]> git.apps.os.sepia.ceph.com Git - ceph.git/log
ceph.git
16 years agomon: dump full pgmap on each state change (for debugging)
Sage Weil [Fri, 5 Dec 2008 22:29:17 +0000 (14:29 -0800)]
mon: dump full pgmap on each state change (for debugging)

16 years agoosd: don't die on stray sub op acks
Sage Weil [Fri, 5 Dec 2008 22:27:25 +0000 (14:27 -0800)]
osd: don't die on stray sub op acks

If a replica drops out of the pg, we force and ack in on_change(), but
may still get it later.  Don't freak out.

16 years agoosd: generate_backlog sanity check
Sage Weil [Fri, 5 Dec 2008 19:48:27 +0000 (11:48 -0800)]
osd: generate_backlog sanity check

If item is on disk and log, then log entry shouldn't be a delete.

16 years agoosd: fix merge_log divergent item detection
Sage Weil [Fri, 5 Dec 2008 19:46:55 +0000 (11:46 -0800)]
osd: fix merge_log divergent item detection

An item in our log isn't divergent if it is below the bottom of
olog.  Using the last_kept item isn't helpful here because
last_kept is in olog, and may be below that log's bottom.

16 years agoosd: clean on ondisklog a bit
Sage Weil [Fri, 5 Dec 2008 19:45:05 +0000 (11:45 -0800)]
osd: clean on ondisklog a bit

Express as extent, not interval.

16 years agoosd: make read_log output a bit more informative
Sage Weil [Fri, 5 Dec 2008 19:18:25 +0000 (11:18 -0800)]
osd: make read_log output a bit more informative

16 years agovstart: only sudo if -e dev/sudo
Sage Weil [Fri, 5 Dec 2008 19:08:53 +0000 (11:08 -0800)]
vstart: only sudo if -e dev/sudo

16 years agoosd: revise missing map adjustment
Sage Weil [Fri, 5 Dec 2008 19:00:47 +0000 (11:00 -0800)]
osd: revise missing map adjustment

Rewrite helpers in terms of how they are actually used.

16 years agoosd: mark backlog events as BACKLOG
Sage Weil [Fri, 5 Dec 2008 18:59:23 +0000 (10:59 -0800)]
osd: mark backlog events as BACKLOG

This is purely to make the logs easier to read.

16 years agoosd: generate_backlog fixes
Sage Weil [Fri, 5 Dec 2008 18:01:28 +0000 (10:01 -0800)]
osd: generate_backlog fixes

Generate backlog records even if the object appears in the log if
the existing entry's prior_version in non-zero and isn't also
in the log.  This allows us to accurately generate the .have field
when we are building the missing map.

16 years agocrush: add include
Sage Weil [Fri, 5 Dec 2008 04:49:11 +0000 (20:49 -0800)]
crush: add include

16 years agokclient: reduce stack usage
Yehuda Sadeh [Fri, 5 Dec 2008 00:50:50 +0000 (16:50 -0800)]
kclient: reduce stack usage

16 years agomon: 'osd scrub \*' to scrub all osds
Sage Weil [Fri, 5 Dec 2008 00:32:53 +0000 (16:32 -0800)]
mon: 'osd scrub \*' to scrub all osds

16 years agofilestore: fix up listxattr buffer management a bit
Sage Weil [Fri, 5 Dec 2008 00:28:33 +0000 (16:28 -0800)]
filestore: fix up listxattr buffer management a bit

16 years agodstart: put debug output on local disk
Sage Weil [Fri, 5 Dec 2008 00:27:28 +0000 (16:27 -0800)]
dstart: put debug output on local disk

16 years agodebug: allow output and output symlinks to go in different directories
Sage Weil [Fri, 5 Dec 2008 00:27:09 +0000 (16:27 -0800)]
debug: allow output and output symlinks to go in different directories

16 years agologmonitor: append all notifications in a single file
Yehuda Sadeh [Fri, 5 Dec 2008 00:14:55 +0000 (16:14 -0800)]
logmonitor: append all notifications in a single file

16 years agoset/check subprotocol versions
Sage Weil [Thu, 4 Dec 2008 22:58:58 +0000 (14:58 -0800)]
set/check subprotocol versions

16 years agomon: clean up paxos service registration a bit. rev disk format.
Sage Weil [Thu, 4 Dec 2008 22:30:11 +0000 (14:30 -0800)]
mon: clean up paxos service registration a bit.  rev disk format.

16 years agocleanup, whitespace
Yehuda Sadeh [Thu, 4 Dec 2008 22:08:39 +0000 (14:08 -0800)]
cleanup, whitespace

16 years agolog: use of cascading dispatcher for log messages
Yehuda Sadeh [Thu, 4 Dec 2008 21:59:34 +0000 (13:59 -0800)]
log: use of cascading dispatcher for log messages

16 years agodispatcher: cascading dispatch infrastructure
Yehuda Sadeh [Thu, 4 Dec 2008 21:31:00 +0000 (13:31 -0800)]
dispatcher: cascading dispatch infrastructure

16 years agomon: keep pgmap consistent
Sage Weil [Thu, 4 Dec 2008 21:46:19 +0000 (13:46 -0800)]
mon: keep pgmap consistent

We were cutting corners and updating the live map before it
committed to paxos, since pg stats aren't system critical.  This
can lead to problems due to the way "latest" is saved out, though,
and it can be confusing to see things jump backward in time.

16 years agoosd: make replica scrub_map generation a subop
Sage Weil [Thu, 4 Dec 2008 21:41:29 +0000 (13:41 -0800)]
osd: make replica scrub_map generation a subop

This puts build_scrub_map in a worker thread, _and_ ensures it is
serialized wrt any in-progress writes.

16 years agologclient: always print log messages to debug output
Sage Weil [Thu, 4 Dec 2008 21:40:30 +0000 (13:40 -0800)]
logclient: always print log messages to debug output

16 years agoosd: fix up scrub error log formatting
Sage Weil [Thu, 4 Dec 2008 21:01:28 +0000 (13:01 -0800)]
osd: fix up scrub error log formatting

16 years agologclient: optionally take a stringstream
Sage Weil [Thu, 4 Dec 2008 21:01:19 +0000 (13:01 -0800)]
logclient: optionally take a stringstream

16 years agoosd: some scrub fixes
Sage Weil [Thu, 4 Dec 2008 20:18:07 +0000 (12:18 -0800)]
osd: some scrub fixes

Don't drop locks just yet; atm this leaves the dout() prefix
exposed to concurrent modifications of pg state.

Don't requeue for scrub if already scrubbing.

Fix missing object detection bugs.

16 years agoosd: fix pg_stats.reported value
Sage Weil [Thu, 4 Dec 2008 19:57:18 +0000 (11:57 -0800)]
osd: fix pg_stats.reported value

16 years agoosd: drop lock during most of scrub; only disallow concurrent writes
Sage Weil [Thu, 4 Dec 2008 19:17:58 +0000 (11:17 -0800)]
osd: drop lock during most of scrub; only disallow concurrent writes

Make the PG go read-only during a scrub.  Only take the pg lock
when absolutely necessary.  Wait for any pending writes to
complete before starting the scrub.

16 years agoosd: ignore dup scrub maps
Sage Weil [Thu, 4 Dec 2008 19:16:25 +0000 (11:16 -0800)]
osd: ignore dup scrub maps

16 years agoosd: take pg ref on scrub_wq
Sage Weil [Thu, 4 Dec 2008 19:15:46 +0000 (11:15 -0800)]
osd: take pg ref on scrub_wq

16 years agoosd: make scrub verify replica object attrs match
Sage Weil [Thu, 4 Dec 2008 18:58:23 +0000 (10:58 -0800)]
osd: make scrub verify replica object attrs match

16 years agoosd: fix pg stat acking in osd
Sage Weil [Thu, 4 Dec 2008 18:57:13 +0000 (10:57 -0800)]
osd: fix pg stat acking in osd

16 years agolog: logclient uses log types instead of log level
Yehuda Sadeh [Thu, 4 Dec 2008 18:47:35 +0000 (10:47 -0800)]
log: logclient uses log types instead of log level

16 years agoosd: check for missing clones in pick_read_snap
Sage Weil [Thu, 4 Dec 2008 18:08:22 +0000 (10:08 -0800)]
osd: check for missing clones in pick_read_snap

We may need to wait from op_read if we are missing a specific
clone.

16 years agoosd: clear waiting_for_head when we pull the head; set skipped if we do
Sage Weil [Thu, 4 Dec 2008 18:07:43 +0000 (10:07 -0800)]
osd: clear waiting_for_head when we pull the head; set skipped if we do

Need to ++skipped if we skip because we're waiting for the head
or else we'll incorrectly advanced requested_to.

Clear waiting_for_head entry when we pull a head we're waiting for.

16 years agoosd: fix missing.add_event
Sage Weil [Thu, 4 Dec 2008 18:06:09 +0000 (10:06 -0800)]
osd: fix missing.add_event

We only should set have to prior_version if we aren't missing the
prior_version too!

16 years agokclient: fix NULL dereferencing oops
Yehuda Sadeh [Thu, 4 Dec 2008 17:46:24 +0000 (09:46 -0800)]
kclient: fix NULL dereferencing oops

16 years agoosd: version pg_stats_t with <epoch,version> pair; clean up pgmon a bit
Sage Weil [Thu, 4 Dec 2008 05:12:11 +0000 (21:12 -0800)]
osd: version pg_stats_t with <epoch,version> pair; clean up pgmon a bit

16 years agoosd: keep projected info on in-progress object modifications in memory
Sage Weil [Thu, 4 Dec 2008 03:46:08 +0000 (19:46 -0800)]
osd: keep projected info on in-progress object modifications in memory

Since the primary delays its writes until after replicas ack, we need to
keep projected object info in memory for the duration, because the
semantics very much depend on whether the object exists and what its size
is (well, mainly the pg_stats do).

This can avoid re-parsing SnapSet et al for certain workloads hitting
the same objects repeatedly (e.g., mds journal objects).

16 years agoosd: fix problems with propagation of info.stats during recovery
Sage Weil [Thu, 4 Dec 2008 00:42:38 +0000 (16:42 -0800)]
osd: fix problems with propagation of info.stats during recovery

merge_log() is called on replicas, do don't use peer_info (which
is primary-only)!

16 years agoosd: keep tabs on total object copies vs missing/degraded
Sage Weil [Thu, 4 Dec 2008 00:41:53 +0000 (16:41 -0800)]
osd: keep tabs on total object copies vs missing/degraded

Define degraded as an object copy that is not present in the
proper location.

16 years agoosd: fix uninitialized var use
Sage Weil [Thu, 4 Dec 2008 00:26:18 +0000 (16:26 -0800)]
osd: fix uninitialized var use

16 years agoadd missing header file declaration
Yehuda Sadeh [Thu, 4 Dec 2008 00:53:38 +0000 (16:53 -0800)]
add missing header file declaration

16 years agoosd: move logging messages to a common infrastructure
Yehuda Sadeh [Thu, 4 Dec 2008 00:36:19 +0000 (16:36 -0800)]
osd: move logging messages to a common infrastructure

16 years agomon: clean up pg dump
Sage Weil [Wed, 3 Dec 2008 23:42:02 +0000 (15:42 -0800)]
mon: clean up pg dump

16 years agoosd: update stats on primary pull
Sage Weil [Wed, 3 Dec 2008 23:41:56 +0000 (15:41 -0800)]
osd: update stats on primary pull

16 years agoosd: fix stupid no-op typo
Sage Weil [Wed, 3 Dec 2008 23:32:44 +0000 (15:32 -0800)]
osd: fix stupid no-op typo

Everything was showing up as a no-op.

16 years agoosd: log scrub errors to central log
Sage Weil [Wed, 3 Dec 2008 23:14:41 +0000 (15:14 -0800)]
osd: log scrub errors to central log

16 years agoMerge branch 'diskformat' into unstable
Sage Weil [Wed, 3 Dec 2008 23:06:25 +0000 (15:06 -0800)]
Merge branch 'diskformat' into unstable

16 years agomon: include last_scrub info in pg dump
Sage Weil [Wed, 3 Dec 2008 22:57:23 +0000 (14:57 -0800)]
mon: include last_scrub info in pg dump

16 years agoosd: remove pg from recovery_wq with clear_primary_state
Sage Weil [Wed, 3 Dec 2008 22:57:09 +0000 (14:57 -0800)]
osd: remove pg from recovery_wq with clear_primary_state

16 years agoosd: make pg refcounting vs work queues constent
Sage Weil [Wed, 3 Dec 2008 22:51:26 +0000 (14:51 -0800)]
osd: make pg refcounting vs work queues constent

Either refcount items in queue, or don't.

16 years agoosd: do clone scrub based on our generated scrub map
Sage Weil [Wed, 3 Dec 2008 22:37:10 +0000 (14:37 -0800)]
osd: do clone scrub based on our generated scrub map

16 years agoosd: scrub info in pg_stat_t. scrub states.
Sage Weil [Wed, 3 Dec 2008 22:17:54 +0000 (14:17 -0800)]
osd: scrub info in pg_stat_t.  scrub states.

16 years agoosd: fix small quirk read_log missing generation
Sage Weil [Wed, 3 Dec 2008 21:55:48 +0000 (13:55 -0800)]
osd: fix small quirk read_log missing generation

The missing entry .have field was probably wrong due to the use
of missing.add_event (which assumes missing is up to date wrt
the previous log entry).  Use the prior version we just pulled off
disk instead.

Also, be a bit more verbose.

16 years agomon: always discard pending on election completion
Sage Weil [Wed, 3 Dec 2008 21:41:12 +0000 (13:41 -0800)]
mon: always discard pending on election completion

Previously we tried to save the pending if we were still the
leader.  The problem is that while we were not leader, we may have
missed out on some updates, in which case the pending may no longer
be based on the current state.

In the future, we could make the commit waiters smart about callback
return codes so that they try to reapply.  For now, don't worry
about it.

16 years agomds: update segment on ETableServer replay
Sage Weil [Wed, 3 Dec 2008 20:26:24 +0000 (12:26 -0800)]
mds: update segment on ETableServer replay

Otherwise we may forget to flush table changes to disk before
trimming.

Also, clean up code a bit to use update_segment() whenever
possible (instead of duplicating the specific LogSegment update).

16 years agomds: print table version loaded during log replay
Sage Weil [Wed, 3 Dec 2008 20:20:35 +0000 (12:20 -0800)]
mds: print table version loaded during log replay

16 years agoosd: default 2x pg only for now
Sage Weil [Wed, 3 Dec 2008 20:20:22 +0000 (12:20 -0800)]
osd: default 2x pg only for now

16 years agoosd: distributed scrub compares primary vs replica contents
Sage Weil [Wed, 3 Dec 2008 20:19:59 +0000 (12:19 -0800)]
osd: distributed scrub compares primary vs replica contents

The checks are still pretty trivial at this point.

16 years agoosd: do not clear ops vector to indicate noop (protocol change)
Sage Weil [Wed, 3 Dec 2008 18:35:10 +0000 (10:35 -0800)]
osd: do not clear ops vector to indicate noop (protocol change)

The reply needs to include the full ops vector.  Use a separate
flag to indicate a noop.

16 years agoosd: rewrite pg_stats queueing
Sage Weil [Wed, 3 Dec 2008 00:02:43 +0000 (16:02 -0800)]
osd: rewrite pg_stats queueing

Use an xlist instead of a separate map.  Avoid inefficient
requeueing and external map overhead.

16 years agoosd: remove useless raid4pg from build
Sage Weil [Tue, 2 Dec 2008 22:38:12 +0000 (14:38 -0800)]
osd: remove useless raid4pg from build

16 years agokclient: fix oops in case written size doesn't match request
Yehuda Sadeh [Wed, 3 Dec 2008 18:10:55 +0000 (10:10 -0800)]
kclient: fix oops in case written size doesn't match request

16 years agokclient: some logs revision
Yehuda Sadeh [Wed, 3 Dec 2008 00:16:41 +0000 (16:16 -0800)]
kclient: some logs revision

16 years agoosd: don't forget about skipped clones during recover_primary
Sage Weil [Tue, 2 Dec 2008 22:27:30 +0000 (14:27 -0800)]
osd: don't forget about skipped clones during recover_primary

Only advance requested_to if we haven't skipped any items.

16 years agoalways print snapids in hex
Sage Weil [Tue, 2 Dec 2008 22:26:06 +0000 (14:26 -0800)]
always print snapids in hex

16 years agoosd: log scrub errors to system log
Sage Weil [Tue, 2 Dec 2008 22:25:54 +0000 (14:25 -0800)]
osd: log scrub errors to system log

16 years agomon: system-wide log
Sage Weil [Tue, 2 Dec 2008 22:28:30 +0000 (14:28 -0800)]
mon: system-wide log

Pretty rudimentary still.

16 years agomds: pick_inode_snap should consider follows==0 valid
Sage Weil [Tue, 2 Dec 2008 20:03:15 +0000 (12:03 -0800)]
mds: pick_inode_snap should consider follows==0 valid

I suspect a larger audit of 'follows' semantics may be necessary..
but this fixes the bug I was seeing with:

echo asdf > a
mkdir .snap/1
echo qwer > b
mkdir .snap/2
echo zxcv > a
mkdir .snap/3
sync
cat .snap/1/a   # empty
ll .snap/*      # hangs

16 years agokclient: initialize some protocol fields
Yehuda Sadeh [Tue, 2 Dec 2008 21:37:02 +0000 (13:37 -0800)]
kclient: initialize some protocol fields

16 years agoosd: log stats for push and pull bytes
Sage Weil [Tue, 2 Dec 2008 03:00:58 +0000 (19:00 -0800)]
osd: log stats for push and pull bytes

16 years agomds: avoid overlapping release attempts
Sage Weil [Tue, 2 Dec 2008 19:45:37 +0000 (11:45 -0800)]
mds: avoid overlapping release attempts

If the first release attempt is waiting for the log to flush, we
should avoid sending any RELEASED ack until all releases have
flushed.  That is, only the last release will ack.  Keep a counter
in Capability to do this.

Otherwise, we may close out a capability from under a release
that is flushing, and our seq # will be meaningless later in
_finish_release_cap() when we're trying to decide what to do.

16 years agokclient: avoid queueing cap_snap when nothing is dirty or writing
Sage Weil [Tue, 2 Dec 2008 19:09:16 +0000 (11:09 -0800)]
kclient: avoid queueing cap_snap when nothing is dirty or writing

Save ourselves the trouble when there is nothing to flush

16 years agoosd: send old_version to replicas
Sage Weil [Tue, 2 Dec 2008 18:49:22 +0000 (10:49 -0800)]
osd: send old_version to replicas

Otherwise CLONE entries in the PG log on replicas have 0'0 for
prior_version, and everything goes to hell.

16 years agoosd: always get old_version; include in debug output
Sage Weil [Tue, 2 Dec 2008 18:43:36 +0000 (10:43 -0800)]
osd: always get old_version; include in debug output

16 years agoosd: show prior_version in log dump, output recovery_primary debug output
Sage Weil [Tue, 2 Dec 2008 18:00:18 +0000 (10:00 -0800)]
osd: show prior_version in log dump, output recovery_primary debug output

16 years agomds: suspend instead of suicide on beacon timeout
Sage Weil [Tue, 2 Dec 2008 05:13:34 +0000 (21:13 -0800)]
mds: suspend instead of suicide on beacon timeout

If we don't hear from the monitor, suspend doing any useful work instead
of just committing suicide.  If the monitor comes back and hasn't killed
us off, then we're fine.  If we've been marked as failed, we will shut
down as before.

16 years agofilestore: show return codes in debug output
Sage Weil [Tue, 2 Dec 2008 00:55:40 +0000 (16:55 -0800)]
filestore: show return codes in debug output

16 years agoobject: print snapid in hex
Sage Weil [Tue, 2 Dec 2008 00:53:16 +0000 (16:53 -0800)]
object: print snapid in hex

16 years agoosd: allow admin to mark osd lost to kickstart recovery (disk format change)
Sage Weil [Wed, 26 Nov 2008 22:48:52 +0000 (14:48 -0800)]
osd: allow admin to mark osd lost to kickstart recovery (disk format change)

This is important when an osd (or osds) may contain modifications
but is offline.  If the data is truly lost, we can kickstart
recovery.

Note that if the osd was storing metadata, this could be
especially dangerous!

16 years agoosd: clean up pg_stat, osd_stat summation, fields a bit
Sage Weil [Tue, 2 Dec 2008 00:31:16 +0000 (16:31 -0800)]
osd: clean up pg_stat, osd_stat summation, fields a bit

16 years agokclient: no page cache for write without wrbuf cap
Yehuda Sadeh [Tue, 2 Dec 2008 00:00:35 +0000 (16:00 -0800)]
kclient: no page cache for write without wrbuf cap

16 years agoMerge branch 'unstable' of ssh://ceph.newdream.net/git/ceph into unstable
Yehuda Sadeh [Mon, 1 Dec 2008 23:51:58 +0000 (15:51 -0800)]
Merge branch 'unstable' of ssh://ceph.newdream.net/git/ceph into unstable

16 years agokclient: sync writes use page cache
Yehuda Sadeh [Mon, 1 Dec 2008 23:50:36 +0000 (15:50 -0800)]
kclient: sync writes use page cache

16 years agokclient: clean out old BACKOFF bit flag, comments
Sage Weil [Mon, 1 Dec 2008 23:29:56 +0000 (15:29 -0800)]
kclient: clean out old BACKOFF bit flag, comments

16 years agoosdmaptool: print new osd_info fields
Sage Weil [Mon, 1 Dec 2008 22:24:35 +0000 (14:24 -0800)]
osdmaptool: print new osd_info fields

down_at, last_clean interval, etc.

16 years agoosd: make pg log dump obey debug levels
Sage Weil [Mon, 1 Dec 2008 23:16:24 +0000 (15:16 -0800)]
osd: make pg log dump obey debug levels

16 years agoosd: rebuild past intervals when needed; tolerate partial info
Sage Weil [Mon, 1 Dec 2008 23:13:25 +0000 (15:13 -0800)]
osd: rebuild past intervals when needed; tolerate partial info

Tolerate missing past_intervals attr.

Also, if only some past intervals are missing, rebuild them all; don't
assume that if any are there then all are there.

16 years agoosd: send and process heartbeats in separate thread, channel
Sage Weil [Mon, 1 Dec 2008 19:03:52 +0000 (11:03 -0800)]
osd: send and process heartbeats in separate thread, channel

Use a separate dispatch thread to process heartbeats.  Use a
separate thread to send them.  This ensures something slow
(e.g. a map update) does not make an osd appear to be down.

This also means a spearate entity_addr for heartbeats, which puts
them over a separate TCP stream.

16 years agomds: do not purge until leases expire
Sage Weil [Mon, 1 Dec 2008 22:01:45 +0000 (14:01 -0800)]
mds: do not purge until leases expire

16 years agolockdep: lockdep_dump_locks()
Sage Weil [Mon, 1 Dec 2008 18:41:29 +0000 (10:41 -0800)]
lockdep: lockdep_dump_locks()

Handy for dumping held locks in gdb w/ 'p lockdep_dump_lock()'

16 years agorwlock: try_get_read, try_get_write
Sage Weil [Mon, 1 Dec 2008 18:13:00 +0000 (10:13 -0800)]
rwlock: try_get_read, try_get_write

16 years agoosd: optionally avoid zeroing trimmed log on disk
Sage Weil [Mon, 1 Dec 2008 15:13:30 +0000 (07:13 -0800)]
osd: optionally avoid zeroing trimmed log on disk

This is a half-hearted attempt to keep old PG log content around.  It'll
still be lost if a PG moves to another node or the entire log is written
to disk for some other reason.

16 years agokclient: fix bad check
Yehuda Sadeh [Mon, 1 Dec 2008 20:36:37 +0000 (12:36 -0800)]
kclient: fix bad check

16 years agokclient: slient down some log message
Yehuda Sadeh [Mon, 1 Dec 2008 19:39:00 +0000 (11:39 -0800)]
kclient: slient down some log message

16 years agoosd: skip peer_info on down osds
Sage Weil [Mon, 1 Dec 2008 04:42:44 +0000 (20:42 -0800)]
osd: skip peer_info on down osds

We don't clean old/down OSDs out of peer_info map, since we may not
restart peering when strays go up/down.  That's fine... just make sure
we ignore them later.

16 years agoosd: remove bad PG::put() assertion
Sage Weil [Mon, 1 Dec 2008 04:41:50 +0000 (20:41 -0800)]
osd: remove bad PG::put() assertion