Line data Source code
1 : /*
2 : * SPDX-License-Identifier: MPL-2.0
3 : *
4 : * This Source Code Form is subject to the terms of the Mozilla Public
5 : * License, v. 2.0. If a copy of the MPL was not distributed with this
6 : * file, You can obtain one at http://mozilla.org/MPL/2.0/.
7 : *
8 : * Copyright 2024 MonetDB Foundation;
9 : * Copyright August 2008 - 2023 MonetDB B.V.;
10 : * Copyright 1997 - July 2008 CWI.
11 : */
12 :
13 : #include "monetdb_config.h"
14 : #include "sql_catalog.h"
15 : #include "sql_storage.h"
16 :
17 : #include "gdk_atoms.h"
18 :
19 : struct versionhead ;
20 :
21 : #define active (0)
22 : #define under_destruction (1<<1)
23 : #define block_destruction (1<<2)
24 : #define deleted (1<<3)
25 : #define rollbacked (1<<4)
26 :
27 : /* This objectversion owns its associated versionhead.
28 : * When this objectversion gets destroyed,
29 : * the cleanup procedure should also destroy the associated (name|id) based versionhead.*/
30 : #define name_based_versionhead_owner (1<<5)
31 : #define id_based_versionhead_owner (1<<6)
32 :
33 : typedef struct objectversion {
34 : ulng ts;
35 : ATOMIC_TYPE state;
36 : sql_base *b; // base of underlying sql object
37 : struct objectset* os;
38 : struct objectversion *name_based_older;
39 : struct objectversion *name_based_newer;
40 : struct versionhead *name_based_head;
41 :
42 : struct objectversion *id_based_older;
43 : struct objectversion *id_based_newer;
44 : struct versionhead *id_based_head;
45 : } objectversion;
46 :
47 : typedef struct versionhead {
48 : struct versionhead * prev;
49 : struct versionhead * next;
50 : objectversion* ov;
51 : } versionhead ;
52 :
53 : typedef struct objectset {
54 : ATOMIC_TYPE refcnt;
55 : allocator *sa;
56 : destroy_fptr destroy;
57 : MT_RWLock rw_lock; /*readers-writer lock to protect the links (chains) in the objectversion chain.*/
58 : MT_Lock lock; /* global objectset lock for os_add/del */
59 : versionhead *name_based_h;
60 : versionhead *name_based_t;
61 : versionhead *id_based_h;
62 : versionhead *id_based_t;
63 : int name_based_cnt;
64 : int id_based_cnt;
65 : struct sql_hash *name_map;
66 : struct sql_hash *id_map;
67 : bool
68 : temporary:1,
69 : unique:1, /* names are unique */
70 : concurrent:1, /* concurrent inserts are allowed */
71 : nested:1;
72 : sql_store store;
73 : } objectset;
74 :
75 : static int
76 316350 : os_id_key(versionhead *n)
77 : {
78 245248 : return (int) BATatoms[TYPE_int].atomHash(&n->ov->b->id);
79 : }
80 :
81 : static inline void
82 21617497 : lock_reader(objectset* os)
83 : {
84 21617497 : MT_rwlock_rdlock(&os->rw_lock);
85 1085696 : }
86 :
87 : static inline void
88 21613238 : unlock_reader(objectset* os)
89 : {
90 21613238 : MT_rwlock_rdunlock(&os->rw_lock);
91 4039871 : }
92 :
93 : static inline void
94 580009 : lock_writer(objectset* os)
95 : {
96 580009 : MT_rwlock_wrlock(&os->rw_lock);
97 : }
98 :
99 : static inline void
100 580009 : unlock_writer(objectset* os)
101 : {
102 580009 : MT_rwlock_wrunlock(&os->rw_lock);
103 131 : }
104 :
105 18469574 : static bte os_atmc_get_state(objectversion *ov) {
106 18469574 : bte state = (bte) ATOMIC_GET(&ov->state);
107 18469574 : return state;
108 : }
109 :
110 64564 : static void os_atmc_set_state(objectversion *ov, bte state) {
111 64564 : ATOMIC_SET(&ov->state, state);
112 0 : }
113 :
114 : static versionhead *
115 432677 : find_id(objectset *os, sqlid id)
116 : {
117 432677 : if (os) {
118 432677 : lock_reader(os);
119 432677 : if (os->id_map) {
120 401909 : int key = (int) BATatoms[TYPE_int].atomHash(&id);
121 401909 : sql_hash_e *he = os->id_map->buckets[key&(os->id_map->size-1)];
122 :
123 1336483 : for (; he; he = he->chain) {
124 1109321 : versionhead *n = he->value;
125 :
126 1109321 : if (n && n->ov->b->id == id) {
127 174747 : unlock_reader(os);
128 174747 : return n;
129 : }
130 : }
131 227162 : unlock_reader(os);
132 227162 : return NULL;
133 : }
134 :
135 92892 : for (versionhead *n = os->id_based_h; n; n = n->next) {
136 64666 : objectversion *ov = n->ov;
137 :
138 : /* check if ids match */
139 64666 : if (id == ov->b->id) {
140 2542 : unlock_reader(os);
141 2542 : return n;
142 : }
143 : }
144 28226 : unlock_reader(os);
145 : }
146 :
147 : return NULL;
148 : }
149 :
150 : // TODO copy of static function from sql_list.c. Needs to be made external
151 : static void
152 29585 : hash_delete(sql_hash *h, void *data)
153 : {
154 29585 : int key = h->key(data);
155 29585 : sql_hash_e *e, *p = h->buckets[key&(h->size-1)];
156 :
157 29585 : e = p;
158 45938 : for (; p && p->value != data ; p = p->chain)
159 16353 : e = p;
160 29585 : if (p && p->value == data) {
161 29585 : if (p == e)
162 25282 : h->buckets[key&(h->size-1)] = p->chain;
163 : else
164 4303 : e->chain = p->chain;
165 29585 : if (!h->sa)
166 29585 : _DELETE(p);
167 : }
168 29585 : h->entries--;
169 29585 : }
170 :
171 : static void
172 484464 : node_destroy(objectset *os, sqlstore *store, versionhead *n)
173 : {
174 484464 : if (!os->sa)
175 484464 : _DELETE(n);
176 484464 : (void)store;
177 484464 : }
178 :
179 : static versionhead *
180 15531 : os_remove_name_based_chain(objectset *os, objectversion* ov)
181 : {
182 15531 : lock_writer(os);
183 15531 : versionhead *n = ov->name_based_head;
184 15531 : versionhead *p = os->name_based_h;
185 15531 : if (p != n)
186 1152590 : while (p && p->next != n)
187 : p = p->next;
188 15531 : assert(p==n||(p && p->next == n));
189 15531 : if (p == n) {
190 1332 : os->name_based_h = n->next;
191 1332 : if (os->name_based_h) // i.e. non-empty os
192 818 : os->name_based_h->prev = NULL;
193 : p = NULL;
194 14199 : } else if ( p != NULL) {
195 14199 : p->next = n->next;
196 14199 : if (p->next) // node in the middle
197 5412 : p->next->prev = p;
198 : }
199 15531 : if (n == os->name_based_t)
200 9301 : os->name_based_t = p;
201 :
202 15531 : if (os->name_map && n)
203 14509 : hash_delete(os->name_map, n);
204 :
205 15531 : os->name_based_cnt--;
206 15531 : unlock_writer(os);
207 :
208 15531 : bte state = os_atmc_get_state(ov);
209 15531 : state |= name_based_versionhead_owner;
210 15531 : os_atmc_set_state(ov, state);
211 15531 : return p;
212 : }
213 :
214 : static versionhead *
215 15603 : os_remove_id_based_chain(objectset *os, objectversion* ov)
216 : {
217 15603 : lock_writer(os);
218 15603 : versionhead *n = ov->id_based_head;
219 15603 : versionhead *p = os->id_based_h;
220 :
221 15603 : if (p != n)
222 1160049 : while (p && p->next != n)
223 : p = p->next;
224 15603 : assert(p==n||(p && p->next == n));
225 15603 : if (p == n) {
226 1332 : os->id_based_h = n->next;
227 1332 : if (os->id_based_h) // i.e. non-empty os
228 818 : os->id_based_h->prev = NULL;
229 : p = NULL;
230 14271 : } else if ( p != NULL) {
231 14271 : p->next = n->next;
232 14271 : if (p->next) // node in the middle
233 5484 : p->next->prev = p;
234 : }
235 15603 : if (n == os->id_based_t)
236 9301 : os->id_based_t = p;
237 :
238 15603 : if (os->id_map && n)
239 15076 : hash_delete(os->id_map, n);
240 :
241 15603 : os->name_based_cnt--;
242 15603 : unlock_writer(os);
243 :
244 15603 : bte state = os_atmc_get_state(ov);
245 15603 : state |= id_based_versionhead_owner;
246 15603 : os_atmc_set_state(ov, state);
247 15603 : return p;
248 : }
249 :
250 : static versionhead *
251 509662 : node_create(allocator *sa, objectversion *ov)
252 : {
253 509662 : versionhead *n = SA_NEW(sa, versionhead );
254 :
255 509662 : if (n == NULL)
256 : return NULL;
257 509662 : *n = (versionhead ) {
258 : .ov = ov,
259 : };
260 509662 : return n;
261 : }
262 :
263 : static inline int
264 307623 : os_name_key(versionhead *n)
265 : {
266 307623 : return hash_key(n->ov->b->name);
267 : }
268 :
269 : static objectset *
270 254795 : os_append_node_name(objectset *os, versionhead *n)
271 : {
272 254795 : lock_writer(os);
273 254795 : if ((!os->name_map || os->name_map->size*16 < os->name_based_cnt) && os->name_based_cnt > HASH_MIN_SIZE) {
274 3647 : hash_destroy(os->name_map);
275 3647 : os->name_map = hash_new(os->sa, os->name_based_cnt, (fkeyvalue)& os_name_key);
276 3647 : if (os->name_map == NULL) {
277 0 : unlock_writer(os);
278 0 : return NULL;
279 : }
280 :
281 67152 : for (versionhead *n = os->name_based_h; n; n = n->next ) {
282 63505 : int key = os_name_key(n);
283 :
284 63505 : if (hash_add(os->name_map, key, n) == NULL) {
285 0 : unlock_writer(os);
286 0 : return NULL;
287 : }
288 : }
289 : }
290 :
291 254795 : if (os->name_map) {
292 229609 : int key = os->name_map->key(n);
293 :
294 229609 : if (hash_add(os->name_map, key, n) == NULL) {
295 0 : unlock_writer(os);
296 0 : return NULL;
297 : }
298 : }
299 :
300 254795 : if (os->name_based_t) {
301 246817 : os->name_based_t->next = n;
302 : } else {
303 7978 : os->name_based_h = n;
304 : }
305 254795 : n->prev = os->name_based_t; // aka the double linked list.
306 254795 : os->name_based_t = n;
307 254795 : os->name_based_cnt++;
308 254795 : unlock_writer(os);
309 254795 : return os;
310 : }
311 :
312 : static objectset *
313 254795 : os_append_name(objectset *os, objectversion *ov)
314 : {
315 254795 : versionhead *n = node_create(os->sa, ov);
316 :
317 254795 : if (n == NULL)
318 : return NULL;
319 :
320 254795 : ov->name_based_head = n;
321 254795 : if (!(os = os_append_node_name(os, n))){
322 0 : _DELETE(n);
323 0 : return NULL;
324 : }
325 :
326 : return os;
327 : }
328 :
329 : static void
330 3741 : os_append_id_map(objectset *os)
331 : {
332 3741 : if (os->id_map)
333 432 : hash_destroy(os->id_map);
334 3741 : os->id_map = hash_new(os->sa, os->id_based_cnt, (fkeyvalue)&os_id_key);
335 3741 : if (os->id_map == NULL)
336 : return ;
337 74843 : for (versionhead *n = os->id_based_h; n; n = n->next ) {
338 71102 : int key = os_id_key(n);
339 :
340 71102 : if (hash_add(os->id_map, key, n) == NULL) {
341 0 : hash_destroy(os->id_map);
342 0 : os->id_map = NULL;
343 0 : return ;
344 : }
345 : }
346 : }
347 :
348 : static objectset *
349 254867 : os_append_node_id(objectset *os, versionhead *n)
350 : {
351 254867 : lock_writer(os);
352 254867 : if ((!os->id_map || os->id_map->size*16 < os->id_based_cnt) && os->id_based_cnt > HASH_MIN_SIZE)
353 3741 : os_append_id_map(os); /* on failure just fall back to slow method */
354 :
355 254867 : if (os->id_map) {
356 230172 : int key = os->id_map->key(n);
357 230172 : if (hash_add(os->id_map, key, n) == NULL) {
358 0 : hash_destroy(os->id_map);
359 0 : os->id_map = NULL;
360 : /* fall back to slow search */
361 : }
362 : }
363 :
364 254867 : if (os->id_based_t) {
365 246889 : os->id_based_t->next = n;
366 : } else {
367 7978 : os->id_based_h = n;
368 : }
369 254867 : n->prev = os->id_based_t; // aka the double linked list.
370 254867 : os->id_based_t = n;
371 254867 : os->id_based_cnt++;
372 254867 : unlock_writer(os);
373 254867 : return os;
374 : }
375 :
376 : static objectset *
377 254867 : os_append_id(objectset *os, objectversion *ov)
378 : {
379 254867 : versionhead *n = node_create(os->sa, ov);
380 :
381 254867 : if (n == NULL)
382 : return NULL;
383 254867 : ov->id_based_head = n;
384 254867 : if (!(os = os_append_node_id(os, n))){
385 0 : _DELETE(n);
386 0 : return NULL;
387 : }
388 :
389 : return os;
390 : }
391 :
392 : static versionhead * find_name(objectset *os, const char *name);
393 :
394 : static void
395 261776 : objectversion_destroy(sqlstore *store, objectset* os, objectversion *ov)
396 : {
397 :
398 261776 : bte state = os_atmc_get_state(ov);
399 :
400 261776 : if (state & name_based_versionhead_owner) {
401 15531 : node_destroy(ov->os, store, ov->name_based_head);
402 : }
403 :
404 261776 : if (state & id_based_versionhead_owner) {
405 15603 : node_destroy(ov->os, store, ov->id_based_head);
406 : }
407 :
408 261776 : if (os->destroy && ov->b)
409 236975 : os->destroy(store, ov->b);
410 :
411 261776 : if (os->temporary && (state & deleted || state & under_destruction || state & rollbacked))
412 168 : os_destroy(os, store); // TODO transaction_layer_revamp: embed into refcounting subproject : reference is already dropped by os_cleanup
413 261776 : _DELETE(ov);
414 261776 : }
415 :
416 : static void
417 6101 : _os_rollback(objectversion *ov, sqlstore *store)
418 : {
419 6101 : assert(ov->ts >= TRANSACTION_ID_BASE);
420 :
421 6101 : bte state = os_atmc_get_state(ov);
422 6101 : if (state & rollbacked) {
423 : return;
424 : }
425 :
426 5601 : state |= rollbacked;
427 5601 : os_atmc_set_state(ov, state);
428 :
429 5601 : bte state_older;
430 :
431 : /*
432 : * We have to use the readers-writer lock here,
433 : * since the pointer containing the address of the older objectversion might be concurrently overwritten if the older itself has just been put in the under_destruction state .
434 : */
435 5601 : lock_reader(ov->os);
436 5601 : objectversion* name_based_older = ov->name_based_older;
437 5601 : unlock_reader(ov->os);
438 :
439 5601 : if (name_based_older && !((state_older= os_atmc_get_state(name_based_older)) & rollbacked)) {
440 762 : if (ov->ts != name_based_older->ts) {
441 : // older is last committed state or belongs to parent transaction.
442 : // In any case, we restore versionhead pointer to that.
443 :
444 267 : ATOMIC_BASE_TYPE expected_deleted = deleted;
445 267 : if (state_older == active || (state_older == deleted && ATOMIC_CAS(&name_based_older->state, &expected_deleted, block_destruction))) {
446 267 : ov->name_based_head->ov = name_based_older;
447 267 : name_based_older->name_based_newer=NULL;
448 267 : if (state_older != active && expected_deleted == deleted)
449 0 : os_atmc_set_state(name_based_older, deleted); //Restore the deleted older back to its deleted state.
450 : }
451 : }
452 : else {
453 495 : _os_rollback(name_based_older, store);
454 : }
455 : }
456 4839 : else if (!name_based_older) {
457 : // this is a terminal node. i.e. this objectversion does not have name based committed history
458 4839 : if (ov->name_based_head) // The opposite can happen during an early conflict in os_add or os_del.
459 4833 : os_remove_name_based_chain(ov->os, ov);
460 : }
461 :
462 : /*
463 : * We have to use the readers-writer lock here,
464 : * since the pointer containing the address of the older objectversion might be concurrently overwritten if the older itself has just been put in the under_destruction state .
465 : */
466 5601 : lock_reader(ov->os);
467 5601 : objectversion* id_based_older = ov->id_based_older;
468 5601 : unlock_reader(ov->os);
469 5601 : if (id_based_older && !((state_older= os_atmc_get_state(id_based_older)) & rollbacked)) {
470 272 : if (ov->ts != id_based_older->ts) {
471 : // older is last committed state or belongs to parent transaction.
472 : // In any case, we restore versionhead pointer to that.
473 :
474 267 : ATOMIC_BASE_TYPE expected_deleted = deleted;
475 267 : if (state_older == active || (state_older == deleted && ATOMIC_CAS(&id_based_older->state, &expected_deleted, block_destruction))) {
476 267 : ov->id_based_head->ov = id_based_older;
477 267 : id_based_older->id_based_newer=NULL;
478 267 : if (state_older != active && expected_deleted == deleted)
479 0 : os_atmc_set_state(id_based_older, deleted); //Restore the deleted older back to its deleted state.
480 : }
481 : }
482 5 : else if (id_based_older != name_based_older)
483 5 : _os_rollback(id_based_older, store);
484 : }
485 4876 : else if (!id_based_older) {
486 : // this is a terminal node. i.e. this objectversion does not have id based committed history
487 4876 : os_remove_id_based_chain(ov->os, ov);
488 : }
489 :
490 5601 : if (ov->name_based_newer && !(os_atmc_get_state(ov->name_based_newer) & rollbacked)) {
491 0 : _os_rollback(ov->name_based_newer, store);
492 : }
493 :
494 5601 : if (ov->id_based_newer && ov->id_based_newer != ov->name_based_newer && !(os_atmc_get_state(ov->id_based_newer) & rollbacked)) {
495 0 : _os_rollback(ov->id_based_newer, store);
496 : }
497 : }
498 :
499 : static int
500 5601 : os_rollback(objectversion *ov, sqlstore *store)
501 : {
502 5601 : _os_rollback(ov, store);
503 :
504 5601 : return LOG_OK;
505 : }
506 :
507 : static void
508 24801 : ov_destroy_obj_recursive(sqlstore* store, objectversion *ov)
509 : {
510 24801 : if (ov->id_based_older && ov->id_based_older == ov->name_based_older) {
511 14217 : ov_destroy_obj_recursive(store, ov->id_based_older);
512 : }
513 24801 : if (ov->os->destroy && ov->b) {
514 24801 : ov->os->destroy(store, ov->b);
515 24801 : ov->b = NULL;
516 : }
517 24801 : }
518 :
519 : static inline void
520 10778 : try_to_mark_deleted_for_destruction(sqlstore* store, objectversion *ov)
521 : {
522 10778 : ATOMIC_BASE_TYPE expected_deleted = deleted;
523 10778 : if (ATOMIC_CAS(&ov->state, &expected_deleted, under_destruction)) {
524 :
525 10778 : if (!ov->name_based_newer || (os_atmc_get_state(ov->name_based_newer) & rollbacked)) {
526 10698 : os_remove_name_based_chain(ov->os, ov);
527 : }
528 : else {
529 80 : lock_writer(ov->os);
530 80 : ov->name_based_newer->name_based_older = NULL;
531 80 : unlock_writer(ov->os);
532 : }
533 :
534 10778 : if (!ov->id_based_newer || (os_atmc_get_state(ov->id_based_newer) & rollbacked)) {
535 10727 : os_remove_id_based_chain(ov->os, ov);
536 : }
537 : else {
538 51 : lock_writer(ov->os);
539 51 : ov->id_based_newer->id_based_older = NULL;
540 51 : unlock_writer(ov->os);
541 : }
542 :
543 10778 : ov->ts = store_get_timestamp(store)+1;
544 10778 : if (!ov->os->nested)
545 10584 : ov_destroy_obj_recursive(store, ov);
546 : }
547 10778 : }
548 :
549 : static void
550 29019 : objectversion_destroy_recursive(sqlstore* store, objectversion *ov)
551 : {
552 29019 : if (ov->id_based_older && ov->id_based_older == ov->name_based_older) {
553 14411 : objectversion_destroy_recursive(store, ov->id_based_older);
554 : }
555 29019 : objectversion_destroy(store, ov->os, ov);
556 29019 : }
557 :
558 : static int
559 301043 : os_cleanup(sqlstore* store, objectversion *ov, ulng oldest)
560 : {
561 301043 : if (os_atmc_get_state(ov) & under_destruction) {
562 14670 : if (ov->ts < oldest) {
563 : // This one is ready to be freed
564 10778 : objectversion_destroy_recursive(store, ov);
565 10778 : return LOG_ERR;
566 : }
567 :
568 : // not yet old enough to be safely removed. Try later.
569 : return LOG_OK;
570 : }
571 :
572 286373 : if (os_atmc_get_state(ov) & rollbacked) {
573 17507 : if (ov->ts < oldest) {
574 : // This one is ready to be freed
575 5601 : if (ov->name_based_older && ov->name_based_older->name_based_newer == ov)
576 495 : ov->name_based_older->name_based_newer=NULL;
577 5601 : if (ov->id_based_older && ov->id_based_older->id_based_newer == ov)
578 458 : ov->id_based_older->id_based_newer=NULL;
579 5601 : objectversion_destroy(store, ov->os, ov);
580 5601 : return LOG_ERR;
581 : }
582 :
583 11906 : if (ov->ts > TRANSACTION_ID_BASE) {
584 : /* We mark it with the latest possible starttime and reinsert it into the cleanup list.
585 : * This will cause a safe eventual destruction of this rollbacked ov.
586 : */
587 5601 : ov->ts = store_get_timestamp(store)+1;
588 : }
589 :
590 : // not yet old enough to be safely removed. Try later.
591 11906 : return LOG_OK;
592 : }
593 :
594 268866 : if (os_atmc_get_state(ov) == deleted) {
595 10870 : if (ov->ts <= oldest) {
596 : // the oldest relevant state is deleted so lets try to mark it as destroyed
597 10778 : try_to_mark_deleted_for_destruction(store, ov);
598 10778 : return LOG_OK+1;
599 : }
600 :
601 : // Keep it inplace on the cleanup list, either because it is now marked for destruction or
602 : // we want to retry marking it for destruction later.
603 : return LOG_OK;
604 : }
605 :
606 257996 : assert(os_atmc_get_state(ov) != deleted && os_atmc_get_state(ov) != under_destruction && os_atmc_get_state(ov) != rollbacked);
607 257996 : if (ov->os->temporary) os_destroy(ov->os, store); // TODO transaction_layer_revamp: embed into refcounting subproject: (old) live versions should drop their reference to the os
608 :
609 265950 : while (ov->id_based_older && ov->id_based_older == ov->name_based_older && ov->ts >= oldest) {
610 : ov = ov->id_based_older;
611 : }
612 :
613 257996 : if (ov->id_based_older && ov->id_based_older == ov->name_based_older) {
614 : // Destroy everything older then the oldest possibly relevant objectversion.
615 3830 : objectversion_destroy_recursive(store, ov->id_based_older);
616 3830 : ov->id_based_older = NULL;
617 : }
618 :
619 : return LOG_ERR;
620 : }
621 :
622 : static int
623 301050 : tc_gc_objectversion(sql_store store, sql_change *change, ulng oldest)
624 : {
625 : // assert(!change->handled);
626 301050 : objectversion *ov = (objectversion*)change->data;
627 :
628 301050 : if (oldest >= TRANSACTION_ID_BASE)
629 : return 0;
630 301043 : int res = os_cleanup( (sqlstore*) store, ov, oldest);
631 301043 : change->handled = (res)?true:false;
632 301043 : return res>=0?LOG_OK:LOG_ERR;
633 : }
634 :
635 : static int
636 274382 : tc_commit_objectversion(sql_trans *tr, sql_change *change, ulng commit_ts, ulng oldest)
637 : {
638 274382 : objectversion *ov = (objectversion*)change->data;
639 274382 : if (commit_ts) {
640 268781 : assert(ov->ts == tr->tid);
641 268781 : ov->ts = commit_ts;
642 268781 : change->committed = commit_ts < TRANSACTION_ID_BASE ? true: false;
643 268781 : (void)oldest;
644 268781 : if (!tr->parent)
645 268774 : change->obj->new = 0;
646 268781 : if (!ov->os->temporary)
647 264388 : ATOMIC_INC(&tr->cat->schema_version);
648 : } else {
649 5601 : os_rollback(ov, tr->store);
650 : }
651 :
652 274382 : return LOG_OK;
653 : }
654 :
655 : objectset *
656 58335 : os_new(allocator *sa, destroy_fptr destroy, bool temporary, bool unique, bool concurrent, bool nested, sql_store store)
657 : {
658 58335 : assert(!sa);
659 58335 : objectset *os = SA_NEW(sa, objectset);
660 58335 : if (os) {
661 58335 : *os = (objectset) {
662 : .refcnt = ATOMIC_VAR_INIT(1),
663 : .sa = sa,
664 : .destroy = destroy,
665 : .temporary = temporary,
666 : .unique = unique,
667 : .concurrent = concurrent,
668 : .nested = nested,
669 : .store = store
670 : };
671 58335 : os->destroy = destroy;
672 58335 : MT_rwlock_init(&os->rw_lock, "sa_readers_lock");
673 58335 : MT_lock_init(&os->lock, "single_writer_lock");
674 : }
675 :
676 58335 : return os;
677 : }
678 :
679 : objectset *
680 4632 : os_dup(objectset *os)
681 : {
682 4632 : ATOMIC_INC(&os->refcnt);
683 4632 : return os;
684 : }
685 :
686 : void
687 61869 : os_destroy(objectset *os, sql_store store)
688 : {
689 61869 : if (ATOMIC_DEC(&os->refcnt) > 0)
690 : return;
691 57237 : MT_lock_destroy(&os->lock);
692 57237 : MT_rwlock_destroy(&os->rw_lock);
693 57237 : versionhead* n=os->id_based_h;
694 283902 : while(n) {
695 226665 : objectversion *ov = n->ov;
696 453821 : while(ov) {
697 227156 : objectversion *older = ov->id_based_older;
698 227156 : objectversion_destroy(store, os, ov);
699 227156 : ov = older;
700 : }
701 226665 : versionhead* hn =n->next;
702 226665 : node_destroy(os, store, n);
703 226665 : n = hn;
704 : }
705 :
706 57237 : n=os->name_based_h;
707 283902 : while(n) {
708 226665 : versionhead* hn =n->next;
709 226665 : node_destroy(os, store, n);
710 226665 : n = hn;
711 : }
712 :
713 57237 : if (os->id_map)
714 3113 : hash_destroy(os->id_map);
715 :
716 57237 : if (os->name_map)
717 3087 : hash_destroy(os->name_map);
718 :
719 57237 : if (!os->sa)
720 57237 : _DELETE(os);
721 : }
722 :
723 : static versionhead *
724 14955228 : find_name(objectset *os, const char *name)
725 : {
726 14955228 : lock_reader(os);
727 14951034 : if (os->name_map) {
728 11502597 : int key = hash_key(name);
729 11502597 : sql_hash_e *he = os->name_map->buckets[key&(os->name_map->size-1)];
730 :
731 57398322 : for (; he; he = he->chain) {
732 57034408 : versionhead *n = he->value;
733 :
734 57034408 : if (n && n->ov->b->name && strcmp(n->ov->b->name, name) == 0) {
735 11138683 : unlock_reader(os);
736 11138683 : return n;
737 : }
738 : }
739 363914 : unlock_reader(os);
740 363914 : return NULL;
741 : }
742 :
743 4128841 : for (versionhead *n = os->name_based_h; n; n = n->next) {
744 3828639 : objectversion *ov = n->ov;
745 :
746 : /* check if names match */
747 3828639 : if (name[0] == ov->b->name[0] && strcmp(name, ov->b->name) == 0) {
748 3148235 : unlock_reader(os);
749 3148235 : return n;
750 : }
751 : }
752 :
753 300202 : unlock_reader(os);
754 300202 : return NULL;
755 : }
756 :
757 : static objectversion*
758 15869211 : get_valid_object_name(sql_trans *tr, objectversion *ov)
759 : {
760 15869393 : while(ov) {
761 15869361 : if (ov->ts == tr->tid || (tr->parent && tr_version_of_parent(tr, ov->ts)) || ov->ts < tr->ts)
762 15869116 : return ov;
763 : else {
764 245 : lock_reader(ov->os);
765 170 : objectversion* name_based_older = ov->name_based_older;
766 170 : unlock_reader(ov->os);
767 170 : ov = name_based_older;
768 : }
769 : }
770 : return ov;
771 : }
772 :
773 : static objectversion*
774 499543 : get_valid_object_id(sql_trans *tr, objectversion *ov)
775 : {
776 499724 : while(ov) {
777 499603 : if (ov->ts == tr->tid || (tr->parent && tr_version_of_parent(tr, ov->ts)) || ov->ts < tr->ts)
778 499422 : return ov;
779 : else {
780 181 : lock_reader(ov->os);
781 181 : objectversion* id_based_older = ov->id_based_older;
782 181 : unlock_reader(ov->os);
783 181 : ov = id_based_older;
784 : }
785 : }
786 : return ov;
787 : }
788 :
789 : static int
790 263120 : os_add_name_based(objectset *os, struct sql_trans *tr, const char *name, objectversion *ov) {
791 263120 : versionhead *name_based_node = NULL;
792 263120 : if (ov->id_based_older && strcmp(ov->id_based_older->b->name, name) == 0)
793 8229 : name_based_node = ov->id_based_older->name_based_head;
794 254891 : else if (os->unique) // Previous name based objectversion is of a different id, so now we do have to perform an extensive look up
795 78504 : name_based_node = find_name(os, name);
796 : // else names are not unique and each id based version head maps to its own name based version head.
797 :
798 86733 : if (name_based_node) {
799 8325 : objectversion *co = name_based_node->ov;
800 8325 : objectversion *oo = get_valid_object_name(tr, co);
801 8325 : if (co != oo) { /* conflict ? */
802 6 : TRC_WARNING(SQL_STORE, "%s" "if (co != oo) { /* conflict ? */", __func__);
803 6 : return -3;
804 : }
805 :
806 8319 : assert(ov != oo); // Time loops are not allowed
807 :
808 8319 : bte state = os_atmc_get_state(oo);
809 8319 : if (state != active) {
810 : // This can only happen if the parent oo was a committed deleted at some point.
811 123 : assert(state == deleted || state == under_destruction || state == block_destruction);
812 : /* Since our parent oo is committed deleted objectversion, we might have a conflict with
813 : * another transaction that tries to clean up oo or also wants to add a new objectversion.
814 : */
815 123 : ATOMIC_BASE_TYPE expected_deleted = deleted;
816 123 : if (!ATOMIC_CAS(&oo->state, &expected_deleted, block_destruction)) {
817 0 : TRC_WARNING(SQL_STORE, "%s: " "if (!ATOMIC_CAS(&oo->state, &expected_deleted, block_destruction)) { /*conflict with cleaner or write-write conflict*/ ", __func__);
818 0 : return -3; /*conflict with cleaner or write-write conflict*/
819 : }
820 : }
821 :
822 : /* new object with same name within transaction, should have a delete in between */
823 8196 : assert(!(state == active && oo->ts == ov->ts && !(os_atmc_get_state(ov) & deleted)));
824 :
825 8319 : lock_writer(os);
826 8319 : ov->name_based_head = oo->name_based_head;
827 8319 : ov->name_based_older = oo;
828 :
829 8319 : name_based_node->ov = ov;
830 8319 : if (oo) {
831 8319 : oo->name_based_newer = ov;
832 : // if the parent was originally deleted, we restore it to that state.
833 8319 : os_atmc_set_state(oo, state);
834 : }
835 8319 : unlock_writer(os);
836 8319 : return 0;
837 : } else { /* new */
838 254795 : if (os_append_name(os, ov) == NULL)
839 : return -1; // MALLOC_FAIL
840 : return 0;
841 : }
842 : }
843 :
844 : static int
845 263130 : os_add_id_based(objectset *os, struct sql_trans *tr, sqlid id, objectversion *ov) {
846 263130 : versionhead *id_based_node;
847 :
848 263130 : id_based_node = find_id(os, id);
849 :
850 263130 : if (id_based_node) {
851 8263 : objectversion *co = id_based_node->ov;
852 8263 : objectversion *oo = get_valid_object_id(tr, co);
853 8263 : if (co != oo) { /* conflict ? */
854 10 : TRC_WARNING(SQL_STORE, "%s" "if (co != oo) { /* conflict ? */", __func__);
855 10 : return -3;
856 : }
857 :
858 8253 : assert(ov != oo); // Time loops are not allowed
859 :
860 8253 : bte state = os_atmc_get_state(oo);
861 8253 : if (state != active) {
862 : // This can only happen if the parent oo was a committed deleted at some point.
863 57 : assert(state == deleted || state == under_destruction || state == block_destruction);
864 : /* Since our parent oo is committed deleted objectversion, we might have a conflict with
865 : * another transaction that tries to clean up oo or also wants to add a new objectversion.
866 : */
867 57 : ATOMIC_BASE_TYPE expected_deleted = deleted;
868 57 : if (!ATOMIC_CAS(&oo->state, &expected_deleted, block_destruction)) {
869 0 : TRC_WARNING(SQL_STORE, "%s" "if (!ATOMIC_CAS(&oo->state, &expected_deleted, block_destruction)) { /*conflict with cleaner or write-write conflict*/", __func__);
870 0 : return -3; /*conflict with cleaner or write-write conflict*/
871 : }
872 : }
873 :
874 8253 : lock_writer(os);
875 8253 : ov->id_based_head = oo->id_based_head;
876 8253 : ov->id_based_older = oo;
877 :
878 8253 : id_based_node->ov = ov;
879 8253 : if (oo) {
880 8253 : oo->id_based_newer = ov;
881 : // if the parent was originally deleted, we restore it to that state.
882 8253 : os_atmc_set_state(oo, state);
883 : }
884 8253 : unlock_writer(os);
885 8253 : return 0;
886 : } else { /* new */
887 254867 : if (os_append_id(os, ov) == NULL)
888 : return -1; // MALLOC_FAIL
889 :
890 : return 0;
891 : }
892 : }
893 :
894 : static int /*ok, error (name existed) and conflict (added before) */
895 263131 : os_add_(objectset *os, struct sql_trans *tr, const char *name, sql_base *b)
896 : {
897 263131 : int res = 0;
898 263131 : objectversion *ov = SA_NEW(os->sa, objectversion);
899 :
900 263131 : *ov = (objectversion) {
901 263131 : .ts = tr->tid,
902 : .b = b,
903 : .os = os,
904 : };
905 :
906 263131 : if (!os->concurrent && os_has_changes(os, tr)) { /* for object sets without concurrent support, conflict if concurrent changes are there */
907 1 : if (os->destroy)
908 1 : os->destroy(os->store, ov->b);
909 1 : _DELETE(ov);
910 1 : TRC_WARNING(SQL_STORE, "%s" "if (!os->concurrent && os_has_changes(os, tr)) { /* for object sets without concurrent support, conflict if concurrent changes are there */", __func__);
911 1 : return -3; /* conflict */
912 : }
913 :
914 263130 : if ((res = os_add_id_based(os, tr, b->id, ov))) {
915 10 : if (os->destroy)
916 10 : os->destroy(os->store, ov->b);
917 10 : _DELETE(ov);
918 10 : return res;
919 : }
920 :
921 263120 : if ((res = os_add_name_based(os, tr, name, ov))) {
922 6 : trans_add(tr, b, ov, &tc_gc_objectversion, &tc_commit_objectversion, NULL);
923 6 : return res;
924 : }
925 :
926 263114 : if (os->temporary) (void) os_dup(os); // TODO transaction_layer_revamp: embed into refcounting subproject
927 263114 : trans_add(tr, b, ov, &tc_gc_objectversion, &tc_commit_objectversion, NULL);
928 263114 : return res;
929 : }
930 :
931 : int
932 263131 : os_add(objectset *os, struct sql_trans *tr, const char *name, sql_base *b)
933 : {
934 263131 : MT_lock_set(&os->lock);
935 263131 : int res = os_add_(os, tr, name, b);
936 263131 : MT_lock_unset(&os->lock);
937 263131 : return res;
938 : }
939 :
940 : static int
941 11255 : os_del_name_based(objectset *os, struct sql_trans *tr, const char *name, objectversion *ov) {
942 11255 : versionhead *name_based_node = NULL;
943 11255 : if (ov->id_based_older && strcmp(ov->id_based_older->b->name, name) == 0)
944 11255 : name_based_node = ov->id_based_older->name_based_head;
945 0 : else if (os->unique) // Previous name based objectversion is of a different id, so now we do have to perform an extensive look up
946 0 : name_based_node = find_name(os, name);
947 :
948 11255 : if (name_based_node) {
949 11255 : objectversion *co = name_based_node->ov;
950 11255 : objectversion *oo = get_valid_object_name(tr, co);
951 11255 : ov->name_based_head = oo->name_based_head;
952 11255 : if (co != oo) { /* conflict ? */
953 0 : TRC_WARNING(SQL_STORE, "%s: " "if (co != oo) { /* conflict ? */", __func__);
954 0 : return -3;
955 : }
956 11255 : ov->name_based_older = oo;
957 :
958 11255 : lock_writer(os);
959 11255 : if (oo) {
960 11255 : oo->name_based_newer = ov;
961 11255 : assert(os_atmc_get_state(oo) == active);
962 : }
963 11255 : name_based_node->ov = ov;
964 11255 : unlock_writer(os);
965 11255 : return 0;
966 : } else {
967 : /* missing */
968 0 : return -1;
969 : }
970 : }
971 :
972 : static int
973 11257 : os_del_id_based(objectset *os, struct sql_trans *tr, sqlid id, objectversion *ov) {
974 :
975 11257 : versionhead *id_based_node;
976 :
977 11257 : if (ov->name_based_older && ov->name_based_older->b->id == id)
978 0 : id_based_node = ov->name_based_older->id_based_head;
979 : else // Previous id based objectversion is of a different name, so now we do have to perform an extensive look up
980 11257 : id_based_node = find_id(os, id);
981 :
982 11257 : if (id_based_node) {
983 11257 : objectversion *co = id_based_node->ov;
984 11257 : objectversion *oo = get_valid_object_id(tr, co);
985 11257 : ov->id_based_head = oo->id_based_head;
986 11257 : if (co != oo) { /* conflict ? */
987 2 : TRC_WARNING(SQL_STORE, "%s" "if (co != oo) { /* conflict ? */", __func__);
988 2 : return -3;
989 : }
990 11255 : ov->id_based_older = oo;
991 :
992 11255 : lock_writer(os);
993 11255 : if (oo) {
994 11255 : oo->id_based_newer = ov;
995 11255 : assert(os_atmc_get_state(oo) == active);
996 : }
997 11255 : id_based_node->ov = ov;
998 11255 : unlock_writer(os);
999 11255 : return 0;
1000 : } else {
1001 : /* missing */
1002 : return -1;
1003 : }
1004 : }
1005 :
1006 : static int
1007 11257 : os_del_(objectset *os, struct sql_trans *tr, const char *name, sql_base *b)
1008 : {
1009 11257 : int res = 0;
1010 11257 : objectversion *ov = SA_NEW(os->sa, objectversion);
1011 :
1012 11257 : *ov = (objectversion) {
1013 11257 : .ts = tr->tid,
1014 : .b = b,
1015 : .os = os,
1016 : };
1017 11257 : os_atmc_set_state(ov, deleted);
1018 :
1019 11257 : if ((res = os_del_id_based(os, tr, b->id, ov))) {
1020 2 : if (os->destroy)
1021 2 : os->destroy(os->store, ov->b);
1022 2 : _DELETE(ov);
1023 2 : return res;
1024 : }
1025 :
1026 11255 : if ((res = os_del_name_based(os, tr, name, ov))) {
1027 0 : trans_add(tr, b, ov, &tc_gc_objectversion, &tc_commit_objectversion, NULL);
1028 0 : return res;
1029 : }
1030 :
1031 11255 : if (os->temporary) (void) os_dup(os); // TODO transaction_layer_revamp: embed into refcounting subproject
1032 11255 : trans_add(tr, b, ov, &tc_gc_objectversion, &tc_commit_objectversion, NULL);
1033 11255 : return res;
1034 : }
1035 :
1036 : int
1037 11257 : os_del(objectset *os, sql_trans *tr, const char *name, sql_base *b)
1038 : {
1039 11257 : MT_lock_set(&os->lock);
1040 11257 : int res = os_del_(os, tr, name, b);
1041 11257 : MT_lock_unset(&os->lock);
1042 11257 : return res;
1043 : }
1044 :
1045 : int
1046 315 : os_size(objectset *os, struct sql_trans *tr)
1047 : {
1048 315 : int cnt = 0;
1049 315 : if (os) {
1050 315 : lock_reader(os);
1051 324 : for(versionhead *n = os->name_based_h; n; n=n->next) {
1052 9 : objectversion *ov = n->ov;
1053 9 : if ((ov=get_valid_object_name(tr, ov)) && os_atmc_get_state(ov) == active)
1054 6 : cnt++;
1055 : }
1056 315 : unlock_reader(os);
1057 : }
1058 315 : return cnt;
1059 : }
1060 :
1061 : int
1062 0 : os_empty(objectset *os, struct sql_trans *tr)
1063 : {
1064 0 : return os_size(os, tr)==0;
1065 : }
1066 :
1067 : int
1068 0 : os_remove(objectset *os, sql_trans *tr, const char *name)
1069 : {
1070 0 : (void) os;
1071 0 : (void) tr;
1072 0 : (void) name;
1073 : // TODO remove entire versionhead corresponding to this name.
1074 :
1075 : // TODO assert os->unique?s
1076 0 : return LOG_OK;
1077 : }
1078 :
1079 : sql_base *
1080 14877774 : os_find_name(objectset *os, struct sql_trans *tr, const char *name)
1081 : {
1082 14877774 : if (!os)
1083 : return NULL;
1084 :
1085 14877772 : assert(os->unique);
1086 14877772 : versionhead *n = find_name(os, name);
1087 :
1088 14874498 : if (n) {
1089 14288793 : objectversion *ov = get_valid_object_name(tr, n->ov);
1090 14287292 : if (ov && os_atmc_get_state(ov) == active)
1091 14287185 : return ov->b;
1092 : }
1093 : return NULL;
1094 : }
1095 :
1096 : sql_base *
1097 154745 : os_find_id(objectset *os, struct sql_trans *tr, sqlid id)
1098 : {
1099 154745 : if (!os)
1100 : return NULL;
1101 154745 : versionhead *n = find_id(os, id);
1102 :
1103 154745 : if (n) {
1104 154224 : objectversion *ov = get_valid_object_id(tr, n->ov);
1105 154224 : if (ov && os_atmc_get_state(ov) == active)
1106 154220 : return ov->b;
1107 : }
1108 : return NULL;
1109 : }
1110 :
1111 : void
1112 2206307 : os_iterator(struct os_iter *oi, struct objectset *os, struct sql_trans *tr, const char *name /*optional*/)
1113 : {
1114 2206307 : *oi = (struct os_iter) {
1115 : .os = os,
1116 : .tr = tr,
1117 : .name = name,
1118 : };
1119 :
1120 2206307 : lock_reader(os);
1121 2206314 : if (name && os->name_map) {
1122 1021045 : int key = hash_key(name);
1123 1021045 : oi->n = (void*)os->name_map->buckets[key&(os->name_map->size-1)];
1124 : } else
1125 1185269 : oi->n = os->name_based_h;
1126 2206314 : unlock_reader(os);
1127 2206312 : }
1128 :
1129 : sql_base *
1130 4011342 : oi_next(struct os_iter *oi)
1131 : {
1132 4011342 : sql_base *b = NULL;
1133 :
1134 4011342 : if (oi->name) {
1135 2925649 : lock_reader(oi->os); /* intentionally outside of while loop */
1136 2925649 : if (oi->os->name_map) {
1137 2565093 : sql_hash_e *he = (void*)oi->n;
1138 :
1139 6351046 : for (; he && !b; he = he->chain) {
1140 3785953 : versionhead *n = he->value;
1141 :
1142 3785953 : if (n->ov->b->name && strcmp(n->ov->b->name, oi->name) == 0) {
1143 1549249 : objectversion *ov = n->ov;
1144 :
1145 1549249 : ov = get_valid_object_name(oi->tr, ov);
1146 1549249 : if (ov && os_atmc_get_state(ov) == active)
1147 1549059 : b = ov->b;
1148 : }
1149 : }
1150 2565093 : oi->n = (void*)he;
1151 : } else {
1152 360556 : versionhead *n = oi->n;
1153 :
1154 406543 : while (n && !b) {
1155 :
1156 45987 : if (n->ov->b->name && strcmp(n->ov->b->name, oi->name) == 0) {
1157 11970 : objectversion *ov = n->ov;
1158 :
1159 11970 : n = oi->n = n->next;
1160 11970 : ov = get_valid_object_name(oi->tr, ov);
1161 11970 : if (ov && os_atmc_get_state(ov) == active)
1162 11970 : b = ov->b;
1163 : } else {
1164 34017 : n = oi->n = n->next;
1165 : }
1166 : }
1167 : }
1168 2925649 : unlock_reader(oi->os);
1169 : } else {
1170 1085693 : versionhead *n = oi->n;
1171 :
1172 1085693 : lock_reader(oi->os); /* intentionally outside of while loop */
1173 2497188 : while (n && !b) {
1174 325799 : objectversion *ov = n->ov;
1175 325799 : n = oi->n = n->next;
1176 :
1177 325799 : ov = get_valid_object_id(oi->tr, ov);
1178 325799 : if (ov && os_atmc_get_state(ov) == active)
1179 301350 : b = ov->b;
1180 : }
1181 1085696 : unlock_reader(oi->os);
1182 : }
1183 4011330 : return b;
1184 : }
1185 :
1186 : bool
1187 3545 : os_obj_intransaction(objectset *os, struct sql_trans *tr, sql_base *b)
1188 : {
1189 3545 : versionhead *n = find_id(os, b->id);
1190 :
1191 3545 : if (n) {
1192 3545 : objectversion *ov = n->ov;
1193 :
1194 3545 : if (ov && os_atmc_get_state(ov) == active && ov->ts == tr->tid)
1195 : return true;
1196 : }
1197 : return false;
1198 : }
1199 :
1200 : /* return true if this object set has changes pending for an other transaction */
1201 : bool
1202 171418 : os_has_changes(objectset *os, struct sql_trans *tr)
1203 : {
1204 171418 : versionhead *n = os->id_based_t;
1205 :
1206 171418 : if (n) {
1207 167152 : objectversion *ov = n->ov;
1208 :
1209 167152 : if (ov && os_atmc_get_state(ov) == active && ov->ts != tr->tid && ov->ts > TRANSACTION_ID_BASE)
1210 : return true;
1211 : }
1212 : return false;
1213 : }
|
