Changeset: 92c641b07606 for MonetDB
URL: https://dev.monetdb.org/hg/MonetDB/rev/92c641b07606
Modified Files:
sql/storage/bat/bat_storage.c
sql/storage/bat/bat_storage.h
Branch: Jun2023
Log Message:
Turn segment next pointer into an atomic.
This seems to fix the occasional problem that we get a crash when
concurrent work (changes and lookups) are done on segment chains that
are due to the compiler/CPU doing things out of the written order (which
they may).
diffs (truncated from 371 to 300 lines):
diff --git a/sql/storage/bat/bat_storage.c b/sql/storage/bat/bat_storage.c
--- a/sql/storage/bat/bat_storage.c
+++ b/sql/storage/bat/bat_storage.c
@@ -121,6 +121,7 @@ tc_gc_seg( sql_store Store, sql_change *
if (s->ts <= oldest) {
while(s) {
segment *n = s->prev;
+ ATOMIC_PTR_DESTROY(&s->next);
_DELETE(s);
s = n;
}
@@ -158,10 +159,10 @@ new_segment(segment *o, sql_trans *tr, s
n->start = 0;
n->end = cnt;
}
- n->next = NULL;
+ ATOMIC_PTR_INIT(&n->next, NULL);
n->prev = NULL;
if (o)
- o->next = n;
+ ATOMIC_PTR_SET(&o->next, n);
}
return n;
}
@@ -197,19 +198,19 @@ split_segment(segments *segs, segment *o
* inserted before */
n->start = o->start;
n->end = n->start + cnt;
- n->next = o;
+ ATOMIC_PTR_INIT(&n->next, o);
if (segs->h == o)
segs->h = n;
if (p)
- p->next = n;
+ ATOMIC_PTR_SET(&p->next, n);
o->start = n->end;
} else if (start+cnt == o->end) {
/* 2-way split: o remains first part of segment, new one is
* added after */
n->start = o->end - cnt;
n->end = o->end;
- n->next = o->next;
- o->next = n;
+ ATOMIC_PTR_INIT(&n->next, ATOMIC_PTR_GET(&o->next));
+ ATOMIC_PTR_SET(&o->next, n);
if (segs->t == o)
segs->t = n;
o->end = n->start;
@@ -221,15 +222,16 @@ split_segment(segments *segs, segment *o
GDKfree(n);
return NULL;
}
- n->next = n2;
+ ATOMIC_PTR_INIT(&n->next, n2);
n->start = start;
n->end = start + cnt;
*n2 = *o;
+ ATOMIC_PTR_INIT(&n2->next, ATOMIC_PTR_GET(&o->next));
n2->start = n->end;
n2->prev = NULL;
if (segs->t == o)
segs->t = n2;
- o->next = n;
+ ATOMIC_PTR_SET(&o->next, n);
o->end = start;
}
return n;
@@ -239,7 +241,7 @@ static void
rollback_segments(segments *segs, sql_trans *tr, sql_change *change, ulng
oldest)
{
segment *cur = segs->h, *seg = NULL;
- for (; cur; cur = cur->next) {
+ for (; cur; cur = ATOMIC_PTR_GET(&cur->next)) {
if (cur->ts == tr->tid) { /* revert */
cur->deleted = !cur->deleted || (cur->ts == cur->oldts);
cur->ts = cur->oldts==tr->tid?0:cur->oldts; /* need old
ts */
@@ -251,7 +253,7 @@ rollback_segments(segments *segs, sql_tr
} else if (seg->end == cur->start && seg->deleted ==
cur->deleted) {
/* merge with previous */
seg->end = cur->end;
- seg->next = cur->next;
+ ATOMIC_PTR_SET(&seg->next,
ATOMIC_PTR_GET(&cur->next));
if (cur == segs->t)
segs->t = seg;
mark4destroy(cur, change,
store_get_timestamp(tr->store));
@@ -269,7 +271,7 @@ segs_end_include_deleted( segments *segs
size_t cnt = 0;
segment *s = segs->h, *l = NULL;
- for(;s; s = s->next) {
+ for(;s; s = ATOMIC_PTR_GET(&s->next)) {
if (s->ts == tr->tid || SEG_IS_VALID(s, tr))
l = s;
}
@@ -312,7 +314,7 @@ segments2cs(sql_trans *tr, segments *seg
uint32_t *restrict dst;
/* why hashlock ?? */
MT_rwlock_wrlock(&b->thashlock);
- for (; s ; s=s->next) {
+ for (; s ; s=ATOMIC_PTR_GET(&s->next)) {
if (s->start >= nr)
break;
if (s->ts == tr->tid && s->end != s->start) {
@@ -395,7 +397,7 @@ merge_segments(storage *s, sql_trans *tr
{
sqlstore* store = tr->store;
segment *cur = s->segs->h, *seg = NULL;
- for (; cur; cur = cur->next) {
+ for (; cur; cur = ATOMIC_PTR_GET(&cur->next)) {
if (cur->ts == tr->tid) {
if (!cur->deleted)
cur->oldts = 0;
@@ -430,12 +432,13 @@ merge_segments(storage *s, sql_trans *tr
/* merge segments */
if (merge) {
seg->end = cur->end;
- seg->next = cur->next;
+ ATOMIC_PTR_SET(&seg->next,
ATOMIC_PTR_GET(&cur->next));
if (cur == s->segs->t)
s->segs->t = seg;
- if (commit_ts == oldest)
+ if (commit_ts == oldest) {
+ ATOMIC_PTR_DESTROY(&cur->next);
_DELETE(cur);
- else
+ } else
mark4destroy(cur, change,
commit_ts);
cur = seg;
continue;
@@ -454,7 +457,7 @@ segments_in_transaction(sql_trans *tr, s
if (seg && s->segs->t->ts == tr->tid)
return 1;
- for (; seg ; seg=seg->next) {
+ for (; seg ; seg=ATOMIC_PTR_GET(&seg->next)) {
if (seg->ts == tr->tid)
return 1;
}
@@ -472,7 +475,7 @@ segs_end( segments *segs, sql_trans *tr,
if (segs->t && SEG_IS_VALID(segs->t, tr))
l = s = segs->t;
- for(;s; s = s->next) {
+ for(;s; s = ATOMIC_PTR_GET(&s->next)) {
if (SEG_IS_VALID(s, tr))
l = s;
}
@@ -573,7 +576,7 @@ count_inserts( segment *s, sql_trans *tr
{
size_t cnt = 0;
- for(;s; s = s->next) {
+ for(;s; s = ATOMIC_PTR_GET(&s->next)) {
if (!s->deleted && s->ts == tr->tid)
cnt += s->end - s->start;
}
@@ -585,10 +588,10 @@ count_deletes_in_range( segment *s, sql_
{
size_t cnt = 0;
- for(;s && s->end <= start; s = s->next)
+ for(;s && s->end <= start; s = ATOMIC_PTR_GET(&s->next))
;
- for(;s && s->start < end; s = s->next) {
+ for(;s && s->start < end; s = ATOMIC_PTR_GET(&s->next)) {
if (SEG_IS_DELETED(s, tr)) /* assume aligned s->end and end */
cnt += s->end - s->start;
}
@@ -600,7 +603,7 @@ count_deletes( segment *s, sql_trans *tr
{
size_t cnt = 0;
- for(;s; s = s->next) {
+ for(;s; s = ATOMIC_PTR_GET(&s->next)) {
if (SEG_IS_DELETED(s, tr))
cnt += s->end - s->start;
}
@@ -1043,7 +1046,7 @@ cs_real_update_bats( column_storage *cs,
static int
segments_is_append(segment *s, sql_trans *tr, oid rid)
{
- for(; s; s=s->next) {
+ for(; s; s=ATOMIC_PTR_GET(&s->next)) {
if (s->start <= rid && s->end > rid) {
if (s->ts == tr->tid && !s->deleted) {
return 1;
@@ -1057,7 +1060,7 @@ segments_is_append(segment *s, sql_trans
static int
segments_is_deleted(segment *s, sql_trans *tr, oid rid)
{
- for(; s; s=s->next) {
+ for(; s; s=ATOMIC_PTR_GET(&s->next)) {
if (s->start <= rid && s->end > rid) {
if (s->ts >= tr->ts && s->deleted) {
return 1;
@@ -1350,7 +1353,7 @@ cs_update_bat( sql_trans *tr, sql_delta
oid start = tids->tseqbase, offset = start;
oid end = start + ucnt;
- for(segment *seg = s->segs->h; seg && res == LOG_OK ;
seg=seg->next) {
+ for(segment *seg = s->segs->h; seg && res == LOG_OK ;
seg=ATOMIC_PTR_GET(&seg->next)) {
if (seg->start <= start && seg->end > start) {
/* check for delete conflicts */
if (seg->ts >= tr->ts && seg->deleted) {
@@ -1395,7 +1398,7 @@ cs_update_bat( sql_trans *tr, sql_delta
while ( seg && res == LOG_OK && i < ucnt) {
oid rid = canditer_next(&ci);
if (seg->end <= rid)
- seg = seg->next;
+ seg = ATOMIC_PTR_GET(&seg->next);
else if (seg->start <= rid && seg->end > rid) {
/* check for delete conflicts */
if (seg->ts >= tr->ts && seg->deleted) {
@@ -1435,7 +1438,7 @@ cs_update_bat( sql_trans *tr, sql_delta
segment *seg = s->segs->h;
while ( seg && res == LOG_OK && i < ucnt) {
if (seg->end <= rid[i])
- seg = seg->next;
+ seg = ATOMIC_PTR_GET(&seg->next);
else if (seg->start <= rid[i] && seg->end >
rid[i]) {
/* check for delete conflicts */
if (seg->ts >= tr->ts && seg->deleted) {
@@ -2340,7 +2343,7 @@ storage_delete_val(sql_trans *tr, sql_ta
lock_table(tr->store, t->base.id);
/* find segment of rid, split, mark new segment deleted (for tr->tid) */
segment *seg = s->segs->h, *p = NULL;
- for (; seg; p = seg, seg = seg->next) {
+ for (; seg; p = seg, seg = ATOMIC_PTR_GET(&seg->next)) {
if (seg->start <= rid && seg->end > rid) {
if (!SEG_VALID_4_DELETE(seg,tr)) {
unlock_table(tr->store, t->base.id);
@@ -2367,7 +2370,7 @@ static int
seg_delete_range(sql_trans *tr, sql_table *t, storage *s, segment **Seg,
size_t start, size_t cnt)
{
segment *seg = *Seg, *p = NULL;
- for (; seg; p = seg, seg = seg->next) {
+ for (; seg; p = seg, seg = ATOMIC_PTR_GET(&seg->next)) {
if (seg->start <= start && seg->end > start) {
size_t lcnt = cnt;
if (start+lcnt > seg->end)
@@ -2489,7 +2492,8 @@ destroy_segments(segments *s)
return;
segment *seg = s->h;
while(seg) {
- segment *n = seg->next;
+ segment *n = ATOMIC_PTR_GET(&seg->next);
+ ATOMIC_PTR_DESTROY(&seg->next);
_DELETE(seg);
seg = n;
}
@@ -2518,11 +2522,11 @@ static int
segments_conflict(sql_trans *tr, segments *segs, int uncommitted)
{
if (uncommitted) {
- for (segment *s = segs->h; s; s = s->next)
+ for (segment *s = segs->h; s; s = ATOMIC_PTR_GET(&s->next))
if (!VALID_4_READ(s->ts,tr))
return 1;
} else {
- for (segment *s = segs->h; s; s = s->next)
+ for (segment *s = segs->h; s; s = ATOMIC_PTR_GET(&s->next))
if (s->ts < TRANSACTION_ID_BASE &&
!VALID_4_READ(s->ts,tr))
return 1;
}
@@ -2736,7 +2740,7 @@ count_segs(segment *s)
{
size_t nr = 0;
- for( ; s; s = s->next)
+ for( ; s; s = ATOMIC_PTR_GET(&s->next))
nr++;
return nr;
}
@@ -3272,7 +3276,7 @@ load_storage(sql_trans *tr, sql_table *t
bat_iterator_end(&bi);
}
if (ok == LOG_OK)
- for (segment *seg = s->segs->h; seg; seg = seg->next)
+ for (segment *seg = s->segs->h; seg; seg =
ATOMIC_PTR_GET(&seg->next))
if (seg->ts == tr->tid)
seg->ts = 1;
} else {
@@ -3349,7 +3353,7 @@ log_segments(sql_trans *tr, segments *se
{
/* log segments */
lock_table(tr->store, id);
- for (segment *seg = segs->h; seg; seg=seg->next) {
+ for (segment *seg = segs->h; seg; seg=ATOMIC_PTR_GET(&seg->next)) {
unlock_table(tr->store, id);
if (seg->ts == tr->tid && seg->end-seg->start) {
if (log_segment(tr, seg, id) != LOG_OK) {
@@ -3870,7 +3874,7 @@ log_table_append(sql_trans *tr, sql_tabl
size_t nr_appends = 0;
lock_table(tr->store, t->base.id);
- for (segment *seg = segs->h; seg; seg=seg->next) {
+ for (segment *seg = segs->h; seg; seg=ATOMIC_PTR_GET(&seg->next)) {
unlock_table(tr->store, t->base.id);
_______________________________________________
checkin-list mailing list -- checkin-list@monetdb.org
To unsubscribe send an email to checkin-list-le...@monetdb.org
