  This is necessary
because the may have previously been in use and then closed, which could
result in a racing workqueue running con_work().

Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Yehuda Sadeh <yehuda@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
12 years agoceph: close old con before reopening on mds reconnect
Sage Weil [Mon, 30 Jul 2012 23:21:17 +0000 (16:21 -0700)]
ceph: close old con before reopening on mds reconnect

When we detect a mds session reset, close the old ceph_connection before
reopening it.  This ensures we clean up the old socket properly and keep
the ceph_connection state correct.

Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
Reviewed-by: Yehuda Sadeh <yehuda@inktank.com>
12 years agolibceph: (re)initialize bio_iter on start of message receive
Sage Weil [Mon, 30 Jul 2012 23:20:25 +0000 (16:20 -0700)]
libceph: (re)initialize bio_iter on start of message receive

Previously, we were opportunistically initializing the bio_iter if it
appeared to be uninitialized in the middle of the read path.  The problem
is that a sequence like:

 - start reading message
 - initialize bio_iter
 - read half a message
 - messenger fault, reconnect
 - restart reading message
 - ** bio_iter now non-NULL, not reinitialized **
 - read past end of bio, crash

Instead, initialize the bio_iter unconditionally when we allocate/claim
the message for read.

Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
Reviewed-by: Yehuda Sadeh <yehuda@inktank.com>
12 years agolibceph: resubmit linger ops when pg mapping changes
Sage Weil [Mon, 30 Jul 2012 23:19:28 +0000 (16:19 -0700)]
libceph: resubmit linger ops when pg mapping changes

The linger op registration (i.e., watch) modifies the object state.  As
such, the OSD will reply with success if it has already applied without
doing the associated side-effects (setting up the watch session state).
If we lose the ACK and resubmit, we will see success but the watch will not
be correctly registered and we won't get notifies.

To fix this, always resubmit the linger op with a new tid.  We accomplish
this by re-registering as a linger (i.e., 'registered') if we are not yet
registered.  Then the second loop will treat this just like a normal
case of re-registering.

This mirrors a similar fix on the userland ceph.git, commit 5dd68b95, and
ceph bug #2796.

Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
Reviewed-by: Yehuda Sadeh <yehuda@inktank.com>
12 years agolibceph: fix mutex coverage for ceph_con_close
Sage Weil [Mon, 30 Jul 2012 23:24:37 +0000 (16:24 -0700)]
libceph: fix mutex coverage for ceph_con_close

Hold the mutex while twiddling all of the state bits to avoid possible
races.  While we're here, make not of why we cannot close the socket
directly.

Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
Reviewed-by: Yehuda Sadeh <yehuda@inktank.com>
12 years agolibceph: report socket read/write error message
Sage Weil [Mon, 30 Jul 2012 23:24:21 +0000 (16:24 -0700)]
libceph: report socket read/write error message

We need to set error_msg to something useful before calling ceph_fault();
do so here for try_{read,write}().  This is more informative than

libceph: osd0 192.168.106.220:6801 (null)

Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
Reviewed-by: Yehuda Sadeh <yehuda@inktank.com>
12 years agolibceph: support crush tunables
Sage Weil [Tue, 31 Jul 2012 01:15:23 +0000 (18:15 -0700)]
libceph: support crush tunables

The server side recently added support for tuning some magic
crush variables. Decode these variables if they are present, or use the
default values if they are not present.

Corresponds to ceph.git commit 89af369c25f274fe62ef730e5e8aad0c54f1e5a5.

Signed-off-by: caleb miles <caleb.miles@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
Reviewed-by: Yehuda Sadeh <yehuda@inktank.com>
12 years agolibceph: move feature bits to separate header
Sage Weil [Mon, 30 Jul 2012 23:23:22 +0000 (16:23 -0700)]
libceph: move feature bits to separate header

This is simply cleanup that will keep things more closely synced with the
userland code.

Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
Reviewed-by: Yehuda Sadeh <yehuda@inktank.com>
12 years agorbd: kill num_reply parameters
Alex Elder [Tue, 26 Jun 2012 19:57:03 +0000 (12:57 -0700)]
rbd: kill num_reply parameters

Several functions include a num_reply parameter, but it is never
used.  Just get rid of it everywhere--it seems to be something
that never got fully implemented.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: option symbol renames
Alex Elder [Tue, 3 Jul 2012 21:01:18 +0000 (16:01 -0500)]
rbd: option symbol renames

Use the name "ceph_opts" consistently (rather than just "opt") for
pointers to a ceph_options structure.

Change the few spots that don't use "rbd_opts" for a rbd_options
pointer to match the rest.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: more symbol renames
Alex Elder [Tue, 3 Jul 2012 21:01:18 +0000 (16:01 -0500)]
rbd: more symbol renames

Rename variables named "obj" which represent object names so they're
consistently named "object_name".

Rename the "cls" and "method" parameters in rbd_req_sync_exec()
to be "class_name" and "method_name", and make similar changes
to the names of local variables in that function representing
the lengths of those names.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: rename some fields in struct rbd_dev
Alex Elder [Tue, 3 Jul 2012 21:01:18 +0000 (16:01 -0500)]
rbd: rename some fields in struct rbd_dev

An rbd image is not a single object, but a logical construct made up
of an aggregation of objects.

Rename some fields in struct rbd_dev, in hopes of reinforcing this.
    obj         --> image_name
    obj_len     --> image_name_len
    obj_md_name --> header_name

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: use rbd_dev consistently
Alex Elder [Tue, 3 Jul 2012 21:01:18 +0000 (16:01 -0500)]
rbd: use rbd_dev consistently

Most variables that represent a struct rbd_device are named
"rbd_dev", but in some cases "dev" is used instead.  Change all the
"dev" references so they use "rbd_dev" consistently, to make it
clear from the name that we're working with an RBD device (as
opposed to, for example, a struct device).  Similarly, change the
name of the "dev" field in struct rbd_notify_info to be "rbd_dev".

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: dynamically allocate snapshot name
Alex Elder [Tue, 10 Jul 2012 02:04:24 +0000 (21:04 -0500)]
rbd: dynamically allocate snapshot name

There is no need to impose a small limit the length of the snapshot
name recorded for an rbd image in a struct rbd_dev.  Remove the
limitation by allocating space for the snapshot name dynamically.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: dynamically allocate image name
Alex Elder [Tue, 10 Jul 2012 02:04:23 +0000 (21:04 -0500)]
rbd: dynamically allocate image name

There is no need to impose a small limit the length of the rbd image
name recorded in a struct rbd_dev.  Remove the limitation by
allocating space for the image name dynamically.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: dynamically allocate image header name
Alex Elder [Tue, 10 Jul 2012 02:04:23 +0000 (21:04 -0500)]
rbd: dynamically allocate image header name

There is no need to impose a small limit the length of the header
name recorded for an rbd image in a struct rbd_dev.  Remove the
limitation by allocating space for the header name dynamically.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: dynamically allocate object prefix
Alex Elder [Tue, 10 Jul 2012 02:04:24 +0000 (21:04 -0500)]
rbd: dynamically allocate object prefix

There is no need to impose a small limit the length of the object
prefix recorded for an rbd image in a struct rbd_image_header.
Remove the limitation by allocating space for the object prefix
dynamically.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: dynamically allocate pool name
Alex Elder [Thu, 12 Jul 2012 15:46:35 +0000 (10:46 -0500)]
rbd: dynamically allocate pool name

There is no need to impose a small limit the length of the pool name
recorded for an rbd image in a struct rbd_device.  Remove the
limitation by allocating space for the pool name ynamically.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: create pool_id device attribute
Alex Elder [Thu, 12 Jul 2012 15:46:35 +0000 (10:46 -0500)]
rbd: create pool_id device attribute

Add an entry under /sys/bus/rbd/devices/<N>/ named "pool_id" that
provides the id for the pool the rbd image is assocatied with.  This
is in addition to the pool name already provided.

Rename the "poolid" field in struct rbd_device  to be "pool_id".

Update the documentation to reflect the addition of this new entry.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: rename rbd_dev->block_name
Alex Elder [Wed, 11 Jul 2012 01:30:09 +0000 (20:30 -0500)]
rbd: rename rbd_dev->block_name

Each rbd image has a name that forms the basis of all data objects
backing the device.  Old (format 1) images refer to this name as the
"block name," while new (format 2) images use the term "object
prefix" for this.

Change the field name in the in-core rbd image header structure to
reflect the more modern usage.  We intentionally keep the the name
"block_name" in the on-disk definition for format 1 image headers.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Yehuda Sadeh <yehuda@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agorbd: define dup_token()
Alex Elder [Tue, 10 Jul 2012 02:04:23 +0000 (21:04 -0500)]
rbd: define dup_token()

Define a new function dup_token(), to be used during argument
parsing for making dynamically-allocated copies of tokens being
parsed.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Yehuda Sadeh <yehuda@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agolibceph: define ceph_extract_encoded_string()
Alex Elder [Wed, 11 Jul 2012 13:24:45 +0000 (08:24 -0500)]
libceph: define ceph_extract_encoded_string()

This adds a new utility routine which will return a dynamically-
allocated buffer containing a string that has been decoded from ceph
over-the-wire format.  It also returns the length of the string
if the address of a size variable is supplied to receive it.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
12 years agorbd: drop a useless local variable
Alex Elder [Tue, 3 Jul 2012 21:01:19 +0000 (16:01 -0500)]
rbd: drop a useless local variable

In rbd_req_sync_notify_ack(), a local variable was needlessly being
used to hold a null pointer.  Just pass NULL instead.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Yehuda Sadeh <yehuda@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agolibceph: fix off-by-one bug in ceph_encode_filepath()
Alex Elder [Tue, 3 Jul 2012 21:01:18 +0000 (16:01 -0500)]
libceph: fix off-by-one bug in ceph_encode_filepath()

There is a BUG_ON() call that doesn't account for the single byte
structure version at the start of an encoded filepath in
ceph_encode_filepath().  Fix that.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Yehuda Sadeh <yehuda@inktank.com>
Reviewed-by: Josh Durgin <josh.durgin@inktank.com>
12 years agoceph: clean up useless d_parent checks
Sage Weil [Thu, 7 Jun 2012 20:43:35 +0000 (13:43 -0700)]
ceph: clean up useless d_parent checks

d_parent is never NULL, and IS_ROOT() is the proper way to check for a
(non-self-referential) parent.

Reported-by: Al Viro <viro@ZenIV.linux.org.uk>
Signed-off-by: Sage Weil <sage@inktank.com>
12 years agolibceph: prevent the race of incoming work during teardown
Guanjun He [Mon, 9 Jul 2012 02:50:33 +0000 (19:50 -0700)]
libceph: prevent the race of incoming work during teardown

Add an atomic variable 'stopping' as flag in struct ceph_messenger,
set this flag to 1 in function ceph_destroy_client(), and add the condition code
in function ceph_data_ready() to test the flag value, if true(1), just return.

Signed-off-by: Guanjun He <gjhe@suse.com>
Reviewed-by: Sage Weil <sage@inktank.com>
12 years agolibceph: fix messenger retry
Sage Weil [Tue, 10 Jul 2012 18:53:34 +0000 (11:53 -0700)]
libceph: fix messenger retry

In ancient times, the messenger could both initiate and accept connections.
An artifact if that was data structures to store/process an incoming
ceph_msg_connect request and send an outgoing ceph_msg_connect_reply.
Sadly, the negotiation code was referencing those structures and ignoring
important information (like the peer's connect_seq) from the correct ones.

Among other things, this fixes tight reconnect loops where the server sends
RETRY_SESSION and we (the client) retries with the same connect_seq as last
time.  This bug pretty easily triggered by injecting socket failures on the
MDS and running some fs workload like workunits/direct_io/test_sync_io.

Signed-off-by: Sage Weil <sage@inktank.com>
12 years agolibceph: initialize rb, list nodes in ceph_osd_request
Sage Weil [Mon, 9 Jul 2012 21:31:41 +0000 (14:31 -0700)]
libceph: initialize rb, list nodes in ceph_osd_request

These don't strictly need to be initialized based on how they are used, but
it is good practice to do so.

Reported-by: Alex Elder <elder@inktank.com>
Signed-off-by: Sage Weil <sage@inktank.com>
12 years agolibceph: initialize msgpool message types
Sage Weil [Mon, 9 Jul 2012 21:22:34 +0000 (14:22 -0700)]
libceph: initialize msgpool message types

Initialize the type field for messages in a msgpool.  The caller was doing
this for osd ops, but not for the reply messages.

Reported-by: Alex Elder <elder@inktank.com>
Signed-off-by: Sage Weil <sage@inktank.com>
12 years agolibceph: allow sock transition from CONNECTING to CLOSED
Sage Weil [Wed, 27 Jun 2012 19:31:02 +0000 (12:31 -0700)]
libceph: allow sock transition from CONNECTING to CLOSED

It is possible to close a socket that is in the OPENING state.  For
example, it can happen if ceph_con_close() is called on the con before
the TCP connection is established.  con_work() will come around and shut
down the socket.

Signed-off-by: Sage Weil <sage@inktank.com>
12 years agolibceph: initialize mon_client con only once
Sage Weil [Wed, 27 Jun 2012 19:24:34 +0000 (12:24 -0700)]
libceph: initialize mon_client con only once

Do not re-initialize the con on every connection attempt.  When we
ceph_con_close, there may still be work queued on the socket (e.g., to
close it), and re-initializing will clobber the work_struct state.

Signed-off-by: Sage Weil <sage@inktank.com>
12 years agolibceph: set peer name on con_open, not init
Sage Weil [Wed, 27 Jun 2012 19:24:08 +0000 (12:24 -0700)]
libceph: set peer name on con_open, not init

The peer name may change on each open attempt, even when the connection is
reused.

Signed-off-by: Sage Weil <sage@inktank.com>
12 years agolibceph: drop declaration of ceph_con_get()
Alex Elder [Thu, 21 Jun 2012 19:49:23 +0000 (12:49 -0700)]
libceph: drop declaration of ceph_con_get()

For some reason the declaration of ceph_con_get() and
ceph_con_put() did not get deleted in this commit:
    d59315ca libceph: drop ceph_con_get/put helpers and nref member

Clean that up.

Signed-off-by: Alex Elder <elder@inktank.com>
12 years agolibceph: add some fine ASCII art
Alex Elder [Thu, 21 Jun 2012 02:53:53 +0000 (21:53 -0500)]
libceph: add some fine ASCII art

Sage liked the state diagram I put in my commit description so
I'm putting it in with the code.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
12 years agolibceph: small changes to messenger.c
Alex Elder [Mon, 11 Jun 2012 19:57:13 +0000 (14:57 -0500)]
libceph: small changes to messenger.c

This patch gathers a few small changes in "net/ceph/messenger.c":
  out_msg_pos_next()
    - small logic change that mostly affects indentation
  write_partial_msg_pages().
    - use a local variable trail_off to represent the offset into
      a message of the trail portion of the data (if present)
    - once we are in the trail portion we will always be there, so we
      don't always need to check against our data position
    - avoid computing len twice after we've reached the trail
    - get rid of the variable tmpcrc, which is not needed
    - trail_off and trail_len never change so mark them const
    - update some comments
  read_partial_message_bio()
    - bio_iovec_idx() will never return an error, so don't bother
      checking for it

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
12 years agolibceph: distinguish two phases of connect sequence
Alex Elder [Thu, 24 May 2012 16:55:03 +0000 (11:55 -0500)]
libceph: distinguish two phases of connect sequence

Currently a ceph connection enters a "CONNECTING" state when it
begins the process of (re-)connecting with its peer.  Once the two
ends have successfully exchanged their banner and addresses, an
additional NEGOTIATING bit is set in the ceph connection's state to
indicate the connection information exhange has begun.  The
CONNECTING bit/state continues to be set during this phase.

Rather than have the CONNECTING state continue while the NEGOTIATING
bit is set, interpret these two phases as distinct states.  In other
words, when NEGOTIATING is set, clear CONNECTING.  That way only
one of them will be active at a time.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
12 years agolibceph: separate banner and connect writes
Alex Elder [Thu, 31 May 2012 16:37:29 +0000 (11:37 -0500)]
libceph: separate banner and connect writes

There are two phases in the process of linking together the two ends
of a ceph connection.  The first involves exchanging a banner and
IP addresses, and if that is successful a second phase exchanges
some detail about each side's connection capabilities.

When initiating a connection, the client side now queues to send
its information for both phases of this process at the same time.
This is probably a bit more efficient, but it is slightly messier
from a layering perspective in the code.

So rearrange things so that the client doesn't send the connection
information until it has received and processed the response in the
initial banner phase (in process_banner()).

Move the code (in the (con->sock == NULL) case in try_write()) that
prepares for writing the connection information, delaying doing that
until the banner exchange has completed.  Move the code that begins
the transition to this second "NEGOTIATING" phase out of
process_banner() and into its caller, so preparing to write the
connection information and preparing to read the response are
adjacent to each other.

Finally, preparing to write the connection information now requires
the output kvec to be reset in all cases, so move that into the
prepare_write_connect() and delete it from all callers.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
12 years agolibceph: define and use an explicit CONNECTED state
Alex Elder [Wed, 23 May 2012 19:35:23 +0000 (14:35 -0500)]
libceph: define and use an explicit CONNECTED state

There is no state explicitly defined when a ceph connection is fully
operational.  So define one.

It's set when the connection sequence completes successfully, and is
cleared when the connection gets closed.

Be a little more careful when examining the old state when a socket
disconnect event is reported.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
12 years agolibceph: clear NEGOTIATING when done
Alex Elder [Wed, 23 May 2012 19:35:23 +0000 (14:35 -0500)]
libceph: clear NEGOTIATING when done

A connection state's NEGOTIATING bit gets set while in CONNECTING
state after we have successfully exchanged a ceph banner and IP
addresses with the connection's peer (the server).  But that bit
is not cleared again--at least not until another connection attempt
is initiated.

Instead, clear it as soon as the connection is fully established.
Also, clear it when a socket connection gets prematurely closed
in the midst of establishing a ceph connection (in case we had
reached the point where it was set).

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
12 years agolibceph: clear CONNECTING in ceph_con_close()
Alex Elder [Thu, 21 Jun 2012 02:53:53 +0000 (21:53 -0500)]
libceph: clear CONNECTING in ceph_con_close()

A connection that is closed will no longer be connecting.  So
clear the CONNECTING state bit in ceph_con_close().  Similarly,
if the socket has been closed we no longer are in connecting
state (a new connect sequence will need to be initiated).

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
12 years agolibceph: don't touch con state in con_close_socket()
Alex Elder [Thu, 21 Jun 2012 02:53:53 +0000 (21:53 -0500)]
libceph: don't touch con state in con_close_socket()

In con_close_socket(), a connection's SOCK_CLOSED flag gets set and
then cleared while its shutdown method is called and its reference
gets dropped.

Previously, that flag got set only if it had not already been set,
so setting it in con_close_socket() might have prevented additional
processing being done on a socket being shut down.  We no longer set
SOCK_CLOSED in the socket event routine conditionally, so setting
that bit here no longer provides whatever benefit it might have
provided before.

A race condition could still leave the SOCK_CLOSED bit set even
after we've issued the call to con_close_socket(), so we still clear
that bit after shutting the socket down.  Add a comment explaining
the reason for this.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
12 years agolibceph: just set SOCK_CLOSED when state changes
Alex Elder [Thu, 21 Jun 2012 02:53:53 +0000 (21:53 -0500)]
libceph: just set SOCK_CLOSED when state changes

When a TCP_CLOSE or TCP_CLOSE_WAIT event occurs, the SOCK_CLOSED
connection flag bit is set, and if it had not been previously set
queue_con() is called to ensure con_work() will get a chance to
handle the changed state.

con_work() atomically checks--and if set, clears--the SOCK_CLOSED
bit if it was set.  This means that even if the bit were set
repeatedly, the related processing in con_work() only gets called
once per transition of the bit from 0 to 1.

What's important then is that we ensure con_work() gets called *at
least* once when a socket close event occurs, not that it gets
called *exactly* once.

The work queue mechanism already takes care of queueing work
only if it is not already queued, so there's no need for us
to call queue_con() conditionally.

So this patch just makes it so the SOCK_CLOSED flag gets set
unconditionally in ceph_sock_state_change().

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
12 years agolibceph: don't change socket state on sock event
Alex Elder [Thu, 21 Jun 2012 02:53:53 +0000 (21:53 -0500)]
libceph: don't change socket state on sock event

Currently the socket state change event handler records an error
message on a connection to distinguish a close while connecting from
a close while a connection was already established.

Changing connection information during handling of a socket event is
not very clean, so instead move this assignment inside con_work(),
where it can be done during normal connection-level processing (and
under protection of the connection mutex as well).

Move the handling of a socket closed event up to the top of the
processing loop in con_work(); there's no point in handling backoff
etc. if we have a newly-closed socket to take care of.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
12 years agolibceph: SOCK_CLOSED is a flag, not a state
Alex Elder [Thu, 21 Jun 2012 02:53:53 +0000 (21:53 -0500)]
libceph: SOCK_CLOSED is a flag, not a state

The following commit changed it so SOCK_CLOSED bit was stored in
a connection's new "flags" field rather than its "state" field.

    libceph: start separating connection flags from state
    commit 928443cd

That bit is used in con_close_socket() to protect against setting an
error message more than once in the socket event handler function.

Unfortunately, the field being operated on in that function was not
updated to be "flags" as it should have been.  This fixes that
error.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
12 years agolibceph: don't use bio_iter as a flag
Alex Elder [Mon, 11 Jun 2012 19:57:13 +0000 (14:57 -0500)]
libceph: don't use bio_iter as a flag

Recently a bug was fixed in which the bio_iter field in a ceph
message was not being properly re-initialized when a message got
re-transmitted:
    commit 43643528cce60ca184fe8197efa8e8da7c89a037
    Author: Yan, Zheng <zheng.z.yan@intel.com>
    rbd: Clear ceph_msg->bio_iter for retransmitted message

We are now only initializing the bio_iter field when we are about to
start to write message data (in prepare_write_message_data()),
rather than every time we are attempting to write any portion of the
message data (in write_partial_msg_pages()).  This means we no
longer need to use the msg->bio_iter field as a flag.

So just don't do that any more.  Trust prepare_write_message_data()
to ensure msg->bio_iter is properly initialized, every time we are
about to begin writing (or re-writing) a message's bio data.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
12 years agolibceph: move init of bio_iter
Alex Elder [Mon, 11 Jun 2012 19:57:13 +0000 (14:57 -0500)]
libceph: move init of bio_iter

If a message has a non-null bio pointer, its bio_iter field is
initialized in write_partial_msg_pages() if this has not been done
already.  This is really a one-time setup operation for sending a
message's (bio) data, so move that initialization code into
prepare_write_message_data() which serves that purpose.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
12 years agolibceph: move init_bio_*() functions up
Alex Elder [Mon, 11 Jun 2012 19:57:13 +0000 (14:57 -0500)]
libceph: move init_bio_*() functions up

Move init_bio_iter() and iter_bio_next() up in their source file so
the'll be defined before they're needed.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
12 years agolibceph: don't mark footer complete before it is
Alex Elder [Mon, 11 Jun 2012 19:57:13 +0000 (14:57 -0500)]
libceph: don't mark footer complete before it is

This is a nit, but prepare_write_message() sets the FOOTER_COMPLETE
flag before the CRC for the data portion (recorded in the footer)
has been completely computed.  Hold off setting the complete flag
until we've decided it's ready to send.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
12 years agolibceph: encapsulate advancing msg page
Alex Elder [Mon, 11 Jun 2012 19:57:13 +0000 (14:57 -0500)]
libceph: encapsulate advancing msg page

In write_partial_msg_pages(), once all the data from a page has been
sent we advance to the next one.  Put the code that takes care of
this into its own function.

While modifying write_partial_msg_pages(), make its local variable
"in_trail" be Boolean, and use the local variable "msg" (which is
just the connection's current out_msg pointer) consistently.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
12 years agolibceph: encapsulate out message data setup
Alex Elder [Mon, 11 Jun 2012 19:57:13 +0000 (14:57 -0500)]
libceph: encapsulate out message data setup

Move the code that prepares to write the data portion of a message
into its own function.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
12 years agolibceph: drop ceph_con_get/put helpers and nref member
Sage Weil [Thu, 21 Jun 2012 19:49:23 +0000 (12:49 -0700)]
libceph: drop ceph_con_get/put helpers and nref member

These are no longer used.  Every ceph_connection instance is embedded in
another structure, and refcounts manipulated via the get/put ops.

Signed-off-by: Sage Weil <sage@inktank.com>
12 years agolibceph: use con get/put methods
Sage Weil [Thu, 21 Jun 2012 19:47:08 +0000 (12:47 -0700)]
libceph: use con get/put methods

The ceph_con_get/put() helpers manipulate the embedded con ref
count, which isn't used now that ceph_connections are embedded in
other structures.

Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
12 years agolibceph: fix NULL dereference in reset_connection()
Dan Carpenter [Tue, 19 Jun 2012 13:52:33 +0000 (08:52 -0500)]
libceph: fix NULL dereference in reset_connection()

We dereference "con->in_msg" on the line after it was set to NULL.

Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Reviewed-by: Alex Elder <elder@inktank.com>
12 years agoMerge tag 'v3.5-rc1'
Sage Weil [Fri, 15 Jun 2012 19:32:04 +0000 (12:32 -0700)]
Merge tag 'v3.5-rc1'

Linux 3.5-rc1

Conflicts:
net/ceph/messenger.c

12 years agolibceph: flush msgr queue during mon_client shutdown
Sage Weil [Mon, 11 Jun 2012 03:43:56 +0000 (20:43 -0700)]
libceph: flush msgr queue during mon_client shutdown

We need to flush the msgr workqueue during mon_client shutdown to
ensure that any work affecting our embedded ceph_connection is
finished so that we can be safely destroyed.

Previously, we were flushing the work queue after osd_client
shutdown and before mon_client shutdown to ensure that any osd
connection refs to authorizers are flushed.  Remove the redundant
flush, and document in the comment that the mon_client flush is
needed to cover that case as well.

Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
12 years agolibceph: transition socket state prior to actual connect
Sage Weil [Sat, 9 Jun 2012 21:19:21 +0000 (14:19 -0700)]
libceph: transition socket state prior to actual connect

Once we call ->connect(), we are racing against the actual
connection, and a subsequent transition from CONNECTING ->
CONNECTED.  Set the state to CONNECTING before that, under the
protection of the mutex, to avoid the race.

This was introduced in 928443cd9644e7cfd46f687dbeffda2d1a357ff9,
with the original socket state code.

Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Alex Elder <elder@inktank.com>
12 years agolibceph: fix overflow in osdmap_apply_incremental()
Xi Wang [Thu, 7 Jun 2012 00:35:55 +0000 (19:35 -0500)]
libceph: fix overflow in osdmap_apply_incremental()

On 32-bit systems, a large `pglen' would overflow `pglen*sizeof(u32)'
and bypass the check ceph_decode_need(p, end, pglen*sizeof(u32), bad).
It would also overflow the subsequent kmalloc() size, leading to
out-of-bounds write.

Signed-off-by: Xi Wang <xi.wang@gmail.com>
Reviewed-by: Alex Elder <elder@inktank.com>
12 years agolibceph: fix overflow in osdmap_decode()
Xi Wang [Thu, 7 Jun 2012 00:35:55 +0000 (19:35 -0500)]
libceph: fix overflow in osdmap_decode()

On 32-bit systems, a large `n' would overflow `n * sizeof(u32)' and bypass
the check ceph_decode_need(p, end, n * sizeof(u32), bad).  It would also
overflow the subsequent kmalloc() size, leading to out-of-bounds write.

Signed-off-by: Xi Wang <xi.wang@gmail.com>
Reviewed-by: Alex Elder <elder@inktank.com>
12 years agolibceph: fix overflow in __decode_pool_names()
Xi Wang [Thu, 7 Jun 2012 00:35:55 +0000 (19:35 -0500)]
libceph: fix overflow in __decode_pool_names()

`len' is read from network and thus needs validation.  Otherwise a
large `len' would cause out-of-bounds access via the memcpy() call.
In addition, len = 0xffffffff would overflow the kmalloc() size,
leading to out-of-bounds write.

This patch adds a check of `len' via ceph_decode_need().  Also use
kstrndup rather than kmalloc/memcpy.

[elder@inktank.com: added -ENOMEM return for null kstrndup() result]

Signed-off-by: Xi Wang <xi.wang@gmail.com>
Reviewed-by: Alex Elder <elder@inktank.com>
12 years agorbd: Clear ceph_msg->bio_iter for retransmitted message
Yan, Zheng [Thu, 7 Jun 2012 00:35:55 +0000 (19:35 -0500)]
rbd: Clear ceph_msg->bio_iter for retransmitted message

The bug can cause NULL pointer dereference in write_partial_msg_pages

Signed-off-by: Zheng Yan <zheng.z.yan@intel.com>
Reviewed-by: Alex Elder <elder@inktank.com>
12 years agolibceph: make ceph_con_revoke_message() a msg op
Alex Elder [Fri, 1 Jun 2012 19:56:43 +0000 (14:56 -0500)]
libceph: make ceph_con_revoke_message() a msg op

ceph_con_revoke_message() is passed both a message and a ceph
connection.  A ceph_msg allocated for incoming messages on a
connection always has a pointer to that connection, so there's no
need to provide the connection when revoking such a message.

Note that the existing logic does not preclude the message supplied
being a null/bogus message pointer.  The only user of this interface
is the OSD client, and the only value an osd client passes is a
request's r_reply field.  That is always non-null (except briefly in
an error path in ceph_osdc_alloc_request(), and that drops the
only reference so the request won't ever have a reply to revoke).
So we can safely assume the passed-in message is non-null, but add a
BUG_ON() to make it very obvious we are imposing this restriction.

Rename the function ceph_msg_revoke_incoming() to reflect that it is
really an operation on an incoming message.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
12 years agolibceph: make ceph_con_revoke() a msg operation
Alex Elder [Fri, 1 Jun 2012 19:56:43 +0000 (14:56 -0500)]
libceph: make ceph_con_revoke() a msg operation

ceph_con_revoke() is passed both a message and a ceph connection.
Now that any message associated with a connection holds a pointer
to that connection, there's no need to provide the connection when
revoking a message.

This has the added benefit of precluding the possibility of the
providing the wrong connection pointer.  If the message's connection
pointer is null, it is not being tracked by any connection, so
revoking it is a no-op.  This is supported as a convenience for
upper layers, so they can revoke a message that is not actually
"in flight."

Rename the function ceph_msg_revoke() to reflect that it is really
an operation on a message, not a connection.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
12 years agolibceph: have messages take a connection reference
Alex Elder [Mon, 4 Jun 2012 19:43:33 +0000 (14:43 -0500)]
libceph: have messages take a connection reference

There are essentially two types of ceph messages: incoming and
outgoing.  Outgoing messages are always allocated via ceph_msg_new(),
and at the time of their allocation they are not associated with any
particular connection.  Incoming messages are always allocated via
ceph_con_in_msg_alloc(), and they are initially associated with the
connection from which incoming data will be placed into the message.

When an outgoing message gets sent, it becomes associated with a
connection and remains that way until the message is successfully
sent.  The association of an incoming message goes away at the point
it is sent to an upper layer via a con->ops->dispatch method.

This patch implements reference counting for all ceph messages, such
that every message holds a reference (and a pointer) to a connection
if and only if it is associated with that connection (as described
above).

For background, here is an explanation of the ceph message
lifecycle, emphasizing when an association exists between a message
and a connection.

Outgoing Messages
An outgoing message is "owned" by its allocator, from the time it is
allocated in ceph_msg_new() up to the point it gets queued for
sending in ceph_con_send().  Prior to that point the message's
msg->con pointer is null; at the point it is queued for sending its
message pointer is assigned to refer to the connection.  At that
time the message is inserted into a connection's out_queue list.

When a message on the out_queue list has been sent to the socket
layer to be put on the wire, it is transferred out of that list and
into the connection's out_sent list.  At that point it is still owned
by the connection, and will remain so until an acknowledgement is
received from the recipient that indicates the message was
successfully transferred.  When such an acknowledgement is received
(in process_ack()), the message is removed from its list (in
ceph_msg_remove()), at which point it is no longer associated with
the connection.

So basically, any time a message is on one of a connection's lists,
it is associated with that connection.  Reference counting outgoing
messages can thus be done at the points a message is added to the
out_queue (in ceph_con_send()) and the point it is removed from
either its two lists (in ceph_msg_remove())--at which point its
connection pointer becomes null.

Incoming Messages
When an incoming message on a connection is getting read (in
read_partial_message()) and there is no message in con->in_msg,
a new one is allocated using ceph_con_in_msg_alloc().  At that
point the message is associated with the connection.  Once that
message has been completely and successfully read, it is passed to
upper layer code using the connection's con->ops->dispatch method.
At that point the association between the message and the connection
no longer exists.

Reference counting of connections for incoming messages can be done
by taking a reference to the connection when the message gets
allocated, and releasing that reference when it gets handed off
using the dispatch method.

We should never fail to get a connection reference for a
message--the since the caller should already hold one.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
12 years agolibceph: have messages point to their connection
Alex Elder [Fri, 1 Jun 2012 19:56:43 +0000 (14:56 -0500)]
libceph: have messages point to their connection

When a ceph message is queued for sending it is placed on a list of
pending messages (ceph_connection->out_queue).  When they are
actually sent over the wire, they are moved from that list to
another (ceph_connection->out_sent).  When acknowledgement for the
message is received, it is removed from the sent messages list.

During that entire time the message is "in the possession" of a
single ceph connection.  Keep track of that connection in the
message.  This will be used in the next patch (and is a helpful
bit of information for debugging anyway).

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
12 years agolibceph: tweak ceph_alloc_msg()
Alex Elder [Mon, 4 Jun 2012 19:43:32 +0000 (14:43 -0500)]
libceph: tweak ceph_alloc_msg()

The function ceph_alloc_msg() is only used to allocate a message
that will be assigned to a connection's in_msg pointer.  Rename the
function so this implied usage is more clear.

In addition, make that assignment inside the function (again, since
that's precisely what it's intended to be used for).  This allows us
to return what is now provided via the passed-in address of a "skip"
variable.  The return type is now Boolean to be explicit that there
are only two possible outcomes.

Make sure the result of an ->alloc_msg method call always sets the
value of *skip properly.

Signed-off-by: Alex Elder <elder@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>