ISAR/libmdbx
3
0
Fork 0
mirror of https://github.com/isar/libmdbx.git synced 2025-01-06 19:24:13 +08:00
Code Issues Packages Projects Releases Wiki Activity

mdbx: доработка поддержки авто-слияния записей GC внутри page_alloc_slowpath().

Browse Source
This commit is contained in:
Леонид Юрьев (Leonid Yuriev) 2022-11-19 18:17:48 +03:00
parent c46c03e7c8
commit 543e52730d
1 changed files with 72 additions and 38 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

110
src/core.c
View File

@ -6595,7 +6595,8 @@ static pgno_t *scan4seq_resolver(pgno_t *range, const size_t len,
#define MDBX_ALLOC_RESERVE 16 #define MDBX_ALLOC_RESERVE 16
#define MDBX_ALLOC_BACKLOG 32 #define MDBX_ALLOC_BACKLOG 32
#define MDBX_ALLOC_ALL (MDBX_ALLOC_GC | MDBX_ALLOC_NEW) #define MDBX_ALLOC_ALL (MDBX_ALLOC_GC | MDBX_ALLOC_NEW)
#define MDBX_ALLOC_LIFO 128 #define MDBX_ALLOC_SHOULD_SCAN 64 /* internal state */
#define MDBX_ALLOC_LIFO 128 /* internal state */
static __inline bool is_gc_usable(const MDBX_txn *txn) { static __inline bool is_gc_usable(const MDBX_txn *txn) {
/* If txn is updating the GC, then the retired-list cannot play catch-up with /* If txn is updating the GC, then the retired-list cannot play catch-up with
@ -6692,13 +6693,18 @@ static pgr_t page_alloc_slowpath(const MDBX_cursor *mc, const size_t num,
if (unlikely(!is_gc_usable(txn))) if (unlikely(!is_gc_usable(txn)))
goto no_gc; goto no_gc;
eASSERT(env, (flags & (MDBX_ALLOC_COALESCE | MDBX_ALLOC_LIFO)) == 0); eASSERT(env, (flags & (MDBX_ALLOC_COALESCE | MDBX_ALLOC_LIFO |
MDBX_ALLOC_SHOULD_SCAN)) == 0);
flags += (env->me_flags & MDBX_LIFORECLAIM) ? MDBX_ALLOC_LIFO : 0; flags += (env->me_flags & MDBX_LIFORECLAIM) ? MDBX_ALLOC_LIFO : 0;
const unsigned coalesce_threshold = env->me_maxgc_ov1page >> 2; if (/* Не коагулируем записи при подготовке резерва для обновления GC.
if (txn->mt_dbs[FREE_DBI].md_branch_pages && * Иначе попытка увеличить резерв может приводить к необходимости ещё
MDBX_PNL_GETSIZE(txn->tw.relist) < coalesce_threshold && num) * большего резерва из-за увеличения списка переработанных страниц. */
flags += MDBX_ALLOC_COALESCE; flags < MDBX_ALLOC_COALESCE) {
if (txn->mt_dbs[FREE_DBI].md_branch_pages &&
re_len < env->me_maxgc_ov1page / 2)
flags += MDBX_ALLOC_COALESCE;
}
MDBX_cursor recur; MDBX_cursor recur;
ret.err = gc_cursor_init(&recur, txn); ret.err = gc_cursor_init(&recur, txn);
@ -6718,7 +6724,6 @@ retry_gc_have_oldest:
const txnid_t detent = oldest + 1; const txnid_t detent = oldest + 1;
txnid_t id = 0; txnid_t id = 0;
bool should_scan = false;
MDBX_cursor_op op = MDBX_FIRST; MDBX_cursor_op op = MDBX_FIRST;
if (flags & MDBX_ALLOC_LIFO) { if (flags & MDBX_ALLOC_LIFO) {
if (!txn->tw.lifo_reclaimed) { if (!txn->tw.lifo_reclaimed) {
@ -6790,24 +6795,54 @@ next_gc:;
ret.err = MDBX_CORRUPTED; ret.err = MDBX_CORRUPTED;
goto fail; goto fail;
} }
const size_t gc_len = MDBX_PNL_GETSIZE(gc_pnl); const size_t gc_len = MDBX_PNL_GETSIZE(gc_pnl);
if (unlikely(/* list is too long already */ MDBX_PNL_GETSIZE( TRACE("gc-read: id #%" PRIaTXN " len %zu, re-list will %zu ", id, gc_len,
txn->tw.relist) >= env->me_options.rp_augment_limit) && gc_len + re_len);
((/* not a slot-request from gc-update */
(flags & MDBX_ALLOC_SLOT) == 0 && eASSERT(env, re_len == MDBX_PNL_GETSIZE(txn->tw.relist));
/* have enough unallocated space */ txn->mt_geo.upper >= if (unlikely(gc_len + re_len >= env->me_maxgc_ov1page)) {
txn->mt_next_pgno + num) || /* Don't try to coalesce too much. */
gc_len + MDBX_PNL_GETSIZE(txn->tw.relist) >= MDBX_PGL_LIMIT)) { if (flags & MDBX_ALLOC_SHOULD_SCAN) {
/* Stop reclaiming to avoid large/overflow the page list. eASSERT(env, flags & MDBX_ALLOC_COALESCE);
* This is a rare case while search for a continuously multi-page region eASSERT(env, num > 0);
* in a large database. #if MDBX_ENABLE_PROFGC
* https://web.archive.org/web/https://github.com/erthink/libmdbx/issues/123 env->me_lck->mti_pgop_stat.gc_prof.coalescences += 1;
*/ #endif /* MDBX_ENABLE_PROFGC */
NOTICE("stop reclaiming to avoid PNL overflow: %zu (current) + %zu " TRACE("clear %s %s", "MDBX_ALLOC_COALESCE", "since got threshold");
"(chunk) -> %zu", if (re_len >= num) {
MDBX_PNL_GETSIZE(txn->tw.relist), gc_len, eASSERT(env, MDBX_PNL_LAST(txn->tw.relist) < txn->mt_next_pgno &&
gc_len + MDBX_PNL_GETSIZE(txn->tw.relist)); MDBX_PNL_FIRST(txn->tw.relist) < txn->mt_next_pgno);
goto depleted_gc; range = txn->tw.relist + (MDBX_PNL_ASCENDING ? 1 : re_len);
pgno = *range;
if (num == 1)
goto done;
range = scan4seq(range, re_len, num - 1);
eASSERT(env, range == scan4range_checker(txn->tw.relist, num - 1));
if (likely(range)) {
pgno = *range;
goto done;
}
}
flags -= MDBX_ALLOC_COALESCE | MDBX_ALLOC_SHOULD_SCAN;
}
if (unlikely(/* list is too long already */ re_len >=
env->me_options.rp_augment_limit) &&
((/* not a slot-request from gc-update */
(flags & MDBX_ALLOC_SLOT) == 0 &&
/* have enough unallocated space */ txn->mt_geo.upper >=
txn->mt_next_pgno + num) ||
gc_len + re_len >= MDBX_PGL_LIMIT)) {
/* Stop reclaiming to avoid large/overflow the page list.
* This is a rare case while search for a continuously multi-page region
* in a large database.
* https://web.archive.org/web/https://github.com/erthink/libmdbx/issues/123
*/
NOTICE("stop reclaiming to avoid PNL overflow: %zu (current) + %zu "
"(chunk) -> %zu",
re_len, gc_len, gc_len + re_len);
goto depleted_gc;
}
} }
/* Remember ID of readed GC record */ /* Remember ID of readed GC record */
@ -6834,7 +6869,7 @@ next_gc:;
/* Merge in descending sorted order */ /* Merge in descending sorted order */
re_len = pnl_merge(txn->tw.relist, gc_pnl); re_len = pnl_merge(txn->tw.relist, gc_pnl);
should_scan = true; flags |= MDBX_ALLOC_SHOULD_SCAN;
if (AUDIT_ENABLED()) { if (AUDIT_ENABLED()) {
if (unlikely(!pnl_check(txn->tw.relist, txn->mt_next_pgno))) { if (unlikely(!pnl_check(txn->tw.relist, txn->mt_next_pgno))) {
ret.err = MDBX_CORRUPTED; ret.err = MDBX_CORRUPTED;
@ -6860,26 +6895,22 @@ next_gc:;
/* Done for a kick-reclaim mode, actually no page needed */ /* Done for a kick-reclaim mode, actually no page needed */
if (unlikely(flags & MDBX_ALLOC_SLOT)) { if (unlikely(flags & MDBX_ALLOC_SLOT)) {
eASSERT(env, ret.err == MDBX_SUCCESS); eASSERT(env, ret.err == MDBX_SUCCESS);
TRACE("%s: last id #%" PRIaTXN ", re-len %zu", "early-exit for slot", id,
re_len);
goto early_exit; goto early_exit;
} }
/* TODO: delete reclaimed records */ /* TODO: delete reclaimed records */
/* Don't try to coalesce too much. */
eASSERT(env, op == MDBX_PREV || op == MDBX_NEXT); eASSERT(env, op == MDBX_PREV || op == MDBX_NEXT);
if (flags & MDBX_ALLOC_COALESCE) { if (flags & MDBX_ALLOC_COALESCE) {
if (re_len /* current size */ < coalesce_threshold) { TRACE("%s: last id #%" PRIaTXN ", re-len %zu", "coalesce-continue", id,
#if MDBX_ENABLE_PROFGC re_len);
env->me_lck->mti_pgop_stat.gc_prof.coalescences += 1; goto next_gc;
#endif /* MDBX_ENABLE_PROFGC */
goto next_gc;
}
TRACE("clear %s %s", "MDBX_ALLOC_COALESCE", "since got threshold");
flags &= ~MDBX_ALLOC_COALESCE;
} }
scan: scan:
eASSERT(env, should_scan); eASSERT(env, flags & MDBX_ALLOC_SHOULD_SCAN);
if (re_len >= num) { if (re_len >= num) {
eASSERT(env, MDBX_PNL_LAST(txn->tw.relist) < txn->mt_next_pgno && eASSERT(env, MDBX_PNL_LAST(txn->tw.relist) < txn->mt_next_pgno &&
MDBX_PNL_FIRST(txn->tw.relist) < txn->mt_next_pgno); MDBX_PNL_FIRST(txn->tw.relist) < txn->mt_next_pgno);
@ -6894,13 +6925,16 @@ scan:
goto done; goto done;
} }
} }
should_scan = false; flags -= MDBX_ALLOC_SHOULD_SCAN;
if (ret.err == MDBX_SUCCESS) if (ret.err == MDBX_SUCCESS) {
TRACE("%s: last id #%" PRIaTXN ", re-len %zu", "continue-search", id,
re_len);
goto next_gc; goto next_gc;
}
depleted_gc: depleted_gc:
ret.err = MDBX_NOTFOUND; ret.err = MDBX_NOTFOUND;
if (should_scan) if (flags & MDBX_ALLOC_SHOULD_SCAN)
goto scan; goto scan;
//------------------------------------------------------------------------- //-------------------------------------------------------------------------