test/clie/ceph-conf: fix loop detection output test

Signed-off-by: Sage Weil <sage@redhat.com>

common/config: use cerr for --show-config-value

Signed-off-by: Sage Weil <sage@redhat.com>

common/config: normalize the default value

Signed-off-by: Sage Weil <sage@redhat.com>

qa/tasks/rebuild_mondb: --no-mon-config for ceph-monstore-tool

Signed-off-by: Sage Weil <sage@redhat.com>

qa/workunits/objectstore/test_fuse.sh: no mon config for ceph-objectstore-tool

There isn't a running cluster for this test.

Signed-off-by: Sage Weil <sage@redhat.com>

qa/workunits/cephtool/test.sh: streamline test_mon_injectargs

Be less sensitive to formatting of output string on set commands.

Signed-off-by: Sage Weil <sage@redhat.com>

osd: implement tell/daemon 'config get'

Signed-off-by: Sage Weil <sage@redhat.com>

rados: only suggest changing mon_allow_pool_delete on EPERM

Signed-off-by: Sage Weil <sage@redhat.com>

common/ceph_context: only common_init_finish() once per cct

Signed-off-by: Sage Weil <sage@redhat.com>

common/options: 'developer' -> 'dev'

I like shorter.

Signed-off-by: Sage Weil <sage@redhat.com>

doc/rados/configuration: big update

- remove a bunch of outdated info
- update to reflect new monitor-based configuration
- move a bunch of common settings to another document/page

Signed-off-by: Sage Weil <sage@redhat.com>

qa: test config CLI interface

Signed-off-by: Sage Weil <sage@redhat.com>

common/options: fix stringification

There is a generic operator<< for boost::variant--we can't overload it
for Option::value_t.  Make a to_str() method instead.

Add a parse_option() output arg for the normalized value for convenience.

Signed-off-by: Sage Weil <sage@redhat.com>

mon/ConfigMonitor: indent 'config dump'

Signed-off-by: Sage Weil <sage@redhat.com>

doc/dev/config-key: document how we structure things in config-key

Signed-off-by: Sage Weil <sage@redhat.com>

common/config,mgr: fix config for new encode/decode syntax

Signed-off-by: Sage Weil <sage@redhat.com>

messages/MMgrOpen: fix decode

Signed-off-by: Sage Weil <sage@redhat.com>

mon/ConfigMonitor: validate and normalize option on 'config set'

Signed-off-by: Sage Weil <sage@redhat.com>

common/options: flag a bunch of pool settings for RUNTIME

Signed-off-by: Sage Weil <sage@redhat.com>

mon/ConfigMonitor: 'config (get|show)' RW -> RO

Signed-off-by: Sage Weil <sage@redhat.com>

osdmaptool: eliminiate global osd_pg[p]_bits option

osdmaptool is the only user.

Signed-off-by: Sage Weil <sage@redhat.com>

common/options: some osd startup-only options

Signed-off-by: Sage Weil <sage@redhat.com>

common/options: a few more [CLUSTER_]CREATE options

Signed-off-by: Sage Weil <sage@redhat.com>

common/options: replace 'mkfs' tag with FLAG_[CLUSTER_]CREATE

Signed-off-by: Sage Weil <sage@redhat.com>

common/options,ceph_test_msgr: unmark misc ms options as RUNTIME

These were marked RUNTIME for the benefit of the test.  Replace with a
no-op observer.

Signed-off-by: Sage Weil <sage@redhat.com>

common/config_obs: NoopConfigObserver

Signed-off-by: Sage Weil <sage@redhat.com>

common/options: replace FLAG_SAFE with FLAG_{RUNTIME,STARTUP,CLUSTER_CREATE,DAEMON_CREATE}

Signed-off-by: Sage Weil <sage@redhat.com>

common/options: is_safe() -> can_update_at_runtime()

Signed-off-by: Sage Weil <sage@redhat.com>

mon/ConfigMonitor: show changeable for 'config get|show'

Signed-off-by: Sage Weil <sage@redhat.com>

common/options: 'can update at runtime' in 'config help'

Signed-off-by: Sage Weil <sage@redhat.com>

mgr/DaemonServer: add IGNORES column to 'config show'

If the daemon is ignoring a mon config, show that.  Note that this
column only ever says 'mon', which might be confusing to users.

Signed-off-by: Sage Weil <sage@redhat.com>

common/config: report failed mon config values to mgr

If we can't set a mon config value, report that back to the mgr.

Signed-off-by: Sage Weil <sage@redhat.com>

common/config: move unsafe runtime change into _set_val

We want this to apply to other callers, set_mon_vals() in particular.

Signed-off-by: Sage Weil <sage@redhat.com>

common/options: drop 'safe' flag on plugin dir options

These were only marked safe to accomodate the unit test CEPH_LIB hackery,
which is now removed.

Signed-off-by: Sage Weil <sage@redhat.com>

mgr: 'config show <who> <key>'

Signed-off-by: Sage Weil <sage@redhat.com>

mon: 'ceph config get <who> <key>'

Simple command to get a config value for a given entity (either the set value or the
default), similar to ceph-conf or --show-config-value.

Signed-off-by: Sage Weil <sage@redhat.com>

test: CINIT_FLAG_NO_MON_CONFIG or NO_DEFAULT_CONFIG_FILE

Usually we use NO_DEFAULT_CONFIG_FILE, but for those tests that need the
plugin directory overrides in order to work properly when run manually
(e.g., "bin/unittest_compression"), read the config and only skip the
mon config.

The NO_DFEAULT_CONFIG_FILE (which also skips the mon) would work for
everything when run via ctest, but as a dev one often runs the unittest
manually and having to pass CEPH_LIB=lib is annoying.  Note that
do_cmake.sh already writes out a trivial ceph.conf with these options
set for us.

Signed-off-by: Sage Weil <sage@redhat.com>

unittest_daemon_config: add needed rm_val()

Otherwise we polluate later tests.

Signed-off-by: Sage Weil <sage@redhat.com>

ceph-disk/tests: simplify CEPH_LIB hack

Signed-off-by: Sage Weil <sage@redhat.com>

qa/standalone: drop CEPH_LIB hacks

Signed-off-by: Sage Weil <sage@redhat.com>

test: drop hackery for CEPH_LIB

Generic config code parses this now as part of global_init.

Signed-off-by: Sage Weil <sage@redhat.com>

common/config: interpret CEPH_LIB in parse_env()

Apply to the three plugin directory options.

Signed-off-by: Sage Weil <sage@redhat.com>

doc/man/8/ceph-conf: make not about ceph-conf's lack of mon config support

Steer people toward the mon explicitly or the --show-config-value option.

Signed-off-by: Sage Weil <sage@redhat.com>

common/config: comment config levels

Signed-off-by: Sage Weil <sage@redhat.com>

drop unnecessary env_to_vec calls

global[_pre]_init does this for us now.

Signed-off-by: Sage Weil <sage@redhat.com>

librados,libcephfs,...: use parse_env()

Signed-off-by: Sage Weil <sage@redhat.com>

common/global_init: parse environment

Signed-off-by: Sage Weil <sage@redhat.com>

common/config: take level to parse_argv

A few callers are explicitly parsing args from the environment.

Signed-off-by: Sage Weil <sage@redhat.com>

common/config: distinguish between cmdline and override (injectargs)

This makes for a more informative 'ceph config show ...'.

Signed-off-by: Sage Weil <sage@redhat.com>

global/global_init: make --show-config and --show-config-val reflect mon configs

Pull down all configs, including both argv and mon, before we do the
--show-config[-val] step.

Signed-off-by: Sage Weil <sage@redhat.com>

mon/ConfigMonitor: "ceph config assimilate-conf"

Inject as much of a ceph.conf file as we can; spit we can't take back out
to stdout.

Signed-off-by: Sage Weil <sage@redhat.com>

common/options: mark many options NO_MON_UPDATE

- things that identify monitors during startup
- things that identify auth key(s)
- things that only happy during mon mkfs
- things that are very specific to daemon forking/chdir/daemonize behavior

Signed-off-by: Sage Weil <sage@redhat.com>

mgr/DaemonServer: make 'config show' indicate duplicate overrides

If we override an option with the same value, show the OVERRIDES item
in parenthesis.  (I tried "=$src" and "(=$src)" but neither looked any
better.)

Signed-off-by: Sage Weil <sage@redhat.com>

mgr/DaemonServer: 'ceph config show-with-defaults'

Signed-off-by: Sage Weil <sage@redhat.com>

mgr: get config defaults on session open

Leave them encoded until they are used, since usually we won't use them.
This is about 50k currently.

Signed-off-by: Sage Weil <sage@redhat.com>

mgr/DaemonServer: fix daemon registration

We want to take in the initial config for non-service daemons, too!
Restructure this so it makes more sense.

Signed-off-by: Sage Weil <sage@redhat.com>

common/config: get_defaults_bl()

Signed-off-by: Sage Weil <sage@redhat.com>

osd,mds: 'config unset' tell command

Signed-off-by: Sage Weil <sage@redhat.com>

common/ceph_context: implement 'config unset <var>'

Clear an override value (if any).  This works for 'ceph daemon ...'
asok commands.

Signed-off-by: Sage Weil <sage@redhat.com>

mon/ConfigMonitor: show fabricated options as level 'unknown'

Signed-off-by: Sage Weil <sage@redhat.com>

mon/ConfigMonitor: include LEVEL column in config dump|get

Signed-off-by: Sage Weil <sage@redhat.com>

mon/ConfigMap: const * to Option, so we don't have to copy it

These are biggish structs (lots of strings), so it's expensive to make
copies of Option when it is an immutable struct.  However, we need to
handle arbitrary strings that we don't recognize, too.

Signed-off-by: Sage Weil <sage@redhat.com>

mgr: have MgrStandby send mgr reports to active mgr (even self!)

This allows the mgr daemons to register state with the active mgr just
like the rest of the cluster, including perfcounters and current config.

Signed-off-by: Sage Weil <sage@redhat.com>

mon/ConfigMonitor: 'config help'

Signed-off-by: Sage Weil <sage@redhat.com>

common/options: fix dump()

By convention dump() does not include the containing object section.

Signed-off-by: Sage Weil <sage@redhat.com>

common/options: operator<< for value_t

Signed-off-by: Sage Weil <sage@redhat.com>

vstart.sh: move debug values over to mon config

Signed-off-by: Sage Weil <sage@redhat.com>

ceph-osd: do not set/override osd_objectstore config option

We use the one on disk, and nobody consumes the option except ceph-volume/
ceph-disk during mkfs.  Do not pollute the config space with a value that
isn't consumed.

Signed-off-by: Sage Weil <sage@redhat.com>

common/config: "conffile" -> "file"

Easier on the eyes

Signed-off-by: Sage Weil <sage@redhat.com>

common/config: exclude host, fsid from config reproted to mgr

Signed-off-by: Sage Weil <sage@redhat.com>

mon/ConfigMonitor: load_config() a bit less often

We could get very clever and avoid reparsing the full config every time
too, but that is a bit more work.  Leave that as a TODO.

Signed-off-by: Sage Weil <sage@redhat.com>

mon/ConfigMonitor: implement 'config {rm,get,dump}'

Signed-off-by: Sage Weil <sage@redhat.com>

mgr: remove old 'config set' command

This is a cli command to set a config option on the mgr.  It doesn't
really belong because it only sets it on the mgr that happens to be
active.  It was probably mixed up with the asok or tell command.

Added in 057b73d641decb9403aba50caae9d139f3a34dd4

Signed-off-by: Sage Weil <sage@redhat.com>

mon/MonClient: one-shot mon connection on started to fetch config

This is not particularly efficient, but it works:

 - connect to the monitor to get the monmap and config
 - tear it all down
 - proceed with normal startup (which presumably involves reconnecting
   to the mon all over again).

This allows us to set config options that may affect the mon communication
itself, like ms_type.

Signed-off-by: Sage Weil <sage@redhat.com>

mon/ConfigMonitor: always send initial config, even if empty

Signed-off-by: Sage Weil <sage@redhat.com>

common/ceph_context: count crypto/shutdown cycles

We need to allow crypto to be set up and torn down before a fork.  Count
the init/shutdown cycles to make that safe.

Signed-off-by: Sage Weil <sage@redhat.com>

common/ceph_context: whitespace

Signed-off-by: Sage Weil <sage@redhat.com>

mon: handle MGetConfig

Signed-off-by: Sage Weil <sage@redhat.com>

vstart.sh: use mon config for many things

We also trim out some obsolete options while we're at it.

Signed-off-by: Sage Weil <sage@redhat.com>

mon: 'config set <who> <name> <value>'

Signed-off-by: Sage Weil <sage@redhat.com>

mon/ConfigMonitor: refresh mon configs too

Signed-off-by: Sage Weil <sage@redhat.com>

mon/ConfigMap: some improvements

Signed-off-by: Sage Weil <sage@redhat.com>

mon/ConfigMap: move option mask parsing into helper

Signed-off-by: Sage Weil <sage@redhat.com>

mgr/DaemonServer: 'config show <who>'

Signed-off-by: Sage Weil <sage@redhat.com>

mgr/MgrClient: report running config to mgr

Signed-off-by: Sage Weil <sage@redhat.com>

mon/ConfigKeyService: add 'prefix' arg to 'config-key dump ...'

Signed-off-by: Sage Weil <sage@redhat.com>

vstart.sh: less config noise

Signed-off-by: Sage Weil <sage@redhat.com>

common/options: remove obsolete 'crushtool' option

Signed-off-by: Sage Weil <sage@redhat.com>

mon/MonClient: subscribe to config, handle updates

Signed-off-by: Sage Weil <sage@redhat.com>

mon/ConfigMonitor: add monitor service for distributing configs

Add 'config' subscribe option to get MConfigs on auth and when config is
updated.

Signed-off-by: Sage Weil <sage@redhat.com>

mon/ConfigMap: parsed cluster-wide config

Breaks down options by global, type, daemon, and can filter options by
crush location.

Signed-off-by: Sage Weil <sage@redhat.com>

messages/MConfig: send config map

Signed-off-by: Sage Weil <sage@redhat.com>

common/config: set_mon_vals()

Signed-off-by: Sage Weil <sage@redhat.com>

common/config: some cleanup; rm_val

Signed-off-by: Sage Weil <sage@redhat.com>

cleanup: src/.libs -> build/lib

Clean up remaining traces of automake's .libs.  This makes unittest_*
tests easier to run by hand (mostly).

Signed-off-by: Sage Weil <sage@redhat.com>

common/config: rewrite (╯°□°)╯︵ ┻━┻

This is too complete a rewrite to reasonably break down into small steps,
and even if I could, it would be harder to review that way than to simply
review the new implementation.  The semantics of the old one were so weird
that it's harder to reason about the change in behavior than to simply
review the new behavior.

That's my story, at least, and I'm sticking to it!

So, here are the highlights:

 - $foo meta expansions are evaluated at get_val() time.  This means the
weird bool arguments to set_val specifying whether things were expanded
are removed (they didn't make any sense unless you were thinking about the
old implementation).
 - for every option, we track values from any inputs (config, mon,
   override).  At get_val() time, we pick the highest priority one.
 - diff() is rewritten to be simple and to show you all of the above.
 - internal interfaces are simplified, and in terms of Option::value_t
   whenever possible.
 - unit tests simplified somewhat based on the above.

Known issues:

 - legacy values get pushed out in select cases.  Notably if foo=$bar
   and bar is updated, we do not update $foo (there is no dependency
   tracking to do this efficiently).

Signed-off-by: Sage Weil <sage@redhat.com>

common/config: whitespace

Signed-off-by: Sage Weil <sage@redhat.com>

common/config: streamline get_val_generic

Lookup value first; avoid a schema lookup entirely.

Signed-off-by: Sage Weil <sage@redhat.com>

common: refactor default config value handling

- pass an optional map<string,string> instead of an arg list
- remember default values in md_config_t
- some nicer helpers

Signed-off-by: Sage Weil <sage@redhat.com>

os/bluestore: pass rocksdb options via map, not global config

Signed-off-by: Sage Weil <sage@redhat.com>