Jason Dillaman [Wed, 8 Apr 2015 23:06:52 +0000 (19:06 -0400)]
librbd: avoid blocking AIO API methods
Enqueue all AIO API methods within the new librbd thread pool to
reduce the possibility of any blocking operations. To maintain
backwards compatibility with the legacy return codes of the API's
AIO methods, it's still possible to block attempting to acquire
the snap_lock.
Ken Dreyer [Fri, 13 Mar 2015 22:08:35 +0000 (16:08 -0600)]
packaging: include ceph_perf_objectstore
The /usr/bin/ceph_perf_objectstore file is installed by default. Prior
to this commit it was missing from the packaging. This caused the RPM to
fail to build in mock.
Add ceph_perf_objectstore to the "ceph-test" RPM and Debian package.
If we end up developing further ceph_perf_* utilities, it would make
sense to glob them all with a wildcard, similar to what we are doing
with all the ceph_test_* utilities in ceph-test.
Yehuda Sadeh [Thu, 14 May 2015 00:05:22 +0000 (17:05 -0700)]
rgw: merge manifests correctly when there's prefix override
Fixes: #11622
Backport: hammer, firefly
Prefix override happens in a manifest when a rados object does not
conform to the generic prefix set on the manifest. When merging
manifests (specifically being used in multipart objects upload), we need
to check if the rule that we try to merge has a prefix that is the same
as the previous rule. Beforehand we checked if both had the same
override_prefix setting, but that might not apply as both manifests
might have different prefixes.
Jon Bernard [Fri, 8 May 2015 15:54:06 +0000 (11:54 -0400)]
common/admin_socket: close socket descriptor in destructor
Long-running processes that do not reuse a single client connection will
see accumulating file descriptors as a result of not closing the
listening socket. In this case, eventually the system will reach
file-max and subsequent connections will fail.
Samuel Just [Tue, 21 Apr 2015 06:45:57 +0000 (23:45 -0700)]
OSD: handle the case where we resurrected an old, deleted pg
Prior to giant, we would skip pgs in load_pgs which were not present in
the current osdmap. Those pgs would eventually refer to very old
osdmaps, which we no longer have causing the assertion failure in 11429
once the osd is finally upgraded to a version which does not skip the
pgs. Instead, if we do not have the map for the pg epoch, complain to
the osd log and skip the pg.
Objects that start with underscore need to have an object locator,
this is due to an old behavior that we need to retain. Some objects
might have been created without the locator. This tool creates a new
rados object with the appropriate locator.
Yehuda Sadeh [Fri, 27 Mar 2015 23:32:48 +0000 (16:32 -0700)]
rgw: generate new tag for object when setting object attrs
Fixes: #11256
Backport: firefly, hammer
Beforehand we were reusing the object's tag, which is problematic as
this tag is used for bucket index updates, and we might be clobbering a
racing update (like object removal).
Ken Dreyer [Wed, 22 Apr 2015 22:36:42 +0000 (16:36 -0600)]
init-radosgw: run RGW as root
The ceph-radosgw service fails to start if the httpd package is not
installed. This is because the init.d file attempts to start the RGW
process with the "apache" UID. If a user is running civetweb, there is
no reason for the httpd or apache2 package to be present on the system.
Switch the init scripts to use "root" as is done on Ubuntu.
http://tracker.ceph.com/issues/11453 Refs: #11453
Reported-by: Vickey Singh <vickey.singh22693@gmail.com> Signed-off-by: Ken Dreyer <kdreyer@redhat.com>
(cherry picked from commit 47339c5ac352d305e68a58f3d744c3ce0fd3a2ac)
Boris Ranto [Mon, 13 Apr 2015 10:36:07 +0000 (12:36 +0200)]
Rework mds/Makefile.am to support a dencoder client build
The patch adds all the mds sources to DENCODER_SOURCES to allow a
dencoder client build. The patch also splits the Makefile.am file to
better accomodate the change.
Dencoder is built if ENABLE_CLIENT is set. However, the rgw/Makefile.am
populated DENCODER_SOURCES only if WITH_RADOSGW was set. The patch fixes
this and populates DENCODER_SOURES if ENABLE_CLIENT is set.
Sage Weil [Fri, 10 Apr 2015 15:43:45 +0000 (08:43 -0700)]
crush: fix has_v4_buckets()
alg, not type!
This bug made us incorrectly think we were using v4 features when user type
5 was being used. That's currently 'rack' with recent crush maps, but
was other types for clusters that were created with older versions. This
is clearly problematic as it will lock out non-hammer clients incorrectly,
breaking deployments on upgrade.
Guang Yang [Thu, 26 Feb 2015 08:13:12 +0000 (08:13 +0000)]
osd: fix negative degraded objects during backfilling
When there is deleting requests during backfilling, the reported number of degraded
objects could be negative, as the primary's num_objects is the latest (locally) but
the number for replicas might not reflect the deletings. A simple fix is to ignore
the negative subtracted value.
This can be done better in a separate script, which puts these in
CEPH_EXTRA_CONFIGURE_ARGS. In particular, this lets us enable
lttng for gitbuilder builds, but not release builds.
Jason Dillaman [Mon, 16 Mar 2015 22:40:49 +0000 (18:40 -0400)]
librbd: snap_remove should ignore -ENOENT errors
If the attempt to deregister the snapshot from the parent
image fails with -ENOENT, ignore the error as it is safe
to assume that the child is not associated with the parent.
Samuel Just [Thu, 26 Mar 2015 17:26:48 +0000 (10:26 -0700)]
ReplicatedPG::cancel_pull: requeue waiters as well
If we are in recovery_wait, we might not recover that object as part of
recover_primary for some time. Worse, if we are waiting on a backfill
which is blocked waiting on a copy_from on the missing object in
question, it can become a dead lock.
Fixes: 11244
Backport: firefly Signed-off-by: Samuel Just <sjust@redhat.com>
Sage Weil [Fri, 27 Mar 2015 22:35:21 +0000 (15:35 -0700)]
common: send cluster log messages to 'cluster' channel by default
The CLOG_CHANNEL_DEFAULT constant was being abused for two purposes:
- the default channel to log messages to
- the name of the config option key in the key/value pair string that is
used for the default option, e.g. "default=true foo=false bar=false"
Fix this by making the config option key CLOG_CONFIG_DEFAULT_KEY and
replacing throughout, and changing CLOG_CHANNEL_DEFAULT to "cluster" (as
it should be and has been historically).
Fixes: #11177 Signed-off-by: Sage Weil <sage@redhat.com>
Samuel Just [Tue, 24 Mar 2015 17:48:02 +0000 (10:48 -0700)]
PG: set/clear CREATING in Primary state entry/exit
Previously, we did not actually set it when we got a pg creation message from
the mon. It would actually get set on the first start_peering_interval after
that point. If we don't get that far, but do send a stat update to the mon, we
can end up with 11197. Instead, let's just set it and clear it upon entry into
and exit from the Primary state.
Fixes: 11197 Signed-off-by: Samuel Just <sjust@redhat.com>
Samuel Just [Tue, 24 Mar 2015 22:14:34 +0000 (15:14 -0700)]
ReplicatedPG: trim backfill intervals based on peer's last_backfill_started
Otherwise, we fail to trim the peer's last_backfill_started and get bug 11199.
1) osd 4 backfills up to 31bccdb2/mira01213209-286/head (henceforth: foo)
2) Interval change happens
3) osd 0 now finds itself backfilling to 4 (lb=foo) and osd.5
(lb=b6670ba2/mira01213209-160/snapdir//1, henceforth: bar)
4) recover_backfill causes both 4 and 5 to scan forward, so 4 has an interval
starting at foo, 5 has an interval starting at bar.
5) Once those have come back, recover_backfill attempts to trim off the
last_backfill_started, but 4's interval starts after that, so foo remains in
osd 4's interval (this is the bug)
7) We serve a copyfrom on foo (sent to 4 as well).
8) We eventually get to foo in the backfilling. Normally, they would have the
same version, but of course we don't update osd.4's interval from the log since
it should not have received writes in that interval. Thus, we end up trying to
recover foo on osd.4 anyway.
9) But, an interval change happens between removing foo from osd.4 and
completing the recovery, leaving osd.4 without foo, but with lb >= foo
Fixes: #11199
Backport: firefly Signed-off-by: Samuel Just <sjust@redhat.com>
Samuel Just [Fri, 20 Mar 2015 22:28:15 +0000 (15:28 -0700)]
ReplicatedPG::promote_object: check scrubber and block if necessary
Otherwise, we might attempt to promote into an in-progress scrub
interval causing 11156. I would have added a return value to
promote_object(), but could not find an existing user which
cared to distinguish the cases, even with a null op passed.
All existing users are in maybe_handle_cache. The ones which
pass a null op are for promoting the object in parallel
with a proxy -- a case where not actually performing the promote
does not really matter.
Fixes: #11156 Signed-off-by: Samuel Just <sjust@redhat.com>
Currently, this method also returns true if the object is backfilling.
This commit was reverted earlier in the branch in order to make the
other reverts clean. It's actually a nice rename though, so I'm
re-cherry-picking it.
Signed-off-by: Samuel Just <sjust@redhat.com>
Conflicts:
src/osd/ReplicatedPG.cc
projects / ceph.git / log