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.
16 * Placement Group Map. Placement Groups are logical sets of objects
17 * that are replicated by the same set of devices. pgid=(r,hash(o)&m)
18 * where & is a bit-wise AND and m=2^k-1
24 #include "include/health.h"
25 #include "common/debug.h"
26 #include "common/TextTable.h"
27 #include "osd/osd_types.h"
28 #include "include/mempool.h"
29 #include "mon/health_check.h"
31 #include "mon/PGStatService.h"
33 // FIXME: don't like including this here to get OSDMap::Incremental, maybe
34 // PGMapUpdater needs its own header.
35 #include "osd/OSDMap.h"
37 namespace ceph
{ class Formatter
; }
41 MEMPOOL_CLASS_HELPERS();
42 virtual ~PGMapDigest() {}
44 mempool::pgmap::vector
<uint64_t> osd_last_seq
;
46 mutable std::map
<int, int64_t> avail_space_by_rule
;
48 // aggregate state, populated by PGMap child
49 int64_t num_pg
= 0, num_osd
= 0;
50 int64_t num_pg_active
= 0;
51 int64_t num_pg_unknown
= 0;
52 mempool::pgmap::unordered_map
<int32_t,pool_stat_t
> pg_pool_sum
;
53 mempool::pgmap::map
<int64_t,int64_t> num_pg_by_pool
;
56 mempool::pgmap::unordered_map
<int32_t,int32_t> num_pg_by_state
;
61 void encode(bufferlist
& bl
) const {
64 ::encode(primary
, bl
);
66 void decode(bufferlist::iterator
& p
) {
72 mempool::pgmap::unordered_map
<int32_t,pg_count
> num_pg_by_osd
;
74 // recent deltas, and summation
76 * keep track of last deltas for each pool, calculated using
77 * @p pg_pool_sum as baseline.
79 mempool::pgmap::unordered_map
<uint64_t, mempool::pgmap::list
< pair
<pool_stat_t
, utime_t
> > > per_pool_sum_deltas
;
81 * keep track of per-pool timestamp deltas, according to last update on
84 mempool::pgmap::unordered_map
<uint64_t, utime_t
> per_pool_sum_deltas_stamps
;
86 * keep track of sum deltas, per-pool, taking into account any previous
87 * deltas existing in @p per_pool_sum_deltas. The utime_t as second member
88 * of the pair is the timestamp refering to the last update (i.e., the first
89 * member of the pair) for a given pool.
91 mempool::pgmap::unordered_map
<uint64_t, pair
<pool_stat_t
,utime_t
> > per_pool_sum_delta
;
93 pool_stat_t pg_sum_delta
;
97 void print_summary(Formatter
*f
, ostream
*out
) const;
98 void print_oneline_summary(Formatter
*f
, ostream
*out
) const;
100 void recovery_summary(Formatter
*f
, list
<string
> *psl
,
101 const pool_stat_t
& pool_sum
) const;
102 void overall_recovery_summary(Formatter
*f
, list
<string
> *psl
) const;
103 void pool_recovery_summary(Formatter
*f
, list
<string
> *psl
,
104 uint64_t poolid
) const;
105 void recovery_rate_summary(Formatter
*f
, ostream
*out
,
106 const pool_stat_t
& delta_sum
,
107 utime_t delta_stamp
) const;
108 void overall_recovery_rate_summary(Formatter
*f
, ostream
*out
) const;
109 void pool_recovery_rate_summary(Formatter
*f
, ostream
*out
,
110 uint64_t poolid
) const;
112 * Obtain a formatted/plain output for client I/O, source from stats for a
113 * given @p delta_sum pool over a given @p delta_stamp period of time.
115 void client_io_rate_summary(Formatter
*f
, ostream
*out
,
116 const pool_stat_t
& delta_sum
,
117 utime_t delta_stamp
) const;
119 * Obtain a formatted/plain output for the overall client I/O, which is
120 * calculated resorting to @p pg_sum_delta and @p stamp_delta.
122 void overall_client_io_rate_summary(Formatter
*f
, ostream
*out
) const;
124 * Obtain a formatted/plain output for client I/O over a given pool
125 * with id @p pool_id. We will then obtain pool-specific data
126 * from @p per_pool_sum_delta.
128 void pool_client_io_rate_summary(Formatter
*f
, ostream
*out
,
129 uint64_t poolid
) const;
131 * Obtain a formatted/plain output for cache tier IO, source from stats for a
132 * given @p delta_sum pool over a given @p delta_stamp period of time.
134 void cache_io_rate_summary(Formatter
*f
, ostream
*out
,
135 const pool_stat_t
& delta_sum
,
136 utime_t delta_stamp
) const;
138 * Obtain a formatted/plain output for the overall cache tier IO, which is
139 * calculated resorting to @p pg_sum_delta and @p stamp_delta.
141 void overall_cache_io_rate_summary(Formatter
*f
, ostream
*out
) const;
143 * Obtain a formatted/plain output for cache tier IO over a given pool
144 * with id @p pool_id. We will then obtain pool-specific data
145 * from @p per_pool_sum_delta.
147 void pool_cache_io_rate_summary(Formatter
*f
, ostream
*out
,
148 uint64_t poolid
) const;
151 * Return the number of additional bytes that can be stored in this
152 * pool before the first OSD fills up, accounting for PG overhead.
154 int64_t get_pool_free_space(const OSDMap
&osd_map
, int64_t poolid
) const;
156 virtual void dump_pool_stats_full(const OSDMap
&osd_map
, stringstream
*ss
,
157 Formatter
*f
, bool verbose
) const;
158 void dump_fs_stats(stringstream
*ss
, Formatter
*f
, bool verbose
) const;
159 static void dump_object_stat_sum(TextTable
&tbl
, Formatter
*f
,
160 const object_stat_sum_t
&sum
,
163 bool verbose
, const pg_pool_t
*pool
);
165 size_t get_num_pg_by_osd(int osd
) const {
166 auto p
= num_pg_by_osd
.find(osd
);
167 if (p
== num_pg_by_osd
.end())
170 return p
->second
.acting
;
172 int get_num_primary_pg_by_osd(int osd
) const {
173 auto p
= num_pg_by_osd
.find(osd
);
174 if (p
== num_pg_by_osd
.end())
177 return p
->second
.primary
;
180 ceph_statfs
get_statfs(OSDMap
&osdmap
,
181 boost::optional
<int64_t> data_pool
) const;
183 int64_t get_rule_avail(int ruleno
) const {
184 auto i
= avail_space_by_rule
.find(ruleno
);
185 if (i
!= avail_space_by_rule
.end())
186 return avail_space_by_rule
[ruleno
];
191 // kill me post-mimic or -nautilus
192 bool definitely_converted_snapsets() const {
193 // false negative is okay; false positive is not!
196 num_pg_unknown
== 0 &&
197 pg_sum
.stats
.sum
.num_legacy_snapsets
== 0;
200 // kill me post-luminous:
201 virtual float get_fallback_full_ratio() const {
205 uint64_t get_last_osd_stat_seq(int osd
) {
206 if (osd
< (int)osd_last_seq
.size())
207 return osd_last_seq
[osd
];
211 void encode(bufferlist
& bl
, uint64_t features
) const;
212 void decode(bufferlist::iterator
& p
);
213 void dump(Formatter
*f
) const;
214 static void generate_test_instances(list
<PGMapDigest
*>& ls
);
216 WRITE_CLASS_ENCODER(PGMapDigest::pg_count
);
217 WRITE_CLASS_ENCODER_FEATURES(PGMapDigest
);
219 class PGMap
: public PGMapDigest
{
221 MEMPOOL_CLASS_HELPERS();
225 epoch_t last_osdmap_epoch
; // last osdmap epoch i applied to the pgmap
226 epoch_t last_pg_scan
; // osdmap epoch
227 mempool::pgmap::unordered_map
<int32_t,osd_stat_t
> osd_stat
;
228 mempool::pgmap::unordered_map
<pg_t
,pg_stat_t
> pg_stat
;
229 mempool::pgmap::set
<int32_t> full_osds
; // for pre-luminous only
230 mempool::pgmap::set
<int32_t> nearfull_osds
; // for pre-luminous only
232 float nearfull_ratio
;
234 // mapping of osd to most recently reported osdmap epoch
235 mempool::pgmap::unordered_map
<int32_t,epoch_t
> osd_epochs
;
239 MEMPOOL_CLASS_HELPERS();
241 mempool::pgmap::map
<pg_t
,pg_stat_t
> pg_stat_updates
;
242 epoch_t osdmap_epoch
;
243 epoch_t pg_scan
; // osdmap epoch
244 mempool::pgmap::set
<pg_t
> pg_remove
;
246 float nearfull_ratio
;
250 mempool::pgmap::map
<int32_t,osd_stat_t
> osd_stat_updates
;
251 mempool::pgmap::set
<int32_t> osd_stat_rm
;
253 // mapping of osd to most recently reported osdmap epoch.
254 // 1:1 with osd_stat_updates.
255 mempool::pgmap::map
<int32_t,epoch_t
> osd_epochs
;
258 const mempool::pgmap::map
<int32_t, osd_stat_t
> &get_osd_stat_updates() const {
259 return osd_stat_updates
;
261 const mempool::pgmap::set
<int32_t> &get_osd_stat_rm() const {
264 const mempool::pgmap::map
<int32_t, epoch_t
> &get_osd_epochs() const {
268 template<typename OsdStat
>
269 void update_stat(int32_t osd
, epoch_t epoch
, OsdStat
&& stat
) {
270 osd_stat_updates
[osd
] = std::forward
<OsdStat
>(stat
);
271 osd_epochs
[osd
] = epoch
;
272 assert(osd_epochs
.size() == osd_stat_updates
.size());
274 void stat_osd_out(int32_t osd
, epoch_t epoch
) {
275 // 0 the stats for the osd
276 osd_stat_updates
[osd
] = osd_stat_t();
277 // only fill in the epoch if the osd didn't already report htis
278 // epoch. that way we zero the stat but still preserve a reported
280 if (!osd_epochs
.count(osd
))
281 osd_epochs
[osd
] = epoch
;
282 // ...and maintain our invariant.
283 assert(osd_epochs
.size() == osd_stat_updates
.size());
285 void stat_osd_down_up(int32_t osd
, epoch_t epoch
, const PGMap
& pg_map
) {
286 // 0 the op_queue_age_hist for this osd
287 auto p
= osd_stat_updates
.find(osd
);
288 if (p
!= osd_stat_updates
.end()) {
289 p
->second
.op_queue_age_hist
.clear();
292 auto q
= pg_map
.osd_stat
.find(osd
);
293 if (q
!= pg_map
.osd_stat
.end()) {
294 osd_stat_t
& t
= osd_stat_updates
[osd
] = q
->second
;
295 t
.op_queue_age_hist
.clear();
296 osd_epochs
[osd
] = epoch
;
299 void rm_stat(int32_t osd
) {
300 osd_stat_rm
.insert(osd
);
301 osd_epochs
.erase(osd
);
302 osd_stat_updates
.erase(osd
);
304 void encode(bufferlist
&bl
, uint64_t features
=-1) const;
305 void decode(bufferlist::iterator
&bl
);
306 void dump(Formatter
*f
) const;
307 static void generate_test_instances(list
<Incremental
*>& o
);
309 Incremental() : version(0), osdmap_epoch(0), pg_scan(0),
310 full_ratio(0), nearfull_ratio(0) {}
314 // aggregate stats (soft state), generated by calc_stats()
315 mutable epoch_t min_last_epoch_clean
= 0;
316 mempool::pgmap::unordered_map
<int,set
<pg_t
> > pg_by_osd
;
317 mempool::pgmap::unordered_map
<int,int> blocked_by_sum
;
318 mempool::pgmap::list
< pair
<pool_stat_t
, utime_t
> > pg_sum_deltas
;
322 void update_global_delta(
324 const utime_t ts
, const pool_stat_t
& pg_sum_old
);
325 void update_pool_deltas(
328 const mempool::pgmap::unordered_map
<uint64_t, pool_stat_t
>& pg_pool_sum_old
);
331 void deleted_pool(int64_t pool
) {
332 pg_pool_sum
.erase(pool
);
333 num_pg_by_pool
.erase(pool
);
334 per_pool_sum_deltas
.erase(pool
);
335 per_pool_sum_deltas_stamps
.erase(pool
);
336 per_pool_sum_delta
.erase(pool
);
343 const pool_stat_t
& old_pool_sum
,
345 const pool_stat_t
& current_pool_sum
,
346 pool_stat_t
*result_pool_delta
,
347 utime_t
*result_ts_delta
,
348 mempool::pgmap::list
<pair
<pool_stat_t
,utime_t
> > *delta_avg_list
);
350 void update_one_pool_delta(CephContext
*cct
,
353 const pool_stat_t
& old_pool_sum
);
355 epoch_t
calc_min_last_epoch_clean() const;
359 mempool::pgmap::set
<pg_t
> creating_pgs
;
360 mempool::pgmap::map
<int,map
<epoch_t
,set
<pg_t
> > > creating_pgs_by_osd_epoch
;
362 // Bits that use to be enum StuckPG
363 static const int STUCK_INACTIVE
= (1<<0);
364 static const int STUCK_UNCLEAN
= (1<<1);
365 static const int STUCK_UNDERSIZED
= (1<<2);
366 static const int STUCK_DEGRADED
= (1<<3);
367 static const int STUCK_STALE
= (1<<4);
371 last_osdmap_epoch(0), last_pg_scan(0),
372 full_ratio(0), nearfull_ratio(0)
375 void set_full_ratios(float full
, float nearfull
) {
376 if (full_ratio
== full
&& nearfull_ratio
== nearfull
)
379 nearfull_ratio
= nearfull
;
383 version_t
get_version() const {
386 void set_version(version_t v
) {
389 epoch_t
get_last_osdmap_epoch() const {
390 return last_osdmap_epoch
;
392 void set_last_osdmap_epoch(epoch_t e
) {
393 last_osdmap_epoch
= e
;
395 epoch_t
get_last_pg_scan() const {
398 void set_last_pg_scan(epoch_t e
) {
401 utime_t
get_stamp() const {
404 void set_stamp(utime_t s
) {
408 pool_stat_t
get_pg_pool_sum_stat(int64_t pool
) const {
409 auto p
= pg_pool_sum
.find(pool
);
410 if (p
!= pg_pool_sum
.end())
412 return pool_stat_t();
416 void update_pg(pg_t pgid
, bufferlist
& bl
);
417 void remove_pg(pg_t pgid
);
418 void update_osd(int osd
, bufferlist
& bl
);
419 void remove_osd(int osd
);
421 void apply_incremental(CephContext
*cct
, const Incremental
& inc
);
422 void redo_full_sets();
423 void register_nearfull_status(int osd
, const osd_stat_t
& s
);
425 void stat_pg_add(const pg_t
&pgid
, const pg_stat_t
&s
,
426 bool sameosds
=false);
427 void stat_pg_sub(const pg_t
&pgid
, const pg_stat_t
&s
,
428 bool sameosds
=false);
429 void stat_pg_update(const pg_t pgid
, pg_stat_t
&prev
, bufferlist::iterator
& blp
);
430 void stat_osd_add(int osd
, const osd_stat_t
&s
);
431 void stat_osd_sub(int osd
, const osd_stat_t
&s
);
433 void encode(bufferlist
&bl
, uint64_t features
=-1) const;
434 void decode(bufferlist::iterator
&bl
);
436 /// encode subset of our data to a PGMapDigest
437 void encode_digest(const OSDMap
& osdmap
,
438 bufferlist
& bl
, uint64_t features
) const;
440 void dirty_all(Incremental
& inc
);
442 int64_t get_rule_avail(const OSDMap
& osdmap
, int ruleno
) const;
443 void get_rules_avail(const OSDMap
& osdmap
,
444 std::map
<int,int64_t> *avail_map
) const;
445 void dump(Formatter
*f
) const;
446 void dump_basic(Formatter
*f
) const;
447 void dump_pg_stats(Formatter
*f
, bool brief
) const;
448 void dump_pool_stats(Formatter
*f
) const;
449 void dump_osd_stats(Formatter
*f
) const;
450 void dump_delta(Formatter
*f
) const;
451 void dump_filtered_pg_stats(Formatter
*f
, set
<pg_t
>& pgs
) const;
452 void dump_pool_stats_full(const OSDMap
&osd_map
, stringstream
*ss
,
453 Formatter
*f
, bool verbose
) const override
{
454 get_rules_avail(osd_map
, &avail_space_by_rule
);
455 PGMapDigest::dump_pool_stats_full(osd_map
, ss
, f
, verbose
);
458 void dump_pg_stats_plain(
460 const mempool::pgmap::unordered_map
<pg_t
, pg_stat_t
>& pg_stats
,
462 void get_stuck_stats(
463 int types
, const utime_t cutoff
,
464 mempool::pgmap::unordered_map
<pg_t
, pg_stat_t
>& stuck_pgs
) const;
465 bool get_stuck_counts(const utime_t cutoff
, map
<string
, int>& note
) const;
466 void dump_stuck(Formatter
*f
, int types
, utime_t cutoff
) const;
467 void dump_stuck_plain(ostream
& ss
, int types
, utime_t cutoff
) const;
468 int dump_stuck_pg_stats(stringstream
&ds
,
471 vector
<string
>& args
) const;
472 void dump(ostream
& ss
) const;
473 void dump_basic(ostream
& ss
) const;
474 void dump_pg_stats(ostream
& ss
, bool brief
) const;
475 void dump_pg_sum_stats(ostream
& ss
, bool header
) const;
476 void dump_pool_stats(ostream
& ss
, bool header
) const;
477 void dump_osd_stats(ostream
& ss
) const;
478 void dump_osd_sum_stats(ostream
& ss
) const;
479 void dump_filtered_pg_stats(ostream
& ss
, set
<pg_t
>& pgs
) const;
481 void dump_osd_perf_stats(Formatter
*f
) const;
482 void print_osd_perf_stats(std::ostream
*ss
) const;
484 void dump_osd_blocked_by_stats(Formatter
*f
) const;
485 void print_osd_blocked_by_stats(std::ostream
*ss
) const;
487 void get_filtered_pg_stats(uint32_t state
, int64_t poolid
, int64_t osdid
,
488 bool primary
, set
<pg_t
>& pgs
) const;
490 epoch_t
get_min_last_epoch_clean() const {
491 if (!min_last_epoch_clean
)
492 min_last_epoch_clean
= calc_min_last_epoch_clean();
493 return min_last_epoch_clean
;
496 float get_fallback_full_ratio() const override
{
497 if (full_ratio
> 0) {
503 void get_health(CephContext
*cct
,
504 const OSDMap
& osdmap
,
505 list
<pair
<health_status_t
,string
> >& summary
,
506 list
<pair
<health_status_t
,string
> > *detail
) const;
508 void get_health_checks(
510 const OSDMap
& osdmap
,
511 health_check_map_t
*checks
) const;
513 static void generate_test_instances(list
<PGMap
*>& o
);
515 WRITE_CLASS_ENCODER_FEATURES(PGMap::Incremental
)
516 WRITE_CLASS_ENCODER_FEATURES(PGMap
)
518 inline ostream
& operator<<(ostream
& out
, const PGMapDigest
& m
) {
519 m
.print_oneline_summary(NULL
, &out
);
523 int process_pg_map_command(
524 const string
& prefix
,
525 const map
<string
,cmd_vartype
>& cmdmap
,
527 const OSDMap
& osdmap
,
535 static void check_osd_map(
536 const OSDMap::Incremental
&osd_inc
,
537 std::set
<int> *need_check_down_pg_osds
,
538 std::map
<int,utime_t
> *last_osd_report
,
540 PGMap::Incremental
*pending_inc
);
542 static void check_osd_map(
544 const OSDMap
&osdmap
,
546 PGMap::Incremental
*pending_inc
);
548 * check latest osdmap for new pgs to register
550 static void register_new_pgs(
551 const OSDMap
&osd_map
,
553 PGMap::Incremental
*pending_inc
);
556 * recalculate creating pg mappings
558 static void update_creating_pgs(
559 const OSDMap
&osd_map
,
561 PGMap::Incremental
*pending_inc
);
563 static void register_pg(
564 const OSDMap
&osd_map
,
565 pg_t pgid
, epoch_t epoch
,
568 PGMap::Incremental
*pending_inc
);
570 // mark pg's state stale if its acting primary osd is down
571 static void check_down_pgs(
572 const OSDMap
&osd_map
,
575 const set
<int>& need_check_down_pg_osds
,
576 PGMap::Incremental
*pending_inc
);
580 /* Assign a lower weight to overloaded OSDs.
582 * The osds that will get a lower weight are those with with a utilization
583 * percentage 'oload' percent greater than the average utilization.
585 int by_utilization(const OSDMap
&osd_map
,
590 bool by_pg
, const set
<int64_t> *pools
,
592 mempool::osdmap::map
<int32_t, uint32_t>* new_weights
,
593 std::stringstream
*ss
,
594 std::string
*out_str
,
599 class PGMapStatService
: virtual public PGStatService
{
603 PGMapStatService(const PGMap
& o
)
606 bool is_readable() const override
{ return true; }
608 const pool_stat_t
* get_pool_stat(int64_t poolid
) const override
{
609 auto i
= pgmap
.pg_pool_sum
.find(poolid
);
610 if (i
!= pgmap
.pg_pool_sum
.end()) {
616 const osd_stat_t
& get_osd_sum() const override
{ return pgmap
.osd_sum
; }
618 const osd_stat_t
*get_osd_stat(int osd
) const override
{
619 auto i
= pgmap
.osd_stat
.find(osd
);
620 if (i
== pgmap
.osd_stat
.end()) {
625 const mempool::pgmap::unordered_map
<int32_t,osd_stat_t
>& get_osd_stat() const override
{
626 return pgmap
.osd_stat
;
628 float get_full_ratio() const override
{ return pgmap
.full_ratio
; }
629 float get_nearfull_ratio() const override
{ return pgmap
.nearfull_ratio
; }
631 bool have_creating_pgs() const override
{
632 return !pgmap
.creating_pgs
.empty();
634 bool is_creating_pg(pg_t pgid
) const override
{
635 return pgmap
.creating_pgs
.count(pgid
);
638 epoch_t
get_min_last_epoch_clean() const override
{
639 return pgmap
.get_min_last_epoch_clean();
642 bool have_full_osds() const override
{ return !pgmap
.full_osds
.empty(); }
643 bool have_nearfull_osds() const override
{
644 return !pgmap
.nearfull_osds
.empty();
647 size_t get_num_pg_by_osd(int osd
) const override
{
648 return pgmap
.get_num_pg_by_osd(osd
);
650 ceph_statfs
get_statfs(OSDMap
& osd_map
,
651 boost::optional
<int64_t> data_pool
) const override
{
653 statfs
.kb
= pgmap
.osd_sum
.kb
;
654 statfs
.kb_used
= pgmap
.osd_sum
.kb_used
;
655 statfs
.kb_avail
= pgmap
.osd_sum
.kb_avail
;
656 statfs
.num_objects
= pgmap
.pg_sum
.stats
.sum
.num_objects
;
659 void print_summary(Formatter
*f
, ostream
*out
) const override
{
660 pgmap
.print_summary(f
, out
);
662 virtual void dump_info(Formatter
*f
) const override
{
663 f
->dump_object("pgmap", pgmap
);
665 void dump_fs_stats(stringstream
*ss
,
667 bool verbose
) const override
{
668 pgmap
.dump_fs_stats(ss
, f
, verbose
);
670 void dump_pool_stats(const OSDMap
& osdm
, stringstream
*ss
, Formatter
*f
,
671 bool verbose
) const override
{
672 pgmap
.dump_pool_stats_full(osdm
, ss
, f
, verbose
);
675 int process_pg_command(const string
& prefix
,
676 const map
<string
,cmd_vartype
>& cmdmap
,
677 const OSDMap
& osdmap
,
680 bufferlist
*odata
) const override
{
681 return process_pg_map_command(prefix
, cmdmap
, pgmap
, osdmap
, f
, ss
, odata
);