Samuel Just [Tue, 16 Nov 2010 23:29:40 +0000 (15:29 -0800)]
Causes the MDSes to switch among a set of stray directories when
switching to a new journal segment.
MDSCache:
The stray member has been replaced with strays, an array of inodes
representing the set of available stray directories, as well as
stray_index indicating the index of the current stray directory.
get_stray() now returns a pointer to the current stray directory
inode.
advance_stray() advances stray_index to the next stray directory.
migrate_stray no longer takes a source argument, the source mds
is inferred from the parent of the dir entry.
stray dir entries are now stray<index> rather than stray.
scan_stray_dir now scans all stray directories.
MDSLog:
start_new_segment now calls advance_stray() on MDSCache to force a new
stray directory.
mdstypes:
NUM_STRAY indicates the number of stray directories to use per MDS
MDS_INO_STRAY now takes an index argument as well as the mds number
MDS_INO_STRAY_OWNER(i) returns the mds owner of the stray directory i
MDS_INO_STRAY_OWNER(i) returns the index of the stray directory i
Signed-off-by: Samuel Just <samuelj@hq.newdream.net>
generate_past_intervals:generate back to lastclean
PG::generate_past_intervals needs to generate all the intervals back to
history.last_epoch_clean, rather than just to
history.last_epoch_started. This is required by
PG::build_might_have_unfound, which needs to examine these intervals
when building the might_have_unfound set.
Move the check for whether past_intervals is up-to-date into
generate_past_intervals itself. Fix the check.
Signed-off-by: Colin McCabe <colinm@hq.newdream.net>
Sage Weil [Mon, 22 Nov 2010 17:49:43 +0000 (09:49 -0800)]
osd: bind to new cluster address when wrongly marked down
If we come back up on the same address, there is a possible race. Other
nodes will mark_down when they see us go down. If we go up first, queue
some messages, and _then_ they see that we're down and mark_down, the
messages we queued will get lost. Since it's stateful on the cluster
backend, we need to introduce an ordering so that closing out the _old_
session doesn't break the new session. We do this by binding to a new
address (just a different port, actually) before marking ourselves back
up.
Sage Weil [Mon, 22 Nov 2010 03:59:43 +0000 (19:59 -0800)]
osd: unconditionally set up separate msgr instance for osd<->osd msgs
Always set up cluster_messenger (before we would only do so if there was
an explicit address configured for it). The overhead to do so is minimal,
it simplifies the code, and will allow us to fix down->up transitions
(later).
The test for unfound objects was reversed, leading us to try to pull
unfound objects and refrain from pulling objects that we knew how to
get. Should fix bug #585.
Signed-off-by: Colin McCabe <colinm@hq.newdream.net>
ReplicatedPG::get_object_context takes three parameters. The last two
are "const object_locator_t& oloc" and "bool can_create".
Unfortunately, booleans can degrade to ints, and ints can be used to
initialize objects of type object_locator_t.
So when you make a call like:
> ctx->snapset_obc = get_object_context(snapoid, true);
What happens is that you actually call:
> get_object_context(snapoid, object_locator(1), false);
So you pass an invalid and *not* blank object_locator_t, and pass false
for can_create. This is not what the caller wanted. This change gets rid
of the default parameters and fixes the callers.
Signed-off-by: Colin McCabe <colinm@hq.newdream.net>
Don't loop in ReplicatedPG::start_recovery_ops. There is already a loop
in both recover_replicas and recover_primary that will try to do as many
recovery ops as it can, there's no need to repeat it. Also, the former
loop provably would never execute more than once because of the way
the code was structured.
If there are no more recovery operations to do, and PG::is_all_uptodate
is true at the end of ReplicatedPG::start_recovery_ops, call
finish_recovery.
Signed-off-by: Colin McCabe <colinm@hq.newdream.net>
Sage Weil [Wed, 17 Nov 2010 22:37:38 +0000 (14:37 -0800)]
osd: rev PG::Info encoding for last_epoch_clean change
This was missed by 184fbf582b27c10b47101735a4495fe8c73ad186, so any fs
created between now and then won't decode properly. It's more important
to make an fs prior to that work, though, so that the upgrade path from
the last stable version works.
Sage Weil [Wed, 17 Nov 2010 19:39:24 +0000 (11:39 -0800)]
mds: wrlock scatterlocks to prevent a gather racing with split/merge logging
We have the dirs split in our cache for some time while journaling it to
disk, before the fragment_notify goes out. Make sure we don't do a
scatterlock gather during that time that will confuse the inode auth (who
has their dirfrags fragmented differently).
Track discover requests by tid. The old system of tracking outstanding
discovers was kludgey and somewhat broken. Also there is a possibility
of getting dup replies if someone does kick_requests().
There is still room for improvement with the logic detemrining when a
discover is sent: we may want to discover multiple dirfrags in parallel,
but the current code will only do one at a time.
Signed-off-by: Sage Weil <sage@newdream.net>
comment
Jim Schutt [Wed, 17 Nov 2010 20:39:52 +0000 (13:39 -0700)]
Detect broken system linux/fiemap.h
RedHat 5.5 has a /usr/include/linux/fiemap.h, but it is
broken because it does not itself include linux/types.h.
As a result, __u64 and friends are not defined.
We have a Ceph-local copy of fiemap.h, so use it
if the system version is broken.
While we're at it, fix up the configure message to
note we're using a local copy.
Signed-off-by: Jim Schutt <jaschut@sandia.gov> Signed-off-by: Sage Weil <sage@newdream.net>
There are two phases in recovery: one where we get all the right objects
on to the primary, and another where we push all those objects out to
the replicas. Formerly, we would not start the second phase until there
were no missing objects at all on the primary.
This change modifies that so that we will start the second phase even if
there are unfound objects. However, we will still wait for all findable
missing objects to be brought to us, of course.
Get rid of uptodate_set. We can find the same information by looking at
the missing and missing_loc sets directly. Keeping the uptodate_set...
er... up-to-date would be very difficult in the presence of all the things
that can modify the missing and missing_loc sets.
Signed-off-by: Colin McCabe <colinm@hq.newdream.net>
Greg Farnum [Mon, 15 Nov 2010 20:59:29 +0000 (12:59 -0800)]
mds: Add bloom filter to CDir.
You can now add items to a bloom filter and check for their existence.
This is intended to be used when trimming items out of the cache; the
filter is cleared when you mark_complete and is not transferred between
nodes. Neither does it change how you set or remove the STATE_COMPLETE flag.
You must explicitly check the bloom filter as appropriate; likewise, if
you start to fill it in you must always continue filling it in until
you delete the current instance of the filter.
Add discover_all_missing. This function makes sure that we have messages
en route to any OSD that we think might have information that could help
us discover where our unfound objects lie.
We call this function:
* In do_peer, right after activating the PG
* In _process_pg_info, when we're the primary of an active PG
* From handle_pg_notify, when we're the primary of an active PG
Signed-off-by: Colin McCabe <colinm@hq.newdream.net>
PG::search_for_missing: fix a bug with the handling of MSG_OSD_PG_INFO
messages. Formerly, when processing these messages, we erroneously
assumed that there was nothing missing on the peer at all even in cases
where there were missing objects.
PG::merge_log: drop unused Missing parameter.
_process_pg_info: Don't assume that just because we requested a Log
message at some point, that that is the message we're prcessing.
Correctly handle cases where we didn't get the peer's Missing set or
Log.
Signed-off-by: Colin McCabe <colinm@hq.newdream.net>
Add MOSDPGMissing, a message which just contains the missing objects
information for a PG. We will request messages like this one in order to
locate all of our unfound objects.
Signed-off-by: Colin McCabe <colinm@hq.newdream.net>
Sage Weil [Mon, 15 Nov 2010 04:26:52 +0000 (20:26 -0800)]
msgr: use provided rx buffer if present
This changes the read path so that we hold the Connection::lock mutex while
reading data off the socket. This ensures that we are reading into a
buffer we are allowed to use, and allows users to revoke a previously
posted buffer. If that happens, switch over to a newly allocated buffer.
Note that currently the final result bufferlist may contain part of the
provided buffer and part of a newly allocated buffer. This is okay as long
as we will always read the same data into the buffer. And in practice, if
the rx buffer is revoked then the message itself will be thrown out anyway.
We have to explictly shut down the timer in Objecter::shutdown.
Otherwise, we are relying on the destructor of SafeTimer to do it.
Unfortunately, that destructor gets called after the mutex the timer is
using has already been destroyed.
Signed-off-by: Colin McCabe <colinm@hq.newdream.net>
Sage Weil [Fri, 12 Nov 2010 23:56:54 +0000 (15:56 -0800)]
msgr: do not clear halt_delivery
We need to keep the halt_delivery plug set on failure/shutdown in order to
prevent a racing reader from queuing new messages. The only time we clear
it is when we discard messages due to a session reset.
projects / ceph.git / log