mp_sync.c

这是国外的resip协议栈

		mutexp = &hp->hash_mutex;
		MUTEX_LOCK(dbenv, mutexp);
		for (bhp = SH_TAILQ_FIRST(&hp->hash_bucket, __bh);
		    bhp != NULL; bhp = SH_TAILQ_NEXT(bhp, hq, __bh))
			if (bhp->pgno == bharray[i].track_pgno &&
			    bhp->mf_offset == bharray[i].track_off)
				break;

		/*
		 * If we can't find the buffer we're done, somebody else had
		 * to have written it.
		 *
		 * If the buffer isn't pinned or dirty, we're done, there's
		 * no work needed.
		 */
		if (bhp == NULL || (bhp->ref == 0 && !F_ISSET(bhp, BH_DIRTY))) {
			MUTEX_UNLOCK(dbenv, mutexp);
			--remaining;
			bharray[i].track_hp = NULL;
			continue;
		}

		/*
		 * If the buffer is locked by another thread, ignore it, we'll
		 * come back to it.
		 *
		 * If the buffer is pinned and it's only the first or second
		 * time we have looked at it, ignore it, we'll come back to
		 * it.
		 *
		 * In either case, skip the buffer if we're not required to
		 * write it.
		 */
		if (F_ISSET(bhp, BH_LOCKED) || (bhp->ref != 0 && pass < 2)) {
			MUTEX_UNLOCK(dbenv, mutexp);
			if (op != DB_SYNC_CACHE && op != DB_SYNC_FILE) {
				--remaining;
				bharray[i].track_hp = NULL;
			}
			continue;
		}

		/*
		 * The buffer is either pinned or dirty.
		 *
		 * Set the sync wait-for count, used to count down outstanding
		 * references to this buffer as they are returned to the cache.
		 */
		bhp->ref_sync = bhp->ref;

		/* Pin the buffer into memory and lock it. */
		++bhp->ref;
		F_SET(bhp, BH_LOCKED);
		MUTEX_LOCK(dbenv, &bhp->mutex);

		/*
		 * Unlock the hash bucket and wait for the wait-for count to
		 * go to 0.   No new thread can acquire the buffer because we
		 * have it locked.
		 *
		 * If a thread attempts to re-pin a page, the wait-for count
		 * will never go to 0 (the thread spins on our buffer lock,
		 * while we spin on the thread's ref count).  Give up if we
		 * don't get the buffer in 3 seconds, we can try again later.
		 *
		 * If, when the wait-for count goes to 0, the buffer is found
		 * to be dirty, write it.
		 */
		MUTEX_UNLOCK(dbenv, mutexp);
		for (wait_cnt = 1;
		    bhp->ref_sync != 0 && wait_cnt < 4; ++wait_cnt)
			__os_sleep(dbenv, 1, 0);
		MUTEX_LOCK(dbenv, mutexp);
		hb_lock = 1;

		/*
		 * If we've switched files, check to see if we're configured
		 * to close file descriptors.
		 */
		if (maxopenfd != 0 && bhp->mf_offset != last_mf_offset) {
			if (++filecnt >= maxopenfd) {
				filecnt = 0;
				if ((ret = __memp_close_flush_files(
				    dbenv, dbmp, 1)) != 0)
					break;
			}
			last_mf_offset = bhp->mf_offset;
		}

		/*
		 * If the ref_sync count has gone to 0, we're going to be done
		 * with this buffer no matter what happens.
		 */
		if (bhp->ref_sync == 0) {
			--remaining;
			bharray[i].track_hp = NULL;
		}

		/*
		 * If the ref_sync count has gone to 0 and the buffer is still
		 * dirty, we write it.  We only try to write the buffer once.
		 */
		if (bhp->ref_sync == 0 && F_ISSET(bhp, BH_DIRTY)) {
			hb_lock = 0;
			MUTEX_UNLOCK(dbenv, mutexp);

			mfp = R_ADDR(dbmp->reginfo, bhp->mf_offset);
			if ((ret = __memp_bhwrite(dbmp, hp, mfp, bhp, 1)) == 0)
				++wrote;
			else
				__db_err(dbenv, "%s: unable to flush page: %lu",
				    __memp_fns(dbmp, mfp), (u_long)bhp->pgno);

			/*
			 * Avoid saturating the disk, sleep once we've done
			 * some number of writes.
			 */
			if (maxwrite != 0 && ++write_cnt >= maxwrite) {
				write_cnt = 0;
				__os_sleep(dbenv, 0, (u_long)maxwrite_sleep);
			}
		}

		/*
		 * If ref_sync count never went to 0, the buffer was written
		 * by another thread, or the write failed, we still have the
		 * buffer locked.
		 *
		 * We may or may not currently hold the hash bucket mutex.  If
		 * the __memp_bhwrite -> __memp_pgwrite call was successful,
		 * then __memp_pgwrite will have swapped the buffer lock for
		 * the hash lock.  All other call paths will leave us without
		 * the hash bucket lock.
		 *
		 * The order of mutexes above was to acquire the buffer lock
		 * while holding the hash bucket lock.  Don't deadlock here,
		 * release the buffer lock and then acquire the hash bucket
		 * lock.
		 */
		if (F_ISSET(bhp, BH_LOCKED)) {
			F_CLR(bhp, BH_LOCKED);
			MUTEX_UNLOCK(dbenv, &bhp->mutex);

			if (!hb_lock)
				MUTEX_LOCK(dbenv, mutexp);
		}

		/*
		 * Reset the ref_sync count regardless of our success, we're
		 * done with this buffer for now.
		 */
		bhp->ref_sync = 0;

		/* Discard our reference and unlock the bucket. */
		--bhp->ref;
		MUTEX_UNLOCK(dbenv, mutexp);

		if (ret != 0)
			break;
	}

done:	/*
	 * If doing a checkpoint or flushing a file for the application, we
	 * have to force the pages to disk.  We don't do this as we go along
	 * because we want to give the OS as much time as possible to lazily
	 * flush, and because we have to flush files that might not even have
	 * had dirty buffers in the cache, so we have to walk the files list.
	 */
	if (ret == 0 && (op == DB_SYNC_CACHE || op == DB_SYNC_FILE)) {
		if (dbmfp == NULL)
			ret = __memp_sync_files(dbenv, dbmp);
		else
			ret = __os_fsync(dbenv, dbmfp->fhp);
	}

	/* If we've opened files to flush pages, close them. */
	if ((t_ret = __memp_close_flush_files(dbenv, dbmp, 0)) != 0 && ret == 0)
		ret = t_ret;

err:	__os_free(dbenv, bharray);
	if (wrotep != NULL)
		*wrotep = wrote;

	return (ret);
}

/*
 * __memp_sync_files --
 *	Sync all the files in the environment, open or not.
 */
static
int __memp_sync_files(dbenv, dbmp)
	DB_ENV *dbenv;
	DB_MPOOL *dbmp;
{
	DB_MPOOLFILE *dbmfp;
	MPOOL *mp;
	MPOOLFILE *mfp;
	int final_ret, ret;

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

	R_LOCK(dbenv, dbmp->reginfo);
	for (mfp = SH_TAILQ_FIRST(&mp->mpfq, __mpoolfile);
	    mfp != NULL; mfp = SH_TAILQ_NEXT(mfp, q, __mpoolfile)) {
		if (!mfp->file_written ||
		    mfp->deadfile || F_ISSET(mfp, MP_TEMP))
			continue;

		/*
		 * Look for an already open, writeable handle (fsync doesn't
		 * work on read-only Windows handles).
		 */
		ret = 0;
		MUTEX_THREAD_LOCK(dbenv, dbmp->mutexp);
		for (dbmfp = TAILQ_FIRST(&dbmp->dbmfq);
		    dbmfp != NULL; dbmfp = TAILQ_NEXT(dbmfp, q)) {
			if (dbmfp->mfp != mfp || F_ISSET(dbmfp, MP_READONLY))
				continue;
			ret = __os_fsync(dbenv, dbmfp->fhp);
			break;
		}
		MUTEX_THREAD_UNLOCK(dbenv, dbmp->mutexp);

		/* If we don't find one, open one. */
		if (dbmfp == NULL)
			ret = __memp_mf_sync(dbmp, mfp);
		if (ret != 0) {
			__db_err(dbenv, "%s: unable to flush: %s",
			    (char *)R_ADDR(dbmp->reginfo, mfp->path_off),
			    db_strerror(ret));
			if (final_ret == 0)
				final_ret = ret;
			continue;
		}

		/*
		 * If we wrote the file and there are no open handles (or there
		 * is a single open handle, and it's the one we opened to write
		 * buffers during checkpoint), clear the file_written flag.  We
		 * do this so that applications opening thousands of files don't
		 * loop here opening and flushing those files during checkpoint.
		 *
		 * The danger here is if a buffer were to be written as part of
		 * a checkpoint, and then not be flushed to disk.  This cannot
		 * happen because we only clear file_written when there are no
		 * other users of the MPOOLFILE in the system, and, as we hold
		 * the region lock, no possibility of another thread of control
		 * racing with us to open a MPOOLFILE.
		 */
		if (mfp->mpf_cnt == 0 || (mfp->mpf_cnt == 1 &&
		    dbmfp != NULL && F_ISSET(dbmfp, MP_FLUSH)))
			mfp->file_written = 0;
	}
	R_UNLOCK(dbenv, dbmp->reginfo);

	return (final_ret);
}

/*
 * __memp_mf_sync --
 *	 Flush an MPOOLFILE.
 *
 *	Should only be used when the file is not already open in this process.
 *
 * PUBLIC: int __memp_mf_sync __P((DB_MPOOL *, MPOOLFILE *));
 */
int
__memp_mf_sync(dbmp, mfp)
	DB_MPOOL *dbmp;
	MPOOLFILE *mfp;
{
	DB_ENV *dbenv;
	DB_FH *fhp;
	int ret, t_ret;
	char *rpath;

	dbenv = dbmp->dbenv;

	/*
	 * Expects caller to be holding the region lock: we're using the path
	 * name and __memp_nameop might try and rename the file.
	 */
	if ((ret = __db_appname(dbenv, DB_APP_DATA,
	    R_ADDR(dbmp->reginfo, mfp->path_off), 0, NULL,
	    &rpath)) == 0) {
		if ((ret = __os_open(dbenv, rpath, 0, 0, &fhp)) == 0) {
			ret = __os_fsync(dbenv, fhp);
			if ((t_ret =
			    __os_closehandle(dbenv, fhp)) != 0 && ret == 0)
				ret = t_ret;
		}
		__os_free(dbenv, rpath);
	}

	return (ret);
}

/*
 * __memp_close_flush_files --
 *	Close files opened only to flush buffers.
 */
static int
__memp_close_flush_files(dbenv, dbmp, dosync)
	DB_ENV *dbenv;
	DB_MPOOL *dbmp;
	int dosync;
{
	DB_MPOOLFILE *dbmfp;
	MPOOLFILE *mfp;
	int ret;

	/*
	 * The routine exists because we must close files opened by sync to
	 * flush buffers.  There are two cases: first, extent files have to
	 * be closed so they may be removed when empty.  Second, regular
	 * files have to be closed so we don't run out of descriptors (for
	 * example, an application partitioning its data into databases
	 * based on timestamps, so there's a continually increasing set of
	 * files).
	 *
	 * We mark files opened in the __memp_bhwrite() function with the
	 * MP_FLUSH flag.  Here we walk through our file descriptor list,
	 * and, if a file was opened by __memp_bhwrite(), we close it.
	 */
retry:	MUTEX_THREAD_LOCK(dbenv, dbmp->mutexp);
	for (dbmfp = TAILQ_FIRST(&dbmp->dbmfq);
	    dbmfp != NULL; dbmfp = TAILQ_NEXT(dbmfp, q))
		if (F_ISSET(dbmfp, MP_FLUSH)) {
			F_CLR(dbmfp, MP_FLUSH);
			MUTEX_THREAD_UNLOCK(dbenv, dbmp->mutexp);
			if (dosync) {
				if ((ret = __os_fsync(dbenv, dbmfp->fhp)) != 0)
					return (ret);
				/*
				 * If the file is clean and we have the only
				 * open handle on the file, clear the dirty
				 * flag so we don't re-open and sync it again.
				 */
				mfp = dbmfp->mfp;
				if (mfp->mpf_cnt == 1) {
					R_LOCK(dbenv, dbmp->reginfo);
					if (mfp->mpf_cnt == 1)
						mfp->file_written = 0;
					R_UNLOCK(dbenv, dbmp->reginfo);
				}
			}
			if ((ret = __memp_fclose(dbmfp, 0)) != 0)
				return (ret);
			goto retry;
		}
	MUTEX_THREAD_UNLOCK(dbenv, dbmp->mutexp);

	return (0);
}

static int
__bhcmp(p1, p2)
	const void *p1, *p2;
{
	BH_TRACK *bhp1, *bhp2;

	bhp1 = (BH_TRACK *)p1;
	bhp2 = (BH_TRACK *)p2;

	/* Sort by file (shared memory pool offset). */
	if (bhp1->track_off < bhp2->track_off)
		return (-1);
	if (bhp1->track_off > bhp2->track_off)
		return (1);

	/*
	 * !!!
	 * Defend against badly written quicksort code calling the comparison
	 * function with two identical pointers (e.g., WATCOM C++ (Power++)).
	 */
	if (bhp1->track_pgno < bhp2->track_pgno)
		return (-1);
	if (bhp1->track_pgno > bhp2->track_pgno)
		return (1);
	return (0);
}