mp_sync.c

这是国外的resip协议栈

/*-
 * See the file LICENSE for redistribution information.
 *
 * Copyright (c) 1996-2004
 *	Sleepycat Software.  All rights reserved.
 *
 * $Id: mp_sync.c,v 11.98 2004/10/15 16:59:43 bostic Exp $
 */

#include "db_config.h"

#ifndef NO_SYSTEM_INCLUDES
#include <sys/types.h>

#include <stdlib.h>
#endif

#include "db_int.h"
#include "dbinc/db_shash.h"
#include "dbinc/log.h"
#include "dbinc/mp.h"

typedef struct {
	DB_MPOOL_HASH *track_hp;	/* Hash bucket. */

	roff_t	  track_off;		/* Page file offset. */
	db_pgno_t track_pgno;		/* Page number. */
} BH_TRACK;

static int __bhcmp __P((const void *, const void *));
static int __memp_close_flush_files __P((DB_ENV *, DB_MPOOL *, int));
static int __memp_sync_files __P((DB_ENV *, DB_MPOOL *));

/*
 * __memp_sync_pp --
 *	DB_ENV->memp_sync pre/post processing.
 *
 * PUBLIC: int __memp_sync_pp __P((DB_ENV *, DB_LSN *));
 */
int
__memp_sync_pp(dbenv, lsnp)
	DB_ENV *dbenv;
	DB_LSN *lsnp;
{
	int rep_check, ret;

	PANIC_CHECK(dbenv);
	ENV_REQUIRES_CONFIG(dbenv,
	    dbenv->mp_handle, "memp_sync", DB_INIT_MPOOL);

	/*
	 * If no LSN is provided, flush the entire cache (reasonable usage
	 * even if there's no log subsystem configured).
	 */
	if (lsnp != NULL)
		ENV_REQUIRES_CONFIG(dbenv,
		    dbenv->lg_handle, "memp_sync", DB_INIT_LOG);

	rep_check = IS_ENV_REPLICATED(dbenv) ? 1 : 0;
	if (rep_check)
		__env_rep_enter(dbenv);
	ret = __memp_sync(dbenv, lsnp);
	if (rep_check)
		__env_db_rep_exit(dbenv);
	return (ret);
}

/*
 * __memp_sync --
 *	DB_ENV->memp_sync.
 *
 * PUBLIC: int __memp_sync __P((DB_ENV *, DB_LSN *));
 */
int
__memp_sync(dbenv, lsnp)
	DB_ENV *dbenv;
	DB_LSN *lsnp;
{
	DB_MPOOL *dbmp;
	MPOOL *mp;
	int ret;

	dbmp = dbenv->mp_handle;
	mp = dbmp->reginfo[0].primary;

	/* If we've flushed to the requested LSN, return that information. */
	if (lsnp != NULL) {
		R_LOCK(dbenv, dbmp->reginfo);
		if (log_compare(lsnp, &mp->lsn) <= 0) {
			*lsnp = mp->lsn;

			R_UNLOCK(dbenv, dbmp->reginfo);
			return (0);
		}
		R_UNLOCK(dbenv, dbmp->reginfo);
	}

	if ((ret = __memp_sync_int(dbenv, NULL, 0, DB_SYNC_CACHE, NULL)) != 0)
		return (ret);

	if (lsnp != NULL) {
		R_LOCK(dbenv, dbmp->reginfo);
		if (log_compare(lsnp, &mp->lsn) > 0)
			mp->lsn = *lsnp;
		R_UNLOCK(dbenv, dbmp->reginfo);
	}

	return (0);
}

/*
 * __memp_fsync_pp --
 *	DB_MPOOLFILE->sync pre/post processing.
 *
 * PUBLIC: int __memp_fsync_pp __P((DB_MPOOLFILE *));
 */
int
__memp_fsync_pp(dbmfp)
	DB_MPOOLFILE *dbmfp;
{
	DB_ENV *dbenv;
	int rep_check, ret;

	dbenv = dbmfp->dbenv;

	PANIC_CHECK(dbenv);
	MPF_ILLEGAL_BEFORE_OPEN(dbmfp, "DB_MPOOLFILE->sync");

	rep_check = IS_ENV_REPLICATED(dbenv) ? 1 : 0;
	if (rep_check)
		__env_rep_enter(dbenv);
	ret = __memp_fsync(dbmfp);
	if (rep_check)
		__env_db_rep_exit(dbenv);
	return (ret);
}

/*
 * __memp_fsync --
 *	DB_MPOOLFILE->sync.
 *
 * PUBLIC: int __memp_fsync __P((DB_MPOOLFILE *));
 */
int
__memp_fsync(dbmfp)
	DB_MPOOLFILE *dbmfp;
{
	MPOOLFILE *mfp;

	mfp = dbmfp->mfp;

	/*
	 * If this handle doesn't have a file descriptor that's open for
	 * writing, or if the file is a temporary, or if the file hasn't
	 * been written since it was flushed, there's no reason to proceed
	 * further.
	 */
	if (F_ISSET(dbmfp, MP_READONLY))
		return (0);

	if (F_ISSET(mfp, MP_TEMP))
		return (0);

	if (mfp->file_written == 0)
		return (0);

	return (__memp_sync_int(dbmfp->dbenv, dbmfp, 0, DB_SYNC_FILE, NULL));
}

/*
 * __mp_xxx_fh --
 *	Return a file descriptor for DB 1.85 compatibility locking.
 *
 * PUBLIC: int __mp_xxx_fh __P((DB_MPOOLFILE *, DB_FH **));
 */
int
__mp_xxx_fh(dbmfp, fhp)
	DB_MPOOLFILE *dbmfp;
	DB_FH **fhp;
{
	/*
	 * This is a truly spectacular layering violation, intended ONLY to
	 * support compatibility for the DB 1.85 DB->fd call.
	 *
	 * Sync the database file to disk, creating the file as necessary.
	 *
	 * We skip the MP_READONLY and MP_TEMP tests done by memp_fsync(3).
	 * The MP_READONLY test isn't interesting because we will either
	 * already have a file descriptor (we opened the database file for
	 * reading) or we aren't readonly (we created the database which
	 * requires write privileges).  The MP_TEMP test isn't interesting
	 * because we want to write to the backing file regardless so that
	 * we get a file descriptor to return.
	 */
	if ((*fhp = dbmfp->fhp) != NULL)
		return (0);

	return (__memp_sync_int(dbmfp->dbenv, dbmfp, 0, DB_SYNC_FILE, NULL));
}

/*
 * __memp_sync_int --
 *	Mpool sync internal function.
 *
 * PUBLIC: int __memp_sync_int __P((DB_ENV *,
 * PUBLIC:     DB_MPOOLFILE *, u_int32_t, db_sync_op, u_int32_t *));
 */
int
__memp_sync_int(dbenv, dbmfp, trickle_max, op, wrotep)
	DB_ENV *dbenv;
	DB_MPOOLFILE *dbmfp;
	u_int32_t trickle_max, *wrotep;
	db_sync_op op;
{
	BH *bhp;
	BH_TRACK *bharray;
	DB_MPOOL *dbmp;
	DB_MPOOL_HASH *hp;
	DB_MUTEX *mutexp;
	MPOOL *c_mp, *mp;
	MPOOLFILE *mfp;
	roff_t last_mf_offset;
	u_int32_t ar_cnt, ar_max, i, n_cache, remaining, wrote;
	int filecnt, hb_lock, maxopenfd, maxwrite, maxwrite_sleep;
	int pass, ret, t_ret, wait_cnt, write_cnt;

	dbmp = dbenv->mp_handle;
	mp = dbmp->reginfo[0].primary;
	last_mf_offset = INVALID_ROFF;
	filecnt = pass = wrote = 0;

	/* Get shared configuration information. */
	R_LOCK(dbenv, dbmp->reginfo);
	maxopenfd = mp->mp_maxopenfd;
	maxwrite = mp->mp_maxwrite;
	maxwrite_sleep = mp->mp_maxwrite_sleep;
	R_UNLOCK(dbenv, dbmp->reginfo);

	/* Assume one dirty page per bucket. */
	ar_max = mp->nreg * mp->htab_buckets;
	if ((ret =
	    __os_malloc(dbenv, ar_max * sizeof(BH_TRACK), &bharray)) != 0)
		return (ret);

	/*
	 * Walk each cache's list of buffers and mark all dirty buffers to be
	 * written and all pinned buffers to be potentially written, depending
	 * on our flags.
	 */
	for (ar_cnt = 0, n_cache = 0; n_cache < mp->nreg; ++n_cache) {
		c_mp = dbmp->reginfo[n_cache].primary;

		hp = R_ADDR(&dbmp->reginfo[n_cache], c_mp->htab);
		for (i = 0; i < c_mp->htab_buckets; i++, hp++) {
			/*
			 * We can check for empty buckets before locking as we
			 * only care if the pointer is zero or non-zero.  We
			 * can ignore empty buckets because we only need write
			 * buffers that were dirty before we started.
			 */
			if (SH_TAILQ_FIRST(&hp->hash_bucket, __bh) == NULL)
				continue;

			MUTEX_LOCK(dbenv, &hp->hash_mutex);
			for (bhp = SH_TAILQ_FIRST(&hp->hash_bucket, __bh);
			    bhp != NULL; bhp = SH_TAILQ_NEXT(bhp, hq, __bh)) {
				/* Always ignore unreferenced, clean pages. */
				if (bhp->ref == 0 && !F_ISSET(bhp, BH_DIRTY))
					continue;

				/*
				 * Checkpoints have to wait on all pinned pages,
				 * as pages may be marked dirty when returned to
				 * the cache.
				 *
				 * File syncs only wait on pages both pinned and
				 * dirty.  (We don't care if pages are marked
				 * dirty when returned to the cache, that means
				 * there's another writing thread and flushing
				 * the cache for this handle is meaningless.)
				 */
				if (op == DB_SYNC_FILE &&
				    !F_ISSET(bhp, BH_DIRTY))
					continue;

				mfp = R_ADDR(dbmp->reginfo, bhp->mf_offset);

				/*
				 * Ignore temporary files -- this means you
				 * can't even flush temporary files by handle.
				 * (Checkpoint doesn't require temporary files
				 * be flushed and the underlying buffer write
				 * write routine may not be able to write it
				 * anyway.)
				 */
				if (F_ISSET(mfp, MP_TEMP))
					continue;

				/*
				 * If we're flushing a specific file, see if
				 * this page is from that file.
				 */
				if (dbmfp != NULL && mfp != dbmfp->mfp)
					continue;

				/*
				 * Ignore files that aren't involved in DB's
				 * transactional operations during checkpoints.
				 */
				if (dbmfp == NULL && mfp->lsn_off == -1)
					continue;

				/* Track the buffer, we want it. */
				bharray[ar_cnt].track_hp = hp;
				bharray[ar_cnt].track_pgno = bhp->pgno;
				bharray[ar_cnt].track_off = bhp->mf_offset;
				ar_cnt++;

				/*
				 * If we run out of space, double and continue.
				 * Don't stop at trickle_max, we want to sort
				 * as large a sample set as possible in order
				 * to minimize disk seeks.
				 */
				if (ar_cnt >= ar_max) {
					if ((ret = __os_realloc(dbenv,
					    (ar_max * 2) * sizeof(BH_TRACK),
					    &bharray)) != 0)
						break;
					ar_max *= 2;
				}
			}
			MUTEX_UNLOCK(dbenv, &hp->hash_mutex);

			if (ret != 0)
				goto err;
		}
	}

	/* If there no buffers to write, we're done. */
	if (ar_cnt == 0)
		goto done;

	/*
	 * Write the buffers in file/page order, trying to reduce seeks by the
	 * filesystem and, when pages are smaller than filesystem block sizes,
	 * reduce the actual number of writes.
	 */
	if (ar_cnt > 1)
		qsort(bharray, ar_cnt, sizeof(BH_TRACK), __bhcmp);

	/*
	 * If we're trickling buffers, only write enough to reach the correct
	 * percentage.
	 */
	if (op == DB_SYNC_TRICKLE && ar_cnt > trickle_max)
		ar_cnt = trickle_max;

	/*
	 * Flush the log.  We have to ensure the log records reflecting the
	 * changes on the database pages we're writing have already made it
	 * to disk.  We still have to check the log each time we write a page
	 * (because pages we are about to write may be modified after we have
	 * flushed the log), but in general this will at least avoid any I/O
	 * on the log's part.
	 */
	if (LOGGING_ON(dbenv) && (ret = __log_flush(dbenv, NULL)) != 0)
		goto err;

	/*
	 * Walk the array, writing buffers.  When we write a buffer, we NULL
	 * out its hash bucket pointer so we don't process a slot more than
	 * once.
	 */
	for (i = pass = write_cnt = 0, remaining = ar_cnt; remaining > 0; ++i) {
		if (i >= ar_cnt) {
			i = 0;
			++pass;
			__os_sleep(dbenv, 1, 0);
		}
		if ((hp = bharray[i].track_hp) == NULL)
			continue;

		/* Lock the hash bucket and find the buffer. */