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 : sql_allocator *sa;
56 : destroy_fptr destroy;
57 : MT_RWLock rw_lock; /*readers-writer lock to protect the links (chains) in the objectversion chain.*/
58 : versionhead *name_based_h;
59 : versionhead *name_based_t;
60 : versionhead *id_based_h;
61 : versionhead *id_based_t;
62 : int name_based_cnt;
63 : int id_based_cnt;
64 : struct sql_hash *name_map;
65 : struct sql_hash *id_map;
66 : bool
67 : temporary:1,
68 : unique:1, /* names are unique */
69 : concurrent:1, /* concurrent inserts are allowed */
70 : nested:1;
71 : sql_store store;
72 : } objectset;
73 :
74 : static int
75 324232 : os_id_key(versionhead *n)
76 : {
77 250575 : return (int) BATatoms[TYPE_int].atomHash(&n->ov->b->id);
78 : }
79 :
80 : static inline void
81 23420112 : lock_reader(objectset* os)
82 : {
83 23420112 : MT_rwlock_rdlock(&os->rw_lock);
84 4222200 : }
85 :
86 : static inline void
87 23417335 : unlock_reader(objectset* os)
88 : {
89 23417335 : MT_rwlock_rdunlock(&os->rw_lock);
90 4252280 : }
91 :
92 : static inline void
93 592510 : lock_writer(objectset* os)
94 : {
95 592510 : MT_rwlock_wrlock(&os->rw_lock);
96 : }
97 :
98 : static inline void
99 592510 : unlock_writer(objectset* os)
100 : {
101 592510 : MT_rwlock_wrunlock(&os->rw_lock);
102 82 : }
103 :
104 19528971 : static bte os_atmc_get_state(objectversion *ov) {
105 19528971 : bte state = (bte) ATOMIC_GET(&ov->state);
106 19528971 : return state;
107 : }
108 :
109 63411 : static void os_atmc_set_state(objectversion *ov, bte state) {
110 63411 : ATOMIC_SET(&ov->state, state);
111 0 : }
112 :
113 : static versionhead *
114 430228 : find_id(objectset *os, sqlid id)
115 : {
116 430228 : if (os) {
117 430228 : lock_reader(os);
118 430228 : if (os->id_map) {
119 397984 : int key = (int) BATatoms[TYPE_int].atomHash(&id);
120 397984 : sql_hash_e *he = os->id_map->buckets[key&(os->id_map->size-1)];
121 :
122 1398479 : for (; he; he = he->chain) {
123 1165871 : versionhead *n = he->value;
124 :
125 1165871 : if (n && n->ov->b->id == id) {
126 165376 : unlock_reader(os);
127 165376 : return n;
128 : }
129 : }
130 232608 : unlock_reader(os);
131 232608 : return NULL;
132 : }
133 :
134 97980 : for (versionhead *n = os->id_based_h; n; n = n->next) {
135 68241 : objectversion *ov = n->ov;
136 :
137 : /* check if ids match */
138 68241 : if (id == ov->b->id) {
139 2505 : unlock_reader(os);
140 2505 : return n;
141 : }
142 : }
143 29739 : unlock_reader(os);
144 : }
145 :
146 : return NULL;
147 : }
148 :
149 : // TODO copy of static function from sql_list.c. Needs to be made external
150 : static void
151 28842 : hash_delete(sql_hash *h, void *data)
152 : {
153 28842 : int key = h->key(data);
154 28842 : sql_hash_e *e, *p = h->buckets[key&(h->size-1)];
155 :
156 28842 : e = p;
157 42016 : for (; p && p->value != data ; p = p->chain)
158 13174 : e = p;
159 28842 : if (p && p->value == data) {
160 28842 : if (p == e)
161 24764 : h->buckets[key&(h->size-1)] = p->chain;
162 : else
163 4078 : e->chain = p->chain;
164 28842 : if (!h->sa)
165 28842 : _DELETE(p);
166 : }
167 28842 : h->entries--;
168 28842 : }
169 :
170 : static void
171 522638 : node_destroy(objectset *os, sqlstore *store, versionhead *n)
172 : {
173 522638 : if (!os->sa)
174 522638 : _DELETE(n);
175 522638 : (void)store;
176 522638 : }
177 :
178 : static versionhead *
179 15182 : os_remove_name_based_chain(objectset *os, objectversion* ov)
180 : {
181 15182 : lock_writer(os);
182 15182 : versionhead *n = ov->name_based_head;
183 15182 : versionhead *p = os->name_based_h;
184 15182 : if (p != n)
185 1136996 : while (p && p->next != n)
186 : p = p->next;
187 15182 : assert(p==n||(p && p->next == n));
188 15182 : if (p == n) {
189 1335 : os->name_based_h = n->next;
190 1335 : if (os->name_based_h) // i.e. non-empty os
191 810 : os->name_based_h->prev = NULL;
192 : p = NULL;
193 13847 : } else if ( p != NULL) {
194 13847 : p->next = n->next;
195 13847 : if (p->next) // node in the middle
196 5186 : p->next->prev = p;
197 : }
198 15182 : if (n == os->name_based_t)
199 9186 : os->name_based_t = p;
200 :
201 15182 : if (os->name_map && n)
202 14140 : hash_delete(os->name_map, n);
203 :
204 15182 : os->name_based_cnt--;
205 15182 : unlock_writer(os);
206 :
207 15182 : bte state = os_atmc_get_state(ov);
208 15182 : state |= name_based_versionhead_owner;
209 15182 : os_atmc_set_state(ov, state);
210 15182 : return p;
211 : }
212 :
213 : static versionhead *
214 15248 : os_remove_id_based_chain(objectset *os, objectversion* ov)
215 : {
216 15248 : lock_writer(os);
217 15248 : versionhead *n = ov->id_based_head;
218 15248 : versionhead *p = os->id_based_h;
219 :
220 15248 : if (p != n)
221 1144605 : while (p && p->next != n)
222 : p = p->next;
223 15248 : assert(p==n||(p && p->next == n));
224 15248 : if (p == n) {
225 1335 : os->id_based_h = n->next;
226 1335 : if (os->id_based_h) // i.e. non-empty os
227 810 : os->id_based_h->prev = NULL;
228 : p = NULL;
229 13913 : } else if ( p != NULL) {
230 13913 : p->next = n->next;
231 13913 : if (p->next) // node in the middle
232 5253 : p->next->prev = p;
233 : }
234 15248 : if (n == os->id_based_t)
235 9185 : os->id_based_t = p;
236 :
237 15248 : if (os->id_map && n)
238 14702 : hash_delete(os->id_map, n);
239 :
240 15248 : os->name_based_cnt--;
241 15248 : unlock_writer(os);
242 :
243 15248 : bte state = os_atmc_get_state(ov);
244 15248 : state |= id_based_versionhead_owner;
245 15248 : os_atmc_set_state(ov, state);
246 15248 : return p;
247 : }
248 :
249 : static versionhead *
250 523748 : node_create(sql_allocator *sa, objectversion *ov)
251 : {
252 523748 : versionhead *n = SA_NEW(sa, versionhead );
253 :
254 523748 : if (n == NULL)
255 : return NULL;
256 523748 : *n = (versionhead ) {
257 : .ov = ov,
258 : };
259 523748 : return n;
260 : }
261 :
262 : static inline int
263 315752 : os_name_key(versionhead *n)
264 : {
265 315752 : return hash_key(n->ov->b->name);
266 : }
267 :
268 : static objectset *
269 261841 : os_append_node_name(objectset *os, versionhead *n)
270 : {
271 261841 : lock_writer(os);
272 261841 : if ((!os->name_map || os->name_map->size*16 < os->name_based_cnt) && os->name_based_cnt > HASH_MIN_SIZE) {
273 3882 : hash_destroy(os->name_map);
274 3882 : os->name_map = hash_new(os->sa, os->name_based_cnt, (fkeyvalue)& os_name_key);
275 3882 : if (os->name_map == NULL) {
276 0 : unlock_writer(os);
277 0 : return NULL;
278 : }
279 :
280 70178 : for (versionhead *n = os->name_based_h; n; n = n->next ) {
281 66296 : int key = os_name_key(n);
282 :
283 66296 : if (hash_add(os->name_map, key, n) == NULL) {
284 0 : unlock_writer(os);
285 0 : return NULL;
286 : }
287 : }
288 : }
289 :
290 261841 : if (os->name_map) {
291 235316 : int key = os->name_map->key(n);
292 :
293 235316 : if (hash_add(os->name_map, key, n) == NULL) {
294 0 : unlock_writer(os);
295 0 : return NULL;
296 : }
297 : }
298 :
299 261841 : if (os->name_based_t) {
300 253606 : os->name_based_t->next = n;
301 : } else {
302 8235 : os->name_based_h = n;
303 : }
304 261841 : n->prev = os->name_based_t; // aka the double linked list.
305 261841 : os->name_based_t = n;
306 261841 : os->name_based_cnt++;
307 261841 : unlock_writer(os);
308 261841 : return os;
309 : }
310 :
311 : static objectset *
312 261841 : os_append_name(objectset *os, objectversion *ov)
313 : {
314 261841 : versionhead *n = node_create(os->sa, ov);
315 :
316 261841 : if (n == NULL)
317 : return NULL;
318 :
319 261841 : ov->name_based_head = n;
320 261841 : if (!(os = os_append_node_name(os, n))){
321 0 : _DELETE(n);
322 0 : return NULL;
323 : }
324 :
325 : return os;
326 : }
327 :
328 : static void
329 3973 : os_append_id_map(objectset *os)
330 : {
331 3973 : if (os->id_map)
332 442 : hash_destroy(os->id_map);
333 3973 : os->id_map = hash_new(os->sa, os->id_based_cnt, (fkeyvalue)&os_id_key);
334 3973 : if (os->id_map == NULL)
335 : return ;
336 77630 : for (versionhead *n = os->id_based_h; n; n = n->next ) {
337 73657 : int key = os_id_key(n);
338 :
339 73657 : if (hash_add(os->id_map, key, n) == NULL) {
340 0 : hash_destroy(os->id_map);
341 0 : os->id_map = NULL;
342 0 : return ;
343 : }
344 : }
345 : }
346 :
347 : static objectset *
348 261907 : os_append_node_id(objectset *os, versionhead *n)
349 : {
350 261907 : lock_writer(os);
351 261907 : if ((!os->id_map || os->id_map->size*16 < os->id_based_cnt) && os->id_based_cnt > HASH_MIN_SIZE)
352 3973 : os_append_id_map(os); /* on failure just fall back to slow method */
353 :
354 261907 : if (os->id_map) {
355 235873 : int key = os->id_map->key(n);
356 235873 : if (hash_add(os->id_map, key, n) == NULL) {
357 0 : hash_destroy(os->id_map);
358 0 : os->id_map = NULL;
359 : /* fall back to slow search */
360 : }
361 : }
362 :
363 261907 : if (os->id_based_t) {
364 253672 : os->id_based_t->next = n;
365 : } else {
366 8235 : os->id_based_h = n;
367 : }
368 261907 : n->prev = os->id_based_t; // aka the double linked list.
369 261907 : os->id_based_t = n;
370 261907 : os->id_based_cnt++;
371 261907 : unlock_writer(os);
372 261907 : return os;
373 : }
374 :
375 : static objectset *
376 261907 : os_append_id(objectset *os, objectversion *ov)
377 : {
378 261907 : versionhead *n = node_create(os->sa, ov);
379 :
380 261907 : if (n == NULL)
381 : return NULL;
382 261907 : ov->id_based_head = n;
383 261907 : if (!(os = os_append_node_id(os, n))){
384 0 : _DELETE(n);
385 0 : return NULL;
386 : }
387 :
388 : return os;
389 : }
390 :
391 : static versionhead * find_name(objectset *os, const char *name);
392 :
393 : static void
394 280447 : objectversion_destroy(sqlstore *store, objectset* os, objectversion *ov)
395 : {
396 :
397 280447 : bte state = os_atmc_get_state(ov);
398 :
399 280447 : if (state & name_based_versionhead_owner) {
400 15182 : node_destroy(ov->os, store, ov->name_based_head);
401 : }
402 :
403 280447 : if (state & id_based_versionhead_owner) {
404 15248 : node_destroy(ov->os, store, ov->id_based_head);
405 : }
406 :
407 280447 : if (os->destroy && ov->b)
408 256420 : os->destroy(store, ov->b);
409 :
410 280447 : if (os->temporary && (state & deleted || state & under_destruction || state & rollbacked))
411 195 : os_destroy(os, store); // TODO transaction_layer_revamp: embed into refcounting subproject : reference is already dropped by os_cleanup
412 280447 : _DELETE(ov);
413 280447 : }
414 :
415 : static void
416 6098 : _os_rollback(objectversion *ov, sqlstore *store)
417 : {
418 6098 : assert(ov->ts >= TRANSACTION_ID_BASE);
419 :
420 6098 : bte state = os_atmc_get_state(ov);
421 6098 : if (state & rollbacked) {
422 : return;
423 : }
424 :
425 5596 : state |= rollbacked;
426 5596 : os_atmc_set_state(ov, state);
427 :
428 5596 : bte state_older;
429 :
430 : /*
431 : * We have to use the readers-writer lock here,
432 : * since the pointer containing the adress of the older objectversion might be concurrently overwritten if the older itself hass just been put in the under_destruction state .
433 : */
434 5596 : lock_reader(ov->os);
435 5596 : objectversion* name_based_older = ov->name_based_older;
436 5596 : unlock_reader(ov->os);
437 :
438 5596 : if (name_based_older && !((state_older= os_atmc_get_state(name_based_older)) & rollbacked)) {
439 745 : if (ov->ts != name_based_older->ts) {
440 : // older is last committed state or belongs to parent transaction.
441 : // In any case, we restore versionhead pointer to that.
442 :
443 248 : ATOMIC_BASE_TYPE expected_deleted = deleted;
444 248 : if (state_older == active || (state_older == deleted && ATOMIC_CAS(&name_based_older->state, &expected_deleted, block_destruction))) {
445 248 : ov->name_based_head->ov = name_based_older;
446 248 : name_based_older->name_based_newer=NULL;
447 248 : if (state_older != active && expected_deleted == deleted)
448 0 : os_atmc_set_state(name_based_older, deleted); //Restore the deleted older back to its deleted state.
449 : }
450 : }
451 : else {
452 497 : _os_rollback(name_based_older, store);
453 : }
454 : }
455 4851 : else if (!name_based_older) {
456 : // this is a terminal node. i.e. this objectversion does not have name based committed history
457 4851 : if (ov->name_based_head) // The oposite can happen during an early conflict in os_add or os_del.
458 4845 : os_remove_name_based_chain(ov->os, ov);
459 : }
460 :
461 : /*
462 : * We have to use the readers-writer lock here,
463 : * since the pointer containing the adress of the older objectversion might be concurrently overwritten if the older itself hass just been put in the under_destruction state .
464 : */
465 5596 : lock_reader(ov->os);
466 5596 : objectversion* id_based_older = ov->id_based_older;
467 5596 : unlock_reader(ov->os);
468 5596 : if (id_based_older && !((state_older= os_atmc_get_state(id_based_older)) & rollbacked)) {
469 253 : if (ov->ts != id_based_older->ts) {
470 : // older is last committed state or belongs to parent transaction.
471 : // In any case, we restore versionhead pointer to that.
472 :
473 248 : ATOMIC_BASE_TYPE expected_deleted = deleted;
474 248 : if (state_older == active || (state_older == deleted && ATOMIC_CAS(&id_based_older->state, &expected_deleted, block_destruction))) {
475 248 : ov->id_based_head->ov = id_based_older;
476 248 : id_based_older->id_based_newer=NULL;
477 248 : if (state_older != active && expected_deleted == deleted)
478 0 : os_atmc_set_state(id_based_older, deleted); //Restore the deleted older back to its deleted state.
479 : }
480 : }
481 5 : else if (id_based_older != name_based_older)
482 5 : _os_rollback(id_based_older, store);
483 : }
484 4889 : else if (!id_based_older) {
485 : // this is a terminal node. i.e. this objectversion does not have id based committed history
486 4889 : os_remove_id_based_chain(ov->os, ov);
487 : }
488 :
489 5596 : if (ov->name_based_newer && !(os_atmc_get_state(ov->name_based_newer) & rollbacked)) {
490 0 : _os_rollback(ov->name_based_newer, store);
491 : }
492 :
493 5596 : if (ov->id_based_newer && ov->id_based_newer != ov->name_based_newer && !(os_atmc_get_state(ov->id_based_newer) & rollbacked)) {
494 0 : _os_rollback(ov->id_based_newer, store);
495 : }
496 : }
497 :
498 : static int
499 5596 : os_rollback(objectversion *ov, sqlstore *store)
500 : {
501 5596 : _os_rollback(ov, store);
502 :
503 5596 : return LOG_OK;
504 : }
505 :
506 : static void
507 24027 : ov_destroy_obj_recursive(sqlstore* store, objectversion *ov)
508 : {
509 24027 : if (ov->id_based_older && ov->id_based_older == ov->name_based_older) {
510 13805 : ov_destroy_obj_recursive(store, ov->id_based_older);
511 : }
512 24027 : if (ov->os->destroy && ov->b) {
513 24027 : ov->os->destroy(store, ov->b);
514 24027 : ov->b = NULL;
515 : }
516 24027 : }
517 :
518 : static inline void
519 10389 : try_to_mark_deleted_for_destruction(sqlstore* store, objectversion *ov)
520 : {
521 10389 : ATOMIC_BASE_TYPE expected_deleted = deleted;
522 10389 : if (ATOMIC_CAS(&ov->state, &expected_deleted, under_destruction)) {
523 :
524 10389 : if (!ov->name_based_newer || (os_atmc_get_state(ov->name_based_newer) & rollbacked)) {
525 10337 : os_remove_name_based_chain(ov->os, ov);
526 : }
527 : else {
528 52 : lock_writer(ov->os);
529 52 : ov->name_based_newer->name_based_older = NULL;
530 52 : unlock_writer(ov->os);
531 : }
532 :
533 10389 : if (!ov->id_based_newer || (os_atmc_get_state(ov->id_based_newer) & rollbacked)) {
534 10359 : os_remove_id_based_chain(ov->os, ov);
535 : }
536 : else {
537 30 : lock_writer(ov->os);
538 30 : ov->id_based_newer->id_based_older = NULL;
539 30 : unlock_writer(ov->os);
540 : }
541 :
542 10389 : ov->ts = store_get_timestamp(store)+1;
543 10389 : if (!ov->os->nested)
544 10222 : ov_destroy_obj_recursive(store, ov);
545 : }
546 10389 : }
547 :
548 : static void
549 28196 : objectversion_destroy_recursive(sqlstore* store, objectversion *ov)
550 : {
551 28196 : if (ov->id_based_older && ov->id_based_older == ov->name_based_older) {
552 13972 : objectversion_destroy_recursive(store, ov->id_based_older);
553 : }
554 28196 : objectversion_destroy(store, ov->os, ov);
555 28196 : }
556 :
557 : static int
558 307873 : os_cleanup(sqlstore* store, objectversion *ov, ulng oldest)
559 : {
560 307873 : if (os_atmc_get_state(ov) & under_destruction) {
561 14722 : if (ov->ts < oldest) {
562 : // This one is ready to be freed
563 10389 : objectversion_destroy_recursive(store, ov);
564 10389 : return LOG_ERR;
565 : }
566 :
567 : // not yet old enough to be safely removed. Try later.
568 : return LOG_OK;
569 : }
570 :
571 293151 : if (os_atmc_get_state(ov) & rollbacked) {
572 17656 : if (ov->ts < oldest) {
573 : // This one is ready to be freed
574 5596 : if (ov->name_based_older && ov->name_based_older->name_based_newer == ov)
575 497 : ov->name_based_older->name_based_newer=NULL;
576 5596 : if (ov->id_based_older && ov->id_based_older->id_based_newer == ov)
577 459 : ov->id_based_older->id_based_newer=NULL;
578 5596 : objectversion_destroy(store, ov->os, ov);
579 5596 : return LOG_ERR;
580 : }
581 :
582 12060 : if (ov->ts > TRANSACTION_ID_BASE) {
583 : /* We mark it with the latest possible starttime and reinsert it into the cleanup list.
584 : * This will cause a safe eventual destruction of this rollbacked ov.
585 : */
586 5596 : ov->ts = store_get_timestamp(store)+1;
587 : }
588 :
589 : // not yet old enough to be safely removed. Try later.
590 12060 : return LOG_OK;
591 : }
592 :
593 275495 : if (os_atmc_get_state(ov) == deleted) {
594 10478 : if (ov->ts <= oldest) {
595 : // the oldest relevant state is deleted so lets try to mark it as destroyed
596 10389 : try_to_mark_deleted_for_destruction(store, ov);
597 10389 : return LOG_OK+1;
598 : }
599 :
600 : // Keep it inplace on the cleanup list, either because it is now marked for destruction or
601 : // we want to retry marking it for destruction later.
602 : return LOG_OK;
603 : }
604 :
605 265017 : assert(os_atmc_get_state(ov) != deleted && os_atmc_get_state(ov) != under_destruction && os_atmc_get_state(ov) != rollbacked);
606 265017 : 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
607 :
608 272986 : while (ov->id_based_older && ov->id_based_older == ov->name_based_older && ov->ts >= oldest) {
609 : ov = ov->id_based_older;
610 : }
611 :
612 265017 : if (ov->id_based_older && ov->id_based_older == ov->name_based_older) {
613 : // Destroy everything older then the oldest possibly relevant objectversion.
614 3835 : objectversion_destroy_recursive(store, ov->id_based_older);
615 3835 : ov->id_based_older = NULL;
616 : }
617 :
618 : return LOG_ERR;
619 : }
620 :
621 : static int
622 307880 : tc_gc_objectversion(sql_store store, sql_change *change, ulng oldest)
623 : {
624 : // assert(!change->handled);
625 307880 : objectversion *ov = (objectversion*)change->data;
626 :
627 307880 : if (oldest >= TRANSACTION_ID_BASE)
628 : return 0;
629 307873 : int res = os_cleanup( (sqlstore*) store, ov, oldest);
630 307873 : change->handled = (res)?true:false;
631 307873 : return res>=0?LOG_OK:LOG_ERR;
632 : }
633 :
634 : static int
635 281009 : tc_commit_objectversion(sql_trans *tr, sql_change *change, ulng commit_ts, ulng oldest)
636 : {
637 281009 : objectversion *ov = (objectversion*)change->data;
638 281009 : if (commit_ts) {
639 275413 : assert(ov->ts == tr->tid);
640 275413 : ov->ts = commit_ts;
641 275413 : change->committed = commit_ts < TRANSACTION_ID_BASE ? true: false;
642 275413 : (void)oldest;
643 275413 : if (!tr->parent)
644 275406 : change->obj->new = 0;
645 : }
646 : else {
647 5596 : os_rollback(ov, tr->store);
648 : }
649 :
650 281009 : return LOG_OK;
651 : }
652 :
653 : objectset *
654 60235 : os_new(sql_allocator *sa, destroy_fptr destroy, bool temporary, bool unique, bool concurrent, bool nested, sql_store store)
655 : {
656 60235 : assert(!sa);
657 60235 : objectset *os = SA_NEW(sa, objectset);
658 60235 : if (os) {
659 60235 : *os = (objectset) {
660 : .refcnt = ATOMIC_VAR_INIT(1),
661 : .sa = sa,
662 : .destroy = destroy,
663 : .temporary = temporary,
664 : .unique = unique,
665 : .concurrent = concurrent,
666 : .nested = nested,
667 : .store = store
668 : };
669 60235 : os->destroy = destroy;
670 60235 : MT_rwlock_init(&os->rw_lock, "sa_readers_lock");
671 : }
672 :
673 60235 : return os;
674 : }
675 :
676 : objectset *
677 4563 : os_dup(objectset *os)
678 : {
679 4563 : ATOMIC_INC(&os->refcnt);
680 4563 : return os;
681 : }
682 :
683 : void
684 64740 : os_destroy(objectset *os, sql_store store)
685 : {
686 64740 : if (ATOMIC_DEC(&os->refcnt) > 0)
687 : return;
688 60177 : MT_rwlock_destroy(&os->rw_lock);
689 60177 : ATOMIC_DESTROY(&os->refcnt);
690 60177 : versionhead* n=os->id_based_h;
691 306281 : while(n) {
692 246104 : objectversion *ov = n->ov;
693 492759 : while(ov) {
694 246655 : objectversion *older = ov->id_based_older;
695 246655 : objectversion_destroy(store, os, ov);
696 246655 : ov = older;
697 : }
698 246104 : versionhead* hn =n->next;
699 246104 : node_destroy(os, store, n);
700 246104 : n = hn;
701 : }
702 :
703 60177 : n=os->name_based_h;
704 306281 : while(n) {
705 246104 : versionhead* hn =n->next;
706 246104 : node_destroy(os, store, n);
707 246104 : n = hn;
708 : }
709 :
710 60177 : if (os->id_map)
711 3521 : hash_destroy(os->id_map);
712 :
713 60177 : if (os->name_map)
714 3495 : hash_destroy(os->name_map);
715 :
716 60177 : if (!os->sa)
717 60177 : _DELETE(os);
718 : }
719 :
720 : static versionhead *
721 16144063 : find_name(objectset *os, const char *name)
722 : {
723 16144063 : lock_reader(os);
724 16141409 : if (os->name_map) {
725 12562002 : int key = hash_key(name);
726 12562002 : sql_hash_e *he = os->name_map->buckets[key&(os->name_map->size-1)];
727 :
728 69326920 : for (; he; he = he->chain) {
729 68964789 : versionhead *n = he->value;
730 :
731 68964789 : if (n && n->ov->b->name && strcmp(n->ov->b->name, name) == 0) {
732 12199871 : unlock_reader(os);
733 12199871 : return n;
734 : }
735 : }
736 362131 : unlock_reader(os);
737 362131 : return NULL;
738 : }
739 :
740 4261916 : for (versionhead *n = os->name_based_h; n; n = n->next) {
741 3967544 : objectversion *ov = n->ov;
742 :
743 : /* check if names match */
744 3967544 : if (name[0] == ov->b->name[0] && strcmp(name, ov->b->name) == 0) {
745 3285035 : unlock_reader(os);
746 3285035 : return n;
747 : }
748 : }
749 :
750 294372 : unlock_reader(os);
751 294372 : return NULL;
752 : }
753 :
754 : static objectversion*
755 16805808 : get_valid_object_name(sql_trans *tr, objectversion *ov)
756 : {
757 16805882 : while(ov) {
758 16805851 : if (ov->ts == tr->tid || (tr->parent && tr_version_of_parent(tr, ov->ts)) || ov->ts < tr->ts)
759 16805550 : return ov;
760 : else {
761 300 : lock_reader(ov->os);
762 166 : objectversion* name_based_older = ov->name_based_older;
763 166 : unlock_reader(ov->os);
764 166 : ov = name_based_older;
765 : }
766 : }
767 : return ov;
768 : }
769 :
770 : static objectversion*
771 560234 : get_valid_object_id(sql_trans *tr, objectversion *ov)
772 : {
773 560407 : while(ov) {
774 560287 : if (ov->ts == tr->tid || (tr->parent && tr_version_of_parent(tr, ov->ts)) || ov->ts < tr->ts)
775 560114 : return ov;
776 : else {
777 173 : lock_reader(ov->os);
778 173 : objectversion* id_based_older = ov->id_based_older;
779 173 : unlock_reader(ov->os);
780 173 : ov = id_based_older;
781 : }
782 : }
783 : return ov;
784 : }
785 :
786 : static int
787 270135 : os_add_name_based(objectset *os, struct sql_trans *tr, const char *name, objectversion *ov) {
788 270135 : versionhead *name_based_node = NULL;
789 270135 : if (ov->id_based_older && strcmp(ov->id_based_older->b->name, name) == 0)
790 8205 : name_based_node = ov->id_based_older->name_based_head;
791 261930 : else if (os->unique) // Previous name based objectversion is of a different id, so now we do have to perform an extensive look up
792 81851 : name_based_node = find_name(os, name);
793 : // else names are not unique and each id based version head maps to its own name based version head.
794 :
795 90056 : if (name_based_node) {
796 8294 : objectversion *co = name_based_node->ov;
797 8294 : objectversion *oo = get_valid_object_name(tr, co);
798 8294 : if (co != oo) { /* conflict ? */
799 6 : TRC_WARNING(SQL_STORE, "%s" "if (co != oo) { /* conflict ? */", __func__);
800 6 : return -3;
801 : }
802 :
803 8288 : assert(ov != oo); // Time loops are not allowed
804 :
805 8288 : bte state = os_atmc_get_state(oo);
806 8288 : if (state != active) {
807 : // This can only happen if the parent oo was a comitted deleted at some point.
808 96 : assert(state == deleted || state == under_destruction || state == block_destruction);
809 : /* Since our parent oo is comitted deleted objectversion, we might have a conflict with
810 : * another transaction that tries to clean up oo or also wants to add a new objectversion.
811 : */
812 96 : ATOMIC_BASE_TYPE expected_deleted = deleted;
813 96 : if (!ATOMIC_CAS(&oo->state, &expected_deleted, block_destruction)) {
814 0 : TRC_WARNING(SQL_STORE, "%s: " "if (!ATOMIC_CAS(&oo->state, &expected_deleted, block_destruction)) { /*conflict with cleaner or write-write conflict*/ ", __func__);
815 0 : return -3; /*conflict with cleaner or write-write conflict*/
816 : }
817 : }
818 :
819 : /* new object with same name within transaction, should have a delete in between */
820 8192 : assert(!(state == active && oo->ts == ov->ts && !(os_atmc_get_state(ov) & deleted)));
821 :
822 8288 : lock_writer(os);
823 8288 : ov->name_based_head = oo->name_based_head;
824 8288 : ov->name_based_older = oo;
825 :
826 8288 : name_based_node->ov = ov;
827 8288 : if (oo) {
828 8288 : oo->name_based_newer = ov;
829 : // if the parent was originally deleted, we restore it to that state.
830 8288 : os_atmc_set_state(oo, state);
831 : }
832 8288 : unlock_writer(os);
833 8288 : return 0;
834 : } else { /* new */
835 261841 : if (os_append_name(os, ov) == NULL)
836 : return -1; // MALLOC_FAIL
837 : return 0;
838 : }
839 : }
840 :
841 : static int
842 270145 : os_add_id_based(objectset *os, struct sql_trans *tr, sqlid id, objectversion *ov) {
843 270145 : versionhead *id_based_node;
844 :
845 270145 : id_based_node = find_id(os, id);
846 :
847 270145 : if (id_based_node) {
848 8238 : objectversion *co = id_based_node->ov;
849 8238 : objectversion *oo = get_valid_object_id(tr, co);
850 8238 : if (co != oo) { /* conflict ? */
851 10 : TRC_WARNING(SQL_STORE, "%s" "if (co != oo) { /* conflict ? */", __func__);
852 10 : return -3;
853 : }
854 :
855 8228 : assert(ov != oo); // Time loops are not allowed
856 :
857 8228 : bte state = os_atmc_get_state(oo);
858 8228 : if (state != active) {
859 : // This can only happen if the parent oo was a comitted deleted at some point.
860 36 : assert(state == deleted || state == under_destruction || state == block_destruction);
861 : /* Since our parent oo is comitted deleted objectversion, we might have a conflict with
862 : * another transaction that tries to clean up oo or also wants to add a new objectversion.
863 : */
864 36 : ATOMIC_BASE_TYPE expected_deleted = deleted;
865 36 : if (!ATOMIC_CAS(&oo->state, &expected_deleted, block_destruction)) {
866 0 : TRC_WARNING(SQL_STORE, "%s" "if (!ATOMIC_CAS(&oo->state, &expected_deleted, block_destruction)) { /*conflict with cleaner or write-write conflict*/", __func__);
867 0 : return -3; /*conflict with cleaner or write-write conflict*/
868 : }
869 : }
870 :
871 8228 : lock_writer(os);
872 8228 : ov->id_based_head = oo->id_based_head;
873 8228 : ov->id_based_older = oo;
874 :
875 8228 : id_based_node->ov = ov;
876 8228 : if (oo) {
877 8228 : oo->id_based_newer = ov;
878 : // if the parent was originally deleted, we restore it to that state.
879 8228 : os_atmc_set_state(oo, state);
880 : }
881 8228 : unlock_writer(os);
882 8228 : return 0;
883 : } else { /* new */
884 261907 : if (os_append_id(os, ov) == NULL)
885 : return -1; // MALLOC_FAIL
886 :
887 : return 0;
888 : }
889 : }
890 :
891 : static int /*ok, error (name existed) and conflict (added before) */
892 270146 : os_add_(objectset *os, struct sql_trans *tr, const char *name, sql_base *b)
893 : {
894 270146 : int res = 0;
895 270146 : objectversion *ov = SA_NEW(os->sa, objectversion);
896 :
897 270146 : *ov = (objectversion) {
898 270146 : .ts = tr->tid,
899 : .b = b,
900 : .os = os,
901 : };
902 :
903 270146 : if (!os->concurrent && os_has_changes(os, tr)) { /* for object sets without concurrent support, conflict if concurrent changes are there */
904 1 : if (os->destroy)
905 1 : os->destroy(os->store, ov->b);
906 1 : _DELETE(ov);
907 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__);
908 1 : return -3; /* conflict */
909 : }
910 :
911 270145 : if ((res = os_add_id_based(os, tr, b->id, ov))) {
912 10 : if (os->destroy)
913 10 : os->destroy(os->store, ov->b);
914 10 : _DELETE(ov);
915 10 : return res;
916 : }
917 :
918 270135 : if ((res = os_add_name_based(os, tr, name, ov))) {
919 6 : trans_add(tr, b, ov, &tc_gc_objectversion, &tc_commit_objectversion, NULL);
920 6 : return res;
921 : }
922 :
923 270129 : if (os->temporary) (void) os_dup(os); // TODO transaction_layer_revamp: embed into refcounting subproject
924 270129 : trans_add(tr, b, ov, &tc_gc_objectversion, &tc_commit_objectversion, NULL);
925 270129 : return res;
926 : }
927 :
928 : int
929 270146 : os_add(objectset *os, struct sql_trans *tr, const char *name, sql_base *b)
930 : {
931 270146 : store_lock(tr->store);
932 270146 : int res = os_add_(os, tr, name, b);
933 270146 : store_unlock(tr->store);
934 270146 : return res;
935 : }
936 :
937 : static int
938 10867 : os_del_name_based(objectset *os, struct sql_trans *tr, const char *name, objectversion *ov) {
939 10867 : versionhead *name_based_node = NULL;
940 10867 : if (ov->id_based_older && strcmp(ov->id_based_older->b->name, name) == 0)
941 10867 : name_based_node = ov->id_based_older->name_based_head;
942 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
943 0 : name_based_node = find_name(os, name);
944 :
945 10867 : if (name_based_node) {
946 10867 : objectversion *co = name_based_node->ov;
947 10867 : objectversion *oo = get_valid_object_name(tr, co);
948 10867 : ov->name_based_head = oo->name_based_head;
949 10867 : if (co != oo) { /* conflict ? */
950 0 : TRC_WARNING(SQL_STORE, "%s: " "if (co != oo) { /* conflict ? */", __func__);
951 0 : return -3;
952 : }
953 10867 : ov->name_based_older = oo;
954 :
955 10867 : lock_writer(os);
956 10867 : if (oo) {
957 10867 : oo->name_based_newer = ov;
958 10867 : assert(os_atmc_get_state(oo) == active);
959 : }
960 10867 : name_based_node->ov = ov;
961 10867 : unlock_writer(os);
962 10867 : return 0;
963 : } else {
964 : /* missing */
965 0 : return -1;
966 : }
967 : }
968 :
969 : static int
970 10869 : os_del_id_based(objectset *os, struct sql_trans *tr, sqlid id, objectversion *ov) {
971 10869 : versionhead *id_based_node;
972 10869 : if (ov->name_based_older && ov->name_based_older->b->id == id)
973 0 : id_based_node = ov->name_based_older->id_based_head;
974 : else // Previous id based objectversion is of a different name, so now we do have to perform an extensive look up
975 10869 : id_based_node = find_id(os, id);
976 :
977 10869 : if (id_based_node) {
978 10869 : objectversion *co = id_based_node->ov;
979 10869 : objectversion *oo = get_valid_object_id(tr, co);
980 10869 : ov->id_based_head = oo->id_based_head;
981 10869 : if (co != oo) { /* conflict ? */
982 2 : TRC_WARNING(SQL_STORE, "%s" "if (co != oo) { /* conflict ? */", __func__);
983 2 : return -3;
984 : }
985 10867 : ov->id_based_older = oo;
986 :
987 10867 : lock_writer(os);
988 10867 : if (oo) {
989 10867 : oo->id_based_newer = ov;
990 10867 : assert(os_atmc_get_state(oo) == active);
991 : }
992 10867 : id_based_node->ov = ov;
993 10867 : unlock_writer(os);
994 10867 : return 0;
995 : } else {
996 : /* missing */
997 : return -1;
998 : }
999 : }
1000 :
1001 : static int
1002 10869 : os_del_(objectset *os, struct sql_trans *tr, const char *name, sql_base *b)
1003 : {
1004 10869 : int res = 0;
1005 10869 : objectversion *ov = SA_NEW(os->sa, objectversion);
1006 :
1007 10869 : *ov = (objectversion) {
1008 10869 : .ts = tr->tid,
1009 : .b = b,
1010 : .os = os,
1011 : };
1012 10869 : os_atmc_set_state(ov, deleted);
1013 :
1014 10869 : if ((res = os_del_id_based(os, tr, b->id, ov))) {
1015 2 : if (os->destroy)
1016 2 : os->destroy(os->store, ov->b);
1017 2 : _DELETE(ov);
1018 2 : return res;
1019 : }
1020 :
1021 10867 : if ((res = os_del_name_based(os, tr, name, ov))) {
1022 0 : trans_add(tr, b, ov, &tc_gc_objectversion, &tc_commit_objectversion, NULL);
1023 0 : return res;
1024 : }
1025 :
1026 10867 : if (os->temporary) (void) os_dup(os); // TODO transaction_layer_revamp: embed into refcounting subproject
1027 10867 : trans_add(tr, b, ov, &tc_gc_objectversion, &tc_commit_objectversion, NULL);
1028 10867 : return res;
1029 : }
1030 :
1031 : int
1032 10869 : os_del(objectset *os, struct sql_trans *tr, const char *name, sql_base *b)
1033 : {
1034 10869 : store_lock(tr->store);
1035 10869 : int res = os_del_(os, tr, name, b);
1036 10869 : store_unlock(tr->store);
1037 10869 : return res;
1038 : }
1039 :
1040 : int
1041 364 : os_size(objectset *os, struct sql_trans *tr)
1042 : {
1043 364 : int cnt = 0;
1044 364 : if (os) {
1045 364 : lock_reader(os);
1046 374 : for(versionhead *n = os->name_based_h; n; n=n->next) {
1047 10 : objectversion *ov = n->ov;
1048 10 : if ((ov=get_valid_object_name(tr, ov)) && os_atmc_get_state(ov) == active)
1049 7 : cnt++;
1050 : }
1051 364 : unlock_reader(os);
1052 : }
1053 364 : return cnt;
1054 : }
1055 :
1056 : int
1057 0 : os_empty(objectset *os, struct sql_trans *tr)
1058 : {
1059 0 : return os_size(os, tr)==0;
1060 : }
1061 :
1062 : int
1063 0 : os_remove(objectset *os, sql_trans *tr, const char *name)
1064 : {
1065 0 : (void) os;
1066 0 : (void) tr;
1067 0 : (void) name;
1068 : // TODO remove entire versionhead corresponding to this name.
1069 :
1070 : // TODO assert os->unique?s
1071 0 : return LOG_OK;
1072 : }
1073 :
1074 : sql_base *
1075 16065029 : os_find_name(objectset *os, struct sql_trans *tr, const char *name)
1076 : {
1077 16065029 : if (!os)
1078 : return NULL;
1079 :
1080 16065027 : assert(os->unique);
1081 16065027 : versionhead *n = find_name(os, name);
1082 :
1083 16061463 : if (n) {
1084 15486723 : objectversion *ov = get_valid_object_name(tr, n->ov);
1085 15484312 : if (ov && os_atmc_get_state(ov) == active)
1086 15484157 : return ov->b;
1087 : }
1088 : return NULL;
1089 : }
1090 :
1091 : sql_base *
1092 145680 : os_find_id(objectset *os, struct sql_trans *tr, sqlid id)
1093 : {
1094 145680 : if (!os)
1095 : return NULL;
1096 145680 : versionhead *n = find_id(os, id);
1097 :
1098 145680 : if (n) {
1099 145240 : objectversion *ov = get_valid_object_id(tr, n->ov);
1100 145240 : if (ov && os_atmc_get_state(ov) == active)
1101 145236 : return ov->b;
1102 : }
1103 : return NULL;
1104 : }
1105 :
1106 : void
1107 2611596 : os_iterator(struct os_iter *oi, struct objectset *os, struct sql_trans *tr, const char *name /*optional*/)
1108 : {
1109 2611596 : *oi = (struct os_iter) {
1110 : .os = os,
1111 : .tr = tr,
1112 : .name = name,
1113 : };
1114 :
1115 2611596 : lock_reader(os);
1116 2611606 : oi->n = os->name_based_h;
1117 2611606 : unlock_reader(os);
1118 2611603 : }
1119 :
1120 : sql_base *
1121 4222196 : oi_next(struct os_iter *oi)
1122 : {
1123 4222196 : sql_base *b = NULL;
1124 :
1125 4222196 : if (oi->name) {
1126 3045900 : versionhead *n = oi->n;
1127 :
1128 3045900 : lock_reader(oi->os); /* intentionally outside of while loop */
1129 479736086 : while (n && !b) {
1130 :
1131 473644286 : if (n->ov->b->name && strcmp(n->ov->b->name, oi->name) == 0) {
1132 1300771 : objectversion *ov = n->ov;
1133 :
1134 1300771 : n = oi->n = n->next;
1135 1300771 : ov = get_valid_object_name(oi->tr, ov);
1136 1300771 : if (ov && os_atmc_get_state(ov) == active)
1137 1300198 : b = ov->b;
1138 : } else {
1139 472343515 : n = oi->n = n->next;
1140 : }
1141 : }
1142 3045900 : unlock_reader(oi->os);
1143 : } else {
1144 1176296 : versionhead *n = oi->n;
1145 :
1146 1176296 : lock_reader(oi->os); /* intentionally outside of while loop */
1147 2748483 : while (n && !b) {
1148 395890 : objectversion *ov = n->ov;
1149 395890 : n = oi->n = n->next;
1150 :
1151 395890 : ov = get_valid_object_id(oi->tr, ov);
1152 395887 : if (ov && os_atmc_get_state(ov) == active)
1153 371478 : b = ov->b;
1154 : }
1155 1176297 : unlock_reader(oi->os);
1156 : }
1157 4222177 : return b;
1158 : }
1159 :
1160 : bool
1161 3534 : os_obj_intransaction(objectset *os, struct sql_trans *tr, sql_base *b)
1162 : {
1163 3534 : versionhead *n = find_id(os, b->id);
1164 :
1165 3534 : if (n) {
1166 3534 : objectversion *ov = n->ov;
1167 :
1168 3534 : if (ov && os_atmc_get_state(ov) == active && ov->ts == tr->tid)
1169 : return true;
1170 : }
1171 : return false;
1172 : }
1173 :
1174 : /* return true if this object set has changes pending for an other transaction */
1175 : bool
1176 174846 : os_has_changes(objectset *os, struct sql_trans *tr)
1177 : {
1178 174846 : versionhead *n = os->id_based_t;
1179 :
1180 174846 : if (n) {
1181 170531 : objectversion *ov = n->ov;
1182 :
1183 170531 : if (ov && os_atmc_get_state(ov) == active && ov->ts != tr->tid && ov->ts > TRANSACTION_ID_BASE)
1184 : return true;
1185 : }
1186 : return false;
1187 : }
|
