For new clusters.
This is the last "disruptive" tunable change we've made (where adjusting/
fixing it requires lots of data movement). For new clusters a lower bound
of jewel seems pretty reasonable, and this has been upstream in Linux
for a while now.
Signed-off-by: Sage Weil <sage@redhat.com>
We are explicitly accounting for buffers via arbitrary pools, defaulting
to buffer_anon, so we don't need buffer_data anymore. It leads to
double-counting.
Signed-off-by: Sage Weil <sage@redhat.com>
- default is buffer_anon
- buffer_data includes *all* buffers across all mempools, and is the
*actual* allocation (includes raw_combined).
- buffer_meta is other buffer-related overhead.
Signed-off-by: Sage Weil <sage@redhat.com>
set-redirect test creates more objects than default due to redirect mapping
(let's assume that we create 500 objects, oid 1 -> oid 501, oid 2 -> oid 502)
But the test should use objects within pre-defined range. For example,
If the number of object is 500, only 1~500 oid are allowed for test operations.
In the failure cases, test operation tried to use wrong oids (which means
higher than 500). This is because the bug in the setup phase.
If max_inflights > 1, set-redirect can be finished after setup is completed.
(so wrong oid is included in oid_in_use).
We need to completely remove redirect objects in oid_in_use before start the test.
Fixes: http://tracker.ceph.com/issues/20114
Signed-off-by: Myoungwon Oh <omwmw@sk.com>
The SELinux package now requires the ceph-disk binary but that one was
part of the ceph-osd package. The ceph-disk python library is already
packaged in ceph-base so moving ceph-disk to ceph-base seems like a
reasonable next step.
Signed-off-by: Boris Ranto <branto@redhat.com>
The state.mtx protects state.events accessed both from readdir and gc.
Fixes: http://tracker.ceph.com/issues/20121
Signed-off-by: Gui Hecheng <guihecheng@cmss.chinamobile.com>
When items are added:
- the value in the weight set is set to the target weight. It is assumed
that weight set are updated on a regular basis and will eventually be
set to a value that prevents excessive over/under filling.
- the value in the id list is set to the item id.
When items are removed, their weight / ids in all choose_args are
removed.
Signed-off-by: Loic Dachary <loic@dachary.org>
Instead of using a string, use the defined CEPH_RELEASE_* values. This
is simpler.
We allow decoding of post-kraken pre-luminous maps that use the string
values.
Signed-off-by: Sage Weil <sage@redhat.com>
The assembly file name suffix was changed from .S to .s,
which caused a build break on ppc64le architecture.
To fix this issue, the suffix will be changed back to .S
for ppc64le assembly files only. The x86 assembly files
can retain the .s suffix.
Fixes: http://tracker.ceph.com/issues/20106
Signed-off-by: Andrew Solomon <asolomon@us.ibm.com>
This commit fixes the file ownership for the /usr/bin/ and /etc/ceph
files and skips missing files as some of the files that we do specify
now can be missing from the system (not installed, e.f. radosgw).
Fixes: http://tracker.ceph.com/issues/20077
Signed-off-by: Boris Ranto <branto@redhat.com>
