0 总结
- 三段释放原因:因为如果先释放锁,没有释放一些共享资源(比如pin住的buffer),别人拿到锁后发现我们仍然持有一些资源,就会有问题。所以三阶段释放主要是以锁为分界线,先释放锁保护的资源,在释放锁,在清理私有资源。这样可以保证别人拿到锁后,一定也能拿到对应的资源。
- 三段:先放pinned buffer、relation、dsm这些共享资源;再放锁;所有放其他会话看不到的私有资源。
1 资源随事务释放
三阶段释放是指ResourceOwnerRelease函数在使用时需要调用三次,按固定顺序调用每次删除特定的资源:
- RESOURCE_RELEASE_BEFORE_LOCKS
- RESOURCE_RELEASE_LOCKS
- RESOURCE_RELEASE_AFTER_LOCKS
typedef enum
{
RESOURCE_RELEASE_BEFORE_LOCKS = 1,
RESOURCE_RELEASE_LOCKS,
RESOURCE_RELEASE_AFTER_LOCKS,
} ResourceReleasePhase;
例如事务提交时CommitTransaction:
static void
CommitTransaction(void)
...
...
...
ResourceOwnerRelease(TopTransactionResourceOwner,
RESOURCE_RELEASE_BEFORE_LOCKS,
true, true);
/* Check we've released all buffer pins */
AtEOXact_Buffers(true);
/* Clean up the relation cache */
AtEOXact_RelationCache(true);
/*
* Make catalog changes visible to all backends. This has to happen after
* relcache references are dropped (see comments for
* AtEOXact_RelationCache), but before locks are released (if anyone is
* waiting for lock on a relation we've modified, we want them to know
* about the catalog change before they start using the relation).
*/
AtEOXact_Inval(true);
AtEOXact_MultiXact();
ResourceOwnerRelease(TopTransactionResourceOwner,
RESOURCE_RELEASE_LOCKS,
true, true);
ResourceOwnerRelease(TopTransactionResourceOwner,
RESOURCE_RELEASE_AFTER_LOCKS,
true, true);
...
...
其他几个事务控制函数资源释放时,也是按照相同的顺序分三阶段释放的:
2 为什么要分三阶段释放?
结论
因为如果先释放锁,没有释放一些共享资源(比如pin住的buffer),别人拿到锁后发现我们仍然持有一些资源,就会有问题。所以三阶段释放主要是以锁为分界线,先释放锁保护的资源,在释放锁,在清理私有资源。这样可以保证别人拿到锁后,一定也能拿到对应的资源。
- RESOURCE_RELEASE_BEFORE_LOCKS(预锁定阶段):需要释放对其他后端可见的资源(pinned buffers)。为了确保当我们释放另一个后端可能正在等待的锁时,它会看到我们已经完全退出了我们的事务。这是为了防止在释放锁之后,其他后端仍然看到我们持有的资源,从而可能导致数据不一致或其他问题。
- RESOURCE_RELEASE_LOCKS(锁定阶段):释放持有的所有锁。这是为了让其他可能正在等待这些锁的后端能够继续执行。
- RESOURCE_RELEASE_AFTER_LOCKS(后锁定阶段):后端内部的清理工作。释放那些只有后端自己知道的、不会影响其他后端的资源。
3 代码分析
下面这次提交后对resowner做了扩展性增强,代码逻辑没变但可读性有点差(PG17dev分支)
commit b8bff07daa85c837a2747b4d35cd5a27e73fb7b2
Author: Heikki Linnakangas <heikki.linnakangas@iki.fi>
Date: Wed Nov 8 13:30:50 2023 +0200
Make ResourceOwners more easily extensible.
围绕本篇主题,后面分析这次提交前代码(PG17前的逻辑)
static void
ResourceOwnerReleaseInternal(ResourceOwner owner,
ResourceReleasePhase phase,
bool isCommit,
bool isTopLevel)
{
ResourceOwner child;
ResourceOwner save;
ResourceReleaseCallbackItem *item;
ResourceReleaseCallbackItem *next;
Datum foundres;
resowner维护树形结构,递归释放自己的child资源。
for (child = owner->firstchild; child != NULL; child = child->nextchild)
ResourceOwnerReleaseInternal(child, phase, isCommit, isTopLevel);
...
3.1 第一阶段:RESOURCE_RELEASE_BEFORE_LOCKS
- 事务提交时不应该有正在进行中的io了,这里需要把标志位清理干净,避免影响后面事务使用这个buffer。
- 清理pinned的buffer
- 提交时不应该有pinned的了,会告警,回滚时就无所谓了,因为错误不一定出在哪,不一定给机会清理。
- 关闭打开的relation
- 提交时不应该有打开的了,会告警;回滚时同上。
- 关闭dsm段
- 关闭jit context
if (phase == RESOURCE_RELEASE_BEFORE_LOCKS)
{
while (ResourceArrayGetAny(&(owner->bufferioarr), &foundres))
{
Buffer res = DatumGetBuffer(foundres);
if (isCommit)
elog(PANIC, "lost track of buffer IO on buffer %d", res);
AbortBufferIO(res);
}
while (ResourceArrayGetAny(&(owner->bufferarr), &foundres))
{
Buffer res = DatumGetBuffer(foundres);
if (isCommit)
PrintBufferLeakWarning(res);
ReleaseBuffer(res);
}
while (ResourceArrayGetAny(&(owner->relrefarr), &foundres))
{
Relation res = (Relation) DatumGetPointer(foundres);
if (isCommit)
PrintRelCacheLeakWarning(res);
RelationClose(res);
}
while (ResourceArrayGetAny(&(owner->dsmarr), &foundres))
{
dsm_segment *res = (dsm_segment *) DatumGetPointer(foundres);
if (isCommit)
PrintDSMLeakWarning(res);
dsm_detach(res);
}
while (ResourceArrayGetAny(&(owner->jitarr), &foundres))
{
JitContext *context = (JitContext *) DatumGetPointer(foundres);
jit_release_context(context);
}
...
...
}
}
3.2 第二阶段:RESOURCE_RELEASE_LOCKS
- 如果是顶层事务,直接释放所有锁,具体:
- 提交时要保留会话锁,释放事务锁。事务没了会话锁还需要继续生效,生命周期比事务长。
- 回滚时要释放所有锁。
- 会话锁:咨询锁。
- 事务锁:行锁、表锁等。
- 如果是子事务,按提交回滚做出不同行为。
- 提交:转移锁到parent resowner。
- 回滚:释放锁。
else if (phase == RESOURCE_RELEASE_LOCKS)
{
if (isTopLevel)
{
if (owner == TopTransactionResourceOwner)
{
ProcReleaseLocks(isCommit);
ReleasePredicateLocks(isCommit, false);
}
}
else
{
LOCALLOCK **locks;
int nlocks;
...
if (isCommit)
LockReassignCurrentOwner(locks, nlocks);
else
LockReleaseCurrentOwner(locks, nlocks);
}
}
3.3 第三阶段:RESOURCE_RELEASE_AFTER_LOCKS
- 释放catcache系统表缓存。
- 释放cached plan。
- 释放tuple desc,注意tuple desc在构造好后,会用IncrTupleDescRefCount函数,在resowner中记录,按引用计数控制删除。引用计数的机制主要是为了处理TupleDesc在多个地方共享使用的情况。例如一个查询的多个部分可能都需要引用同一个TupleDesc。如果在一个部分结束时就直接删除TupleDesc,那么其他部分就无法继续使用这个TupleDesc了。
- 释放快照。
- 释放fd。
else if (phase == RESOURCE_RELEASE_AFTER_LOCKS)
{
while (ResourceArrayGetAny(&(owner->catrefarr), &foundres))
{
HeapTuple res = (HeapTuple) DatumGetPointer(foundres);
if (isCommit)
PrintCatCacheLeakWarning(res);
ReleaseCatCache(res);
}
while (ResourceArrayGetAny(&(owner->catlistrefarr), &foundres))
{
CatCList *res = (CatCList *) DatumGetPointer(foundres);
if (isCommit)
PrintCatCacheListLeakWarning(res);
ReleaseCatCacheList(res);
}
while (ResourceArrayGetAny(&(owner->planrefarr), &foundres))
{
CachedPlan *res = (CachedPlan *) DatumGetPointer(foundres);
if (isCommit)
PrintPlanCacheLeakWarning(res);
ReleaseCachedPlan(res, owner);
}
while (ResourceArrayGetAny(&(owner->tupdescarr), &foundres))
{
TupleDesc res = (TupleDesc) DatumGetPointer(foundres);
if (isCommit)
PrintTupleDescLeakWarning(res);
DecrTupleDescRefCount(res);
}
while (ResourceArrayGetAny(&(owner->snapshotarr), &foundres))
{
Snapshot res = (Snapshot) DatumGetPointer(foundres);
if (isCommit)
PrintSnapshotLeakWarning(res);
UnregisterSnapshot(res);
}
/* Ditto for temporary files */
while (ResourceArrayGetAny(&(owner->filearr), &foundres))
{
File res = DatumGetFile(foundres);
if (isCommit)
PrintFileLeakWarning(res);
FileClose(res);
}
}
for (item = ResourceRelease_callbacks; item; item = next)
{
next = item->next;
item->callback(phase, isCommit, isTopLevel, item->arg);
}
CurrentResourceOwner = save;
}