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 328513 : os_id_key(versionhead *n)
77 : {
78 254071 : return (int) BATatoms[TYPE_int].atomHash(&n->ov->b->id);
79 : }
80 :
81 : static inline void
82 25444463 : lock_reader(objectset* os)
83 : {
84 25444463 : MT_rwlock_rdlock(&os->rw_lock);
85 1249245 : }
86 :
87 : static inline void
88 25525696 : unlock_reader(objectset* os)
89 : {
90 25525696 : MT_rwlock_rdunlock(&os->rw_lock);
91 4255017 : }
92 :
93 : static inline void
94 600341 : lock_writer(objectset* os)
95 : {
96 600341 : MT_rwlock_wrlock(&os->rw_lock);
97 : }
98 :
99 : static inline void
100 600341 : unlock_writer(objectset* os)
101 : {
102 600341 : MT_rwlock_wrunlock(&os->rw_lock);
103 251 : }
104 :
105 22281210 : static bte os_atmc_get_state(objectversion *ov) {
106 22281210 : bte state = (bte) ATOMIC_GET(&ov->state);
107 22281210 : return state;
108 : }
109 :
110 65374 : static void os_atmc_set_state(objectversion *ov, bte state) {
111 65374 : ATOMIC_SET(&ov->state, state);
112 0 : }
113 :
114 : static versionhead *
115 443103 : find_id(objectset *os, sqlid id)
116 : {
117 443103 : if (os) {
118 443103 : lock_reader(os);
119 443103 : if (os->id_map) {
120 411096 : int key = (int) BATatoms[TYPE_int].atomHash(&id);
121 411096 : sql_hash_e *he = os->id_map->buckets[key&(os->id_map->size-1)];
122 :
123 1388861 : for (; he; he = he->chain) {
124 1153278 : versionhead *n = he->value;
125 :
126 1153278 : if (n && n->ov->b->id == id) {
127 175513 : unlock_reader(os);
128 175513 : return n;
129 : }
130 : }
131 235583 : unlock_reader(os);
132 235583 : return NULL;
133 : }
134 :
135 96857 : for (versionhead *n = os->id_based_h; n; n = n->next) {
136 67405 : objectversion *ov = n->ov;
137 :
138 : /* check if ids match */
139 67405 : if (id == ov->b->id) {
140 2555 : unlock_reader(os);
141 2555 : return n;
142 : }
143 : }
144 29452 : 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 29954 : hash_delete(sql_hash *h, void *data)
153 : {
154 29954 : int key = h->key(data);
155 29954 : sql_hash_e *e, *p = h->buckets[key&(h->size-1)];
156 :
157 29954 : e = p;
158 46824 : for (; p && p->value != data ; p = p->chain)
159 16870 : e = p;
160 29954 : if (p && p->value == data) {
161 29954 : if (p == e)
162 25458 : h->buckets[key&(h->size-1)] = p->chain;
163 : else
164 4496 : e->chain = p->chain;
165 29954 : if (!h->sa)
166 29954 : _DELETE(p);
167 : }
168 29954 : h->entries--;
169 29954 : }
170 :
171 : static void
172 527686 : node_destroy(objectset *os, sqlstore *store, versionhead *n)
173 : {
174 527686 : if (!os->sa)
175 527686 : _DELETE(n);
176 527686 : (void)store;
177 527686 : }
178 :
179 : static versionhead *
180 15692 : os_remove_name_based_chain(objectset *os, objectversion* ov)
181 : {
182 15692 : lock_writer(os);
183 15692 : versionhead *n = ov->name_based_head;
184 15692 : versionhead *p = os->name_based_h;
185 15692 : if (p != n)
186 1175248 : while (p && p->next != n)
187 : p = p->next;
188 15692 : assert(p==n||(p && p->next == n));
189 15692 : if (p == n) {
190 1348 : os->name_based_h = n->next;
191 1348 : if (os->name_based_h) // i.e. non-empty os
192 811 : os->name_based_h->prev = NULL;
193 : p = NULL;
194 14344 : } else if ( p != NULL) {
195 14344 : p->next = n->next;
196 14344 : if (p->next) // node in the middle
197 5504 : p->next->prev = p;
198 : }
199 15692 : if (n == os->name_based_t)
200 9377 : os->name_based_t = p;
201 :
202 15692 : if (os->name_map && n)
203 14631 : hash_delete(os->name_map, n);
204 :
205 15692 : os->name_based_cnt--;
206 15692 : unlock_writer(os);
207 :
208 15692 : bte state = os_atmc_get_state(ov);
209 15692 : state |= name_based_versionhead_owner;
210 15692 : os_atmc_set_state(ov, state);
211 15692 : return p;
212 : }
213 :
214 : static versionhead *
215 15884 : os_remove_id_based_chain(objectset *os, objectversion* ov)
216 : {
217 15884 : lock_writer(os);
218 15884 : versionhead *n = ov->id_based_head;
219 15884 : versionhead *p = os->id_based_h;
220 :
221 15884 : if (p != n)
222 1189754 : while (p && p->next != n)
223 : p = p->next;
224 15884 : assert(p==n||(p && p->next == n));
225 15884 : if (p == n) {
226 1348 : os->id_based_h = n->next;
227 1348 : if (os->id_based_h) // i.e. non-empty os
228 811 : os->id_based_h->prev = NULL;
229 : p = NULL;
230 14536 : } else if ( p != NULL) {
231 14536 : p->next = n->next;
232 14536 : if (p->next) // node in the middle
233 5697 : p->next->prev = p;
234 : }
235 15884 : if (n == os->id_based_t)
236 9376 : os->id_based_t = p;
237 :
238 15884 : if (os->id_map && n)
239 15323 : hash_delete(os->id_map, n);
240 :
241 15884 : os->name_based_cnt--;
242 15884 : unlock_writer(os);
243 :
244 15884 : bte state = os_atmc_get_state(ov);
245 15884 : state |= id_based_versionhead_owner;
246 15884 : os_atmc_set_state(ov, state);
247 15884 : return p;
248 : }
249 :
250 : static versionhead *
251 528836 : node_create(allocator *sa, objectversion *ov)
252 : {
253 528836 : versionhead *n = SA_NEW(sa, versionhead );
254 :
255 528836 : if (n == NULL)
256 : return NULL;
257 528836 : *n = (versionhead ) {
258 : .ov = ov,
259 : };
260 528836 : return n;
261 : }
262 :
263 : static inline int
264 318524 : os_name_key(versionhead *n)
265 : {
266 318524 : return hash_key(n->ov->b->name);
267 : }
268 :
269 : static objectset *
270 264322 : os_append_node_name(objectset *os, versionhead *n)
271 : {
272 264322 : lock_writer(os);
273 264322 : if ((!os->name_map || os->name_map->size*16 < os->name_based_cnt) && os->name_based_cnt > HASH_MIN_SIZE) {
274 3814 : hash_destroy(os->name_map);
275 3814 : os->name_map = hash_new(os->sa, os->name_based_cnt, (fkeyvalue)& os_name_key);
276 3814 : if (os->name_map == NULL) {
277 0 : unlock_writer(os);
278 0 : return NULL;
279 : }
280 :
281 69642 : for (versionhead *n = os->name_based_h; n; n = n->next ) {
282 65828 : int key = os_name_key(n);
283 :
284 65828 : if (hash_add(os->name_map, key, n) == NULL) {
285 0 : unlock_writer(os);
286 0 : return NULL;
287 : }
288 : }
289 : }
290 :
291 264322 : if (os->name_map) {
292 238065 : int key = os->name_map->key(n);
293 :
294 238065 : if (hash_add(os->name_map, key, n) == NULL) {
295 0 : unlock_writer(os);
296 0 : return NULL;
297 : }
298 : }
299 :
300 264322 : if (os->name_based_t) {
301 256072 : os->name_based_t->next = n;
302 : } else {
303 8250 : os->name_based_h = n;
304 : }
305 264322 : n->prev = os->name_based_t; // aka the double linked list.
306 264322 : os->name_based_t = n;
307 264322 : os->name_based_cnt++;
308 264322 : unlock_writer(os);
309 264322 : return os;
310 : }
311 :
312 : static objectset *
313 264322 : os_append_name(objectset *os, objectversion *ov)
314 : {
315 264322 : versionhead *n = node_create(os->sa, ov);
316 :
317 264322 : if (n == NULL)
318 : return NULL;
319 :
320 264322 : ov->name_based_head = n;
321 264322 : 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 3916 : os_append_id_map(objectset *os)
331 : {
332 3916 : if (os->id_map)
333 452 : hash_destroy(os->id_map);
334 3916 : os->id_map = hash_new(os->sa, os->id_based_cnt, (fkeyvalue)&os_id_key);
335 3916 : if (os->id_map == NULL)
336 : return ;
337 78358 : for (versionhead *n = os->id_based_h; n; n = n->next ) {
338 74442 : int key = os_id_key(n);
339 :
340 74442 : 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 264514 : os_append_node_id(objectset *os, versionhead *n)
350 : {
351 264514 : lock_writer(os);
352 264514 : if ((!os->id_map || os->id_map->size*16 < os->id_based_cnt) && os->id_based_cnt > HASH_MIN_SIZE)
353 3916 : os_append_id_map(os); /* on failure just fall back to slow method */
354 :
355 264514 : if (os->id_map) {
356 238748 : int key = os->id_map->key(n);
357 238748 : 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 264514 : if (os->id_based_t) {
365 256264 : os->id_based_t->next = n;
366 : } else {
367 8250 : os->id_based_h = n;
368 : }
369 264514 : n->prev = os->id_based_t; // aka the double linked list.
370 264514 : os->id_based_t = n;
371 264514 : os->id_based_cnt++;
372 264514 : unlock_writer(os);
373 264514 : return os;
374 : }
375 :
376 : static objectset *
377 264514 : os_append_id(objectset *os, objectversion *ov)
378 : {
379 264514 : versionhead *n = node_create(os->sa, ov);
380 :
381 264514 : if (n == NULL)
382 : return NULL;
383 264514 : ov->id_based_head = n;
384 264514 : 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 283685 : objectversion_destroy(sqlstore *store, objectset* os, objectversion *ov)
396 : {
397 :
398 283685 : bte state = os_atmc_get_state(ov);
399 :
400 283685 : if (state & name_based_versionhead_owner) {
401 15692 : node_destroy(ov->os, store, ov->name_based_head);
402 : }
403 :
404 283685 : if (state & id_based_versionhead_owner) {
405 15884 : node_destroy(ov->os, store, ov->id_based_head);
406 : }
407 :
408 283685 : if (os->destroy && ov->b)
409 258383 : os->destroy(store, ov->b);
410 :
411 283685 : if (os->temporary && (state & deleted || state & under_destruction || state & rollbacked))
412 202 : os_destroy(os, store); // TODO transaction_layer_revamp: embed into refcounting subproject : reference is already dropped by os_cleanup
413 283685 : _DELETE(ov);
414 283685 : }
415 :
416 : static void
417 6125 : _os_rollback(objectversion *ov, sqlstore *store)
418 : {
419 6125 : assert(ov->ts >= TRANSACTION_ID_BASE);
420 :
421 6125 : bte state = os_atmc_get_state(ov);
422 6125 : if (state & rollbacked) {
423 : return;
424 : }
425 :
426 5625 : state |= rollbacked;
427 5625 : os_atmc_set_state(ov, state);
428 :
429 5625 : 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 5625 : lock_reader(ov->os);
436 5625 : objectversion* name_based_older = ov->name_based_older;
437 5625 : unlock_reader(ov->os);
438 :
439 5625 : if (name_based_older && !((state_older= os_atmc_get_state(name_based_older)) & rollbacked)) {
440 757 : 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 :
444 262 : ATOMIC_BASE_TYPE expected_deleted = deleted;
445 262 : if (state_older == active || (state_older == deleted && ATOMIC_CAS(&name_based_older->state, &expected_deleted, block_destruction))) {
446 262 : ov->name_based_head->ov = name_based_older;
447 262 : name_based_older->name_based_newer=NULL;
448 262 : 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 4868 : else if (!name_based_older) {
457 : // this is a terminal node. i.e. this objectversion does not have name based committed history
458 4868 : if (ov->name_based_head) // The opposite can happen during an early conflict in os_add or os_del.
459 4862 : 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 5625 : lock_reader(ov->os);
467 5625 : objectversion* id_based_older = ov->id_based_older;
468 5625 : unlock_reader(ov->os);
469 5625 : if (id_based_older && !((state_older= os_atmc_get_state(id_based_older)) & rollbacked)) {
470 267 : 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 :
474 262 : ATOMIC_BASE_TYPE expected_deleted = deleted;
475 262 : if (state_older == active || (state_older == deleted && ATOMIC_CAS(&id_based_older->state, &expected_deleted, block_destruction))) {
476 262 : ov->id_based_head->ov = id_based_older;
477 262 : id_based_older->id_based_newer=NULL;
478 262 : 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 4905 : else if (!id_based_older) {
486 : // this is a terminal node. i.e. this objectversion does not have id based committed history
487 4905 : os_remove_id_based_chain(ov->os, ov);
488 : }
489 :
490 5625 : 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 5625 : 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 5625 : os_rollback(objectversion *ov, sqlstore *store)
501 : {
502 5625 : _os_rollback(ov, store);
503 :
504 5625 : return LOG_OK;
505 : }
506 :
507 : static void
508 25302 : ov_destroy_obj_recursive(sqlstore* store, objectversion *ov)
509 : {
510 25302 : if (ov->id_based_older && ov->id_based_older == ov->name_based_older) {
511 14466 : ov_destroy_obj_recursive(store, ov->id_based_older);
512 : }
513 25302 : if (ov->os->destroy && ov->b) {
514 25302 : ov->os->destroy(store, ov->b);
515 25302 : ov->b = NULL;
516 : }
517 25302 : }
518 :
519 : static inline void
520 11030 : try_to_mark_deleted_for_destruction(sqlstore* store, objectversion *ov)
521 : {
522 11030 : ATOMIC_BASE_TYPE expected_deleted = deleted;
523 11030 : if (ATOMIC_CAS(&ov->state, &expected_deleted, under_destruction)) {
524 :
525 11030 : if (!ov->name_based_newer || (os_atmc_get_state(ov->name_based_newer) & rollbacked)) {
526 10830 : os_remove_name_based_chain(ov->os, ov);
527 : }
528 : else {
529 200 : lock_writer(ov->os);
530 200 : ov->name_based_newer->name_based_older = NULL;
531 200 : unlock_writer(ov->os);
532 : }
533 :
534 11030 : if (!ov->id_based_newer || (os_atmc_get_state(ov->id_based_newer) & rollbacked)) {
535 10979 : 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 11030 : ov->ts = store_get_timestamp(store)+1;
544 11030 : if (!ov->os->nested)
545 10836 : ov_destroy_obj_recursive(store, ov);
546 : }
547 11030 : }
548 :
549 : static void
550 29514 : objectversion_destroy_recursive(sqlstore* store, objectversion *ov)
551 : {
552 29514 : if (ov->id_based_older && ov->id_based_older == ov->name_based_older) {
553 14660 : objectversion_destroy_recursive(store, ov->id_based_older);
554 : }
555 29514 : objectversion_destroy(store, ov->os, ov);
556 29514 : }
557 :
558 : static int
559 311751 : os_cleanup(sqlstore* store, objectversion *ov, ulng oldest)
560 : {
561 311751 : if (os_atmc_get_state(ov) & under_destruction) {
562 15238 : if (ov->ts < oldest) {
563 : // This one is ready to be freed
564 11030 : objectversion_destroy_recursive(store, ov);
565 11030 : return LOG_ERR;
566 : }
567 :
568 : // not yet old enough to be safely removed. Try later.
569 : return LOG_OK;
570 : }
571 :
572 296513 : if (os_atmc_get_state(ov) & rollbacked) {
573 17793 : if (ov->ts < oldest) {
574 : // This one is ready to be freed
575 5625 : if (ov->name_based_older && ov->name_based_older->name_based_newer == ov)
576 495 : ov->name_based_older->name_based_newer=NULL;
577 5625 : if (ov->id_based_older && ov->id_based_older->id_based_newer == ov)
578 458 : ov->id_based_older->id_based_newer=NULL;
579 5625 : objectversion_destroy(store, ov->os, ov);
580 5625 : return LOG_ERR;
581 : }
582 :
583 12168 : 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 5625 : ov->ts = store_get_timestamp(store)+1;
588 : }
589 :
590 : // not yet old enough to be safely removed. Try later.
591 12168 : return LOG_OK;
592 : }
593 :
594 278720 : if (os_atmc_get_state(ov) == deleted) {
595 11115 : if (ov->ts <= oldest) {
596 : // the oldest relevant state is deleted so lets try to mark it as destroyed
597 11030 : try_to_mark_deleted_for_destruction(store, ov);
598 11030 : 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 267605 : assert(os_atmc_get_state(ov) != deleted && os_atmc_get_state(ov) != under_destruction && os_atmc_get_state(ov) != rollbacked);
607 267605 : 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 275550 : 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 267605 : if (ov->id_based_older && ov->id_based_older == ov->name_based_older) {
614 : // Destroy everything older then the oldest possibly relevant objectversion.
615 3824 : objectversion_destroy_recursive(store, ov->id_based_older);
616 3824 : ov->id_based_older = NULL;
617 : }
618 :
619 : return LOG_ERR;
620 : }
621 :
622 : static int
623 311758 : tc_gc_objectversion(sql_store store, sql_change *change, ulng oldest)
624 : {
625 : // assert(!change->handled);
626 311758 : objectversion *ov = (objectversion*)change->data;
627 :
628 311758 : if (oldest >= TRANSACTION_ID_BASE)
629 : return 0;
630 311751 : int res = os_cleanup( (sqlstore*) store, ov, oldest);
631 311751 : change->handled = (res)?true:false;
632 311751 : return res>=0?LOG_OK:LOG_ERR;
633 : }
634 :
635 : static int
636 284267 : tc_commit_objectversion(sql_trans *tr, sql_change *change, ulng commit_ts, ulng oldest)
637 : {
638 284267 : objectversion *ov = (objectversion*)change->data;
639 284267 : if (commit_ts) {
640 278642 : assert(ov->ts == tr->tid);
641 278642 : ov->ts = commit_ts;
642 278642 : change->committed = commit_ts < TRANSACTION_ID_BASE ? true: false;
643 278642 : (void)oldest;
644 278642 : if (!tr->parent)
645 278635 : change->obj->new = 0;
646 278642 : if (!ov->os->temporary)
647 274181 : ATOMIC_INC(&tr->cat->schema_version);
648 : } else {
649 5625 : os_rollback(ov, tr->store);
650 : }
651 :
652 284267 : return LOG_OK;
653 : }
654 :
655 : objectset *
656 59255 : os_new(allocator *sa, destroy_fptr destroy, bool temporary, bool unique, bool concurrent, bool nested, sql_store store)
657 : {
658 59255 : assert(!sa);
659 59255 : objectset *os = SA_NEW(sa, objectset);
660 59255 : if (os) {
661 59255 : *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 59255 : os->destroy = destroy;
672 59255 : MT_rwlock_init(&os->rw_lock, "sa_readers_lock");
673 59255 : MT_lock_init(&os->lock, "single_writer_lock");
674 : }
675 :
676 59255 : return os;
677 : }
678 :
679 : objectset *
680 4734 : os_dup(objectset *os)
681 : {
682 4734 : ATOMIC_INC(&os->refcnt);
683 4734 : return os;
684 : }
685 :
686 : void
687 63931 : os_destroy(objectset *os, sql_store store)
688 : {
689 63931 : if (ATOMIC_DEC(&os->refcnt) > 0)
690 : return;
691 59197 : MT_lock_destroy(&os->lock);
692 59197 : MT_rwlock_destroy(&os->rw_lock);
693 59197 : versionhead* n=os->id_based_h;
694 307252 : while(n) {
695 248055 : objectversion *ov = n->ov;
696 496601 : while(ov) {
697 248546 : objectversion *older = ov->id_based_older;
698 248546 : objectversion_destroy(store, os, ov);
699 248546 : ov = older;
700 : }
701 248055 : versionhead* hn =n->next;
702 248055 : node_destroy(os, store, n);
703 248055 : n = hn;
704 : }
705 :
706 59197 : n=os->name_based_h;
707 307252 : while(n) {
708 248055 : versionhead* hn =n->next;
709 248055 : node_destroy(os, store, n);
710 248055 : n = hn;
711 : }
712 :
713 59197 : if (os->id_map)
714 3454 : hash_destroy(os->id_map);
715 :
716 59197 : if (os->name_map)
717 3428 : hash_destroy(os->name_map);
718 :
719 59197 : if (!os->sa)
720 59197 : _DELETE(os);
721 : }
722 :
723 : static versionhead *
724 18474319 : find_name(objectset *os, const char *name)
725 : {
726 18474319 : lock_reader(os);
727 18538425 : if (os->name_map) {
728 15044718 : int key = hash_key(name);
729 15044718 : sql_hash_e *he = os->name_map->buckets[key&(os->name_map->size-1)];
730 :
731 81674579 : for (; he; he = he->chain) {
732 81307593 : versionhead *n = he->value;
733 :
734 81307593 : if (n && n->ov->b->name && strcmp(n->ov->b->name, name) == 0) {
735 14677732 : unlock_reader(os);
736 14677732 : return n;
737 : }
738 : }
739 366986 : unlock_reader(os);
740 366986 : return NULL;
741 : }
742 :
743 4179800 : for (versionhead *n = os->name_based_h; n; n = n->next) {
744 3878400 : objectversion *ov = n->ov;
745 :
746 : /* check if names match */
747 3878400 : if (name[0] == ov->b->name[0] && strcmp(name, ov->b->name) == 0) {
748 3192307 : unlock_reader(os);
749 3192307 : return n;
750 : }
751 : }
752 :
753 301400 : unlock_reader(os);
754 301400 : return NULL;
755 : }
756 :
757 : static objectversion*
758 19482016 : get_valid_object_name(sql_trans *tr, objectversion *ov)
759 : {
760 19505070 : while(ov) {
761 19505038 : if (ov->ts == tr->tid || (tr->parent && tr_version_of_parent(tr, ov->ts)) || ov->ts < tr->ts)
762 19520845 : return ov;
763 : else {
764 0 : 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 579763 : get_valid_object_id(sql_trans *tr, objectversion *ov)
775 : {
776 579931 : while(ov) {
777 579818 : if (ov->ts == tr->tid || (tr->parent && tr_version_of_parent(tr, ov->ts)) || ov->ts < tr->ts)
778 579650 : return ov;
779 : else {
780 168 : lock_reader(ov->os);
781 168 : objectversion* id_based_older = ov->id_based_older;
782 168 : unlock_reader(ov->os);
783 168 : ov = id_based_older;
784 : }
785 : }
786 : return ov;
787 : }
788 :
789 : static int
790 272753 : os_add_name_based(objectset *os, struct sql_trans *tr, const char *name, objectversion *ov) {
791 272753 : versionhead *name_based_node = NULL;
792 272753 : if (ov->id_based_older && strcmp(ov->id_based_older->b->name, name) == 0)
793 8215 : name_based_node = ov->id_based_older->name_based_head;
794 264538 : 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 81573 : 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 89788 : if (name_based_node) {
799 8431 : objectversion *co = name_based_node->ov;
800 8431 : objectversion *oo = get_valid_object_name(tr, co);
801 8431 : if (co != oo) { /* conflict ? */
802 6 : TRC_WARNING(SQL_STORE, "%s" "if (co != oo) { /* conflict ? */", __func__);
803 6 : return -3;
804 : }
805 :
806 8425 : assert(ov != oo); // Time loops are not allowed
807 :
808 8425 : bte state = os_atmc_get_state(oo);
809 8425 : if (state != active) {
810 : // This can only happen if the parent oo was a committed deleted at some point.
811 243 : 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 : */
815 243 : ATOMIC_BASE_TYPE expected_deleted = deleted;
816 243 : 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 8182 : assert(!(state == active && oo->ts == ov->ts && !(os_atmc_get_state(ov) & deleted)));
824 :
825 8425 : lock_writer(os);
826 8425 : ov->name_based_head = oo->name_based_head;
827 8425 : ov->name_based_older = oo;
828 :
829 8425 : name_based_node->ov = ov;
830 8425 : if (oo) {
831 8425 : oo->name_based_newer = ov;
832 : // if the parent was originally deleted, we restore it to that state.
833 8425 : os_atmc_set_state(oo, state);
834 : }
835 8425 : unlock_writer(os);
836 8425 : return 0;
837 : } else { /* new */
838 264322 : if (os_append_name(os, ov) == NULL)
839 : return -1; // MALLOC_FAIL
840 : return 0;
841 : }
842 : }
843 :
844 : static int
845 272763 : os_add_id_based(objectset *os, struct sql_trans *tr, sqlid id, objectversion *ov) {
846 272763 : versionhead *id_based_node;
847 :
848 272763 : id_based_node = find_id(os, id);
849 :
850 272763 : if (id_based_node) {
851 8249 : objectversion *co = id_based_node->ov;
852 8249 : objectversion *oo = get_valid_object_id(tr, co);
853 8249 : if (co != oo) { /* conflict ? */
854 10 : TRC_WARNING(SQL_STORE, "%s" "if (co != oo) { /* conflict ? */", __func__);
855 10 : return -3;
856 : }
857 :
858 8239 : assert(ov != oo); // Time loops are not allowed
859 :
860 8239 : bte state = os_atmc_get_state(oo);
861 8239 : 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 8239 : lock_writer(os);
875 8239 : ov->id_based_head = oo->id_based_head;
876 8239 : ov->id_based_older = oo;
877 :
878 8239 : id_based_node->ov = ov;
879 8239 : if (oo) {
880 8239 : oo->id_based_newer = ov;
881 : // if the parent was originally deleted, we restore it to that state.
882 8239 : os_atmc_set_state(oo, state);
883 : }
884 8239 : unlock_writer(os);
885 8239 : return 0;
886 : } else { /* new */
887 264514 : 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 272764 : os_add_(objectset *os, struct sql_trans *tr, const char *name, sql_base *b)
896 : {
897 272764 : int res = 0;
898 272764 : objectversion *ov = SA_NEW(os->sa, objectversion);
899 :
900 272764 : *ov = (objectversion) {
901 272764 : .ts = tr->tid,
902 : .b = b,
903 : .os = os,
904 : };
905 :
906 272764 : 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 272763 : 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 272753 : 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 272747 : if (os->temporary) (void) os_dup(os); // TODO transaction_layer_revamp: embed into refcounting subproject
927 272747 : trans_add(tr, b, ov, &tc_gc_objectversion, &tc_commit_objectversion, NULL);
928 272747 : return res;
929 : }
930 :
931 : int
932 272764 : os_add(objectset *os, struct sql_trans *tr, const char *name, sql_base *b)
933 : {
934 272764 : MT_lock_set(&os->lock);
935 272764 : int res = os_add_(os, tr, name, b);
936 272764 : MT_lock_unset(&os->lock);
937 272764 : return res;
938 : }
939 :
940 : static int
941 11507 : os_del_name_based(objectset *os, struct sql_trans *tr, const char *name, objectversion *ov) {
942 11507 : versionhead *name_based_node = NULL;
943 11507 : if (ov->id_based_older && strcmp(ov->id_based_older->b->name, name) == 0)
944 11507 : 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 11507 : if (name_based_node) {
949 11507 : objectversion *co = name_based_node->ov;
950 11507 : objectversion *oo = get_valid_object_name(tr, co);
951 11507 : ov->name_based_head = oo->name_based_head;
952 11507 : if (co != oo) { /* conflict ? */
953 0 : TRC_WARNING(SQL_STORE, "%s: " "if (co != oo) { /* conflict ? */", __func__);
954 0 : return -3;
955 : }
956 11507 : ov->name_based_older = oo;
957 :
958 11507 : lock_writer(os);
959 11507 : if (oo) {
960 11507 : oo->name_based_newer = ov;
961 11507 : assert(os_atmc_get_state(oo) == active);
962 : }
963 11507 : name_based_node->ov = ov;
964 11507 : unlock_writer(os);
965 11507 : return 0;
966 : } else {
967 : /* missing */
968 0 : return -1;
969 : }
970 : }
971 :
972 : static int
973 11509 : os_del_id_based(objectset *os, struct sql_trans *tr, sqlid id, objectversion *ov) {
974 :
975 11509 : versionhead *id_based_node;
976 :
977 11509 : 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 11509 : id_based_node = find_id(os, id);
981 :
982 11509 : if (id_based_node) {
983 11509 : objectversion *co = id_based_node->ov;
984 11509 : objectversion *oo = get_valid_object_id(tr, co);
985 11509 : ov->id_based_head = oo->id_based_head;
986 11509 : if (co != oo) { /* conflict ? */
987 2 : TRC_WARNING(SQL_STORE, "%s" "if (co != oo) { /* conflict ? */", __func__);
988 2 : return -3;
989 : }
990 11507 : ov->id_based_older = oo;
991 :
992 11507 : lock_writer(os);
993 11507 : if (oo) {
994 11507 : oo->id_based_newer = ov;
995 11507 : assert(os_atmc_get_state(oo) == active);
996 : }
997 11507 : id_based_node->ov = ov;
998 11507 : unlock_writer(os);
999 11507 : return 0;
1000 : } else {
1001 : /* missing */
1002 : return -1;
1003 : }
1004 : }
1005 :
1006 : static int
1007 11509 : os_del_(objectset *os, struct sql_trans *tr, const char *name, sql_base *b)
1008 : {
1009 11509 : int res = 0;
1010 11509 : objectversion *ov = SA_NEW(os->sa, objectversion);
1011 :
1012 11509 : *ov = (objectversion) {
1013 11509 : .ts = tr->tid,
1014 : .b = b,
1015 : .os = os,
1016 : };
1017 11509 : os_atmc_set_state(ov, deleted);
1018 :
1019 11509 : 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 11507 : 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 11507 : if (os->temporary) (void) os_dup(os); // TODO transaction_layer_revamp: embed into refcounting subproject
1032 11507 : trans_add(tr, b, ov, &tc_gc_objectversion, &tc_commit_objectversion, NULL);
1033 11507 : return res;
1034 : }
1035 :
1036 : int
1037 11509 : os_del(objectset *os, sql_trans *tr, const char *name, sql_base *b)
1038 : {
1039 11509 : MT_lock_set(&os->lock);
1040 11509 : int res = os_del_(os, tr, name, b);
1041 11509 : MT_lock_unset(&os->lock);
1042 11509 : 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 18400790 : os_find_name(objectset *os, struct sql_trans *tr, const char *name)
1081 : {
1082 18400790 : if (!os)
1083 : return NULL;
1084 :
1085 18400788 : assert(os->unique);
1086 18400788 : versionhead *n = find_name(os, name);
1087 :
1088 18503517 : if (n) {
1089 17916557 : objectversion *ov = get_valid_object_name(tr, n->ov);
1090 17894816 : if (ov && os_atmc_get_state(ov) == active)
1091 17894469 : return ov->b;
1092 : }
1093 : return NULL;
1094 : }
1095 :
1096 : sql_base *
1097 155290 : os_find_id(objectset *os, struct sql_trans *tr, sqlid id)
1098 : {
1099 155290 : if (!os)
1100 : return NULL;
1101 155290 : versionhead *n = find_id(os, id);
1102 :
1103 155290 : if (n) {
1104 154769 : objectversion *ov = get_valid_object_id(tr, n->ov);
1105 154769 : if (ov && os_atmc_get_state(ov) == active)
1106 154765 : return ov->b;
1107 : }
1108 : return NULL;
1109 : }
1110 :
1111 : void
1112 2306379 : os_iterator(struct os_iter *oi, struct objectset *os, struct sql_trans *tr, const char *name /*optional*/)
1113 : {
1114 2306379 : *oi = (struct os_iter) {
1115 : .os = os,
1116 : .tr = tr,
1117 : .name = name,
1118 : };
1119 :
1120 2306379 : lock_reader(os);
1121 2307618 : if (name && os->name_map) {
1122 1032584 : int key = hash_key(name);
1123 1032584 : oi->n = (void*)os->name_map->buckets[key&(os->name_map->size-1)];
1124 : } else
1125 1275034 : oi->n = os->name_based_h;
1126 2307618 : unlock_reader(os);
1127 2307653 : }
1128 :
1129 : sql_base *
1130 4224738 : oi_next(struct os_iter *oi)
1131 : {
1132 4224738 : sql_base *b = NULL;
1133 :
1134 4224738 : if (oi->name) {
1135 2975438 : lock_reader(oi->os); /* intentionally outside of while loop */
1136 2975438 : if (oi->os->name_map) {
1137 2609338 : sql_hash_e *he = (void*)oi->n;
1138 :
1139 6465059 : for (; he && !b; he = he->chain) {
1140 3855721 : versionhead *n = he->value;
1141 :
1142 3855721 : if (n->ov->b->name && strcmp(n->ov->b->name, oi->name) == 0) {
1143 1583227 : objectversion *ov = n->ov;
1144 :
1145 1583227 : ov = get_valid_object_name(oi->tr, ov);
1146 1583227 : if (ov && os_atmc_get_state(ov) == active)
1147 1581837 : b = ov->b;
1148 : }
1149 : }
1150 2609338 : oi->n = (void*)he;
1151 : } else {
1152 366100 : versionhead *n = oi->n;
1153 :
1154 410962 : while (n && !b) {
1155 :
1156 44862 : if (n->ov->b->name && strcmp(n->ov->b->name, oi->name) == 0) {
1157 13320 : objectversion *ov = n->ov;
1158 :
1159 13320 : n = oi->n = n->next;
1160 13320 : ov = get_valid_object_name(oi->tr, ov);
1161 13320 : if (ov && os_atmc_get_state(ov) == active)
1162 13320 : b = ov->b;
1163 : } else {
1164 31542 : n = oi->n = n->next;
1165 : }
1166 : }
1167 : }
1168 2975438 : unlock_reader(oi->os);
1169 : } else {
1170 1249300 : versionhead *n = oi->n;
1171 :
1172 1249300 : lock_reader(oi->os); /* intentionally outside of while loop */
1173 2903781 : while (n && !b) {
1174 405272 : objectversion *ov = n->ov;
1175 405272 : n = oi->n = n->next;
1176 :
1177 405272 : ov = get_valid_object_id(oi->tr, ov);
1178 405236 : if (ov && os_atmc_get_state(ov) == active)
1179 380806 : b = ov->b;
1180 : }
1181 1249209 : unlock_reader(oi->os);
1182 : }
1183 4225250 : return b;
1184 : }
1185 :
1186 : bool
1187 3541 : os_obj_intransaction(objectset *os, struct sql_trans *tr, sql_base *b)
1188 : {
1189 3541 : versionhead *n = find_id(os, b->id);
1190 :
1191 3541 : if (n) {
1192 3541 : objectversion *ov = n->ov;
1193 :
1194 3541 : 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 177711 : os_has_changes(objectset *os, struct sql_trans *tr)
1203 : {
1204 177711 : versionhead *n = os->id_based_t;
1205 :
1206 177711 : if (n) {
1207 173340 : objectversion *ov = n->ov;
1208 :
1209 173340 : 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 : }
|