]>
Commit | Line | Data |
---|---|---|
7c673cae FG |
1 | // -*- mode:C++; tab-width:8; c-basic-offset:2; indent-tabs-mode:t -*- |
2 | // vim: ts=8 sw=2 smarttab | |
3 | /* | |
4 | * Ceph - scalable distributed file system | |
5 | * | |
6 | * Copyright (C) 2004-2006 Sage Weil <sage@newdream.net> | |
7 | * | |
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. | |
12 | * | |
13 | */ | |
14 | ||
15 | ||
7c673cae FG |
16 | #ifndef CEPH_CDIR_H |
17 | #define CEPH_CDIR_H | |
18 | ||
7c673cae | 19 | #include <iosfwd> |
7c673cae | 20 | #include <list> |
7c673cae | 21 | #include <map> |
94b18763 | 22 | #include <set> |
7c673cae | 23 | #include <string> |
11fdf7f2 | 24 | #include <string_view> |
7c673cae | 25 | |
94b18763 FG |
26 | #include "common/bloom_filter.hpp" |
27 | #include "common/config.h" | |
28 | #include "include/buffer_fwd.h" | |
29 | #include "include/counter.h" | |
30 | #include "include/types.h" | |
7c673cae FG |
31 | |
32 | #include "CInode.h" | |
94b18763 | 33 | #include "MDSCacheObject.h" |
11fdf7f2 TL |
34 | #include "MDSContext.h" |
35 | #include "cephfs_features.h" | |
36 | #include "SessionMap.h" | |
37 | #include "messages/MClientReply.h" | |
7c673cae FG |
38 | |
39 | class CDentry; | |
40 | class MDCache; | |
41 | ||
42 | struct ObjectOperation; | |
43 | ||
44 | ostream& operator<<(ostream& out, const class CDir& dir); | |
7c673cae | 45 | |
9f95a23c | 46 | class CDir : public MDSCacheObject, public Counter<CDir> { |
7c673cae | 47 | public: |
181888fb | 48 | MEMPOOL_CLASS_HELPERS(); |
9f95a23c TL |
49 | |
50 | typedef mempool::mds_co::map<dentry_key_t, CDentry*> dentry_key_map; | |
51 | typedef mempool::mds_co::set<dentry_key_t> dentry_key_set; | |
52 | ||
53 | // -- freezing -- | |
54 | struct freeze_tree_state_t { | |
55 | CDir *dir; // freezing/frozen tree root | |
56 | int auth_pins = 0; | |
57 | bool frozen = false; | |
58 | freeze_tree_state_t(CDir *d) : dir(d) {} | |
59 | }; | |
60 | ||
61 | class scrub_info_t { | |
62 | public: | |
63 | MEMPOOL_CLASS_HELPERS(); | |
64 | struct scrub_stamps { | |
65 | version_t version; | |
66 | utime_t time; | |
67 | scrub_stamps() : version(0) {} | |
68 | void operator=(const scrub_stamps &o) { | |
69 | version = o.version; | |
70 | time = o.time; | |
71 | } | |
72 | }; | |
73 | ||
74 | scrub_info_t() : | |
75 | directory_scrubbing(false), | |
76 | need_scrub_local(false), | |
77 | last_scrub_dirty(false), | |
78 | pending_scrub_error(false) {} | |
79 | ||
80 | /// inodes we contain with dirty scrub stamps | |
81 | dentry_key_map dirty_scrub_stamps; // TODO: make use of this! | |
82 | ||
83 | scrub_stamps recursive_start; // when we last started a recursive scrub | |
84 | scrub_stamps last_recursive; // when we last finished a recursive scrub | |
85 | scrub_stamps last_local; // when we last did a local scrub | |
86 | ||
87 | bool directory_scrubbing; /// safety check | |
88 | bool need_scrub_local; | |
89 | bool last_scrub_dirty; /// is scrub info dirty or is it flushed to fnode? | |
90 | bool pending_scrub_error; | |
91 | ||
92 | /// these are lists of children in each stage of scrubbing | |
93 | dentry_key_set directories_to_scrub; | |
94 | dentry_key_set directories_scrubbing; | |
95 | dentry_key_set directories_scrubbed; | |
96 | dentry_key_set others_to_scrub; | |
97 | dentry_key_set others_scrubbing; | |
98 | dentry_key_set others_scrubbed; | |
99 | ||
100 | ScrubHeaderRefConst header; | |
101 | }; | |
102 | ||
7c673cae FG |
103 | // -- pins -- |
104 | static const int PIN_DNWAITER = 1; | |
105 | static const int PIN_INOWAITER = 2; | |
106 | static const int PIN_CHILD = 3; | |
107 | static const int PIN_FROZEN = 4; | |
108 | static const int PIN_SUBTREE = 5; | |
109 | static const int PIN_IMPORTING = 7; | |
110 | static const int PIN_IMPORTBOUND = 9; | |
111 | static const int PIN_EXPORTBOUND = 10; | |
112 | static const int PIN_STICKY = 11; | |
113 | static const int PIN_SUBTREETEMP = 12; // used by MDCache::trim_non_auth() | |
7c673cae FG |
114 | |
115 | // -- state -- | |
9f95a23c TL |
116 | static const unsigned STATE_COMPLETE = (1<< 0); // the complete contents are in cache |
117 | static const unsigned STATE_FROZENTREE = (1<< 1); // root of tree (bounded by exports) | |
118 | static const unsigned STATE_FREEZINGTREE = (1<< 2); // in process of freezing | |
11fdf7f2 TL |
119 | static const unsigned STATE_FROZENDIR = (1<< 3); |
120 | static const unsigned STATE_FREEZINGDIR = (1<< 4); | |
9f95a23c TL |
121 | static const unsigned STATE_COMMITTING = (1<< 5); // mid-commit |
122 | static const unsigned STATE_FETCHING = (1<< 6); // currenting fetching | |
11fdf7f2 TL |
123 | static const unsigned STATE_CREATING = (1<< 7); |
124 | static const unsigned STATE_IMPORTBOUND = (1<< 8); | |
125 | static const unsigned STATE_EXPORTBOUND = (1<< 9); | |
126 | static const unsigned STATE_EXPORTING = (1<<10); | |
127 | static const unsigned STATE_IMPORTING = (1<<11); | |
128 | static const unsigned STATE_FRAGMENTING = (1<<12); | |
129 | static const unsigned STATE_STICKY = (1<<13); // sticky pin due to inode stickydirs | |
130 | static const unsigned STATE_DNPINNEDFRAG = (1<<14); // dir is refragmenting | |
131 | static const unsigned STATE_ASSIMRSTAT = (1<<15); // assimilating inode->frag rstats | |
132 | static const unsigned STATE_DIRTYDFT = (1<<16); // dirty dirfragtree | |
133 | static const unsigned STATE_BADFRAG = (1<<17); // bad dirfrag | |
134 | static const unsigned STATE_TRACKEDBYOFT = (1<<18); // tracked by open file table | |
135 | static const unsigned STATE_AUXSUBTREE = (1<<19); // no subtree merge | |
7c673cae FG |
136 | |
137 | // common states | |
138 | static const unsigned STATE_CLEAN = 0; | |
7c673cae FG |
139 | |
140 | // these state bits are preserved by an import/export | |
141 | // ...except if the directory is hashed, in which case none of them are! | |
142 | static const unsigned MASK_STATE_EXPORTED = | |
143 | (STATE_COMPLETE|STATE_DIRTY|STATE_DIRTYDFT|STATE_BADFRAG); | |
144 | static const unsigned MASK_STATE_IMPORT_KEPT = | |
145 | ( | |
11fdf7f2 TL |
146 | STATE_IMPORTING | |
147 | STATE_IMPORTBOUND | | |
148 | STATE_EXPORTBOUND | | |
149 | STATE_FROZENTREE | | |
150 | STATE_STICKY | | |
151 | STATE_TRACKEDBYOFT); | |
7c673cae | 152 | static const unsigned MASK_STATE_EXPORT_KEPT = |
11fdf7f2 TL |
153 | (STATE_EXPORTING | |
154 | STATE_IMPORTBOUND | | |
155 | STATE_EXPORTBOUND | | |
156 | STATE_FROZENTREE | | |
157 | STATE_FROZENDIR | | |
158 | STATE_STICKY | | |
159 | STATE_TRACKEDBYOFT); | |
7c673cae | 160 | static const unsigned MASK_STATE_FRAGMENT_KEPT = |
11fdf7f2 | 161 | (STATE_DIRTY | |
7c673cae FG |
162 | STATE_EXPORTBOUND | |
163 | STATE_IMPORTBOUND | | |
164 | STATE_AUXSUBTREE | | |
165 | STATE_REJOINUNDEF); | |
166 | ||
167 | // -- rep spec -- | |
168 | static const int REP_NONE = 0; | |
169 | static const int REP_ALL = 1; | |
170 | static const int REP_LIST = 2; | |
171 | ||
7c673cae FG |
172 | static const unsigned EXPORT_NONCE = 1; |
173 | ||
7c673cae FG |
174 | // -- wait masks -- |
175 | static const uint64_t WAIT_DENTRY = (1<<0); // wait for item to be in cache | |
176 | static const uint64_t WAIT_COMPLETE = (1<<1); // wait for complete dir contents | |
177 | static const uint64_t WAIT_FROZEN = (1<<2); // auth pins removed | |
178 | static const uint64_t WAIT_CREATED = (1<<3); // new dirfrag is logged | |
179 | ||
180 | static const int WAIT_DNLOCK_OFFSET = 4; | |
181 | ||
182 | static const uint64_t WAIT_ANY_MASK = (uint64_t)(-1); | |
7c673cae FG |
183 | static const uint64_t WAIT_ATSUBTREEROOT = (WAIT_SINGLEAUTH); |
184 | ||
11fdf7f2 TL |
185 | // -- dump flags -- |
186 | static const int DUMP_PATH = (1 << 0); | |
187 | static const int DUMP_DIRFRAG = (1 << 1); | |
188 | static const int DUMP_SNAPID_FIRST = (1 << 2); | |
189 | static const int DUMP_VERSIONS = (1 << 3); | |
190 | static const int DUMP_REP = (1 << 4); | |
191 | static const int DUMP_DIR_AUTH = (1 << 5); | |
192 | static const int DUMP_STATES = (1 << 6); | |
193 | static const int DUMP_MDS_CACHE_OBJECT = (1 << 7); | |
194 | static const int DUMP_ITEMS = (1 << 8); | |
195 | static const int DUMP_ALL = (-1); | |
9f95a23c | 196 | static const int DUMP_DEFAULT = DUMP_ALL & (~DUMP_ITEMS); |
7c673cae | 197 | |
9f95a23c | 198 | CDir(CInode *in, frag_t fg, MDCache *mdcache, bool auth); |
7c673cae | 199 | |
9f95a23c TL |
200 | std::string_view pin_name(int p) const override { |
201 | switch (p) { | |
202 | case PIN_DNWAITER: return "dnwaiter"; | |
203 | case PIN_INOWAITER: return "inowaiter"; | |
204 | case PIN_CHILD: return "child"; | |
205 | case PIN_FROZEN: return "frozen"; | |
206 | case PIN_SUBTREE: return "subtree"; | |
207 | case PIN_IMPORTING: return "importing"; | |
208 | case PIN_IMPORTBOUND: return "importbound"; | |
209 | case PIN_EXPORTBOUND: return "exportbound"; | |
210 | case PIN_STICKY: return "sticky"; | |
211 | case PIN_SUBTREETEMP: return "subtreetemp"; | |
212 | default: return generic_pin_name(p); | |
213 | } | |
214 | } | |
7c673cae FG |
215 | |
216 | bool is_lt(const MDSCacheObject *r) const override { | |
217 | return dirfrag() < (static_cast<const CDir*>(r))->dirfrag(); | |
218 | } | |
219 | ||
7c673cae FG |
220 | void resync_accounted_fragstat(); |
221 | void resync_accounted_rstat(); | |
222 | void assimilate_dirty_rstat_inodes(); | |
223 | void assimilate_dirty_rstat_inodes_finish(MutationRef& mut, EMetaBlob *blob); | |
224 | ||
1adf2230 AA |
225 | void mark_exporting() { |
226 | state_set(CDir::STATE_EXPORTING); | |
227 | inode->num_exporting_dirs++; | |
228 | } | |
229 | void clear_exporting() { | |
230 | state_clear(CDir::STATE_EXPORTING); | |
231 | inode->num_exporting_dirs--; | |
232 | } | |
233 | ||
7c673cae FG |
234 | version_t get_version() const { return fnode.version; } |
235 | void set_version(version_t v) { | |
11fdf7f2 | 236 | ceph_assert(projected_fnode.empty()); |
7c673cae FG |
237 | projected_version = fnode.version = v; |
238 | } | |
239 | version_t get_projected_version() const { return projected_version; } | |
240 | ||
241 | const fnode_t *get_projected_fnode() const { | |
242 | if (projected_fnode.empty()) | |
243 | return &fnode; | |
244 | else | |
94b18763 | 245 | return &projected_fnode.back(); |
7c673cae FG |
246 | } |
247 | ||
248 | fnode_t *get_projected_fnode() { | |
249 | if (projected_fnode.empty()) | |
250 | return &fnode; | |
251 | else | |
94b18763 | 252 | return &projected_fnode.back(); |
7c673cae FG |
253 | } |
254 | fnode_t *project_fnode(); | |
255 | ||
256 | void pop_and_dirty_projected_fnode(LogSegment *ls); | |
257 | bool is_projected() const { return !projected_fnode.empty(); } | |
258 | version_t pre_dirty(version_t min=0); | |
259 | void _mark_dirty(LogSegment *ls); | |
260 | void _set_dirty_flag() { | |
261 | if (!state_test(STATE_DIRTY)) { | |
262 | state_set(STATE_DIRTY); | |
263 | get(PIN_DIRTY); | |
264 | } | |
265 | } | |
266 | void mark_dirty(version_t pv, LogSegment *ls); | |
267 | void mark_clean(); | |
268 | ||
269 | bool is_new() { return item_new.is_on_list(); } | |
270 | void mark_new(LogSegment *ls); | |
271 | ||
272 | bool is_bad() { return state_test(STATE_BADFRAG); } | |
7c673cae | 273 | |
7c673cae FG |
274 | /** |
275 | * Call to start this CDir on a new scrub. | |
276 | * @pre It is not currently scrubbing | |
277 | * @pre The CDir is marked complete. | |
278 | * @post It has set up its internal scrubbing state. | |
279 | */ | |
280 | void scrub_initialize(const ScrubHeaderRefConst& header); | |
281 | /** | |
282 | * Get the next dentry to scrub. Gives you a CDentry* and its meaning. This | |
283 | * function will give you all directory-representing dentries before any | |
284 | * others. | |
285 | * 0: success, you should scrub this CDentry right now | |
286 | * EAGAIN: is currently fetching the next CDentry into memory for you. | |
287 | * It will activate your callback when done; try again when it does! | |
288 | * ENOENT: there are no remaining dentries to scrub | |
289 | * <0: There was an unexpected error | |
290 | * | |
11fdf7f2 | 291 | * @param cb An MDSContext which will be activated only if |
7c673cae FG |
292 | * we return EAGAIN via rcode, or else ignored |
293 | * @param dnout CDentry * which you should next scrub, or NULL | |
294 | * @returns a value as described above | |
295 | */ | |
11fdf7f2 | 296 | int scrub_dentry_next(MDSContext *cb, CDentry **dnout); |
7c673cae FG |
297 | /** |
298 | * Get the currently scrubbing dentries. When returned, the passed-in | |
299 | * list will be filled with all CDentry * which have been returned | |
300 | * from scrub_dentry_next() but not sent back via scrub_dentry_finished(). | |
301 | */ | |
9f95a23c | 302 | std::vector<CDentry*> scrub_dentries_scrubbing(); |
7c673cae FG |
303 | /** |
304 | * Report to the CDir that a CDentry has been scrubbed. Call this | |
305 | * for every CDentry returned from scrub_dentry_next(). | |
306 | * @param dn The CDentry which has been scrubbed. | |
307 | */ | |
308 | void scrub_dentry_finished(CDentry *dn); | |
309 | /** | |
310 | * Call this once all CDentries have been scrubbed, according to | |
311 | * scrub_dentry_next's listing. It finalizes the scrub statistics. | |
312 | */ | |
313 | void scrub_finished(); | |
314 | /** | |
315 | * Tell the CDir to do a local scrub of itself. | |
316 | * @pre The CDir is_complete(). | |
317 | * @returns true if the rstats and directory contents match, false otherwise. | |
318 | */ | |
319 | bool scrub_local(); | |
7c673cae FG |
320 | |
321 | const scrub_info_t *scrub_info() const { | |
322 | if (!scrub_infop) { | |
323 | scrub_info_create(); | |
324 | } | |
325 | return scrub_infop.get(); | |
326 | } | |
327 | ||
7c673cae FG |
328 | // -- accessors -- |
329 | inodeno_t ino() const { return inode->ino(); } // deprecate me? | |
330 | frag_t get_frag() const { return frag; } | |
331 | dirfrag_t dirfrag() const { return dirfrag_t(inode->ino(), frag); } | |
332 | ||
333 | CInode *get_inode() { return inode; } | |
334 | const CInode *get_inode() const { return inode; } | |
335 | CDir *get_parent_dir() { return inode->get_parent_dir(); } | |
336 | ||
94b18763 FG |
337 | dentry_key_map::iterator begin() { return items.begin(); } |
338 | dentry_key_map::iterator end() { return items.end(); } | |
339 | dentry_key_map::iterator lower_bound(dentry_key_t key) { return items.lower_bound(key); } | |
7c673cae FG |
340 | |
341 | unsigned get_num_head_items() const { return num_head_items; } | |
342 | unsigned get_num_head_null() const { return num_head_null; } | |
343 | unsigned get_num_snap_items() const { return num_snap_items; } | |
344 | unsigned get_num_snap_null() const { return num_snap_null; } | |
345 | unsigned get_num_any() const { return num_head_items + num_head_null + num_snap_items + num_snap_null; } | |
346 | ||
347 | bool check_rstats(bool scrub=false); | |
348 | ||
349 | void inc_num_dirty() { num_dirty++; } | |
350 | void dec_num_dirty() { | |
11fdf7f2 | 351 | ceph_assert(num_dirty > 0); |
7c673cae FG |
352 | num_dirty--; |
353 | } | |
354 | int get_num_dirty() const { | |
355 | return num_dirty; | |
356 | } | |
357 | ||
11fdf7f2 TL |
358 | void adjust_num_inodes_with_caps(int d); |
359 | ||
7c673cae FG |
360 | int64_t get_frag_size() const { |
361 | return get_projected_fnode()->fragstat.size(); | |
362 | } | |
363 | ||
364 | // -- dentries and inodes -- | |
11fdf7f2 TL |
365 | CDentry* lookup_exact_snap(std::string_view dname, snapid_t last); |
366 | CDentry* lookup(std::string_view n, snapid_t snap=CEPH_NOSNAP); | |
7c673cae | 367 | |
11fdf7f2 | 368 | CDentry* add_null_dentry(std::string_view dname, |
7c673cae | 369 | snapid_t first=2, snapid_t last=CEPH_NOSNAP); |
11fdf7f2 | 370 | CDentry* add_primary_dentry(std::string_view dname, CInode *in, |
7c673cae | 371 | snapid_t first=2, snapid_t last=CEPH_NOSNAP); |
11fdf7f2 | 372 | CDentry* add_remote_dentry(std::string_view dname, inodeno_t ino, unsigned char d_type, |
7c673cae FG |
373 | snapid_t first=2, snapid_t last=CEPH_NOSNAP); |
374 | void remove_dentry( CDentry *dn ); // delete dentry | |
375 | void link_remote_inode( CDentry *dn, inodeno_t ino, unsigned char d_type); | |
376 | void link_remote_inode( CDentry *dn, CInode *in ); | |
377 | void link_primary_inode( CDentry *dn, CInode *in ); | |
31f18b77 | 378 | void unlink_inode(CDentry *dn, bool adjust_lru=true); |
7c673cae FG |
379 | void try_remove_unlinked_dn(CDentry *dn); |
380 | ||
381 | void add_to_bloom(CDentry *dn); | |
11fdf7f2 | 382 | bool is_in_bloom(std::string_view name); |
7c673cae FG |
383 | bool has_bloom() { return (bloom ? true : false); } |
384 | void remove_bloom() { | |
385 | bloom.reset(); | |
386 | } | |
9f95a23c | 387 | |
7c673cae FG |
388 | void try_remove_dentries_for_stray(); |
389 | bool try_trim_snap_dentry(CDentry *dn, const std::set<snapid_t>& snaps); | |
390 | ||
9f95a23c TL |
391 | void split(int bits, std::vector<CDir*>* subs, MDSContext::vec& waiters, bool replay); |
392 | void merge(const std::vector<CDir*>& subs, MDSContext::vec& waiters, bool replay); | |
7c673cae FG |
393 | |
394 | bool should_split() const { | |
11fdf7f2 | 395 | return (int)get_frag_size() > g_conf()->mds_bal_split_size; |
7c673cae FG |
396 | } |
397 | bool should_split_fast() const; | |
398 | bool should_merge() const { | |
11fdf7f2 | 399 | return (int)get_frag_size() < g_conf()->mds_bal_merge_size; |
7c673cae FG |
400 | } |
401 | ||
7c673cae FG |
402 | mds_authority_t authority() const override; |
403 | mds_authority_t get_dir_auth() const { return dir_auth; } | |
11fdf7f2 | 404 | void set_dir_auth(const mds_authority_t &a); |
7c673cae FG |
405 | void set_dir_auth(mds_rank_t a) { set_dir_auth(mds_authority_t(a, CDIR_AUTH_UNKNOWN)); } |
406 | bool is_ambiguous_dir_auth() const { | |
407 | return dir_auth.second != CDIR_AUTH_UNKNOWN; | |
408 | } | |
409 | bool is_full_dir_auth() const { | |
410 | return is_auth() && !is_ambiguous_dir_auth(); | |
411 | } | |
412 | bool is_full_dir_nonauth() const { | |
413 | return !is_auth() && !is_ambiguous_dir_auth(); | |
414 | } | |
415 | ||
416 | bool is_subtree_root() const { | |
417 | return dir_auth != CDIR_AUTH_DEFAULT; | |
418 | } | |
419 | ||
420 | bool contains(CDir *x); // true if we are x or an ancestor of x | |
421 | ||
7c673cae FG |
422 | // for giving to clients |
423 | void get_dist_spec(std::set<mds_rank_t>& ls, mds_rank_t auth) { | |
f91f0fd5 | 424 | if (is_auth()) { |
7c673cae FG |
425 | list_replicas(ls); |
426 | if (!ls.empty()) | |
427 | ls.insert(auth); | |
428 | } | |
429 | } | |
7c673cae | 430 | |
11fdf7f2 | 431 | static void encode_dirstat(bufferlist& bl, const session_info_t& info, const DirStat& ds); |
7c673cae FG |
432 | |
433 | void _encode_base(bufferlist& bl) { | |
9f95a23c | 434 | ENCODE_START(1, 1, bl); |
11fdf7f2 TL |
435 | encode(first, bl); |
436 | encode(fnode, bl); | |
437 | encode(dir_rep, bl); | |
438 | encode(dir_rep_by, bl); | |
9f95a23c | 439 | ENCODE_FINISH(bl); |
7c673cae | 440 | } |
11fdf7f2 | 441 | void _decode_base(bufferlist::const_iterator& p) { |
9f95a23c | 442 | DECODE_START(1, p); |
11fdf7f2 TL |
443 | decode(first, p); |
444 | decode(fnode, p); | |
445 | decode(dir_rep, p); | |
446 | decode(dir_rep_by, p); | |
9f95a23c | 447 | DECODE_FINISH(p); |
7c673cae | 448 | } |
7c673cae FG |
449 | |
450 | // -- state -- | |
451 | bool is_complete() { return state & STATE_COMPLETE; } | |
452 | bool is_exporting() { return state & STATE_EXPORTING; } | |
453 | bool is_importing() { return state & STATE_IMPORTING; } | |
454 | bool is_dirty_dft() { return state & STATE_DIRTYDFT; } | |
455 | ||
456 | int get_dir_rep() const { return dir_rep; } | |
457 | bool is_rep() const { | |
458 | if (dir_rep == REP_NONE) return false; | |
459 | return true; | |
460 | } | |
461 | ||
462 | // -- fetch -- | |
463 | object_t get_ondisk_object() { | |
464 | return file_object_t(ino(), frag); | |
465 | } | |
11fdf7f2 TL |
466 | void fetch(MDSContext *c, bool ignore_authpinnability=false); |
467 | void fetch(MDSContext *c, std::string_view want_dn, bool ignore_authpinnability=false); | |
468 | void fetch(MDSContext *c, const std::set<dentry_key_t>& keys); | |
7c673cae | 469 | |
7c673cae FG |
470 | #if 0 // unused? |
471 | void wait_for_commit(Context *c, version_t v=0); | |
472 | #endif | |
473 | void commit_to(version_t want); | |
11fdf7f2 | 474 | void commit(version_t want, MDSContext *c, |
7c673cae FG |
475 | bool ignore_authpinnability=false, int op_prio=-1); |
476 | ||
477 | // -- dirtyness -- | |
478 | version_t get_committing_version() const { return committing_version; } | |
479 | version_t get_committed_version() const { return committed_version; } | |
480 | void set_committed_version(version_t v) { committed_version = v; } | |
481 | ||
482 | void mark_complete(); | |
483 | ||
7c673cae FG |
484 | // -- reference counting -- |
485 | void first_get() override; | |
486 | void last_put() override; | |
487 | ||
11fdf7f2 | 488 | bool is_waiting_for_dentry(std::string_view dname, snapid_t snap) { |
7c673cae FG |
489 | return waiting_on_dentry.count(string_snap_t(dname, snap)); |
490 | } | |
11fdf7f2 TL |
491 | void add_dentry_waiter(std::string_view dentry, snapid_t snap, MDSContext *c); |
492 | void take_dentry_waiting(std::string_view dentry, snapid_t first, snapid_t last, MDSContext::vec& ls); | |
493 | void take_sub_waiting(MDSContext::vec& ls); // dentry or ino | |
7c673cae | 494 | |
11fdf7f2 TL |
495 | void add_waiter(uint64_t mask, MDSContext *c) override; |
496 | void take_waiting(uint64_t mask, MDSContext::vec& ls) override; // may include dentry waiters | |
7c673cae | 497 | void finish_waiting(uint64_t mask, int result = 0); // ditto |
7c673cae FG |
498 | |
499 | // -- import/export -- | |
500 | void encode_export(bufferlist& bl); | |
11fdf7f2 | 501 | void finish_export(); |
7c673cae FG |
502 | void abort_export() { |
503 | put(PIN_TEMPEXPORTING); | |
504 | } | |
11fdf7f2 TL |
505 | void decode_import(bufferlist::const_iterator& blp, LogSegment *ls); |
506 | void abort_import(); | |
7c673cae FG |
507 | |
508 | // -- auth pins -- | |
91327a77 | 509 | bool can_auth_pin(int *err_ret=nullptr) const override; |
7c673cae | 510 | int get_auth_pins() const { return auth_pins; } |
7c673cae FG |
511 | int get_dir_auth_pins() const { return dir_auth_pins; } |
512 | void auth_pin(void *who) override; | |
513 | void auth_unpin(void *who) override; | |
514 | ||
11fdf7f2 | 515 | void adjust_nested_auth_pins(int dirinc, void *by); |
7c673cae FG |
516 | void verify_fragstat(); |
517 | ||
11fdf7f2 TL |
518 | void _walk_tree(std::function<bool(CDir*)> cb); |
519 | ||
7c673cae FG |
520 | bool freeze_tree(); |
521 | void _freeze_tree(); | |
522 | void unfreeze_tree(); | |
11fdf7f2 | 523 | void adjust_freeze_after_rename(CDir *dir); |
7c673cae FG |
524 | |
525 | bool freeze_dir(); | |
526 | void _freeze_dir(); | |
527 | void unfreeze_dir(); | |
528 | ||
529 | void maybe_finish_freeze(); | |
530 | ||
11fdf7f2 TL |
531 | pair<bool,bool> is_freezing_or_frozen_tree() const { |
532 | if (freeze_tree_state) { | |
533 | if (freeze_tree_state->frozen) | |
534 | return make_pair(false, true); | |
535 | return make_pair(true, false); | |
536 | } | |
537 | return make_pair(false, false); | |
538 | } | |
91327a77 AA |
539 | |
540 | bool is_freezing() const override { return is_freezing_dir() || is_freezing_tree(); } | |
541 | bool is_freezing_tree() const { | |
542 | if (!num_freezing_trees) | |
543 | return false; | |
544 | return is_freezing_or_frozen_tree().first; | |
545 | } | |
7c673cae FG |
546 | bool is_freezing_tree_root() const { return state & STATE_FREEZINGTREE; } |
547 | bool is_freezing_dir() const { return state & STATE_FREEZINGDIR; } | |
548 | ||
549 | bool is_frozen() const override { return is_frozen_dir() || is_frozen_tree(); } | |
91327a77 AA |
550 | bool is_frozen_tree() const { |
551 | if (!num_frozen_trees) | |
552 | return false; | |
553 | return is_freezing_or_frozen_tree().second; | |
554 | } | |
7c673cae FG |
555 | bool is_frozen_tree_root() const { return state & STATE_FROZENTREE; } |
556 | bool is_frozen_dir() const { return state & STATE_FROZENDIR; } | |
11fdf7f2 | 557 | |
7c673cae FG |
558 | bool is_freezeable(bool freezing=false) const { |
559 | // no nested auth pins. | |
11fdf7f2 TL |
560 | if (auth_pins - (freezing ? 1 : 0) > 0 || |
561 | (freeze_tree_state && freeze_tree_state->auth_pins != auth_pins)) | |
7c673cae FG |
562 | return false; |
563 | ||
564 | // inode must not be frozen. | |
565 | if (!is_subtree_root() && inode->is_frozen()) | |
566 | return false; | |
567 | ||
568 | return true; | |
569 | } | |
11fdf7f2 | 570 | |
7c673cae | 571 | bool is_freezeable_dir(bool freezing=false) const { |
11fdf7f2 | 572 | if ((auth_pins - freezing) > 0 || dir_auth_pins > 0) |
7c673cae FG |
573 | return false; |
574 | ||
575 | // if not subtree root, inode must not be frozen (tree--frozen_dir is okay). | |
576 | if (!is_subtree_root() && inode->is_frozen() && !inode->is_frozen_dir()) | |
577 | return false; | |
578 | ||
579 | return true; | |
580 | } | |
581 | ||
9f95a23c TL |
582 | bool is_any_freezing_or_frozen_inode() const { |
583 | return num_frozen_inodes || !freezing_inodes.empty(); | |
584 | } | |
585 | bool is_auth_pinned_by_lock_cache() const { | |
586 | return frozen_inode_suppressed; | |
587 | } | |
588 | void disable_frozen_inode() { | |
589 | ceph_assert(num_frozen_inodes == 0); | |
590 | frozen_inode_suppressed++; | |
591 | } | |
592 | void enable_frozen_inode(); | |
593 | ||
7c673cae FG |
594 | ostream& print_db_line_prefix(ostream& out) override; |
595 | void print(ostream& out) override; | |
11fdf7f2 TL |
596 | void dump(Formatter *f, int flags = DUMP_DEFAULT) const; |
597 | void dump_load(Formatter *f); | |
9f95a23c TL |
598 | |
599 | // context | |
600 | MDCache *cache; | |
601 | ||
602 | CInode *inode; // my inode | |
603 | frag_t frag; // my frag | |
604 | ||
605 | fnode_t fnode; | |
606 | snapid_t first = 2; | |
607 | mempool::mds_co::compact_map<snapid_t,old_rstat_t> dirty_old_rstat; // [value.first,key] | |
608 | ||
609 | // my inodes with dirty rstat data | |
610 | elist<CInode*> dirty_rstat_inodes; | |
611 | ||
612 | elist<CDentry*> dirty_dentries; | |
613 | elist<CDir*>::item item_dirty, item_new; | |
614 | ||
615 | // lock caches that auth-pin me | |
616 | elist<MDLockCache::DirItem*> lock_caches_with_auth_pins; | |
617 | ||
618 | // all dirfrags within freezing/frozen tree reference the 'state' | |
619 | std::shared_ptr<freeze_tree_state_t> freeze_tree_state; | |
620 | ||
621 | protected: | |
622 | // friends | |
623 | friend class Migrator; | |
624 | friend class CInode; | |
625 | friend class MDCache; | |
626 | friend class MDiscover; | |
627 | friend class MDBalancer; | |
628 | ||
629 | friend class CDirDiscover; | |
630 | friend class CDirExport; | |
631 | friend class C_IO_Dir_TMAP_Fetched; | |
632 | friend class C_IO_Dir_OMAP_Fetched; | |
633 | friend class C_IO_Dir_OMAP_FetchedMore; | |
634 | friend class C_IO_Dir_Committed; | |
635 | ||
636 | void _omap_fetch(MDSContext *fin, const std::set<dentry_key_t>& keys); | |
637 | void _omap_fetch_more( | |
638 | bufferlist& hdrbl, std::map<std::string, bufferlist>& omap, | |
639 | MDSContext *fin); | |
640 | CDentry *_load_dentry( | |
641 | std::string_view key, | |
642 | std::string_view dname, | |
643 | snapid_t last, | |
644 | bufferlist &bl, | |
645 | int pos, | |
646 | const std::set<snapid_t> *snaps, | |
f91f0fd5 | 647 | double rand_threshold, |
9f95a23c TL |
648 | bool *force_dirty); |
649 | ||
9f95a23c TL |
650 | /** |
651 | * Go bad due to a damaged dentry (register with damagetable and go BADFRAG) | |
652 | */ | |
653 | void go_bad_dentry(snapid_t last, std::string_view dname); | |
654 | ||
655 | /** | |
656 | * Go bad due to a damaged header (register with damagetable and go BADFRAG) | |
657 | */ | |
658 | void go_bad(bool complete); | |
659 | ||
660 | void _omap_fetched(bufferlist& hdrbl, std::map<std::string, bufferlist>& omap, | |
661 | bool complete, int r); | |
662 | ||
663 | // -- commit -- | |
664 | void _commit(version_t want, int op_prio); | |
665 | void _omap_commit(int op_prio); | |
666 | void _encode_dentry(CDentry *dn, bufferlist& bl, const std::set<snapid_t> *snaps); | |
667 | void _committed(int r, version_t v); | |
668 | ||
669 | version_t projected_version = 0; | |
670 | mempool::mds_co::list<fnode_t> projected_fnode; | |
671 | ||
672 | std::unique_ptr<scrub_info_t> scrub_infop; | |
673 | ||
674 | // contents of this directory | |
675 | dentry_key_map items; // non-null AND null | |
676 | unsigned num_head_items = 0; | |
677 | unsigned num_head_null = 0; | |
678 | unsigned num_snap_items = 0; | |
679 | unsigned num_snap_null = 0; | |
680 | ||
681 | int num_dirty = 0; | |
682 | ||
683 | int num_inodes_with_caps = 0; | |
684 | ||
685 | // state | |
686 | version_t committing_version = 0; | |
687 | version_t committed_version = 0; | |
688 | ||
689 | mempool::mds_co::compact_set<mempool::mds_co::string> stale_items; | |
690 | ||
691 | // lock nesting, freeze | |
692 | static int num_frozen_trees; | |
693 | static int num_freezing_trees; | |
694 | ||
695 | // freezing/frozen inodes in this dirfrag | |
696 | int num_frozen_inodes = 0; | |
697 | int frozen_inode_suppressed = 0; | |
698 | elist<CInode*> freezing_inodes; | |
699 | ||
700 | int dir_auth_pins = 0; | |
701 | ||
702 | // cache control (defined for authority; hints for replicas) | |
703 | __s32 dir_rep; | |
704 | mempool::mds_co::compact_set<__s32> dir_rep_by; // if dir_rep == REP_LIST | |
705 | ||
706 | // popularity | |
707 | dirfrag_load_vec_t pop_me; | |
708 | dirfrag_load_vec_t pop_nested; | |
709 | dirfrag_load_vec_t pop_auth_subtree; | |
710 | dirfrag_load_vec_t pop_auth_subtree_nested; | |
711 | ||
712 | ceph::coarse_mono_time last_popularity_sample = ceph::coarse_mono_clock::zero(); | |
713 | ||
714 | load_spread_t pop_spread; | |
715 | ||
716 | elist<CInode*> pop_lru_subdirs; | |
717 | ||
718 | std::unique_ptr<bloom_filter> bloom; // XXX not part of mempool::mds_co | |
719 | /* If you set up the bloom filter, you must keep it accurate! | |
720 | * It's deleted when you mark_complete() and is deliberately not serialized.*/ | |
721 | ||
722 | mempool::mds_co::compact_set<mempool::mds_co::string> wanted_items; | |
723 | mempool::mds_co::compact_map<version_t, MDSContext::vec_alloc<mempool::mds_co::pool_allocator> > waiting_for_commit; | |
724 | ||
725 | // -- waiters -- | |
726 | mempool::mds_co::compact_map< string_snap_t, MDSContext::vec_alloc<mempool::mds_co::pool_allocator> > waiting_on_dentry; // FIXME string_snap_t not in mempool | |
727 | ||
728 | private: | |
729 | friend ostream& operator<<(ostream& out, const class CDir& dir); | |
730 | ||
731 | void log_mark_dirty(); | |
732 | ||
733 | /** | |
734 | * Create a scrub_info_t struct for the scrub_infop pointer. | |
735 | */ | |
736 | void scrub_info_create() const; | |
737 | /** | |
738 | * Delete the scrub_infop if it's not got any useful data. | |
739 | */ | |
740 | void scrub_maybe_delete_info(); | |
741 | /** | |
742 | * Check the given set (presumably one of those in scrub_info_t) for the | |
743 | * next key to scrub and look it up (or fail!). | |
744 | */ | |
745 | int _next_dentry_on_set(dentry_key_set &dns, bool missing_okay, | |
746 | MDSContext *cb, CDentry **dnout); | |
747 | ||
748 | void link_inode_work( CDentry *dn, CInode *in ); | |
749 | void unlink_inode_work( CDentry *dn ); | |
750 | void remove_null_dentries(); | |
751 | void purge_stale_snap_data(const std::set<snapid_t>& snaps); | |
752 | ||
753 | void prepare_new_fragment(bool replay); | |
754 | void prepare_old_fragment(map<string_snap_t, MDSContext::vec >& dentry_waiters, bool replay); | |
755 | void steal_dentry(CDentry *dn); // from another dir. used by merge/split. | |
756 | void finish_old_fragment(MDSContext::vec& waiters, bool replay); | |
757 | void init_fragment_pins(); | |
758 | std::string get_path() const; | |
759 | ||
760 | // -- authority -- | |
761 | /* | |
762 | * normal: <parent,unknown> !subtree_root | |
763 | * delegation: <mds,unknown> subtree_root | |
764 | * ambiguous: <mds1,mds2> subtree_root | |
765 | * <parent,mds2> subtree_root | |
766 | */ | |
767 | mds_authority_t dir_auth; | |
7c673cae FG |
768 | }; |
769 | ||
770 | #endif |