Use a separate lock file for prepare and activate to avoid deadlock. This
didn't seem to trigger on all machines, but in many cases, the prepare
process would take the file lock and later trigger a udev event and the
activate would then block on the same lock, either when we explicitly call
'udevadm settle --timeout=10' or when partprobe does it on our behalf
(without a timeout!). Avoid this by using separate locks for prepare
and activate. We only care if multiple activates race; it is
okay for a prepare to be in progress and for an activate to be kicked
off.
Signed-off-by: Sage Weil <sage@inktank.com>
Add installed, but not packaged files to ceph-test (ceph-monstore-tool,
ceph-osdomap-tool) rpm file section.
Signed-off-by: Danny Al-Gaaf <danny.al-gaaf@bisect.de>
We made the common_init_finish and chdir conditional on daemonize in commit
2e0dd5ae6c, breaking init (asok at least)
when -f is specified (as with upstart).
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Greg Farnum <greg@inktank.com>
This is needed when creating new OSDs (via ceph-disk). At least for most
people. Eventually we'll want to include btrfs here.
Signed-off-by: Sage Weil <sage@inktank.com>
This is what the upstart ceph-osd.conf does; we need to do the same so that
new OSDs (e.g., that ceph-deploy creates) get added to the crush map.
Signed-off-by: Sage Weil <sage@inktank.com>
leveldb has static state that prevents it from recreating its worker thread
after our fork(), even when we close and reopen the database (tsk tsk!).
Avoid this by forking early, before we touch leveldb.
Hide the details in a Preforker class. This is modeled after what
ceph-fuse already does; we should convert it later.
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Samuel Just <sam.just@inktank.com>
Reviewed-by: Greg Farnum <greg@inktank.com>
We need to go map-by-map to get the parents right in consume_map()
just as we must in load_pgs().
Fixes: 4884
Signed-off-by: Samuel Just <sam.just@inktank.com>
Reviewed-by: Greg Farnum <greg@inktank.com>
In load_pgs(), we previously called assigned children starting
at the loaded pg created between its stored epoch and the current
osdmap to have that pg as their parent. This is not correct, some
of the children may have been split in subsequent epochs from children
split in earlier epochs. Instead, do each map individually.
Signed-off-by: Samuel Just <sam.just@inktank.com>
expand_pg_num() and load_pgs() may result in a pg with children
in pending_splits which also have children in pending_splits (etc).
Signed-off-by: Samuel Just <sam.just@inktank.com>
Locking expectations changed at some point, and the Dumper wasn't
updated to comply:
1) We need to take the lock for Objecter, as it
doesn't do so on its own any more.
2) We need to drop the lock in several places so that Objecter
can take delivery of messages
Signed-off-by: Greg Farnum <greg@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
This reverts commit e725c3e210.
These inadvertantely got rid of the prefix portion of the key, which
lead to overwriting the wrong keys.
Fixes: #4872
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Samuel Just <sam.just@inktank.com>
The Paxos::trim() -> ::trim_to() path trims old states but does not
update first_committed. This misinforms later paxos rounds such that
peers think they can participate and end up with COMMIT messages
following the COLLECT/LAST exchange that are for future commits they
can't do anything with and then crash out when they get the BEGIN:
mon/Paxos.cc: 557: FAILED assert(begin->last_committed == last_committed)
Fixes: #4879
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Greg Farnum <greg@inktank.com>
We go to the effort of keeping a map of the peer's first/last committed
so that we can send the right commits during the first phase of paxos,
but we forgot to record the first value. This appears to simply be an
oversight. It is mostly harmless; it just means we send extra states
that the peer already has.
Signed-off-by: Sage Weil <sage@inktank.com>
Reviewed-by: Greg Farnum <greg@inktank.com>
There are two cases: 1) The parent pg has not yet initiated the split 2) The
parent pg has initiated the split.
Previously in case 1), _remove_pg left the entry for its children in the
in_progress_splits map blocking subsequent peering attempts.
In case 1), we need to unblock requests on the child pgs for the parent on
parent removal. We don't need to bother waking requests since any requests
received prior to the remove_pg request are necessarily obsolete.
In case 2), we don't need to do anything: the child will complete the split on
its own anyway.
Thus, we now track pending_splits vs in_progress_splits. Children in
pending_splits are in state 1), in_progress_splits in state 2). split_pgs
bumps pgs from pending_splits to in_progress_splits atomically with respect to
_remove_pg since the parent pg lock is held in both places.
Fixes: #4813
Signed-off-by: Samuel Just <sam.just@inktank.com>
Reviewed-by: Greg Farnum <greg@inktank.com>
We should only rely on whether our paxos version is overlap with whatever
they have -- we'll catch up later with them.
Signed-off-by: Joao Eduardo Luis <joao.luis@inktank.com>
We have timeouts that will clean everything up, and this can happen
in some cases that we've decided are legitimate. Hopefully we'll
be able to do something else later.
Signed-off-by: Greg Farnum <greg@inktank.com>
We reverted the gating by paxos sequences, so now we don't
need to look at them at all.
This reverts commit 1e6f02b337.
Signed-off-by: Greg Farnum <greg@inktank.com>
This was somehow broken -- out-of-date leaders were being elected -- and
we've decided smaller band-aids are more appropriate. We don't completely
revert the MMonElection changes, though -- there have been user clusters
running the code which includes these messages so we can't pretend it
never happened. We can make them clearly unused in the code, though.
This reverts commit fcaabf1a22.
Signed-off-by: Greg Farnum <greg@inktank.com>
Stopping the flusher is essentially the shutdown step for the
ObjectCacher - the next thing is actually destroying it.
If we leave any reads outstanding, when they complete they will
attempt to use the now-destroyed ObjectCacher. This is particularly a
problem with rbd images, since an -ENOENT can instantly complete many
readers, so the upper layers don't wait for the other rados-level
reads of that object to finish before trying to shutdown the cache.
Signed-off-by: Josh Durgin <josh.durgin@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
We need to call reset during every election cycle; luckily we
can call it more than once. bump_epoch is (by definition!) only called
once per cycle, and it's called at the beginning, so we put it there.
Fixes#4858.
Signed-off-by: Greg Farnum <greg@inktank.com>
Reviewed-by: Sage Weil <sage@inktank.com>
#leveldb on freenode says > 2MB is nonsense (it might explain the weird
behavior we saw). Riak tuning guide suggests 256KB for large data block
environments. Default is 8KB. 64KB seems sane for us.
Signed-off-by: Sage Weil <sage@inktank.com>
