1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
2 // vim: ts=8 sw=2 smarttab
4 * Ceph - scalable distributed file system
6 * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net>
8 * This is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License version 2.1, as published by the Free Software
11 * Foundation. See file COPYING.
15 #ifndef CEPH_PAXOSSERVICE_H
16 #define CEPH_PAXOSSERVICE_H
18 #include "include/Context.h"
21 #include "MonitorDBStore.h"
27 * A Paxos Service is an abstraction that easily allows one to obtain an
28 * association between a Monitor and a Paxos class, in order to implement any
33 * @defgroup PaxosService_h_class Paxos Service
38 * The Monitor to which this class is associated with
42 * The Paxos instance to which this class is associated with
46 * Our name. This will be associated with the class implementing us, and will
47 * be used mainly for store-related operations.
51 * If we are or have queued anything for proposal, this variable will be true
52 * until our proposal has been finished.
56 bool need_immediate_propose
= false;
60 * Services implementing us used to depend on the Paxos version, back when
61 * each service would have a Paxos instance for itself. However, now we only
62 * have a single Paxos instance, shared by all the services. Each service now
63 * must keep its own version, if so they wish. This variable should be used
66 version_t service_version
;
70 * Event callback responsible for proposing our pending value once a timer
73 Context
*proposal_timer
;
75 * If the implementation class has anything pending to be proposed to Paxos,
76 * then have_pending should be true; otherwise, false.
81 * health checks for this service
83 * Child must populate this during encode_pending() by calling encode_health().
85 health_check_map_t health_checks
;
87 const health_check_map_t
& get_health_checks() {
93 * format of our state in leveldb, 0 for default
95 version_t format_version
;
98 * @defgroup PaxosService_h_callbacks Callback classes
102 * Retry dispatching a given service message
104 * This callback class is used when we had to wait for some condition to
105 * become true while we were dispatching it.
107 * For instance, if the message's version isn't readable, according to Paxos,
108 * then we must wait for it to become readable. So, we just queue an
109 * instance of this class onto the Paxos::wait_for_readable function, and
110 * we will retry the whole dispatch again once the callback is fired.
112 class C_RetryMessage
: public C_MonOp
{
115 C_RetryMessage(PaxosService
*s
, MonOpRequestRef op_
) :
116 C_MonOp(op_
), svc(s
) { }
117 void _finish(int r
) override
{
118 if (r
== -EAGAIN
|| r
>= 0)
120 else if (r
== -ECANCELED
)
123 assert(0 == "bad C_RetryMessage return value");
133 * @param mn A Monitor instance
134 * @param p A Paxos instance
135 * @param name Our service's name.
137 PaxosService(Monitor
*mn
, Paxos
*p
, string name
)
138 : mon(mn
), paxos(p
), service_name(name
),
140 service_version(0), proposal_timer(0), have_pending(false),
142 last_committed_name("last_committed"),
143 first_committed_name("first_committed"),
144 full_prefix_name("full"), full_latest_name("latest"),
145 cached_first_committed(0), cached_last_committed(0)
149 virtual ~PaxosService() {}
152 * Get the service's name.
154 * @returns The service's name.
156 string
get_service_name() { return service_name
; }
159 * Get the store prefixes we utilize
161 virtual void get_store_prefixes(set
<string
>& s
) {
162 s
.insert(service_name
);
165 // i implement and you ignore
167 * Informs this instance that it should consider itself restarted.
169 * This means that we will cancel our proposal_timer event, if any exists.
173 * Informs this instance that an election has finished.
175 * This means that we will invoke a PaxosService::discard_pending while
176 * setting have_pending to false (basically, ignore our pending state) and
177 * we will then make sure we obtain a new state.
179 * Our state shall be updated by PaxosService::_active if the Paxos is
180 * active; otherwise, we will wait for it to become active by adding a
181 * PaxosService::C_Active callback to it.
183 void election_finished();
185 * Informs this instance that it is supposed to shutdown.
187 * Basically, it will instruct Paxos to cancel all events/callbacks and then
188 * will cancel the proposal_timer event if any exists.
194 * Update our state by updating it from Paxos, and then creating a new
195 * pending state if need be.
197 * @remarks We only create a pending state we our Monitor is the Leader.
199 * @pre Paxos is active
200 * @post have_pending is true if our Monitor is the Leader and Paxos is
207 * Propose a new value through Paxos.
209 * This function should be called by the classes implementing
210 * PaxosService, in order to propose a new value through Paxos.
212 * @pre The implementation class implements the encode_pending function.
213 * @pre have_pending is true
214 * @pre Our monitor is the Leader
215 * @pre Paxos is active
216 * @post Cancel the proposal timer, if any
217 * @post have_pending is false
218 * @post propose pending value through Paxos
220 * @note This function depends on the implementation of encode_pending on
221 * the class that is implementing PaxosService
223 void propose_pending();
226 * Let others request us to propose.
228 * At the moment, this is just a wrapper to propose_pending() with an
229 * extra check for is_writeable(), but it's a good practice to dissociate
230 * requests for proposals from direct usage of propose_pending() for
231 * future use -- we might want to perform additional checks or put a
232 * request on hold, for instance.
234 void request_proposal() {
235 assert(is_writeable());
240 * Request service @p other to perform a proposal.
242 * We could simply use the function above, requesting @p other directly,
243 * but we might eventually want to do something to the request -- say,
244 * set a flag stating we're waiting on a cross-proposal to be finished.
246 void request_proposal(PaxosService
*other
) {
247 assert(other
!= NULL
);
248 assert(other
->is_writeable());
250 other
->request_proposal();
254 * Dispatch a message by passing it to several different functions that are
255 * either implemented directly by this service, or that should be implemented
256 * by the class implementing this service.
259 * @returns 'true' on successful dispatch; 'false' otherwise.
261 bool dispatch(MonOpRequestRef op
);
263 void refresh(bool *need_bootstrap
);
267 * @defgroup PaxosService_h_override_funcs Functions that should be
270 * These functions should be overridden at will by the class implementing
275 * Create the initial state for your system.
277 * In some of ours the state is actually set up elsewhere so this does
280 virtual void create_initial() = 0;
283 * Query the Paxos system for the latest state and apply it if it's newer
284 * than the current Monitor state.
286 virtual void update_from_paxos(bool *need_bootstrap
) = 0;
289 * Hook called after all services have refreshed their state from paxos
291 * This is useful for doing any update work that depends on other
292 * service's having up-to-date state.
294 virtual void post_paxos_update() {}
299 * This is called on mon startup, after all of the PaxosService instances'
300 * update_from_paxos() methods have been called
302 virtual void init() {}
305 * Create the pending state.
307 * @invariant This function is only called on a Leader.
308 * @remarks This created state is then modified by incoming messages.
309 * @remarks Called at startup and after every Paxos ratification round.
311 virtual void create_pending() = 0;
314 * Encode the pending state into a bufferlist for ratification and
315 * transmission as the next state.
317 * @invariant This function is only called on a Leader.
319 * @param t The transaction to hold all changes.
321 virtual void encode_pending(MonitorDBStore::TransactionRef t
) = 0;
324 * Discard the pending state
326 * @invariant This function is only called on a Leader.
328 * @remarks This function is NOT overridden in any of our code, but it is
329 * called in PaxosService::election_finished if have_pending is
332 virtual void discard_pending() { }
335 * Look at the query; if the query can be handled without changing state,
338 * @param m A query message
339 * @returns 'true' if the query was handled (e.g., was a read that got
340 * answered, was a state change that has no effect); 'false'
343 virtual bool preprocess_query(MonOpRequestRef op
) = 0;
346 * Apply the message to the pending state.
348 * @invariant This function is only called on a Leader.
350 * @param m An update message
351 * @returns 'true' if the update message was handled (e.g., a command that
352 * went through); 'false' otherwise.
354 virtual bool prepare_update(MonOpRequestRef op
) = 0;
360 * Determine if the Paxos system should vote on pending, and if so how long
361 * it should wait to vote.
363 * @param[out] delay The wait time, used so we can limit the update traffic
365 * @returns 'true' if the Paxos system should propose; 'false' otherwise.
367 virtual bool should_propose(double &delay
);
370 * force an immediate propose.
372 * This is meant to be called from prepare_update(op).
374 void force_immediate_propose() {
375 need_immediate_propose
= true;
379 * @defgroup PaxosService_h_courtesy Courtesy functions
381 * Courtesy functions, in case the class implementing this service has
382 * anything it wants/needs to do at these times.
386 * This is called when the Paxos state goes to active.
388 * On the peon, this is after each election.
389 * On the leader, this is after each election, *and* after each completed
392 * @note This function may get called twice in certain recovery cases.
394 virtual void on_active() { }
397 * This is called when we are shutting down
399 virtual void on_shutdown() {}
402 * this is called when activating on the leader
404 * it should conditionally upgrade the on-disk format by proposing a transaction
406 virtual void upgrade_format() { }
409 * this is called when we detect the store has just upgraded underneath us
411 virtual void on_upgrade() {}
414 * Called when the Paxos system enters a Leader election.
416 * @remarks It's a courtesy method, in case the class implementing this
417 * service has anything it wants/needs to do at that time.
419 virtual void on_restart() { }
427 virtual void tick() {}
430 * Get health information
432 * @param summary list of summary strings and associated severity
433 * @param detail optional list of detailed problem reports; may be NULL
435 virtual void get_health(list
<pair
<health_status_t
,string
> >& summary
,
436 list
<pair
<health_status_t
,string
> > *detail
,
437 CephContext
*cct
) const { }
439 void encode_health(const health_check_map_t
& next
,
440 MonitorDBStore::TransactionRef t
) {
443 t
->put("health", service_name
, bl
);
444 mon
->log_health(next
, health_checks
, t
);
450 * @defgroup PaxosService_h_store_keys Set of keys that are usually used on
451 * all the services implementing this
452 * class, and, being almost the only keys
453 * used, should be standardized to avoid
457 const string last_committed_name
;
458 const string first_committed_name
;
459 const string full_prefix_name
;
460 const string full_latest_name
;
466 * @defgroup PaxosService_h_version_cache Variables holding cached values
467 * for the most used versions (first
468 * and last committed); we only have
469 * to read them when the store is
470 * updated, so in-between updates we
471 * may very well use cached versions
472 * and avoid the overhead.
475 version_t cached_first_committed
;
476 version_t cached_last_committed
;
482 * Callback list to be used whenever we are running a proposal through
483 * Paxos. These callbacks will be awaken whenever the said proposal
486 list
<Context
*> waiting_for_finished_proposal
;
491 * Check if we are proposing a value through Paxos
493 * @returns true if we are proposing; false otherwise.
495 bool is_proposing() {
500 * Check if we are in the Paxos ACTIVE state.
502 * @note This function is a wrapper for Paxos::is_active
504 * @returns true if in state ACTIVE; false otherwise.
509 (paxos
->is_active() || paxos
->is_updating() || paxos
->is_writing());
513 * Check if we are readable.
515 * This mirrors on the paxos check, except that we also verify that
517 * - the client hasn't seen the future relative to this PaxosService
518 * - this service isn't proposing.
519 * - we have committed our initial state (last_committed > 0)
521 * @param ver The version we want to check if is readable
522 * @returns true if it is readable; false otherwise
524 bool is_readable(version_t ver
= 0) {
525 if (ver
> get_last_committed() ||
526 !paxos
->is_readable(0) ||
527 get_last_committed() == 0)
533 * Check if we are writeable.
535 * We consider to be writeable iff:
537 * - we are not proposing a new version;
538 * - we are ready to be written to -- i.e., we have a pending value.
539 * - paxos is (active or updating or writing or refresh)
541 * @returns true if writeable; false otherwise
543 bool is_writeable() {
544 return is_write_ready();
548 * Check if we are ready to be written to. This means we must have a
549 * pending value and be active.
551 * @returns true if we are ready to be written to; false otherwise.
553 bool is_write_ready() {
554 return is_active() && have_pending
;
558 * Wait for a proposal to finish.
560 * Add a callback to be awaken whenever our current proposal finishes being
561 * proposed through Paxos.
563 * @param c The callback to be awaken once the proposal is finished.
565 void wait_for_finished_proposal(MonOpRequestRef op
, Context
*c
) {
567 op
->mark_event_string(service_name
+ ":wait_for_finished_proposal");
568 waiting_for_finished_proposal
.push_back(c
);
570 void wait_for_finished_proposal_ctx(Context
*c
) {
572 wait_for_finished_proposal(o
, c
);
576 * Wait for us to become active
578 * @param c The callback to be awaken once we become active.
580 void wait_for_active(MonOpRequestRef op
, Context
*c
) {
582 op
->mark_event_string(service_name
+ ":wait_for_active");
584 if (!is_proposing()) {
585 paxos
->wait_for_active(op
, c
);
588 wait_for_finished_proposal(op
, c
);
590 void wait_for_active_ctx(Context
*c
) {
592 wait_for_active(o
, c
);
596 * Wait for us to become readable
598 * @param c The callback to be awaken once we become active.
599 * @param ver The version we want to wait on.
601 void wait_for_readable(MonOpRequestRef op
, Context
*c
, version_t ver
= 0) {
602 /* This is somewhat of a hack. We only do check if a version is readable on
603 * PaxosService::dispatch(), but, nonetheless, we must make sure that if that
604 * is why we are not readable, then we must wait on PaxosService and not on
605 * Paxos; otherwise, we may assert on Paxos::wait_for_readable() if it
606 * happens to be readable at that specific point in time.
609 op
->mark_event_string(service_name
+ ":wait_for_readable");
611 if (is_proposing() ||
612 ver
> get_last_committed() ||
613 get_last_committed() == 0)
614 wait_for_finished_proposal(op
, c
);
617 op
->mark_event_string(service_name
+ ":wait_for_readable/paxos");
619 paxos
->wait_for_readable(op
, c
);
623 void wait_for_readable_ctx(Context
*c
, version_t ver
= 0) {
624 MonOpRequestRef o
; // will initialize the shared_ptr to NULL
625 wait_for_readable(o
, c
, ver
);
629 * Wait for us to become writeable
631 * @param c The callback to be awaken once we become writeable.
633 void wait_for_writeable(MonOpRequestRef op
, Context
*c
) {
635 op
->mark_event_string(service_name
+ ":wait_for_writeable");
638 wait_for_finished_proposal(op
, c
);
639 else if (!is_write_ready())
640 wait_for_active(op
, c
);
642 paxos
->wait_for_writeable(op
, c
);
644 void wait_for_writeable_ctx(Context
*c
) {
646 wait_for_writeable(o
, c
);
651 * @defgroup PaxosService_h_Trim Functions for trimming states
655 * trim service states if appropriate
657 * Called at same interval as tick()
662 * Auxiliary function to trim our state from version @p from to version
663 * @p to, not including; i.e., the interval [from, to[
665 * @param t The transaction to which we will add the trim operations.
666 * @param from the lower limit of the interval to be trimmed
667 * @param to the upper limit of the interval to be trimmed (not including)
669 void trim(MonitorDBStore::TransactionRef t
, version_t from
, version_t to
);
672 * encode service-specific extra bits into trim transaction
674 * @param tx transaction
675 * @param first new first_committed value
677 virtual void encode_trim_extra(MonitorDBStore::TransactionRef tx
,
681 * Get the version we should trim to.
683 * Should be overloaded by service if it wants to trim states.
685 * @returns the version we should trim to; if we return zero, it should be
686 * assumed that there's no version to trim to.
688 virtual version_t
get_trim_to() {
696 * @defgroup PaxosService_h_Stash_Full
699 virtual bool should_stash_full();
701 * Encode a full version on @p t
703 * @note We force every service to implement this function, since we strongly
704 * desire the encoding of full versions.
705 * @note Services that do not trim their state, will be bound to only create
706 * one full version. Full version stashing is determined/controled by
707 * trimming: we stash a version each time a trim is bound to erase the
708 * latest full version.
710 * @param t Transaction on which the full version shall be encoded.
712 virtual void encode_full(MonitorDBStore::TransactionRef t
) = 0;
721 * @note This function is a wrapper for Paxos::cancel_events
723 void cancel_events() {
724 paxos
->cancel_events();
728 * @defgroup PaxosService_h_store_funcs Back storage interface functions
732 * @defgroup PaxosService_h_store_modify Wrapper function interface to access
733 * the back store for modification
737 void put_first_committed(MonitorDBStore::TransactionRef t
, version_t ver
) {
738 t
->put(get_service_name(), first_committed_name
, ver
);
741 * Set the last committed version to @p ver
743 * @param t A transaction to which we add this put operation
744 * @param ver The last committed version number being put
746 void put_last_committed(MonitorDBStore::TransactionRef t
, version_t ver
) {
747 t
->put(get_service_name(), last_committed_name
, ver
);
749 /* We only need to do this once, and that is when we are about to make our
750 * first proposal. There are some services that rely on first_committed
751 * being set -- and it should! -- so we need to guarantee that it is,
752 * specially because the services itself do not do it themselves. They do
753 * rely on it, but they expect us to deal with it, and so we shall.
755 if (!get_first_committed())
756 put_first_committed(t
, ver
);
759 * Put the contents of @p bl into version @p ver
761 * @param t A transaction to which we will add this put operation
762 * @param ver The version to which we will add the value
763 * @param bl A bufferlist containing the version's value
765 void put_version(MonitorDBStore::TransactionRef t
, version_t ver
,
767 t
->put(get_service_name(), ver
, bl
);
770 * Put the contents of @p bl into a full version key for this service, that
771 * will be created with @p ver in mind.
773 * @param t The transaction to which we will add this put operation
774 * @param ver A version number
775 * @param bl A bufferlist containing the version's value
777 void put_version_full(MonitorDBStore::TransactionRef t
,
778 version_t ver
, bufferlist
& bl
) {
779 string key
= mon
->store
->combine_strings(full_prefix_name
, ver
);
780 t
->put(get_service_name(), key
, bl
);
783 * Put the version number in @p ver into the key pointing to the latest full
784 * version of this service.
786 * @param t The transaction to which we will add this put operation
787 * @param ver A version number
789 void put_version_latest_full(MonitorDBStore::TransactionRef t
, version_t ver
) {
790 string key
= mon
->store
->combine_strings(full_prefix_name
, full_latest_name
);
791 t
->put(get_service_name(), key
, ver
);
794 * Put the contents of @p bl into the key @p key.
796 * @param t A transaction to which we will add this put operation
797 * @param key The key to which we will add the value
798 * @param bl A bufferlist containing the value
800 void put_value(MonitorDBStore::TransactionRef t
,
801 const string
& key
, bufferlist
& bl
) {
802 t
->put(get_service_name(), key
, bl
);
806 * Put integer value @v into the key @p key.
808 * @param t A transaction to which we will add this put operation
809 * @param key The key to which we will add the value
810 * @param v An integer
812 void put_value(MonitorDBStore::TransactionRef t
,
813 const string
& key
, version_t v
) {
814 t
->put(get_service_name(), key
, v
);
822 * @defgroup PaxosService_h_store_get Wrapper function interface to access
823 * the back store for reading purposes
828 * @defgroup PaxosService_h_version_cache Obtain cached versions for this
833 * Get the first committed version
835 * @returns Our first committed version (that is available)
837 version_t
get_first_committed() const{
838 return cached_first_committed
;
841 * Get the last committed version
843 * @returns Our last committed version
845 version_t
get_last_committed() const{
846 return cached_last_committed
;
854 * Get the contents of a given version @p ver
856 * @param ver The version being obtained
857 * @param bl The bufferlist to be populated
858 * @return 0 on success; <0 otherwise
860 virtual int get_version(version_t ver
, bufferlist
& bl
) {
861 return mon
->store
->get(get_service_name(), ver
, bl
);
864 * Get the contents of a given full version of this service.
866 * @param ver A version number
867 * @param bl The bufferlist to be populated
868 * @returns 0 on success; <0 otherwise
870 virtual int get_version_full(version_t ver
, bufferlist
& bl
) {
871 string key
= mon
->store
->combine_strings(full_prefix_name
, ver
);
872 return mon
->store
->get(get_service_name(), key
, bl
);
875 * Get the latest full version number
877 * @returns A version number
879 version_t
get_version_latest_full() {
880 string key
= mon
->store
->combine_strings(full_prefix_name
, full_latest_name
);
881 return mon
->store
->get(get_service_name(), key
);
885 * Get a value from a given key.
887 * @param[in] key The key
888 * @param[out] bl The bufferlist to be populated with the value
890 int get_value(const string
& key
, bufferlist
& bl
) {
891 return mon
->store
->get(get_service_name(), key
, bl
);
894 * Get an integer value from a given key.
896 * @param[in] key The key
898 version_t
get_value(const string
& key
) {
899 return mon
->store
->get(get_service_name(), key
);