1 // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*-
2 // vim: ts=8 sw=2 smarttab
4 #ifndef CEPH_CRUSH_WRAPPER_H
5 #define CEPH_CRUSH_WRAPPER_H
14 #include "include/types.h"
23 #include "include/assert.h"
24 #include "include/err.h"
25 #include "include/encoding.h"
26 #include "include/mempool.h"
28 #include "common/Mutex.h"
30 #define BUG_ON(x) assert(!(x))
36 namespace CrushTreeDumper
{
37 typedef mempool::osdmap::map
<int64_t,string
> name_map_t
;
40 WRITE_RAW_ENCODER(crush_rule_mask
) // it's all u8's
42 inline static void encode(const crush_rule_step
&s
, bufferlist
&bl
)
48 inline static void decode(crush_rule_step
&s
, bufferlist::iterator
&p
)
58 // magic value used by OSDMap for a "default" fallback choose_args, used if
59 // the choose_arg_map passed to do_rule does not exist. if this also
60 // doesn't exist, fall back to canonical weights.
62 DEFAULT_CHOOSE_ARGS
= -1
65 std::map
<int32_t, string
> type_map
; /* bucket/device type names */
66 std::map
<int32_t, string
> name_map
; /* bucket/device names */
67 std::map
<int32_t, string
> rule_name_map
;
68 std::map
<int32_t, int32_t> class_map
; /* item id -> class id */
69 std::map
<int32_t, string
> class_name
; /* class id -> class name */
70 std::map
<string
, int32_t> class_rname
; /* class name -> class id */
71 std::map
<int32_t, map
<int32_t, int32_t> > class_bucket
; /* bucket[id][class] == id */
72 std::map
<int64_t, crush_choose_arg_map
> choose_args
;
75 struct crush_map
*crush
;
77 bool have_uniform_rules
= false;
80 mutable bool have_rmaps
;
81 mutable std::map
<string
, int> type_rmap
, name_rmap
, rule_name_rmap
;
82 void build_rmaps() const {
83 if (have_rmaps
) return;
84 build_rmap(type_map
, type_rmap
);
85 build_rmap(name_map
, name_rmap
);
86 build_rmap(rule_name_map
, rule_name_rmap
);
89 void build_rmap(const map
<int, string
> &f
, std::map
<string
, int> &r
) const {
91 for (std::map
<int, string
>::const_iterator p
= f
.begin(); p
!= f
.end(); ++p
)
92 r
[p
->second
] = p
->first
;
96 CrushWrapper(const CrushWrapper
& other
);
97 const CrushWrapper
& operator=(const CrushWrapper
& other
);
99 CrushWrapper() : crush(0), have_rmaps(false) {
104 crush_destroy(crush
);
108 crush_map
*get_crush_map() { return crush
; }
113 crush_destroy(crush
);
114 crush
= crush_create();
119 set_tunables_default();
122 /// true if any rule has a ruleset != the rule id
123 bool has_legacy_rulesets() const;
125 /// fix rules whose ruleid != ruleset
126 int renumber_rules_by_ruleset();
128 /// true if any ruleset has more than 1 rule
129 bool has_multirule_rulesets() const;
131 /// true if any buckets that aren't straw2
132 bool has_non_straw2_buckets() const;
135 void set_tunables_argonaut() {
136 crush
->choose_local_tries
= 2;
137 crush
->choose_local_fallback_tries
= 5;
138 crush
->choose_total_tries
= 19;
139 crush
->chooseleaf_descend_once
= 0;
140 crush
->chooseleaf_vary_r
= 0;
141 crush
->chooseleaf_stable
= 0;
142 crush
->allowed_bucket_algs
= CRUSH_LEGACY_ALLOWED_BUCKET_ALGS
;
144 void set_tunables_bobtail() {
145 crush
->choose_local_tries
= 0;
146 crush
->choose_local_fallback_tries
= 0;
147 crush
->choose_total_tries
= 50;
148 crush
->chooseleaf_descend_once
= 1;
149 crush
->chooseleaf_vary_r
= 0;
150 crush
->chooseleaf_stable
= 0;
151 crush
->allowed_bucket_algs
= CRUSH_LEGACY_ALLOWED_BUCKET_ALGS
;
153 void set_tunables_firefly() {
154 crush
->choose_local_tries
= 0;
155 crush
->choose_local_fallback_tries
= 0;
156 crush
->choose_total_tries
= 50;
157 crush
->chooseleaf_descend_once
= 1;
158 crush
->chooseleaf_vary_r
= 1;
159 crush
->chooseleaf_stable
= 0;
160 crush
->allowed_bucket_algs
= CRUSH_LEGACY_ALLOWED_BUCKET_ALGS
;
162 void set_tunables_hammer() {
163 crush
->choose_local_tries
= 0;
164 crush
->choose_local_fallback_tries
= 0;
165 crush
->choose_total_tries
= 50;
166 crush
->chooseleaf_descend_once
= 1;
167 crush
->chooseleaf_vary_r
= 1;
168 crush
->chooseleaf_stable
= 0;
169 crush
->allowed_bucket_algs
=
170 (1 << CRUSH_BUCKET_UNIFORM
) |
171 (1 << CRUSH_BUCKET_LIST
) |
172 (1 << CRUSH_BUCKET_STRAW
) |
173 (1 << CRUSH_BUCKET_STRAW2
);
175 void set_tunables_jewel() {
176 crush
->choose_local_tries
= 0;
177 crush
->choose_local_fallback_tries
= 0;
178 crush
->choose_total_tries
= 50;
179 crush
->chooseleaf_descend_once
= 1;
180 crush
->chooseleaf_vary_r
= 1;
181 crush
->chooseleaf_stable
= 1;
182 crush
->allowed_bucket_algs
=
183 (1 << CRUSH_BUCKET_UNIFORM
) |
184 (1 << CRUSH_BUCKET_LIST
) |
185 (1 << CRUSH_BUCKET_STRAW
) |
186 (1 << CRUSH_BUCKET_STRAW2
);
189 void set_tunables_legacy() {
190 set_tunables_argonaut();
191 crush
->straw_calc_version
= 0;
193 void set_tunables_optimal() {
194 set_tunables_jewel();
195 crush
->straw_calc_version
= 1;
197 void set_tunables_default() {
198 set_tunables_jewel();
199 crush
->straw_calc_version
= 1;
202 int get_choose_local_tries() const {
203 return crush
->choose_local_tries
;
205 void set_choose_local_tries(int n
) {
206 crush
->choose_local_tries
= n
;
209 int get_choose_local_fallback_tries() const {
210 return crush
->choose_local_fallback_tries
;
212 void set_choose_local_fallback_tries(int n
) {
213 crush
->choose_local_fallback_tries
= n
;
216 int get_choose_total_tries() const {
217 return crush
->choose_total_tries
;
219 void set_choose_total_tries(int n
) {
220 crush
->choose_total_tries
= n
;
223 int get_chooseleaf_descend_once() const {
224 return crush
->chooseleaf_descend_once
;
226 void set_chooseleaf_descend_once(int n
) {
227 crush
->chooseleaf_descend_once
= !!n
;
230 int get_chooseleaf_vary_r() const {
231 return crush
->chooseleaf_vary_r
;
233 void set_chooseleaf_vary_r(int n
) {
234 crush
->chooseleaf_vary_r
= n
;
237 int get_chooseleaf_stable() const {
238 return crush
->chooseleaf_stable
;
240 void set_chooseleaf_stable(int n
) {
241 crush
->chooseleaf_stable
= n
;
244 int get_straw_calc_version() const {
245 return crush
->straw_calc_version
;
247 void set_straw_calc_version(int n
) {
248 crush
->straw_calc_version
= n
;
251 unsigned get_allowed_bucket_algs() const {
252 return crush
->allowed_bucket_algs
;
254 void set_allowed_bucket_algs(unsigned n
) {
255 crush
->allowed_bucket_algs
= n
;
258 bool has_argonaut_tunables() const {
260 crush
->choose_local_tries
== 2 &&
261 crush
->choose_local_fallback_tries
== 5 &&
262 crush
->choose_total_tries
== 19 &&
263 crush
->chooseleaf_descend_once
== 0 &&
264 crush
->chooseleaf_vary_r
== 0 &&
265 crush
->chooseleaf_stable
== 0 &&
266 crush
->allowed_bucket_algs
== CRUSH_LEGACY_ALLOWED_BUCKET_ALGS
;
268 bool has_bobtail_tunables() const {
270 crush
->choose_local_tries
== 0 &&
271 crush
->choose_local_fallback_tries
== 0 &&
272 crush
->choose_total_tries
== 50 &&
273 crush
->chooseleaf_descend_once
== 1 &&
274 crush
->chooseleaf_vary_r
== 0 &&
275 crush
->chooseleaf_stable
== 0 &&
276 crush
->allowed_bucket_algs
== CRUSH_LEGACY_ALLOWED_BUCKET_ALGS
;
278 bool has_firefly_tunables() const {
280 crush
->choose_local_tries
== 0 &&
281 crush
->choose_local_fallback_tries
== 0 &&
282 crush
->choose_total_tries
== 50 &&
283 crush
->chooseleaf_descend_once
== 1 &&
284 crush
->chooseleaf_vary_r
== 1 &&
285 crush
->chooseleaf_stable
== 0 &&
286 crush
->allowed_bucket_algs
== CRUSH_LEGACY_ALLOWED_BUCKET_ALGS
;
288 bool has_hammer_tunables() const {
290 crush
->choose_local_tries
== 0 &&
291 crush
->choose_local_fallback_tries
== 0 &&
292 crush
->choose_total_tries
== 50 &&
293 crush
->chooseleaf_descend_once
== 1 &&
294 crush
->chooseleaf_vary_r
== 1 &&
295 crush
->chooseleaf_stable
== 0 &&
296 crush
->allowed_bucket_algs
== ((1 << CRUSH_BUCKET_UNIFORM
) |
297 (1 << CRUSH_BUCKET_LIST
) |
298 (1 << CRUSH_BUCKET_STRAW
) |
299 (1 << CRUSH_BUCKET_STRAW2
));
301 bool has_jewel_tunables() const {
303 crush
->choose_local_tries
== 0 &&
304 crush
->choose_local_fallback_tries
== 0 &&
305 crush
->choose_total_tries
== 50 &&
306 crush
->chooseleaf_descend_once
== 1 &&
307 crush
->chooseleaf_vary_r
== 1 &&
308 crush
->chooseleaf_stable
== 1 &&
309 crush
->allowed_bucket_algs
== ((1 << CRUSH_BUCKET_UNIFORM
) |
310 (1 << CRUSH_BUCKET_LIST
) |
311 (1 << CRUSH_BUCKET_STRAW
) |
312 (1 << CRUSH_BUCKET_STRAW2
));
315 bool has_optimal_tunables() const {
316 return has_jewel_tunables();
318 bool has_legacy_tunables() const {
319 return has_argonaut_tunables();
322 bool has_nondefault_tunables() const {
324 (crush
->choose_local_tries
!= 2 ||
325 crush
->choose_local_fallback_tries
!= 5 ||
326 crush
->choose_total_tries
!= 19);
328 bool has_nondefault_tunables2() const {
330 crush
->chooseleaf_descend_once
!= 0;
332 bool has_nondefault_tunables3() const {
334 crush
->chooseleaf_vary_r
!= 0;
336 bool has_nondefault_tunables5() const {
338 crush
->chooseleaf_stable
!= 0;
341 bool has_v2_rules() const;
342 bool has_v3_rules() const;
343 bool has_v4_buckets() const;
344 bool has_v5_rules() const;
345 bool has_choose_args() const; // any choose_args
346 bool has_incompat_choose_args() const; // choose_args that can't be made compat
348 bool is_v2_rule(unsigned ruleid
) const;
349 bool is_v3_rule(unsigned ruleid
) const;
350 bool is_v5_rule(unsigned ruleid
) const;
352 string
get_min_required_version() const {
353 if (has_v5_rules() || has_nondefault_tunables5())
355 else if (has_v4_buckets())
357 else if (has_nondefault_tunables3())
359 else if (has_nondefault_tunables2() || has_nondefault_tunables())
365 // default bucket types
366 unsigned get_default_bucket_alg() const {
367 // in order of preference
368 if (crush
->allowed_bucket_algs
& (1 << CRUSH_BUCKET_STRAW2
))
369 return CRUSH_BUCKET_STRAW2
;
370 if (crush
->allowed_bucket_algs
& (1 << CRUSH_BUCKET_STRAW
))
371 return CRUSH_BUCKET_STRAW
;
372 if (crush
->allowed_bucket_algs
& (1 << CRUSH_BUCKET_TREE
))
373 return CRUSH_BUCKET_TREE
;
374 if (crush
->allowed_bucket_algs
& (1 << CRUSH_BUCKET_LIST
))
375 return CRUSH_BUCKET_LIST
;
376 if (crush
->allowed_bucket_algs
& (1 << CRUSH_BUCKET_UNIFORM
))
377 return CRUSH_BUCKET_UNIFORM
;
382 int get_num_type_names() const {
383 return type_map
.size();
385 int get_max_type_id() const {
386 if (type_map
.empty())
388 return type_map
.rbegin()->first
;
390 int get_type_id(const string
& name
) const {
392 if (type_rmap
.count(name
))
393 return type_rmap
[name
];
396 const char *get_type_name(int t
) const {
397 std::map
<int,string
>::const_iterator p
= type_map
.find(t
);
398 if (p
!= type_map
.end())
399 return p
->second
.c_str();
402 void set_type_name(int i
, const string
& name
) {
409 bool name_exists(const string
& name
) const {
411 return name_rmap
.count(name
);
413 bool item_exists(int i
) const {
414 return name_map
.count(i
);
416 int get_item_id(const string
& name
) const {
418 if (name_rmap
.count(name
))
419 return name_rmap
[name
];
422 const char *get_item_name(int t
) const {
423 std::map
<int,string
>::const_iterator p
= name_map
.find(t
);
424 if (p
!= name_map
.end())
425 return p
->second
.c_str();
428 int set_item_name(int i
, const string
& name
) {
429 if (!is_valid_crush_name(name
))
436 void swap_names(int a
, int b
) {
437 string an
= name_map
[a
];
438 string bn
= name_map
[b
];
446 int split_id_class(int i
, int *idout
, int *classout
) const;
448 bool class_exists(const string
& name
) const {
449 return class_rname
.count(name
);
451 const char *get_class_name(int i
) const {
452 auto p
= class_name
.find(i
);
453 if (p
!= class_name
.end())
454 return p
->second
.c_str();
457 int get_class_id(const string
& name
) const {
458 auto p
= class_rname
.find(name
);
459 if (p
!= class_rname
.end())
464 int remove_class_name(const string
& name
) {
465 auto p
= class_rname
.find(name
);
466 if (p
== class_rname
.end())
468 int class_id
= p
->second
;
469 auto q
= class_name
.find(class_id
);
470 if (q
== class_name
.end())
472 class_rname
.erase(name
);
473 class_name
.erase(class_id
);
477 int32_t _alloc_class_id() const;
479 int get_or_create_class_id(const string
& name
) {
480 int c
= get_class_id(name
);
482 int i
= _alloc_class_id();
483 class_name
[i
] = name
;
484 class_rname
[name
] = i
;
491 const char *get_item_class(int t
) const {
492 std::map
<int,int>::const_iterator p
= class_map
.find(t
);
493 if (p
== class_map
.end())
495 return get_class_name(p
->second
);
497 int set_item_class(int i
, const string
& name
) {
498 if (!is_valid_crush_name(name
))
500 class_map
[i
] = get_or_create_class_id(name
);
503 int set_item_class(int i
, int c
) {
507 void get_devices_by_class(const string
&name
, set
<int> *devices
) const {
510 if (!class_exists(name
)) {
513 auto cid
= get_class_id(name
);
514 for (auto& p
: class_map
) {
515 if (p
.first
>= 0 && p
.second
== cid
) {
516 devices
->insert(p
.first
);
520 void class_remove_item(int i
) {
521 auto it
= class_map
.find(i
);
522 if (it
== class_map
.end()) {
527 int can_rename_item(const string
& srcname
,
528 const string
& dstname
,
530 int rename_item(const string
& srcname
,
531 const string
& dstname
,
533 int can_rename_bucket(const string
& srcname
,
534 const string
& dstname
,
536 int rename_bucket(const string
& srcname
,
537 const string
& dstname
,
541 bool rule_exists(string name
) const {
543 return rule_name_rmap
.count(name
);
545 int get_rule_id(string name
) const {
547 if (rule_name_rmap
.count(name
))
548 return rule_name_rmap
[name
];
551 const char *get_rule_name(int t
) const {
552 std::map
<int,string
>::const_iterator p
= rule_name_map
.find(t
);
553 if (p
!= rule_name_map
.end())
554 return p
->second
.c_str();
557 void set_rule_name(int i
, const string
& name
) {
558 rule_name_map
[i
] = name
;
560 rule_name_rmap
[name
] = i
;
562 bool is_shadow_item(int id
) const {
563 const char *name
= get_item_name(id
);
564 return name
&& !is_valid_crush_name(name
);
569 * find tree nodes referenced by rules by a 'take' command
571 * Note that these may not be parentless roots.
573 void find_takes(set
<int>& roots
) const;
578 * These are parentless nodes in the map.
580 void find_roots(set
<int>& roots
) const;
584 * find tree roots that contain shadow (device class) items only
586 void find_shadow_roots(set
<int>& roots
) const {
590 if (is_shadow_item(p
)) {
597 * find tree roots that are not shadow (device class) items
599 * These are parentless nodes in the map that are not shadow
600 * items for device classes.
602 void find_nonshadow_roots(set
<int>& roots
) const {
606 if (!is_shadow_item(p
)) {
613 * see if an item is contained within a subtree
615 * @param root haystack
617 * @return true if the item is located beneath the given node
619 bool subtree_contains(int root
, int item
) const;
623 * search for an item in any bucket
626 * @return true if present
628 bool _search_item_exists(int i
) const;
632 * see if item is located where we think it is
634 * This verifies that the given item is located at a particular
635 * location in the hierarchy. However, that check is imprecise; we
636 * are actually verifying that the most specific location key/value
637 * is correct. For example, if loc specifies that rack=foo and
638 * host=bar, it will verify that host=bar is correct; any placement
639 * above that level in the hierarchy is ignored. This matches the
640 * semantics for insert_item().
643 * @param item item id
644 * @param loc location to check (map of type to bucket names)
645 * @param weight optional pointer to weight of item at that location
646 * @return true if item is at specified location
648 bool check_item_loc(CephContext
*cct
, int item
, const map
<string
,string
>& loc
, int *iweight
);
649 bool check_item_loc(CephContext
*cct
, int item
, const map
<string
,string
>& loc
, float *weight
) {
651 bool ret
= check_item_loc(cct
, item
, loc
, &iweight
);
653 *weight
= (float)iweight
/ (float)0x10000;
659 * returns the (type, name) of the parent bucket of id
661 * FIXME: ambiguous for items that occur multiple times in the map
663 pair
<string
,string
> get_immediate_parent(int id
, int *ret
= NULL
);
665 int get_immediate_parent_id(int id
, int *parent
) const;
668 * return ancestor of the given type, or 0 if none
669 * (parent is always a bucket and thus <0)
671 int get_parent_of_type(int id
, int type
) const;
674 * get the fully qualified location of a device by successively finding
675 * parents beginning at ID and ending at highest type number specified in
676 * the CRUSH map which assumes that if device foo is under device bar, the
677 * type_id of foo < bar where type_id is the integer specified in the CRUSH map
679 * returns the location in the form of (type=foo) where type is a type of bucket
680 * specified in the CRUSH map and foo is a name specified in the CRUSH map
682 map
<string
, string
> get_full_location(int id
);
685 * identical to get_full_location(int id) although it returns the type/name
686 * pairs in the order they occur in the hierarchy.
688 * returns -ENOENT if id is not found.
690 int get_full_location_ordered(int id
, vector
<pair
<string
, string
> >& path
);
693 * identical to get_full_location_ordered(int id, vector<pair<string, string> >& path),
694 * although it returns a concatenated string with the type/name pairs in descending
695 * hierarchical order with format key1=val1,key2=val2.
697 * returns the location in descending hierarchy as a string.
699 string
get_full_location_ordered_string(int id
);
702 * returns (type_id, type) of all parent buckets between id and
703 * default, can be used to check for anomolous CRUSH maps
705 map
<int, string
> get_parent_hierarchy(int id
);
708 * enumerate immediate children of given node
710 * @param id parent bucket or device id
711 * @return number of items, or error
713 int get_children(int id
, list
<int> *children
);
716 * enumerate leaves(devices) of given node
718 * @param name parent bucket name
719 * @return 0 on success or a negative errno on error.
721 int get_leaves(const string
&name
, set
<int> *leaves
);
722 int _get_leaves(int id
, list
<int> *leaves
); // worker
725 * insert an item into the map at a specific position
727 * Add an item as a specific location of the hierarchy.
728 * Specifically, we look for the most specific location constraint
729 * for which a bucket already exists, and then create intervening
730 * buckets beneath that in order to place the item.
732 * Note that any location specifiers *above* the most specific match
733 * are ignored. For example, if we specify that osd.12 goes in
734 * host=foo, rack=bar, and row=baz, and rack=bar is the most
735 * specific match, we will create host=foo beneath that point and
736 * put osd.12 inside it. However, we will not verify that rack=bar
737 * is beneath row=baz or move it.
739 * In short, we will build out a hierarchy, and move leaves around,
740 * but not adjust the hierarchy's internal structure. Yet.
742 * If the item is already present in the map, we will return EEXIST.
743 * If the location key/value pairs are nonsensical
744 * (rack=nameofdevice), or location specifies that do not attach us
745 * to any existing part of the hierarchy, we will return EINVAL.
749 * @param weight item weight
750 * @param name item name
751 * @param loc location (map of type to bucket names)
752 * @return 0 for success, negative on error
754 int insert_item(CephContext
*cct
, int id
, float weight
, string name
, const map
<string
,string
>& loc
);
757 * move a bucket in the hierarchy to the given location
759 * This has the same location and ancestor creation behavior as
760 * insert_item(), but will relocate the specified existing bucket.
763 * @param id bucket id
764 * @param loc location (map of type to bucket names)
765 * @return 0 for success, negative on error
767 int move_bucket(CephContext
*cct
, int id
, const map
<string
,string
>& loc
);
770 * swap bucket contents of two buckets without touching bucket ids
773 * @param src bucket a
774 * @param dst bucket b
775 * @return 0 for success, negative on error
777 int swap_bucket(CephContext
*cct
, int src
, int dst
);
780 * add a link to an existing bucket in the hierarchy to the new location
782 * This has the same location and ancestor creation behavior as
783 * insert_item(), but will add a new link to the specified existing
787 * @param id bucket id
788 * @param loc location (map of type to bucket names)
789 * @return 0 for success, negative on error
791 int link_bucket(CephContext
*cct
, int id
, const map
<string
,string
>& loc
);
794 * add or update an item's position in the map
796 * This is analogous to insert_item, except we will move an item if
797 * it is already present.
801 * @param weight item weight
802 * @param name item name
803 * @param loc location (map of type to bucket names)
804 * @return 0 for no change, 1 for successful change, negative on error
806 int update_item(CephContext
*cct
, int id
, float weight
, string name
, const map
<string
,string
>& loc
);
809 * create or move an item, but do not adjust its weight if it already exists
812 * @param item item id
813 * @param weight initial item weight (if we need to create it)
814 * @param name item name
815 * @param loc location (map of type to bucket names)
816 * @return 0 for no change, 1 for successful change, negative on error
818 int create_or_move_item(CephContext
*cct
, int item
, float weight
, string name
,
819 const map
<string
,string
>& loc
);
822 * remove all instances of an item from the map
825 * @param id item id to remove
826 * @param unlink_only unlink but do not remove bucket (useful if multiple links or not empty)
827 * @return 0 on success, negative on error
829 int remove_item(CephContext
*cct
, int id
, bool unlink_only
);
832 * recursively remove buckets starting at item and stop removing
833 * when a bucket is in use.
835 * @param item id to remove
836 * @param unused true if only unused items should be removed
837 * @return 0 on success, negative on error
839 int remove_root(int item
, bool unused
);
842 * remove all instances of an item nested beneath a certain point from the map
845 * @param id item id to remove
846 * @param ancestor ancestor item id under which to search for id
847 * @param unlink_only unlink but do not remove bucket (useful if bucket has multiple links or is not empty)
848 * @return 0 on success, negative on error
851 bool _maybe_remove_last_instance(CephContext
*cct
, int id
, bool unlink_only
);
852 int _remove_item_under(CephContext
*cct
, int id
, int ancestor
, bool unlink_only
);
853 bool _bucket_is_in_use(int id
);
855 int remove_item_under(CephContext
*cct
, int id
, int ancestor
, bool unlink_only
);
858 * calculate the locality/distance from a given id to a crush location map
860 * Specifically, we look for the lowest-valued type for which the
861 * location of id matches that described in loc.
864 * @param id the existing id in the map
865 * @param loc a set of key=value pairs describing a location in the hierarchy
867 int get_common_ancestor_distance(CephContext
*cct
, int id
,
868 const std::multimap
<string
,string
>& loc
);
871 * parse a set of key/value pairs out of a string vector
873 * These are used to describe a location in the CRUSH hierarchy.
875 * @param args list of strings (each key= or key=value)
876 * @param ploc pointer to a resulting location map or multimap
878 static int parse_loc_map(const std::vector
<string
>& args
,
879 std::map
<string
,string
> *ploc
);
880 static int parse_loc_multimap(const std::vector
<string
>& args
,
881 std::multimap
<string
,string
> *ploc
);
884 * get an item's weight
886 * Will return the weight for the first instance it finds.
888 * @param id item id to check
889 * @return weight of item
891 int get_item_weight(int id
) const;
892 float get_item_weightf(int id
) const {
893 return (float)get_item_weight(id
) / (float)0x10000;
895 int get_item_weight_in_loc(int id
, const map
<string
,string
> &loc
);
896 float get_item_weightf_in_loc(int id
, const map
<string
,string
> &loc
) {
897 return (float)get_item_weight_in_loc(id
, loc
) / (float)0x10000;
900 int validate_weightf(float weight
) {
901 uint64_t iweight
= weight
* 0x10000;
902 if (iweight
> std::numeric_limits
<int>::max()) {
907 int adjust_item_weight(CephContext
*cct
, int id
, int weight
);
908 int adjust_item_weightf(CephContext
*cct
, int id
, float weight
) {
909 int r
= validate_weightf(weight
);
913 return adjust_item_weight(cct
, id
, (int)(weight
* (float)0x10000));
915 int adjust_item_weight_in_loc(CephContext
*cct
, int id
, int weight
, const map
<string
,string
>& loc
);
916 int adjust_item_weightf_in_loc(CephContext
*cct
, int id
, float weight
, const map
<string
,string
>& loc
) {
917 int r
= validate_weightf(weight
);
921 return adjust_item_weight_in_loc(cct
, id
, (int)(weight
* (float)0x10000), loc
);
923 void reweight(CephContext
*cct
);
925 int adjust_subtree_weight(CephContext
*cct
, int id
, int weight
);
926 int adjust_subtree_weightf(CephContext
*cct
, int id
, float weight
) {
927 int r
= validate_weightf(weight
);
931 return adjust_subtree_weight(cct
, id
, (int)(weight
* (float)0x10000));
934 /// check if item id is present in the map hierarchy
935 bool check_item_present(int id
) const;
939 int get_max_devices() const {
940 if (!crush
) return 0;
941 return crush
->max_devices
;
947 crush_rule
*get_rule(unsigned ruleno
) const {
948 if (!crush
) return (crush_rule
*)(-ENOENT
);
949 if (ruleno
>= crush
->max_rules
)
951 return crush
->rules
[ruleno
];
953 crush_rule_step
*get_rule_step(unsigned ruleno
, unsigned step
) const {
954 crush_rule
*n
= get_rule(ruleno
);
955 if (IS_ERR(n
)) return (crush_rule_step
*)(-EINVAL
);
956 if (step
>= n
->len
) return (crush_rule_step
*)(-EINVAL
);
957 return &n
->steps
[step
];
962 int get_max_rules() const {
963 if (!crush
) return 0;
964 return crush
->max_rules
;
966 bool rule_exists(unsigned ruleno
) const {
967 if (!crush
) return false;
968 if (ruleno
< crush
->max_rules
&&
969 crush
->rules
[ruleno
] != NULL
)
973 int get_rule_len(unsigned ruleno
) const {
974 crush_rule
*r
= get_rule(ruleno
);
975 if (IS_ERR(r
)) return PTR_ERR(r
);
978 int get_rule_mask_ruleset(unsigned ruleno
) const {
979 crush_rule
*r
= get_rule(ruleno
);
980 if (IS_ERR(r
)) return -1;
981 return r
->mask
.ruleset
;
983 int get_rule_mask_type(unsigned ruleno
) const {
984 crush_rule
*r
= get_rule(ruleno
);
985 if (IS_ERR(r
)) return -1;
988 int get_rule_mask_min_size(unsigned ruleno
) const {
989 crush_rule
*r
= get_rule(ruleno
);
990 if (IS_ERR(r
)) return -1;
991 return r
->mask
.min_size
;
993 int get_rule_mask_max_size(unsigned ruleno
) const {
994 crush_rule
*r
= get_rule(ruleno
);
995 if (IS_ERR(r
)) return -1;
996 return r
->mask
.max_size
;
998 int get_rule_op(unsigned ruleno
, unsigned step
) const {
999 crush_rule_step
*s
= get_rule_step(ruleno
, step
);
1000 if (IS_ERR(s
)) return PTR_ERR(s
);
1003 int get_rule_arg1(unsigned ruleno
, unsigned step
) const {
1004 crush_rule_step
*s
= get_rule_step(ruleno
, step
);
1005 if (IS_ERR(s
)) return PTR_ERR(s
);
1008 int get_rule_arg2(unsigned ruleno
, unsigned step
) const {
1009 crush_rule_step
*s
= get_rule_step(ruleno
, step
);
1010 if (IS_ERR(s
)) return PTR_ERR(s
);
1015 * calculate a map of osds to weights for a given rule
1017 * Generate a map of which OSDs get how much relative weight for a
1020 * @param ruleno [in] rule id
1021 * @param pmap [out] map of osd to weight
1022 * @return 0 for success, or negative error code
1024 int get_rule_weight_osd_map(unsigned ruleno
, map
<int,float> *pmap
);
1028 int add_rule(int ruleno
, int len
, int type
, int minsize
, int maxsize
) {
1029 if (!crush
) return -ENOENT
;
1030 crush_rule
*n
= crush_make_rule(len
, ruleno
, type
, minsize
, maxsize
);
1032 ruleno
= crush_add_rule(crush
, n
, ruleno
);
1035 int set_rule_mask_max_size(unsigned ruleno
, int max_size
) {
1036 crush_rule
*r
= get_rule(ruleno
);
1037 if (IS_ERR(r
)) return -1;
1038 return r
->mask
.max_size
= max_size
;
1040 int set_rule_step(unsigned ruleno
, unsigned step
, int op
, int arg1
, int arg2
) {
1041 if (!crush
) return -ENOENT
;
1042 crush_rule
*n
= get_rule(ruleno
);
1044 crush_rule_set_step(n
, step
, op
, arg1
, arg2
);
1047 int set_rule_step_take(unsigned ruleno
, unsigned step
, int val
) {
1048 return set_rule_step(ruleno
, step
, CRUSH_RULE_TAKE
, val
, 0);
1050 int set_rule_step_set_choose_tries(unsigned ruleno
, unsigned step
, int val
) {
1051 return set_rule_step(ruleno
, step
, CRUSH_RULE_SET_CHOOSE_TRIES
, val
, 0);
1053 int set_rule_step_set_choose_local_tries(unsigned ruleno
, unsigned step
, int val
) {
1054 return set_rule_step(ruleno
, step
, CRUSH_RULE_SET_CHOOSE_LOCAL_TRIES
, val
, 0);
1056 int set_rule_step_set_choose_local_fallback_tries(unsigned ruleno
, unsigned step
, int val
) {
1057 return set_rule_step(ruleno
, step
, CRUSH_RULE_SET_CHOOSE_LOCAL_FALLBACK_TRIES
, val
, 0);
1059 int set_rule_step_set_chooseleaf_tries(unsigned ruleno
, unsigned step
, int val
) {
1060 return set_rule_step(ruleno
, step
, CRUSH_RULE_SET_CHOOSELEAF_TRIES
, val
, 0);
1062 int set_rule_step_set_chooseleaf_vary_r(unsigned ruleno
, unsigned step
, int val
) {
1063 return set_rule_step(ruleno
, step
, CRUSH_RULE_SET_CHOOSELEAF_VARY_R
, val
, 0);
1065 int set_rule_step_set_chooseleaf_stable(unsigned ruleno
, unsigned step
, int val
) {
1066 return set_rule_step(ruleno
, step
, CRUSH_RULE_SET_CHOOSELEAF_STABLE
, val
, 0);
1068 int set_rule_step_choose_firstn(unsigned ruleno
, unsigned step
, int val
, int type
) {
1069 return set_rule_step(ruleno
, step
, CRUSH_RULE_CHOOSE_FIRSTN
, val
, type
);
1071 int set_rule_step_choose_indep(unsigned ruleno
, unsigned step
, int val
, int type
) {
1072 return set_rule_step(ruleno
, step
, CRUSH_RULE_CHOOSE_INDEP
, val
, type
);
1074 int set_rule_step_choose_leaf_firstn(unsigned ruleno
, unsigned step
, int val
, int type
) {
1075 return set_rule_step(ruleno
, step
, CRUSH_RULE_CHOOSELEAF_FIRSTN
, val
, type
);
1077 int set_rule_step_choose_leaf_indep(unsigned ruleno
, unsigned step
, int val
, int type
) {
1078 return set_rule_step(ruleno
, step
, CRUSH_RULE_CHOOSELEAF_INDEP
, val
, type
);
1080 int set_rule_step_emit(unsigned ruleno
, unsigned step
) {
1081 return set_rule_step(ruleno
, step
, CRUSH_RULE_EMIT
, 0, 0);
1084 int add_simple_rule(
1085 string name
, string root_name
, string failure_domain_type
,
1086 string device_class
,
1087 string mode
, int rule_type
, ostream
*err
= 0);
1090 * @param rno rule[set] id to use, -1 to pick the lowest available
1092 int add_simple_rule_at(
1093 string name
, string root_name
,
1094 string failure_domain_type
, string device_class
, string mode
,
1095 int rule_type
, int rno
, ostream
*err
= 0);
1097 int remove_rule(int ruleno
);
1101 const crush_bucket
*get_bucket(int id
) const {
1103 return (crush_bucket
*)(-EINVAL
);
1104 unsigned int pos
= (unsigned int)(-1 - id
);
1105 unsigned int max_buckets
= crush
->max_buckets
;
1106 if (pos
>= max_buckets
)
1107 return (crush_bucket
*)(-ENOENT
);
1108 crush_bucket
*ret
= crush
->buckets
[pos
];
1110 return (crush_bucket
*)(-ENOENT
);
1114 crush_bucket
*get_bucket(int id
) {
1116 return (crush_bucket
*)(-EINVAL
);
1117 unsigned int pos
= (unsigned int)(-1 - id
);
1118 unsigned int max_buckets
= crush
->max_buckets
;
1119 if (pos
>= max_buckets
)
1120 return (crush_bucket
*)(-ENOENT
);
1121 crush_bucket
*ret
= crush
->buckets
[pos
];
1123 return (crush_bucket
*)(-ENOENT
);
1127 * detach a bucket from its parent and adjust the parent weight
1129 * returns the weight of the detached bucket
1131 int detach_bucket(CephContext
*cct
, int item
);
1134 int get_max_buckets() const {
1135 if (!crush
) return -EINVAL
;
1136 return crush
->max_buckets
;
1138 int get_next_bucket_id() const {
1139 if (!crush
) return -EINVAL
;
1140 return crush_get_next_bucket_id(crush
);
1142 bool bucket_exists(int id
) const {
1143 const crush_bucket
*b
= get_bucket(id
);
1148 int get_bucket_weight(int id
) const {
1149 const crush_bucket
*b
= get_bucket(id
);
1150 if (IS_ERR(b
)) return PTR_ERR(b
);
1153 float get_bucket_weightf(int id
) const {
1154 const crush_bucket
*b
= get_bucket(id
);
1155 if (IS_ERR(b
)) return 0;
1156 return b
->weight
/ (float)0x10000;
1158 int get_bucket_type(int id
) const {
1159 const crush_bucket
*b
= get_bucket(id
);
1160 if (IS_ERR(b
)) return PTR_ERR(b
);
1163 int get_bucket_alg(int id
) const {
1164 const crush_bucket
*b
= get_bucket(id
);
1165 if (IS_ERR(b
)) return PTR_ERR(b
);
1168 int get_bucket_hash(int id
) const {
1169 const crush_bucket
*b
= get_bucket(id
);
1170 if (IS_ERR(b
)) return PTR_ERR(b
);
1173 int get_bucket_size(int id
) const {
1174 const crush_bucket
*b
= get_bucket(id
);
1175 if (IS_ERR(b
)) return PTR_ERR(b
);
1178 int get_bucket_item(int id
, int pos
) const {
1179 const crush_bucket
*b
= get_bucket(id
);
1180 if (IS_ERR(b
)) return PTR_ERR(b
);
1181 if ((__u32
)pos
>= b
->size
)
1183 return b
->items
[pos
];
1185 int get_bucket_item_weight(int id
, int pos
) const {
1186 const crush_bucket
*b
= get_bucket(id
);
1187 if (IS_ERR(b
)) return PTR_ERR(b
);
1188 return crush_get_bucket_item_weight(b
, pos
);
1190 float get_bucket_item_weightf(int id
, int pos
) const {
1191 const crush_bucket
*b
= get_bucket(id
);
1192 if (IS_ERR(b
)) return 0;
1193 return (float)crush_get_bucket_item_weight(b
, pos
) / (float)0x10000;
1197 int add_bucket(int bucketno
, int alg
, int hash
, int type
, int size
,
1198 int *items
, int *weights
, int *idout
);
1199 int bucket_add_item(crush_bucket
*bucket
, int item
, int weight
);
1200 int bucket_remove_item(struct crush_bucket
*bucket
, int item
);
1201 int bucket_adjust_item_weight(CephContext
*cct
, struct crush_bucket
*bucket
, int item
, int weight
);
1205 crush_finalize(crush
);
1206 have_uniform_rules
= !has_legacy_rulesets();
1209 int update_device_class(int id
, const string
& class_name
, const string
& name
, ostream
*ss
);
1210 int remove_device_class(CephContext
*cct
, int id
, ostream
*ss
);
1211 int device_class_clone(int original
, int device_class
, int *clone
);
1212 bool class_is_in_use(int class_id
, ostream
*ss
= nullptr);
1213 int populate_classes();
1214 int rebuild_roots_with_classes();
1215 /* remove unused roots generated for class devices */
1216 int trim_roots_with_class(bool unused
);
1217 int cleanup_classes();
1219 void start_choose_profile() {
1220 free(crush
->choose_tries
);
1222 * the original choose_total_tries value was off by one (it
1223 * counted "retries" and not "tries"). add one to alloc.
1225 crush
->choose_tries
= (__u32
*)calloc(sizeof(*crush
->choose_tries
),
1226 (crush
->choose_total_tries
+ 1));
1227 memset(crush
->choose_tries
, 0,
1228 sizeof(*crush
->choose_tries
) * (crush
->choose_total_tries
+ 1));
1230 void stop_choose_profile() {
1231 free(crush
->choose_tries
);
1232 crush
->choose_tries
= 0;
1235 int get_choose_profile(__u32
**vec
) {
1236 if (crush
->choose_tries
) {
1237 *vec
= crush
->choose_tries
;
1238 return crush
->choose_total_tries
;
1244 void set_max_devices(int m
) {
1245 crush
->max_devices
= m
;
1248 int find_rule(int ruleset
, int type
, int size
) const {
1249 if (!crush
) return -1;
1250 if (!have_uniform_rules
) {
1251 return crush_find_rule(crush
, ruleset
, type
, size
);
1253 if (ruleset
< (int)crush
->max_rules
&&
1254 crush
->rules
[ruleset
])
1260 bool ruleset_exists(int const ruleset
) const {
1261 for (size_t i
= 0; i
< crush
->max_rules
; ++i
) {
1262 if (rule_exists(i
) && crush
->rules
[i
]->mask
.ruleset
== ruleset
) {
1271 * Return the lowest numbered ruleset of type `type`
1273 * @returns a ruleset ID, or -1 if no matching rulesets found.
1275 int find_first_ruleset(int type
) const {
1278 for (size_t i
= 0; i
< crush
->max_rules
; ++i
) {
1280 && crush
->rules
[i
]->mask
.type
== type
1281 && (crush
->rules
[i
]->mask
.ruleset
< result
|| result
== -1)) {
1282 result
= crush
->rules
[i
]->mask
.ruleset
;
1289 bool have_choose_args(int64_t choose_args_index
) const {
1290 return choose_args
.count(choose_args_index
);
1293 crush_choose_arg_map
choose_args_get_with_fallback(
1294 int64_t choose_args_index
) const {
1295 auto i
= choose_args
.find(choose_args_index
);
1296 if (i
== choose_args
.end()) {
1297 i
= choose_args
.find(DEFAULT_CHOOSE_ARGS
);
1299 if (i
== choose_args
.end()) {
1300 crush_choose_arg_map arg_map
;
1301 arg_map
.args
= NULL
;
1308 crush_choose_arg_map
choose_args_get(int64_t choose_args_index
) const {
1309 auto i
= choose_args
.find(choose_args_index
);
1310 if (i
== choose_args
.end()) {
1311 crush_choose_arg_map arg_map
;
1312 arg_map
.args
= NULL
;
1320 void destroy_choose_args(crush_choose_arg_map arg_map
) {
1321 for (__u32 i
= 0; i
< arg_map
.size
; i
++) {
1322 crush_choose_arg
*arg
= &arg_map
.args
[i
];
1323 for (__u32 j
= 0; j
< arg
->weight_set_size
; j
++) {
1324 crush_weight_set
*weight_set
= &arg
->weight_set
[j
];
1325 free(weight_set
->weights
);
1327 if (arg
->weight_set
)
1328 free(arg
->weight_set
);
1335 void create_choose_args(int64_t id
, int positions
) {
1336 if (choose_args
.count(id
))
1339 auto &cmap
= choose_args
[id
];
1340 cmap
.args
= (crush_choose_arg
*)calloc(sizeof(crush_choose_arg
),
1341 crush
->max_buckets
);
1342 cmap
.size
= crush
->max_buckets
;
1343 for (int bidx
=0; bidx
< crush
->max_buckets
; ++bidx
) {
1344 crush_bucket
*b
= crush
->buckets
[bidx
];
1345 auto &carg
= cmap
.args
[bidx
];
1348 if (b
&& b
->alg
== CRUSH_BUCKET_STRAW2
) {
1349 crush_bucket_straw2
*sb
= (crush_bucket_straw2
*)b
;
1350 carg
.weight_set_size
= positions
;
1351 carg
.weight_set
= (crush_weight_set
*)calloc(sizeof(crush_weight_set
),
1352 carg
.weight_set_size
);
1353 // initialize with canonical weights
1354 for (int pos
= 0; pos
< positions
; ++pos
) {
1355 carg
.weight_set
[pos
].size
= b
->size
;
1356 carg
.weight_set
[pos
].weights
= (__u32
*)calloc(4, b
->size
);
1357 for (unsigned i
= 0; i
< b
->size
; ++i
) {
1358 carg
.weight_set
[pos
].weights
[i
] = sb
->item_weights
[i
];
1362 carg
.weight_set
= NULL
;
1363 carg
.weight_set_size
= 0;
1368 void rm_choose_args(int64_t id
) {
1369 auto p
= choose_args
.find(id
);
1370 if (p
!= choose_args
.end()) {
1371 destroy_choose_args(p
->second
);
1372 choose_args
.erase(p
);
1376 void choose_args_clear() {
1377 for (auto w
: choose_args
)
1378 destroy_choose_args(w
.second
);
1379 choose_args
.clear();
1382 // adjust choose_args_map weight, preserving the hierarchical summation
1383 // property. used by callers optimizing layouts by tweaking weights.
1384 int _choose_args_adjust_item_weight_in_bucket(
1386 crush_choose_arg_map cmap
,
1389 const vector
<int>& weight
,
1391 int choose_args_adjust_item_weight(
1393 crush_choose_arg_map cmap
,
1394 int id
, const vector
<int>& weight
,
1396 int choose_args_adjust_item_weightf(
1398 crush_choose_arg_map cmap
,
1399 int id
, const vector
<double>& weightf
,
1401 vector
<int> weight(weightf
.size());
1402 for (unsigned i
= 0; i
< weightf
.size(); ++i
) {
1403 weight
[i
] = (int)(weightf
[i
] * (float)0x10000);
1405 return choose_args_adjust_item_weight(cct
, cmap
, id
, weight
, ss
);
1408 int get_choose_args_positions(crush_choose_arg_map cmap
) {
1409 // infer positions from other buckets
1410 for (unsigned j
= 0; j
< cmap
.size
; ++j
) {
1411 if (cmap
.args
[j
].weight_set_size
) {
1412 return cmap
.args
[j
].weight_set_size
;
1418 template<typename WeightVector
>
1419 void do_rule(int rule
, int x
, vector
<int>& out
, int maxout
,
1420 const WeightVector
& weight
,
1421 uint64_t choose_args_index
) const {
1423 char work
[crush_work_size(crush
, maxout
)];
1424 crush_init_workspace(crush
, work
);
1425 crush_choose_arg_map arg_map
= choose_args_get_with_fallback(
1427 int numrep
= crush_do_rule(crush
, rule
, x
, rawout
, maxout
, &weight
[0],
1428 weight
.size(), work
, arg_map
.args
);
1432 for (int i
=0; i
<numrep
; i
++)
1436 int _choose_type_stack(
1438 const vector
<pair
<int,int>>& stack
,
1439 const set
<int>& overfull
,
1440 const vector
<int>& underfull
,
1441 const vector
<int>& orig
,
1442 vector
<int>::const_iterator
& i
,
1444 vector
<int> *pw
) const;
1450 const set
<int>& overfull
,
1451 const vector
<int>& underfull
,
1452 const vector
<int>& orig
,
1453 vector
<int> *out
) const;
1455 bool check_crush_rule(int ruleset
, int type
, int size
, ostream
& ss
) {
1459 for (i
= 0; i
< crush
->max_rules
; i
++) {
1460 if (crush
->rules
[i
] &&
1461 crush
->rules
[i
]->mask
.ruleset
== ruleset
&&
1462 crush
->rules
[i
]->mask
.type
== type
) {
1464 if (crush
->rules
[i
]->mask
.min_size
<= size
&&
1465 crush
->rules
[i
]->mask
.max_size
>= size
) {
1467 } else if (size
< crush
->rules
[i
]->mask
.min_size
) {
1468 ss
<< "pool size is smaller than the crush rule min size";
1471 ss
<< "pool size is bigger than the crush rule max size";
1480 void encode(bufferlist
&bl
, uint64_t features
) const;
1481 void decode(bufferlist::iterator
&blp
);
1482 void decode_crush_bucket(crush_bucket
** bptr
, bufferlist::iterator
&blp
);
1483 void dump(Formatter
*f
) const;
1484 void dump_rules(Formatter
*f
) const;
1485 void dump_rule(int ruleset
, Formatter
*f
) const;
1486 void dump_tunables(Formatter
*f
) const;
1487 void dump_choose_args(Formatter
*f
) const;
1488 void list_rules(Formatter
*f
) const;
1489 void list_rules(ostream
*ss
) const;
1490 void dump_tree(ostream
*out
,
1492 const CrushTreeDumper::name_map_t
& ws
,
1493 bool show_shadow
= false) const;
1494 void dump_tree(ostream
*out
, Formatter
*f
) {
1495 dump_tree(out
, f
, CrushTreeDumper::name_map_t());
1497 void dump_tree(Formatter
*f
,
1498 const CrushTreeDumper::name_map_t
& ws
) const;
1499 static void generate_test_instances(list
<CrushWrapper
*>& o
);
1501 int get_osd_pool_default_crush_replicated_ruleset(CephContext
*cct
);
1503 static bool is_valid_crush_name(const string
& s
);
1504 static bool is_valid_crush_loc(CephContext
*cct
,
1505 const map
<string
,string
>& loc
);
1507 WRITE_CLASS_ENCODER_FEATURES(CrushWrapper
)