inval.c

PostgreSQL 8.1.4的源码 适用于Linux下的开源数据库系统

/*-------------------------------------------------------------------------
 *
 * inval.c
 *	  POSTGRES cache invalidation dispatcher code.
 *
 *	This is subtle stuff, so pay attention:
 *
 *	When a tuple is updated or deleted, our standard time qualification rules
 *	consider that it is *still valid* so long as we are in the same command,
 *	ie, until the next CommandCounterIncrement() or transaction commit.
 *	(See utils/time/tqual.c, and note that system catalogs are generally
 *	scanned under SnapshotNow rules by the system, or plain user snapshots
 *	for user queries.)	At the command boundary, the old tuple stops
 *	being valid and the new version, if any, becomes valid.  Therefore,
 *	we cannot simply flush a tuple from the system caches during heap_update()
 *	or heap_delete().  The tuple is still good at that point; what's more,
 *	even if we did flush it, it might be reloaded into the caches by a later
 *	request in the same command.  So the correct behavior is to keep a list
 *	of outdated (updated/deleted) tuples and then do the required cache
 *	flushes at the next command boundary.  We must also keep track of
 *	inserted tuples so that we can flush "negative" cache entries that match
 *	the new tuples; again, that mustn't happen until end of command.
 *
 *	Once we have finished the command, we still need to remember inserted
 *	tuples (including new versions of updated tuples), so that we can flush
 *	them from the caches if we abort the transaction.  Similarly, we'd better
 *	be able to flush "negative" cache entries that may have been loaded in
 *	place of deleted tuples, so we still need the deleted ones too.
 *
 *	If we successfully complete the transaction, we have to broadcast all
 *	these invalidation events to other backends (via the SI message queue)
 *	so that they can flush obsolete entries from their caches.	Note we have
 *	to record the transaction commit before sending SI messages, otherwise
 *	the other backends won't see our updated tuples as good.
 *
 *	When a subtransaction aborts, we can process and discard any events
 *	it has queued.	When a subtransaction commits, we just add its events
 *	to the pending lists of the parent transaction.
 *
 *	In short, we need to remember until xact end every insert or delete
 *	of a tuple that might be in the system caches.	Updates are treated as
 *	two events, delete + insert, for simplicity.  (There are cases where
 *	it'd be possible to record just one event, but we don't currently try.)
 *
 *	We do not need to register EVERY tuple operation in this way, just those
 *	on tuples in relations that have associated catcaches.	We do, however,
 *	have to register every operation on every tuple that *could* be in a
 *	catcache, whether or not it currently is in our cache.	Also, if the
 *	tuple is in a relation that has multiple catcaches, we need to register
 *	an invalidation message for each such catcache.  catcache.c's
 *	PrepareToInvalidateCacheTuple() routine provides the knowledge of which
 *	catcaches may need invalidation for a given tuple.
 *
 *	Also, whenever we see an operation on a pg_class or pg_attribute tuple,
 *	we register a relcache flush operation for the relation described by that
 *	tuple.	pg_class updates trigger an smgr flush operation as well.
 *
 *	We keep the relcache and smgr flush requests in lists separate from the
 *	catcache tuple flush requests.	This allows us to issue all the pending
 *	catcache flushes before we issue relcache flushes, which saves us from
 *	loading a catcache tuple during relcache load only to flush it again
 *	right away.  Also, we avoid queuing multiple relcache flush requests for
 *	the same relation, since a relcache flush is relatively expensive to do.
 *	(XXX is it worth testing likewise for duplicate catcache flush entries?
 *	Probably not.)
 *
 *	If a relcache flush is issued for a system relation that we preload
 *	from the relcache init file, we must also delete the init file so that
 *	it will be rebuilt during the next backend restart.  The actual work of
 *	manipulating the init file is in relcache.c, but we keep track of the
 *	need for it here.
 *
 *	The request lists proper are kept in CurTransactionContext of their
 *	creating (sub)transaction, since they can be forgotten on abort of that
 *	transaction but must be kept till top-level commit otherwise.  For
 *	simplicity we keep the controlling list-of-lists in TopTransactionContext.
 *
 *
 * Portions Copyright (c) 1996-2005, PostgreSQL Global Development Group
 * Portions Copyright (c) 1994, Regents of the University of California
 *
 * IDENTIFICATION
 *	  $PostgreSQL: pgsql/src/backend/utils/cache/inval.c,v 1.73.2.2 2006/01/19 21:49:30 tgl Exp $
 *
 *-------------------------------------------------------------------------
 */
#include "postgres.h"

#include "access/twophase_rmgr.h"
#include "access/xact.h"
#include "catalog/catalog.h"
#include "miscadmin.h"
#include "storage/sinval.h"
#include "storage/smgr.h"
#include "utils/catcache.h"
#include "utils/inval.h"
#include "utils/memutils.h"
#include "utils/relcache.h"
#include "utils/syscache.h"


/*
 * To minimize palloc traffic, we keep pending requests in successively-
 * larger chunks (a slightly more sophisticated version of an expansible
 * array).	All request types can be stored as SharedInvalidationMessage
 * records.  The ordering of requests within a list is never significant.
 */
typedef struct InvalidationChunk
{
	struct InvalidationChunk *next;		/* list link */
	int			nitems;			/* # items currently stored in chunk */
	int			maxitems;		/* size of allocated array in this chunk */
	SharedInvalidationMessage msgs[1];	/* VARIABLE LENGTH ARRAY */
} InvalidationChunk;			/* VARIABLE LENGTH STRUCTURE */

typedef struct InvalidationListHeader
{
	InvalidationChunk *cclist;	/* list of chunks holding catcache msgs */
	InvalidationChunk *rclist;	/* list of chunks holding relcache/smgr msgs */
} InvalidationListHeader;

/*----------------
 * Invalidation info is divided into two lists:
 *	1) events so far in current command, not yet reflected to caches.
 *	2) events in previous commands of current transaction; these have
 *	   been reflected to local caches, and must be either broadcast to
 *	   other backends or rolled back from local cache when we commit
 *	   or abort the transaction.
 * Actually, we need two such lists for each level of nested transaction,
 * so that we can discard events from an aborted subtransaction.  When
 * a subtransaction commits, we append its lists to the parent's lists.
 *
 * The relcache-file-invalidated flag can just be a simple boolean,
 * since we only act on it at transaction commit; we don't care which
 * command of the transaction set it.
 *----------------
 */

typedef struct TransInvalidationInfo
{
	/* Back link to parent transaction's info */
	struct TransInvalidationInfo *parent;

	/* Subtransaction nesting depth */
	int			my_level;

	/* head of current-command event list */
	InvalidationListHeader CurrentCmdInvalidMsgs;

	/* head of previous-commands event list */
	InvalidationListHeader PriorCmdInvalidMsgs;

	/* init file must be invalidated? */
	bool		RelcacheInitFileInval;
} TransInvalidationInfo;

static TransInvalidationInfo *transInvalInfo = NULL;

/*
 * Dynamically-registered callback functions.  Current implementation
 * assumes there won't be very many of these at once; could improve if needed.
 */

#define MAX_CACHE_CALLBACKS 20

static struct CACHECALLBACK
{
	int16		id;				/* cache number or message type id */
	CacheCallbackFunction function;
	Datum		arg;
}	cache_callback_list[MAX_CACHE_CALLBACKS];

static int	cache_callback_count = 0;

/* info values for 2PC callback */
#define TWOPHASE_INFO_MSG			0	/* SharedInvalidationMessage */
#define TWOPHASE_INFO_FILE_BEFORE	1	/* relcache file inval */
#define TWOPHASE_INFO_FILE_AFTER	2	/* relcache file inval */

static void PersistInvalidationMessage(SharedInvalidationMessage *msg);


/* ----------------------------------------------------------------
 *				Invalidation list support functions
 *
 * These three routines encapsulate processing of the "chunked"
 * representation of what is logically just a list of messages.
 * ----------------------------------------------------------------
 */

/*
 * AddInvalidationMessage
 *		Add an invalidation message to a list (of chunks).
 *
 * Note that we do not pay any great attention to maintaining the original
 * ordering of the messages.
 */
static void
AddInvalidationMessage(InvalidationChunk **listHdr,
					   SharedInvalidationMessage *msg)
{
	InvalidationChunk *chunk = *listHdr;

	if (chunk == NULL)
	{
		/* First time through; create initial chunk */
#define FIRSTCHUNKSIZE 16
		chunk = (InvalidationChunk *)
			MemoryContextAlloc(CurTransactionContext,
							   sizeof(InvalidationChunk) +
					(FIRSTCHUNKSIZE - 1) *sizeof(SharedInvalidationMessage));
		chunk->nitems = 0;
		chunk->maxitems = FIRSTCHUNKSIZE;
		chunk->next = *listHdr;
		*listHdr = chunk;
	}
	else if (chunk->nitems >= chunk->maxitems)
	{
		/* Need another chunk; double size of last chunk */
		int			chunksize = 2 * chunk->maxitems;

		chunk = (InvalidationChunk *)
			MemoryContextAlloc(CurTransactionContext,
							   sizeof(InvalidationChunk) +
						 (chunksize - 1) *sizeof(SharedInvalidationMessage));
		chunk->nitems = 0;
		chunk->maxitems = chunksize;
		chunk->next = *listHdr;
		*listHdr = chunk;
	}
	/* Okay, add message to current chunk */
	chunk->msgs[chunk->nitems] = *msg;
	chunk->nitems++;
}

/*
 * Append one list of invalidation message chunks to another, resetting
 * the source chunk-list pointer to NULL.
 */
static void
AppendInvalidationMessageList(InvalidationChunk **destHdr,
							  InvalidationChunk **srcHdr)
{
	InvalidationChunk *chunk = *srcHdr;

	if (chunk == NULL)
		return;					/* nothing to do */

	while (chunk->next != NULL)
		chunk = chunk->next;

	chunk->next = *destHdr;

	*destHdr = *srcHdr;

	*srcHdr = NULL;
}

/*
 * Process a list of invalidation messages.
 *
 * This is a macro that executes the given code fragment for each message in
 * a message chunk list.  The fragment should refer to the message as *msg.
 */
#define ProcessMessageList(listHdr, codeFragment) \
	do { \
		InvalidationChunk *_chunk; \
		for (_chunk = (listHdr); _chunk != NULL; _chunk = _chunk->next) \
		{ \
			int		_cindex; \
			for (_cindex = 0; _cindex < _chunk->nitems; _cindex++) \
			{ \
				SharedInvalidationMessage *msg = &_chunk->msgs[_cindex]; \
				codeFragment; \
			} \
		} \
	} while (0)


/* ----------------------------------------------------------------
 *				Invalidation set support functions
 *
 * These routines understand about the division of a logical invalidation
 * list into separate physical lists for catcache and relcache/smgr entries.
 * ----------------------------------------------------------------
 */

/*
 * Add a catcache inval entry
 */
static void
AddCatcacheInvalidationMessage(InvalidationListHeader *hdr,
							   int id, uint32 hashValue,
							   ItemPointer tuplePtr, Oid dbId)
{
	SharedInvalidationMessage msg;

	msg.cc.id = (int16) id;
	msg.cc.tuplePtr = *tuplePtr;
	msg.cc.dbId = dbId;
	msg.cc.hashValue = hashValue;
	AddInvalidationMessage(&hdr->cclist, &msg);
}

/*
 * Add a relcache inval entry
 */
static void
AddRelcacheInvalidationMessage(InvalidationListHeader *hdr,
							   Oid dbId, Oid relId)
{
	SharedInvalidationMessage msg;

	/* Don't add a duplicate item */
	/* We assume dbId need not be checked because it will never change */
	ProcessMessageList(hdr->rclist,
					   if (msg->rc.id == SHAREDINVALRELCACHE_ID &&
						   msg->rc.relId == relId)
					   return);

	/* OK, add the item */
	msg.rc.id = SHAREDINVALRELCACHE_ID;
	msg.rc.dbId = dbId;
	msg.rc.relId = relId;
	AddInvalidationMessage(&hdr->rclist, &msg);
}

/*
 * Add an smgr inval entry
 */
static void
AddSmgrInvalidationMessage(InvalidationListHeader *hdr,
						   RelFileNode rnode)
{
	SharedInvalidationMessage msg;

	/* Don't add a duplicate item */
	ProcessMessageList(hdr->rclist,
					   if (msg->sm.id == SHAREDINVALSMGR_ID &&
						   RelFileNodeEquals(msg->sm.rnode, rnode))
					   return);

	/* OK, add the item */
	msg.sm.id = SHAREDINVALSMGR_ID;
	msg.sm.rnode = rnode;
	AddInvalidationMessage(&hdr->rclist, &msg);
}

/*
 * Append one list of invalidation messages to another, resetting
 * the source list to empty.
 */
static void
AppendInvalidationMessages(InvalidationListHeader *dest,
						   InvalidationListHeader *src)
{
	AppendInvalidationMessageList(&dest->cclist, &src->cclist);
	AppendInvalidationMessageList(&dest->rclist, &src->rclist);
}