ISAR/libmdbx
3
0
Fork 0
mirror of https://github.com/isar/libmdbx.git synced 2025-11-07 07:18:56 +08:00
Code Issues Packages Projects Releases Wiki Activity

mdbx: парковка читающих транзакций.

Browse Source
This commit is contained in:
Леонид Юрьев (Leonid Yuriev)
2024-07-09 16:04:01 +03:00
parent f335a16c92
commit ec0ada7b8c
11 changed files with 399 additions and 80 deletions
Download Patch File Download Diff File Expand all files Collapse all files

143
src/txn.c
View File

@@ -453,8 +453,8 @@ static void take_gcprof(MDBX_txn *txn, MDBX_commit_latency *latency) {
}
int mdbx_txn_commit_ex(MDBX_txn *txn, MDBX_commit_latency *latency) {
STATIC_ASSERT(MDBX_TXN_FINISHED ==
MDBX_TXN_BLOCKED - MDBX_TXN_HAS_CHILD - MDBX_TXN_ERROR);
STATIC_ASSERT(MDBX_TXN_FINISHED == MDBX_TXN_BLOCKED - MDBX_TXN_HAS_CHILD -
MDBX_TXN_ERROR - MDBX_TXN_PARKED);
const uint64_t ts_0 = latency ? osal_monotime() : 0;
uint64_t ts_1 = 0, ts_2 = 0, ts_3 = 0, ts_4 = 0, ts_5 = 0, gc_cputime = 0;
@@ -919,7 +919,6 @@ int txn_renew(MDBX_txn *txn, unsigned flags) {
}
#endif /* MDBX_ENV_CHECKPID */
const uintptr_t tid = osal_thread_self();
flags |= env->flags & (MDBX_NOSTICKYTHREADS | MDBX_WRITEMAP);
if (flags & MDBX_TXN_RDONLY) {
eASSERT(env, (flags & ~(txn_ro_begin_flags | MDBX_WRITEMAP |
@@ -949,7 +948,7 @@ int txn_renew(MDBX_txn *txn, unsigned flags) {
r->txnid.weak < SAFE64_INVALID_THRESHOLD))
return MDBX_BAD_RSLOT;
} else if (env->lck_mmap.lck) {
bsr_t brs = mvcc_bind_slot(env, tid);
bsr_t brs = mvcc_bind_slot(env);
if (unlikely(brs.err != MDBX_SUCCESS))
return brs.err;
r = brs.rslot;
@@ -968,7 +967,11 @@ int txn_renew(MDBX_txn *txn, unsigned flags) {
txn->flags = MDBX_TXN_RDONLY | MDBX_TXN_FINISHED;
return MDBX_SUCCESS;
}
txn->owner = tid;
txn->owner = (uintptr_t)r->tid.weak;
if ((env->flags & MDBX_NOSTICKYTHREADS) == 0 && env->txn &&
unlikely(env->basal_txn->owner == txn->owner) &&
(globals.runtime_flags & MDBX_DBG_LEGACY_OVERLAP) == 0)
return MDBX_TXN_OVERLAPPING;
/* Seek & fetch the last meta */
uint64_t timestamp = 0;
@@ -980,7 +983,7 @@ int txn_renew(MDBX_txn *txn, unsigned flags) {
? /* regular */ meta_recent(env, &troika)
: /* recovery mode */ meta_ptr(env, env->stuck_meta);
if (likely(r)) {
safe64_reset(&r->txnid, false);
safe64_reset(&r->txnid, true);
atomic_store32(&r->snapshot_pages_used,
head.ptr_v->geometry.first_unallocated, mo_Relaxed);
atomic_store64(
@@ -1014,7 +1017,7 @@ int txn_renew(MDBX_txn *txn, unsigned flags) {
rc = MDBX_PROBLEM;
txn->txnid = INVALID_TXNID;
if (likely(r))
safe64_reset(&r->txnid, false);
safe64_reset(&r->txnid, true);
goto bailout;
}
timestamp = 0;
@@ -1029,7 +1032,7 @@ int txn_renew(MDBX_txn *txn, unsigned flags) {
if (unlikely(rc != MDBX_RESULT_TRUE)) {
txn->txnid = INVALID_TXNID;
if (likely(r))
safe64_reset(&r->txnid, false);
safe64_reset(&r->txnid, true);
goto bailout;
}
}
@@ -1037,7 +1040,7 @@ int txn_renew(MDBX_txn *txn, unsigned flags) {
if (unlikely(txn->txnid < MIN_TXNID || txn->txnid > MAX_TXNID)) {
ERROR("%s", "environment corrupted by died writer, must shutdown!");
if (likely(r))
safe64_reset(&r->txnid, false);
safe64_reset(&r->txnid, true);
txn->txnid = INVALID_TXNID;
rc = MDBX_CORRUPTED;
goto bailout;
@@ -1050,6 +1053,7 @@ int txn_renew(MDBX_txn *txn, unsigned flags) {
} else {
eASSERT(env, (flags & ~(txn_rw_begin_flags | MDBX_TXN_SPILLS |
MDBX_WRITEMAP | MDBX_NOSTICKYTHREADS)) == 0);
const uintptr_t tid = osal_thread_self();
if (unlikely(txn->owner == tid ||
/* not recovery mode */ env->stuck_meta >= 0))
return MDBX_BUSY;
@@ -1165,7 +1169,8 @@ int txn_renew(MDBX_txn *txn, unsigned flags) {
if (unlikely(env->dbs_flags[MAIN_DBI] !=
(DB_VALID | txn->dbs[MAIN_DBI].flags))) {
const bool need_txn_lock = env->basal_txn && env->basal_txn->owner != tid;
const bool need_txn_lock =
env->basal_txn && env->basal_txn->owner != osal_thread_self();
bool should_unlock = false;
if (need_txn_lock) {
rc = lck_txn_lock(env, true);
@@ -1330,7 +1335,7 @@ bailout:
return rc;
}
int txn_end(MDBX_txn *txn, const unsigned mode) {
int txn_end(MDBX_txn *txn, unsigned mode) {
MDBX_env *env = txn->env;
static const char *const names[] = TXN_END_NAMES;
@@ -1349,14 +1354,27 @@ int txn_end(MDBX_txn *txn, const unsigned mode) {
reader_slot_t *slot = txn->to.reader;
eASSERT(env, slot->pid.weak == env->pid);
if (likely(!(txn->flags & MDBX_TXN_FINISHED))) {
ENSURE(env, txn->txnid >=
/* paranoia is appropriate here */ env->lck
->cached_oldest.weak);
eASSERT(env, txn->txnid == slot->txnid.weak &&
slot->txnid.weak >= env->lck->cached_oldest.weak);
if (likely((txn->flags & MDBX_TXN_PARKED) == 0)) {
ENSURE(env, txn->txnid >=
/* paranoia is appropriate here */ env->lck
->cached_oldest.weak);
eASSERT(env, txn->txnid == slot->txnid.weak &&
slot->txnid.weak >= env->lck->cached_oldest.weak);
} else {
if ((mode & TXN_END_OUSTED) == 0 &&
safe64_read(&slot->tid) == MDBX_TID_TXN_OUSTED)
mode += TXN_END_OUSTED;
do {
safe64_reset(&slot->txnid, false);
atomic_store64(&slot->tid, txn->owner, mo_AcquireRelease);
atomic_yield();
} while (
unlikely(safe64_read(&slot->txnid) < SAFE64_INVALID_THRESHOLD ||
safe64_read(&slot->tid) != txn->owner));
}
dxb_sanitize_tail(env, nullptr);
atomic_store32(&slot->snapshot_pages_used, 0, mo_Relaxed);
safe64_reset(&slot->txnid, false);
safe64_reset(&slot->txnid, true);
atomic_store32(&env->lck->rdt_refresh_flag, true, mo_Relaxed);
} else {
eASSERT(env, slot->pid.weak == env->pid);
@@ -1373,7 +1391,9 @@ int txn_end(MDBX_txn *txn, const unsigned mode) {
imports.srwl_ReleaseShared(&env->remap_guard);
#endif
txn->n_dbi = 0; /* prevent further DBI activity */
txn->flags = MDBX_TXN_RDONLY | MDBX_TXN_FINISHED;
txn->flags = (mode & TXN_END_OUSTED)
? MDBX_TXN_RDONLY | MDBX_TXN_FINISHED | MDBX_TXN_OUSTED
: MDBX_TXN_RDONLY | MDBX_TXN_FINISHED;
txn->owner = 0;
} else if (!(txn->flags & MDBX_TXN_FINISHED)) {
ENSURE(env,
@@ -1483,16 +1503,17 @@ int mdbx_txn_renew(MDBX_txn *txn) {
if (unlikely((txn->flags & MDBX_TXN_RDONLY) == 0))
return MDBX_EINVAL;
int rc;
if (unlikely(txn->owner != 0 || !(txn->flags & MDBX_TXN_FINISHED))) {
rc = mdbx_txn_reset(txn);
int rc = mdbx_txn_reset(txn);
if (unlikely(rc != MDBX_SUCCESS))
return rc;
}
rc = txn_renew(txn, MDBX_TXN_RDONLY);
int rc = txn_renew(txn, MDBX_TXN_RDONLY);
if (rc == MDBX_SUCCESS) {
tASSERT(txn, txn->owner == osal_thread_self());
tASSERT(txn, txn->owner == (txn->flags & MDBX_NOSTICKYTHREADS)
? 0
: osal_thread_self());
DEBUG("renew txn %" PRIaTXN "%c %p on env %p, root page %" PRIaPGNO
"/%" PRIaPGNO,
txn->txnid, (txn->flags & MDBX_TXN_RDONLY) ? 'r' : 'w', (void *)txn,
@@ -1550,12 +1571,7 @@ int mdbx_txn_begin_ex(MDBX_env *env, MDBX_txn *parent, MDBX_txn_flags_t flags,
flags |= parent->flags & (txn_rw_begin_flags | MDBX_TXN_SPILLS |
MDBX_NOSTICKYTHREADS | MDBX_WRITEMAP);
} else if (flags & MDBX_TXN_RDONLY) {
if ((env->flags & MDBX_NOSTICKYTHREADS) == 0 && env->txn &&
unlikely(env->basal_txn->owner == osal_thread_self()) &&
(globals.runtime_flags & MDBX_DBG_LEGACY_OVERLAP) == 0)
return MDBX_TXN_OVERLAPPING;
} else {
} else if ((flags & MDBX_TXN_RDONLY) == 0) {
/* Reuse preallocated write txn. However, do not touch it until
* txn_renew() succeeds, since it currently may be active. */
txn = env->basal_txn;
@@ -1789,8 +1805,10 @@ int mdbx_txn_info(const MDBX_txn *txn, MDBX_txn_info *info, bool scan_rlt) {
info->txn_reader_lag = head.txnid - info->txn_id;
info->txn_space_dirty = info->txn_space_retired = 0;
uint64_t reader_snapshot_pages_retired;
uint64_t reader_snapshot_pages_retired = 0;
if (txn->to.reader &&
((txn->flags & MDBX_TXN_PARKED) == 0 ||
safe64_read(&txn->to.reader->tid) != MDBX_TID_TXN_OUSTED) &&
head_retired >
(reader_snapshot_pages_retired = atomic_load64(
&txn->to.reader->snapshot_pages_retired, mo_Relaxed))) {
@@ -1808,19 +1826,21 @@ int mdbx_txn_info(const MDBX_txn *txn, MDBX_txn_info *info, bool scan_rlt) {
retry:
if (atomic_load32(&lck->rdt[i].pid, mo_AcquireRelease)) {
jitter4testing(true);
const uint64_t snap_tid = safe64_read(&lck->rdt[i].tid);
const txnid_t snap_txnid = safe64_read(&lck->rdt[i].txnid);
const uint64_t snap_retired = atomic_load64(
&lck->rdt[i].snapshot_pages_retired, mo_AcquireRelease);
if (unlikely(snap_retired !=
atomic_load64(&lck->rdt[i].snapshot_pages_retired,
mo_Relaxed)) ||
snap_txnid != safe64_read(&lck->rdt[i].txnid))
snap_txnid != safe64_read(&lck->rdt[i].txnid) ||
snap_tid != safe64_read(&lck->rdt[i].tid))
goto retry;
if (snap_txnid <= txn->txnid) {
retired_next_reader = 0;
break;
}
if (snap_txnid < next_reader) {
if (snap_txnid < next_reader && snap_tid >= MDBX_TID_TXN_OUSTED) {
next_reader = snap_txnid;
retired_next_reader = pgno2bytes(
env, (pgno_t)(snap_retired -
@@ -1885,7 +1905,7 @@ uint64_t mdbx_txn_id(const MDBX_txn *txn) {
return txn->txnid;
}
int mdbx_txn_flags(const MDBX_txn *txn) {
MDBX_txn_flags_t mdbx_txn_flags(const MDBX_txn *txn) {
STATIC_ASSERT(
(MDBX_TXN_INVALID &
(MDBX_TXN_FINISHED | MDBX_TXN_ERROR | MDBX_TXN_DIRTY | MDBX_TXN_SPILLS |
@@ -1894,7 +1914,12 @@ int mdbx_txn_flags(const MDBX_txn *txn) {
if (unlikely(!txn || txn->signature != txn_signature))
return MDBX_TXN_INVALID;
assert(0 == (int)(txn->flags & MDBX_TXN_INVALID));
return txn->flags;
MDBX_txn_flags_t flags = txn->flags;
if (F_ISSET(flags, MDBX_TXN_PARKED | MDBX_TXN_RDONLY) && txn->to.reader &&
safe64_read(&txn->to.reader->tid) == MDBX_TID_TXN_OUSTED)
flags |= MDBX_TXN_OUSTED;
return flags;
}
int mdbx_txn_reset(MDBX_txn *txn) {
@@ -1946,3 +1971,55 @@ int mdbx_txn_abort(MDBX_txn *txn) {
return txn_abort(txn);
}
int mdbx_txn_park(MDBX_txn *txn, bool autounpark) {
STATIC_ASSERT(MDBX_TXN_BLOCKED > MDBX_TXN_ERROR);
int rc = check_txn(txn, MDBX_TXN_BLOCKED - MDBX_TXN_ERROR);
if (unlikely(rc != MDBX_SUCCESS))
return rc;
if (unlikely((txn->flags & MDBX_TXN_RDONLY) == 0))
return MDBX_TXN_INVALID;
if (unlikely((txn->flags & MDBX_TXN_ERROR))) {
rc = txn_end(txn, TXN_END_RESET | TXN_END_UPDATE);
return rc ? rc : MDBX_OUSTED;
}
return txn_park(txn, autounpark);
}
int mdbx_txn_unpark(MDBX_txn *txn, bool restart_if_ousted) {
STATIC_ASSERT(MDBX_TXN_BLOCKED > MDBX_TXN_PARKED + MDBX_TXN_ERROR);
int rc = check_txn(txn, MDBX_TXN_BLOCKED - MDBX_TXN_PARKED - MDBX_TXN_ERROR);
if (unlikely(rc != MDBX_SUCCESS))
return rc;
if (unlikely(!F_ISSET(txn->flags, MDBX_TXN_RDONLY | MDBX_TXN_PARKED)))
return MDBX_SUCCESS;
rc = txn_unpark(txn);
if (likely(rc != MDBX_OUSTED) || !restart_if_ousted)
return rc;
tASSERT(txn, txn->flags & MDBX_TXN_FINISHED);
rc = txn_renew(txn, MDBX_TXN_RDONLY);
return (rc == MDBX_SUCCESS) ? MDBX_RESULT_TRUE : rc;
}
int txn_check_badbits_parked(const MDBX_txn *txn, int bad_bits) {
tASSERT(txn, (bad_bits & MDBX_TXN_PARKED) && (txn->flags & bad_bits));
/* Здесь осознано заложено отличие в поведении припаркованных транзакций:
* - некоторые функции (например mdbx_env_info_ex()), допускают
* использование поломанных транзакций (с флагом MDBX_TXN_ERROR), но
* не могут работать с припаркованными транзакциями (требуют распарковки).
* - но при распарковке поломанные транзакции завершаются.
* - получается что транзакцию можно припарковать, потом поломать вызвав
* mdbx_txn_break(), но далее любое её использование приведет к завершению
* при распарковке. */
if ((txn->flags & (bad_bits | MDBX_TXN_AUTOUNPARK)) !=
(MDBX_TXN_PARKED | MDBX_TXN_AUTOUNPARK))
return MDBX_BAD_TXN;
tASSERT(txn, bad_bits == MDBX_TXN_BLOCKED ||
bad_bits == MDBX_TXN_BLOCKED - MDBX_TXN_ERROR);
return mdbx_txn_unpark((MDBX_txn *)txn, false);
}