hash_rec.c

这是国外的resip协议栈

		else if (cmp_n == 0 && DB_UNDO(op))
			mmeta->lsn = argp->mmetalsn;
	} else
		mmeta = (DBMETA *)hcp->hdr;

#ifdef HAVE_FTRUNCATE
	if (cmp_n == 0 && DB_UNDO(op))
		mmeta->last_pgno = argp->last_pgno;
	else if (DB_REDO(op))
#endif
	if (mmeta->last_pgno < pgno)
		mmeta->last_pgno = pgno;
	mmeta_flags = DB_MPOOL_DIRTY;

	if (argp->mmpgno != argp->mpgno &&
	    (ret = __memp_fput(mpf, mmeta, mmeta_flags)) != 0)
		goto out;
	mmeta = NULL;

	if (did_recover)
		F_SET(hcp, H_DIRTY);

done:	*lsnp = argp->prev_lsn;
	ret = 0;

out:	if (mmeta != NULL)
		(void)__memp_fput(mpf, mmeta, 0);
	if (dbc != NULL)
		(void)__ham_release_meta(dbc);
	if (ret == ENOENT && op == DB_TXN_BACKWARD_ALLOC)
		ret = 0;

	REC_CLOSE;
}

/*
 * __ham_groupalloc_recover --
 *	Recover the batch creation of a set of pages for a new database.
 *
 * PUBLIC: int __ham_groupalloc_recover
 * PUBLIC:   __P((DB_ENV *, DBT *, DB_LSN *, db_recops, void *));
 */
int
__ham_groupalloc_recover(dbenv, dbtp, lsnp, op, info)
	DB_ENV *dbenv;
	DBT *dbtp;
	DB_LSN *lsnp;
	db_recops op;
	void *info;
{
	__ham_groupalloc_args *argp;
	DBMETA *mmeta;
	DB_MPOOLFILE *mpf;
	DB *file_dbp;
	DBC *dbc;
	PAGE *pagep;
	db_pgno_t pgno;
	int cmp_n, cmp_p, modified, ret;

	mmeta = NULL;
	modified = 0;
	REC_PRINT(__ham_groupalloc_print);
	REC_INTRO(__ham_groupalloc_read, 0);

	pgno = PGNO_BASE_MD;
	if ((ret = __memp_fget(mpf, &pgno, 0, &mmeta)) != 0) {
		if (DB_REDO(op)) {
			ret = __db_pgerr(file_dbp, pgno, ret);
			goto out;
		} else
			goto done;
	}

	cmp_n = log_compare(lsnp, &LSN(mmeta));
	cmp_p = log_compare(&LSN(mmeta), &argp->meta_lsn);
	CHECK_LSN(op, cmp_p, &LSN(mmeta), &argp->meta_lsn);

	/*
	 * Basically, we used mpool to allocate a chunk of pages.
	 * We need to either add those to a free list (in the undo
	 * case) or initialize them (in the redo case).
	 *
	 * If we are redoing and this is a hash subdatabase, it's possible
	 * that the pages were never allocated, so we'd better check for
	 * that and handle it here.
	 */
	pgno = argp->start_pgno + argp->num - 1;
	if (DB_REDO(op)) {
		if ((ret = __ham_alloc_pages(file_dbp, argp, lsnp)) != 0)
			goto out;
		if (cmp_p == 0) {
			LSN(mmeta) = *lsnp;
			modified = 1;
		}
	} else if (DB_UNDO(op)) {
		/*
		 * Fetch the last page and determine if it is in
		 * the post allocation state.
		 */
		pagep = NULL;
		if ((ret = __memp_fget(mpf, &pgno, 0, &pagep)) == 0) {
			if (log_compare(&pagep->lsn, lsnp) != 0) {
				if ((ret = __memp_fput(mpf,
				     pagep, DB_MPOOL_DISCARD)) != 0)
					goto out;
				pagep = NULL;
			}
		} else if (ret != DB_PAGE_NOTFOUND)
			goto out;
#ifdef HAVE_FTRUNCATE
		COMPQUIET(info, NULL);
		/*
		 * If the last page was allocated then truncate back
		 * to the first page.
		 */
		if (pagep != NULL) {
			if ((ret =
			    __memp_fput(mpf, pagep, DB_MPOOL_DISCARD)) != 0)
				goto out;
			if ((ret =
			     __memp_ftruncate(mpf, argp->start_pgno, 0)) != 0)
				goto out;
		}

		/*
		 * If we are rolling back the metapage, then make
		 * sure it reflects the the correct last_pgno.
		 */
		if (cmp_n == 0) {
			mmeta->last_pgno = argp->last_pgno;
			modified = 1;
		}
		pgno = 0;
#else
		/*
		 * Reset the last page back to its preallocation state.
		 */
		if (pagep != NULL) {
			if (log_compare(&pagep->lsn, lsnp) == 0)
				ZERO_LSN(pagep->lsn);

			if ((ret =
			    __memp_fput(mpf, pagep, DB_MPOOL_DIRTY)) != 0)
				goto out;
		}
		/*
		 * Put the pages into the limbo list and free them later.
		 */
		if ((ret = __db_add_limbo(dbenv,
		    info, argp->fileid, argp->start_pgno, argp->num)) != 0)
			goto out;
#endif
		if (cmp_n == 0) {
			LSN(mmeta) = argp->meta_lsn;
			modified = 1;
		}
	}

	/*
	 * In both REDO and UNDO, we have grown the file and need to make
	 * sure that last_pgno is correct.  If we HAVE_FTRUNCATE pgno
	 * will only be valid on REDO.
	 */
	if (pgno > mmeta->last_pgno) {
		mmeta->last_pgno = pgno;
		modified = 1;
	}

done:	if (ret == 0)
		*lsnp = argp->prev_lsn;
	ret = 0;

out:	if (mmeta != NULL)
		(void)__memp_fput(mpf, mmeta, modified ? DB_MPOOL_DIRTY : 0);

	if (ret == ENOENT && op == DB_TXN_BACKWARD_ALLOC)
		ret = 0;
	REC_CLOSE;
}

/*
 * __ham_alloc_pages --
 *
 * Called during redo of a file create.  We create new pages in the file
 * using the MPOOL_NEW_GROUP flag.  We then log the meta-data page with a
 * __crdel_metasub message.  If we manage to crash without the newly written
 * pages getting to disk (I'm not sure this can happen anywhere except our
 * test suite?!), then we need to go through a recreate the final pages.
 * Hash normally has holes in its files and handles them appropriately.
 */
static int
__ham_alloc_pages(dbp, argp, lsnp)
	DB *dbp;
	__ham_groupalloc_args *argp;
	DB_LSN *lsnp;
{
	DB_MPOOLFILE *mpf;
	PAGE *pagep;
	db_pgno_t pgno;
	int ret;

	mpf = dbp->mpf;

	/* Read the last page of the allocation. */
	pgno = argp->start_pgno + argp->num - 1;

	/* If the page exists, and it has been initialized, then we're done. */
	if ((ret = __memp_fget(mpf, &pgno, 0, &pagep)) == 0) {
		if (NUM_ENT(pagep) == 0 && IS_ZERO_LSN(pagep->lsn))
			goto reinit_page;
		if ((ret = __memp_fput(mpf, pagep, 0)) != 0)
			return (ret);
		return (0);
	}

	/* Had to create the page. */
	if ((ret = __memp_fget(mpf, &pgno, DB_MPOOL_CREATE, &pagep)) != 0)
		return (__db_pgerr(dbp, pgno, ret));

reinit_page:
	/* Initialize the newly allocated page. */
	P_INIT(pagep, dbp->pgsize, pgno, PGNO_INVALID, PGNO_INVALID, 0, P_HASH);
	pagep->lsn = *lsnp;

	if ((ret = __memp_fput(mpf, pagep, DB_MPOOL_DIRTY)) != 0)
		return (ret);

	return (0);
}

/*
 * __ham_curadj_recover --
 *	Undo cursor adjustments if a subtransaction fails.
 *
 * PUBLIC: int __ham_curadj_recover
 * PUBLIC:   __P((DB_ENV *, DBT *, DB_LSN *, db_recops, void *));
 */
int
__ham_curadj_recover(dbenv, dbtp, lsnp, op, info)
	DB_ENV *dbenv;
	DBT *dbtp;
	DB_LSN *lsnp;
	db_recops op;
	void *info;
{
	__ham_curadj_args *argp;
	DB_MPOOLFILE *mpf;
	DB *file_dbp;
	DBC *dbc;
	int ret;
	HASH_CURSOR *hcp;

	COMPQUIET(info, NULL);
	REC_PRINT(__ham_curadj_print);
	REC_INTRO(__ham_curadj_read, 0);

	if (op != DB_TXN_ABORT)
		goto done;

	/*
	 * Undo the adjustment by reinitializing the the cursor to look like
	 * the one that was used to do the adjustment, then we invert the
	 * add so that undo the adjustment.
	 */
	hcp = (HASH_CURSOR *)dbc->internal;
	hcp->pgno = argp->pgno;
	hcp->indx = argp->indx;
	hcp->dup_off = argp->dup_off;
	hcp->order = argp->order;
	if (!argp->add)
		F_SET(hcp, H_DELETED);
	(void)__ham_c_update(dbc, argp->len, !argp->add, argp->is_dup);

done:	*lsnp = argp->prev_lsn;
out:	REC_CLOSE;
}

/*
 * __ham_chgpg_recover --
 *	Undo cursor adjustments if a subtransaction fails.
 *
 * PUBLIC: int __ham_chgpg_recover
 * PUBLIC:   __P((DB_ENV *, DBT *, DB_LSN *, db_recops, void *));
 */
int
__ham_chgpg_recover(dbenv, dbtp, lsnp, op, info)
	DB_ENV *dbenv;
	DBT *dbtp;
	DB_LSN *lsnp;
	db_recops op;
	void *info;
{
	__ham_chgpg_args *argp;
	BTREE_CURSOR *opdcp;
	DB_MPOOLFILE *mpf;
	DB *file_dbp, *ldbp;
	DBC *dbc;
	int ret;
	DBC *cp;
	HASH_CURSOR *lcp;
	u_int32_t order, indx;

	COMPQUIET(info, NULL);
	REC_PRINT(__ham_chgpg_print);
	REC_INTRO(__ham_chgpg_read, 0);

	if (op != DB_TXN_ABORT)
		goto done;

	/* Overloaded fields for DB_HAM_DEL*PG */
	indx = argp->old_indx;
	order = argp->new_indx;

	MUTEX_THREAD_LOCK(dbenv, dbenv->dblist_mutexp);
	for (ldbp = __dblist_get(dbenv, file_dbp->adj_fileid);
	    ldbp != NULL && ldbp->adj_fileid == file_dbp->adj_fileid;
	    ldbp = LIST_NEXT(ldbp, dblistlinks)) {
		MUTEX_THREAD_LOCK(dbenv, file_dbp->mutexp);

		for (cp = TAILQ_FIRST(&ldbp->active_queue); cp != NULL;
		    cp = TAILQ_NEXT(cp, links)) {
			lcp = (HASH_CURSOR *)cp->internal;

			switch (argp->mode) {
			case DB_HAM_DELFIRSTPG:
				if (lcp->pgno != argp->new_pgno)
					break;
				if (lcp->indx != indx ||
				    !F_ISSET(lcp, H_DELETED) ||
				    lcp->order >= order) {
					lcp->pgno = argp->old_pgno;
					if (lcp->indx == indx)
						lcp->order -= order;
				}
				break;
			case DB_HAM_DELMIDPG:
			case DB_HAM_DELLASTPG:
				if (lcp->pgno == argp->new_pgno &&
				    lcp->indx == indx &&
				    F_ISSET(lcp, H_DELETED) &&
				    lcp->order >= order) {
					lcp->pgno = argp->old_pgno;
					lcp->order -= order;
					lcp->indx = 0;
				}
				break;
			case DB_HAM_CHGPG:
				/*
				 * If we're doing a CHGPG, we're undoing
				 * the move of a non-deleted item to a
				 * new page.  Any cursors with the deleted
				 * flag set do not belong to this item;
				 * don't touch them.
				 */
				if (F_ISSET(lcp, H_DELETED))
					break;
				/* FALLTHROUGH */
			case DB_HAM_SPLIT:
				if (lcp->pgno == argp->new_pgno &&
				    lcp->indx == argp->new_indx) {
					lcp->indx = argp->old_indx;
					lcp->pgno = argp->old_pgno;
				}
				break;
			case DB_HAM_DUP:
				if (lcp->opd == NULL)
					break;
				opdcp = (BTREE_CURSOR *)lcp->opd->internal;
				if (opdcp->pgno != argp->new_pgno ||
				    opdcp->indx != argp->new_indx)
					break;

				if (F_ISSET(opdcp, C_DELETED))
					F_SET(lcp, H_DELETED);
				/*
				 * We can't close a cursor while we have the
				 * dbp mutex locked, since c_close reacquires
				 * it.  It should be safe to drop the mutex
				 * here, though, since newly opened cursors
				 * are put only at the end of the tailq and
				 * the cursor we're adjusting can't be closed
				 * under us.
				 */
				MUTEX_THREAD_UNLOCK(dbenv, file_dbp->mutexp);
				if ((ret = __db_c_close(lcp->opd)) != 0)
					goto out;
				MUTEX_THREAD_LOCK(dbenv, file_dbp->mutexp);
				lcp->opd = NULL;
				break;
			}
		}
		MUTEX_THREAD_UNLOCK(dbenv, file_dbp->mutexp);
	}
	MUTEX_THREAD_UNLOCK(dbenv, dbenv->dblist_mutexp);

done:	*lsnp = argp->prev_lsn;
out:	REC_CLOSE;
}