This is overkill, but a first step before pushing caps enforcement down
into each subfunction that processes monitor commands.
Fixes: #2443
Signed-off-by: Sage Weil <sage@inktank.com>
If current exists, leave it be, but:
* make sure it's a dir
* check if it's a btrfs snapshot, and set btrfs_stable_commits if so
Signed-off-by: Sage Weil <sage@inktank.com>
doc/config-cluster/deploying-ceph-conf.rst <-added sudo. required.
doc/config-cluster/deploying-ceph-with-mkcephfs.rst <-cd to /etc/ceph so keyring goes there.
doc/install/chef.rst <-added update and install opscode-keyring, and upgrade.
doc/install/debian.rst <-added ceph-common to the install
Signed-off-by: John Wilkins <john.wilkins@dreamhost.com>
The addition of accounting for simultaneous writers in
f3043fee3e could lead to an infinite
loop. This could happen because starting a flush doesn't decrease
dirty_waiters, only the number of dirty bytes (they change to tx). If
dirty_waiters is more than target_dirty, the loop would continue
forever while holding the object cacher lock, since the object cacher
lock is required by the write callbacks.
The extra while (!flusher_stop) loop is unnecessary, so remove it.
This means the flusher thread always releases the lock after starting
flushing, so progress can be made.
Signed-off-by: Sage Weil <sage@inktank.com>
Signed-off-by: Josh Durgin <josh.durgin@inktank.com>
This lets us maintain a user's existing Connections even if the underlying
Pipe object disappears (ie, it's lossy and they don't drop the Connection
object).
This is a little distressing for me because it's again emphasizing the
Connection over the Pipe, but that seems to be an ongoing process.
Maybe the next Messenger implementation will be able to do away with
the (fairly messy) distinction between them.
Signed-off-by: Greg Farnum <greg@inktank.com>
This isn't quite complete, and should be integrated with the info in the
ops section on mon cluster expansion, but the raw content is here at least.
Signed-off-by: Sage Weil <sage@inktank.com>
doc/architecture.rst - removed broken reference.
doc/config-cluster - cleanup and added chef
doc/install - Made generic to add Chef, OpenStack and libvert installs
doc/init - Created light start | stop and health section
doc/source - Removed $ from code examples. Trimmed paras to 80 char
doc/images - Added preliminary diagram for Chef.
doc/rec - Added reference to hardware. Added filesystem info.
Signed-off-by: John Wilkins <john.wilkins@dreamhost.com>
It is valid to start with no initial monitors even when the initial set is
not defined; it just means that we are only able to join a cluster, and
never able to form a new one.
Signed-off-by: Sage Weil <sage@inktank.com>
