MDB_VL32 support 64 bit DBs on 32 bit processors

libraries/liblmdb/mdb.c

@@ -1157,6 +1157,9 @@ struct MDB_txn {
 	/** Nested txn under this txn, set together with flag #MDB_TXN_HAS_CHILD */
 	MDB_txn		*mt_child;
 	pgno_t		mt_next_pgno;	/**< next unallocated page */
+#ifdef MDB_VL32
+	pgno_t		mt_last_pgno;	/**< last written page */
+#endif
 	/** The ID of this transaction. IDs are integers incrementing from 1.
 	 *	Only committed write transactions increment the ID. If a transaction
 	 *	aborts, the ID may be re-used by the next writer.
@@ -1203,6 +1206,19 @@ struct MDB_txn {
 	MDB_cursor	**mt_cursors;
 	/** Array of flags for each DB */
 	unsigned char	*mt_dbflags;
+#ifdef MDB_VL32
+	/** List of read-only pages (actually chunks) */
+	MDB_ID3L	mt_rpages;
+	/** We map chunks of 16 pages. Even though Windows uses 4KB pages, all
+	 * mappings must begin on 64KB boundaries. So we round off all pgnos to
+	 * a chunk boundary. We do the same on Linux for symmetry, and also to
+	 * reduce the frequency of mmap/munmap calls.
+	 */
+#define MDB_RPAGE_CHUNK	16
+#define MDB_TRPAGE_SIZE	4096	/**< size of #mt_rpages array of chunks */
+#define MDB_TRPAGE_MAX	(MDB_TRPAGE_SIZE-1)	/**< maximum chunk index */
+	unsigned int mt_rpcheck;	/**< threshold for reclaiming unref'd chunks */
+#endif
 	/**	Number of DB records in use, or 0 when the txn is finished.
 	 *	This number only ever increments until the txn finishes; we
 	 *	don't decrement it when individual DB handles are closed.
@@ -1315,6 +1331,9 @@ struct MDB_env {
 	HANDLE		me_fd;		/**< The main data file */
 	HANDLE		me_lfd;		/**< The lock file */
 	HANDLE		me_mfd;			/**< just for writing the meta pages */
+#if defined(MDB_VL32) && defined(_WIN32)
+	HANDLE		me_fmh;		/**< File Mapping handle */
+#endif
 	/** Failed to update the meta page. Probably an I/O error. */
 #define	MDB_FATAL_ERROR	0x80000000U
 	/** Some fields are initialized. */
@@ -1372,6 +1391,13 @@ struct MDB_env {
 #else
 	mdb_mutex_t	me_rmutex;
 	mdb_mutex_t	me_wmutex;
+#endif
+#ifdef MDB_VL32
+	MDB_ID3L	me_rpages;	/**< like #mt_rpages, but global to env */
+	mdb_mutex_t	me_rpmutex;	/**< control access to #me_rpages */
+#define MDB_ERPAGE_SIZE	16384
+#define MDB_ERPAGE_MAX	(MDB_ERPAGE_SIZE-1)
+	unsigned int me_rpcheck;
 #endif
 	void		*me_userctx;	 /**< User-settable context */
 	MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
@@ -1894,6 +1920,41 @@ mdb_dlist_free(MDB_txn *txn)
 	dl[0].mid = 0;
 }
 
+#ifdef MDB_VL32
+static void
+mdb_page_unref(MDB_txn *txn, MDB_page *mp)
+{
+	pgno_t pgno;
+	MDB_ID3L tl = txn->mt_rpages;
+	unsigned x, rem;
+	if (mp->mp_flags & (P_SUBP|P_DIRTY))
+		return;
+	rem = mp->mp_pgno & (MDB_RPAGE_CHUNK-1);
+	pgno = mp->mp_pgno ^ rem;
+	x = mdb_mid3l_search(tl, pgno);
+	if (x != tl[0].mid && tl[x+1].mid == mp->mp_pgno)
+		x++;
+	if (tl[x].mref)
+		tl[x].mref--;
+}
+#define MDB_PAGE_UNREF(txn, mp)	mdb_page_unref(txn, mp)
+
+static void
+mdb_cursor_unref(MDB_cursor *mc)
+{
+	int i;
+	if (!mc->mc_pg[0] || IS_SUBP(mc->mc_pg[0]))
+		return;
+	for (i=0; i<mc->mc_snum; i++)
+		mdb_page_unref(mc->mc_txn, mc->mc_pg[i]);
+	mc->mc_snum = mc->mc_top = 0;
+	mc->mc_pg[0] = NULL;
+	mc->mc_flags &= ~C_INITIALIZED;
+}
+#else
+#define MDB_PAGE_UNREF(txn, mp)
+#endif /* MDB_VL32 */
+
 /** Loosen or free a single page.
  * Saves single pages to a list for future reuse
  * in this same txn. It has been pulled from the freeDB
@@ -2573,6 +2634,7 @@ done:
 			}
 		}
 	}
+	MDB_PAGE_UNREF(mc->mc_txn, mp);
 	return 0;
 
 fail:
@@ -2854,6 +2916,9 @@ mdb_txn_renew0(MDB_txn *txn)
 
 	/* Moved to here to avoid a data race in read TXNs */
 	txn->mt_next_pgno = meta->mm_last_pg+1;
+#ifdef MDB_VL32
+	txn->mt_last_pgno = txn->mt_next_pgno - 1;
+#endif
 
 	txn->mt_flags = flags;
 
@@ -2932,6 +2997,17 @@ mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
 		DPRINTF(("calloc: %s", strerror(errno)));
 		return ENOMEM;
 	}
+#ifdef MDB_VL32
+	if (!parent) {
+		txn->mt_rpages = malloc(MDB_TRPAGE_SIZE * sizeof(MDB_ID3));
+		if (!txn->mt_rpages) {
+			free(txn);
+			return ENOMEM;
+		}
+		txn->mt_rpages[0].mid = 0;
+		txn->mt_rpcheck = MDB_TRPAGE_SIZE/2;
+	}
+#endif
 	txn->mt_dbxs = env->me_dbxs;	/* static */
 	txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
 	txn->mt_dbflags = (unsigned char *)txn + size - env->me_maxdbs;
@@ -2959,6 +3035,9 @@ mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
 		parent->mt_child = txn;
 		txn->mt_parent = parent;
 		txn->mt_numdbs = parent->mt_numdbs;
+#ifdef MDB_VL32
+		txn->mt_rpages = parent->mt_rpages;
+#endif
 		memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
 		/* Copy parent's mt_dbflags, but clear DB_NEW */
 		for (i=0; i<txn->mt_numdbs; i++)
@@ -2984,8 +3063,12 @@ renew:
 		rc = mdb_txn_renew0(txn);
 	}
 	if (rc) {
-		if (txn != env->me_txn0)
+		if (txn != env->me_txn0) {
+#ifdef MDB_VL32
+			free(txn->mt_rpages);
+#endif
 			free(txn);
+		}
 	} else {
 		txn->mt_flags |= flags;	/* could not change txn=me_txn0 earlier */
 		*ret = txn;
@@ -3110,7 +3193,31 @@ mdb_txn_end(MDB_txn *txn, unsigned mode)
 
 		mdb_midl_free(pghead);
 	}
-
+#ifdef MDB_VL32
+	if (!txn->mt_parent) {
+		MDB_ID3L el = env->me_rpages, tl = txn->mt_rpages;
+		unsigned i, x, n = tl[0].mid;
+		LOCK_MUTEX0(env->me_rpmutex);
+		for (i = 1; i <= n; i++) {
+			if (tl[i].mid & (MDB_RPAGE_CHUNK-1)) {
+				/* tmp overflow pages that we didn't share in env */
+				munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
+			} else {
+				x = mdb_mid3l_search(el, tl[i].mid);
+				if (tl[i].mptr == el[x].mptr) {
+					el[x].mref--;
+				} else {
+					/* another tmp overflow page */
+					munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
+				}
+			}
+		}
+		UNLOCK_MUTEX(env->me_rpmutex);
+		tl[0].mid = 0;
+		if (mode & MDB_END_FREE)
+			free(tl);
+	}
+#endif
 	if (mode & MDB_END_FREE)
 		free(txn);
 }
@@ -3460,6 +3567,10 @@ retry_seek:
 		n++;
 #endif	/* _WIN32 */
 	}
+#ifdef MDB_VL32
+	if (pgno > txn->mt_last_pgno)
+		txn->mt_last_pgno = pgno;
+#endif
 
 	/* MIPS has cache coherency issues, this is a no-op everywhere else
 	 * Note: for any size >= on-chip cache size, entire on-chip cache is
@@ -4046,11 +4157,27 @@ mdb_env_map(MDB_env *env, void *addr)
 	if (rc)
 		return rc;
 	map = addr;
+#ifdef MDB_VL32
+	msize = NUM_METAS * env->me_psize;
+#endif
 	rc = NtMapViewOfSection(mh, GetCurrentProcess(), &map, 0, 0, NULL, &msize, ViewUnmap, alloctype, pageprot);
+#ifdef MDB_VL32
+	env->me_fmh = mh;
+#else
 	NtClose(mh);
+#endif
 	if (rc)
 		return rc;
 	env->me_map = map;
+#else
+#ifdef MDB_VL32
+	(void) flags;
+	env->me_map = mmap(addr, NUM_METAS * env->me_psize, PROT_READ, MAP_SHARED,
+		env->me_fd, 0);
+	if (env->me_map == MAP_FAILED) {
+		env->me_map = NULL;
+		return ErrCode();
+	}
 #else
 	int prot = PROT_READ;
 	if (flags & MDB_WRITEMAP) {
@@ -4084,6 +4211,7 @@ mdb_env_map(MDB_env *env, void *addr)
 	 */
 	if (addr && env->me_map != addr)
 		return EBUSY;	/* TODO: Make a new MDB_* error code? */
+#endif
 
 	p = (MDB_page *)env->me_map;
 	env->me_metas[0] = METADATA(p);
@@ -4099,9 +4227,11 @@ mdb_env_set_mapsize(MDB_env *env, mdb_size_t size)
 	 * sure there are no active txns.
 	 */
 	if (env->me_map) {
-		int rc;
 		MDB_meta *meta;
+#ifndef MDB_VL32
 		void *old;
+		int rc;
+#endif
 		if (env->me_txn)
 			return EINVAL;
 		meta = mdb_env_pick_meta(env);
@@ -4113,12 +4243,17 @@ mdb_env_set_mapsize(MDB_env *env, mdb_size_t size)
 			if (size < minsize)
 				size = minsize;
 		}
+#ifndef MDB_VL32
+		/* For MDB_VL32 this bit is a noop since we dynamically remap
+		 * chunks of the DB anyway.
+		 */
 		munmap(env->me_map, env->me_mapsize);
 		env->me_mapsize = size;
 		old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
 		rc = mdb_env_map(env, old);
 		if (rc)
 			return rc;
+#endif /* !MDB_VL32 */
 	}
 	env->me_mapsize = size;
 	if (env->me_psize)
@@ -4843,6 +4978,13 @@ mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
 	env->me_wmutex->semnum = 1;
 	env->me_rmutex->locked = &env->me_txns->mti_rlocked;
 	env->me_wmutex->locked = &env->me_txns->mti_wlocked;
+#endif
+#ifdef MDB_VL32
+#ifdef _WIN32
+	env->me_rpmutex = CreateMutex(NULL, FALSE, NULL);
+#else
+	pthread_mutex_init(env->me_rpmutex, NULL);
+#endif
 #endif
 
 	return MDB_SUCCESS;
@@ -4883,6 +5025,17 @@ mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode
 	if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
 		return EINVAL;
 
+#ifdef MDB_VL32
+	if (flags & MDB_WRITEMAP) {
+		/* silently ignore WRITEMAP in 32 bit mode */
+		flags ^= MDB_WRITEMAP;
+	}
+	if (flags & MDB_FIXEDMAP) {
+		/* cannot support FIXEDMAP */
+		return EINVAL;
+	}
+#endif
+
 	len = strlen(path);
 	if (flags & MDB_NOSUBDIR) {
 		rc = len + sizeof(LOCKSUFF) + len + 1;
@@ -4912,6 +5065,17 @@ mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode
 			  (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
 			rc = ENOMEM;
 	}
+#ifdef MDB_VL32
+	if (!rc) {
+		env->me_rpages = malloc(MDB_ERPAGE_SIZE * sizeof(MDB_ID3));
+		if (!env->me_rpages) {
+			rc = ENOMEM;
+			goto leave;
+		}
+		env->me_rpages[0].mid = 0;
+		env->me_rpcheck = MDB_ERPAGE_SIZE/2;
+	}
+#endif
 	env->me_flags = flags |= MDB_ENV_ACTIVE;
 	if (rc)
 		goto leave;
@@ -5010,6 +5174,16 @@ mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode
 				txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
 				txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
 				txn->mt_env = env;
+#ifdef MDB_VL32
+				txn->mt_rpages = malloc(MDB_TRPAGE_SIZE * sizeof(MDB_ID3));
+				if (!txn->mt_rpages) {
+					free(txn);
+					rc = ENOMEM;
+					goto leave;
+				}
+				txn->mt_rpages[0].mid = 0;
+				txn->mt_rpcheck = MDB_TRPAGE_SIZE/2;
+#endif
 				txn->mt_dbxs = env->me_dbxs;
 				txn->mt_flags = MDB_TXN_FINISHED;
 				env->me_txn0 = txn;
@@ -5049,6 +5223,13 @@ mdb_env_close0(MDB_env *env, int excl)
 	free(env->me_path);
 	free(env->me_dirty_list);
 	free(env->me_txn0);
+#ifdef MDB_VL32
+	{ unsigned int x;
+		for (x=1; x<=env->me_rpages[0].mid; x++)
+		munmap(env->me_rpages[x].mptr, env->me_rpages[x].mcnt * env->me_psize);
+	}
+	free(env->me_rpages);
+#endif
 	mdb_midl_free(env->me_free_pgs);
 
 	if (env->me_flags & MDB_ENV_TXKEY) {
@@ -5065,7 +5246,11 @@ mdb_env_close0(MDB_env *env, int excl)
 	}
 
 	if (env->me_map) {
+#ifdef MDB_VL32
+		munmap(env->me_map, NUM_METAS*env->me_psize);
+#else
 		munmap(env->me_map, env->me_mapsize);
+#endif
 	}
 	if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
 		(void) close(env->me_mfd);
@@ -5130,6 +5315,13 @@ mdb_env_close0(MDB_env *env, int excl)
 #endif
 		(void) close(env->me_lfd);
 	}
+#ifdef MDB_VL32
+#ifdef _WIN32
+	if (env->me_rpmutex) CloseHandle(env->me_rpmutex);
+#else
+	pthread_mutex_destroy(env->me_rpmutex);
+#endif
+#endif
 
 	env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
 }
@@ -5396,6 +5588,309 @@ mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
 	return MDB_SUCCESS;
 }
 
+#ifdef MDB_VL32
+/** Map a read-only page.
+ * There are two levels of tracking in use, a per-txn list and a per-env list.
+ * ref'ing and unref'ing the per-txn list is faster since it requires no
+ * locking. Pages are cached in the per-env list for global reuse, and a lock
+ * is required. Pages are not immediately unmapped when their refcnt goes to
+ * zero; they hang around in case they will be reused again soon.
+ *
+ * When the per-txn list gets full, all pages with refcnt=0 are purged from the
+ * list and their refcnts in the per-env list are decremented.
+ *
+ * When the per-env list gets full, all pages with refcnt=0 are purged from the
+ * list and their pages are unmapped.
+ *
+ * @note "full" means the list has reached its respective rpcheck threshold.
+ * This threshold slowly raises if no pages could be purged on a given check,
+ * and returns to its original value when enough pages were purged.
+ *
+ * If purging doesn't free any slots, filling the per-txn list will return
+ * MDB_TXN_FULL, and filling the per-env list returns MDB_MAP_FULL.
+ *
+ * Reference tracking in a txn is imperfect, pages can linger with non-zero
+ * refcnt even without active references. It was deemed to be too invasive
+ * to add unrefs in every required location. However, all pages are unref'd
+ * at the end of the transaction. This guarantees that no stale references
+ * linger in the per-env list.
+ *
+ * Usually we map chunks of 16 pages at a time, but if an overflow page begins
+ * at the tail of the chunk we extend the chunk to include the entire overflow
+ * page. Unfortunately, pages can be turned into overflow pages after their
+ * chunk was already mapped. In that case we must remap the chunk if the
+ * overflow page is referenced. If the chunk's refcnt is 0 we can just remap
+ * it, otherwise we temporarily map a new chunk just for the overflow page.
+ *
+ * @note this chunk handling means we cannot guarantee that a data item
+ * returned from the DB will stay alive for the duration of the transaction:
+ *   We unref pages as soon as a cursor moves away from the page
+ *   A subsequent op may cause a purge, which may unmap any unref'd chunks
+ * The caller must copy the data if it must be used later in the same txn.
+ *
+ * Also - our reference counting revolves around cursors, but overflow pages
+ * aren't ever pointed to by a cursor. As such, they always remain referenced
+ * in a txn until it ends.
+ *
+ * @param[in] txn the transaction for this access.
+ * @param[in] pgno the page number for the page to retrieve.
+ * @param[out] ret address of a pointer where the page's address will be stored.
+ * @return 0 on success, non-zero on failure.
+ */
+static int
+mdb_rpage_get(MDB_txn *txn, pgno_t pg0, MDB_page **ret)
+{
+	MDB_env *env = txn->mt_env;
+	MDB_page *p;
+	MDB_ID3L tl = txn->mt_rpages;
+	MDB_ID3L el = env->me_rpages;
+	MDB_ID3 id3;
+	unsigned x, rem;
+	pgno_t pgno;
+	int rc, retries = 1;
+#ifdef _WIN32
+	LARGE_INTEGER off;
+	SIZE_T len;
+#define SET_OFF(off,val)	off.QuadPart = val
+#define MAP(rc,env,addr,len,off)	\
+	addr = NULL; \
+	rc = NtMapViewOfSection(env->me_fmh, GetCurrentProcess(), &addr, 0, \
+		len, &off, &len, ViewUnmap, MEM_RESERVE, PAGE_READONLY)
+#else
+	off_t off;
+	size_t len;
+#define SET_OFF(off,val)	off = val
+#define MAP(rc,env,addr,len,off)	\
+	addr = mmap(NULL, len, PROT_READ, MAP_SHARED, env->me_fd, off); \
+	rc = (addr == MAP_FAILED) ? errno : 0
+#endif
+
+	/* remember the offset of the actual page number, so we can
+	 * return the correct pointer at the end.
+	 */
+	rem = pg0 & (MDB_RPAGE_CHUNK-1);
+	pgno = pg0 ^ rem;
+
+	id3.mid = 0;
+	x = mdb_mid3l_search(tl, pgno);
+	if (x <= tl[0].mid && tl[x].mid == pgno) {
+		if (x != tl[0].mid && tl[x+1].mid == pg0)
+			x++;
+		/* check for overflow size */
+		p = (MDB_page *)((char *)tl[x].mptr + rem * env->me_psize);
+		if (IS_OVERFLOW(p) && p->mp_pages + rem > tl[x].mcnt) {
+			id3.mcnt = p->mp_pages + rem;
+			len = id3.mcnt * env->me_psize;
+			SET_OFF(off, pgno * env->me_psize);
+			MAP(rc, env, id3.mptr, len, off);
+			if (rc)
+				return rc;
+			/* check for local-only page */
+			if (rem) {
+				mdb_tassert(txn, tl[x].mid != pg0);
+				/* hope there's room to insert this locally.
+				 * setting mid here tells later code to just insert
+				 * this id3 instead of searching for a match.
+				 */
+				id3.mid = pg0;
+				goto notlocal;
+			} else {
+				/* ignore the mapping we got from env, use new one */
+				tl[x].mptr = id3.mptr;
+				tl[x].mcnt = id3.mcnt;
+				/* if no active ref, see if we can replace in env */
+				if (!tl[x].mref) {
+					unsigned i;
+					LOCK_MUTEX0(env->me_rpmutex);
+					i = mdb_mid3l_search(el, tl[x].mid);
+					if (el[i].mref == 1) {
+						/* just us, replace it */
+						munmap(el[i].mptr, el[i].mcnt * env->me_psize);
+						el[i].mptr = tl[x].mptr;
+						el[i].mcnt = tl[x].mcnt;
+					} else {
+						/* there are others, remove ourself */
+						el[i].mref--;
+					}
+					UNLOCK_MUTEX(env->me_rpmutex);
+				}
+			}
+		}
+		id3.mptr = tl[x].mptr;
+		id3.mcnt = tl[x].mcnt;
+		tl[x].mref++;
+		goto ok;
+	}
+
+notlocal:
+	if (tl[0].mid >= MDB_TRPAGE_MAX - txn->mt_rpcheck) {
+		unsigned i, y;
+		/* purge unref'd pages from our list and unref in env */
+		LOCK_MUTEX0(env->me_rpmutex);
+retry:
+		y = 0;
+		for (i=1; i<tl[0].mid; i++) {
+			if (!tl[i].mref) {
+				if (!y) y = i;
+				/* tmp overflow pages don't go to env */
+				if (tl[i].mid & (MDB_RPAGE_CHUNK-1)) {
+					munmap(tl[i].mptr, tl[i].mcnt * env->me_psize);
+					continue;
+				}
+				x = mdb_mid3l_search(el, tl[i].mid);
+				el[x].mref--;
+			}
+		}
+		UNLOCK_MUTEX(env->me_rpmutex);
+		if (!y) {
+			/* we didn't find any unref'd chunks.
+			 * if we're out of room, fail.
+			 */
+			if (tl[0].mid >= MDB_TRPAGE_MAX)
+				return MDB_TXN_FULL;
+			/* otherwise, raise threshold for next time around
+			 * and let this go.
+			 */
+			txn->mt_rpcheck /= 2;
+		} else {
+			/* we found some unused; consolidate the list */
+			for (i=y+1; i<= tl[0].mid; i++)
+				if (tl[i].mref)
+					tl[y++] = tl[i];
+			tl[0].mid = y-1;
+			/* decrease the check threshold toward its original value */
+			if (!txn->mt_rpcheck)
+				txn->mt_rpcheck = 1;
+			while (txn->mt_rpcheck < tl[0].mid && txn->mt_rpcheck < MDB_TRPAGE_SIZE/2)
+				txn->mt_rpcheck *= 2;
+		}
+	}
+	if (tl[0].mid < MDB_TRPAGE_SIZE) {
+		id3.mref = 1;
+		if (id3.mid)
+			goto found;
+		len = env->me_psize * MDB_RPAGE_CHUNK;
+		id3.mid = pgno;
+		id3.mcnt = MDB_RPAGE_CHUNK;
+
+		/* search for page in env */
+		LOCK_MUTEX0(env->me_rpmutex);
+		x = mdb_mid3l_search(el, pgno);
+		if (x <= el[0].mid && el[x].mid == pgno) {
+			id3.mptr = el[x].mptr;
+			id3.mcnt = el[x].mcnt;
+			/* check for overflow size */
+			p = (MDB_page *)((char *)id3.mptr + rem * env->me_psize);
+			if (IS_OVERFLOW(p) && p->mp_pages + rem > id3.mcnt) {
+				id3.mcnt = p->mp_pages + rem;
+				len = id3.mcnt * env->me_psize;
+				SET_OFF(off, pgno * env->me_psize);
+				MAP(rc, env, id3.mptr, len, off);
+				if (rc)
+					goto fail;
+				if (!el[x].mref) {
+					munmap(el[x].mptr, el[x].mcnt);
+					el[x].mptr = id3.mptr;
+					el[x].mcnt = id3.mcnt;
+				} else {
+					id3.mid = pg0;
+					UNLOCK_MUTEX(env->me_rpmutex);
+					goto found;
+				}
+			}
+			el[x].mref++;
+			UNLOCK_MUTEX(env->me_rpmutex);
+			goto found;
+		}
+		if (el[0].mid >= MDB_ERPAGE_MAX - env->me_rpcheck) {
+			/* purge unref'd pages */
+			unsigned i, y = 0;
+			for (i=1; i<el[0].mid; i++) {
+				if (!el[i].mref) {
+					if (!y) y = i;
+					munmap(el[i].mptr, env->me_psize * el[i].mcnt);
+				}
+			}
+			if (!y) {
+				if (retries) {
+					/* see if we can unref some local pages */
+					retries--;
+					id3.mid = 0;
+					goto retry;
+				}
+				if (el[0].mid >= MDB_ERPAGE_MAX) {
+					UNLOCK_MUTEX(env->me_rpmutex);
+					return MDB_MAP_FULL;
+				}
+				env->me_rpcheck /= 2;
+			} else {
+				for (i=y+1; i<= el[0].mid; i++)
+					if (el[i].mref)
+						el[y++] = el[i];
+				el[0].mid = y-1;
+				if (!env->me_rpcheck)
+					env->me_rpcheck = 1;
+				while (env->me_rpcheck < el[0].mid && env->me_rpcheck < MDB_ERPAGE_SIZE/2)
+					env->me_rpcheck *= 2;
+			}
+		}
+		SET_OFF(off, pgno * env->me_psize);
+		MAP(rc, env, id3.mptr, len, off);
+		if (rc) {
+fail:
+			UNLOCK_MUTEX(env->me_rpmutex);
+			return rc;
+		}
+		/* If this page is far enough from the end of the env, scan for
+		 * any overflow pages that would spill onto another block.
+		 * Note we must compare against mt_last_pgno, the last written
+		 * page in the environment. Not mt_next_pgno, which increases
+		 * for every newly allocated (but not yet written) page. If
+		 * we scanned beyond the last written page we'd get a bus error.
+		 */
+		if (pgno + MDB_RPAGE_CHUNK <= txn->mt_last_pgno) {
+			int i;
+			char *cp = (char *)id3.mptr + rem * env->me_psize;
+			for (i=rem; i<MDB_RPAGE_CHUNK;) {
+				p = (MDB_page *)cp;
+				if (IS_OVERFLOW(p)) {
+					int nop = p->mp_pages;
+					if (nop + i > MDB_RPAGE_CHUNK) {
+						munmap(id3.mptr, len);
+						id3.mcnt = nop + i;
+						len = id3.mcnt * env->me_psize;
+						MAP(rc, env, id3.mptr, len, off);
+						if (rc)
+							goto fail;
+						break;
+					}
+					i += nop;
+					cp += nop * env->me_psize;
+				} else {
+					i++;
+					cp += env->me_psize;
+				}
+			}
+		}
+		mdb_mid3l_insert(el, &id3);
+		UNLOCK_MUTEX(env->me_rpmutex);
+found:
+		mdb_mid3l_insert(tl, &id3);
+	} else {
+		return MDB_TXN_FULL;
+	}
+ok:
+	p = (MDB_page *)((char *)id3.mptr + rem * env->me_psize);
+#if MDB_DEBUG	/* we don't need this check any more */
+	if (IS_OVERFLOW(p)) {
+		mdb_tassert(txn, p->mp_pages + rem <= id3.mcnt);
+	}
+#endif
+	*ret = p;
+	return MDB_SUCCESS;
+}
+#endif
+
 /** Find the address of the page corresponding to a given page number.
  * @param[in] txn the transaction for this access.
  * @param[in] pgno the page number for the page to retrieve.
@@ -5406,7 +5901,9 @@ mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
 static int
 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
 {
+#ifndef MDB_VL32
 	MDB_env *env = txn->mt_env;
+#endif
 	MDB_page *p = NULL;
 	int level;
 
@@ -5425,7 +5922,13 @@ mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
 				MDB_ID pn = pgno << 1;
 				x = mdb_midl_search(tx2->mt_spill_pgs, pn);
 				if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
+#ifdef MDB_VL32
+					int rc = mdb_rpage_get(txn, pgno, &p);
+					if (rc)
+						return rc;
+#else
 					p = (MDB_page *)(env->me_map + env->me_psize * pgno);
+#endif
 					goto done;
 				}
 			}
@@ -5442,7 +5945,15 @@ mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
 
 	if (pgno < txn->mt_next_pgno) {
 		level = 0;
+#ifdef MDB_VL32
+		{
+			int rc = mdb_rpage_get(txn, pgno, &p);
+			if (rc)
+				return rc;
+		}
+#else
 		p = (MDB_page *)(env->me_map + env->me_psize * pgno);
+#endif
 	} else {
 		DPRINTF(("page %"Y"u not found", pgno));
 		txn->mt_flags |= MDB_TXN_ERROR;
@@ -5617,10 +6128,22 @@ mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
 	}
 
 	mdb_cassert(mc, root > 1);
-	if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
+	if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root) {
+#ifdef MDB_VL32
+		if (mc->mc_pg[0])
+			MDB_PAGE_UNREF(mc->mc_txn, mc->mc_pg[0]);
+#endif
 		if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
 			return rc;
+	}
 
+#ifdef MDB_VL32
+	{
+		int i;
+		for (i=1; i<mc->mc_snum; i++)
+			MDB_PAGE_UNREF(mc->mc_txn, mc->mc_pg[i]);
+	}
+#endif
 	mc->mc_snum = 1;
 	mc->mc_top = 0;
 
@@ -5752,7 +6275,7 @@ mdb_get(MDB_txn *txn, MDB_dbi dbi,
 {
 	MDB_cursor	mc;
 	MDB_xcursor	mx;
-	int exact = 0;
+	int exact = 0, rc;
 	DKBUF;
 
 	DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
@@ -5764,7 +6287,16 @@ mdb_get(MDB_txn *txn, MDB_dbi dbi,
 		return MDB_BAD_TXN;
 
 	mdb_cursor_init(&mc, txn, dbi, &mx);
-	return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
+	rc = mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
+#ifdef MDB_VL32
+	{
+		/* unref all the pages - caller must copy the data
+		 * before doing anything else
+		 */
+		mdb_cursor_unref(&mc);
+	}
+#endif
+	return rc;
 }
 
 /** Find a sibling for a page.
@@ -5781,11 +6313,17 @@ mdb_cursor_sibling(MDB_cursor *mc, int move_right)
 	int		 rc;
 	MDB_node	*indx;
 	MDB_page	*mp;
+#ifdef MDB_VL32
+	MDB_page	*op;
+#endif
 
 	if (mc->mc_snum < 2) {
 		return MDB_NOTFOUND;		/* root has no siblings */
 	}
 
+#ifdef MDB_VL32
+	op = mc->mc_pg[mc->mc_top];
+#endif
 	mdb_cursor_pop(mc);
 	DPRINTF(("parent page is page %"Y"u, index %u",
 		mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
@@ -5810,6 +6348,8 @@ mdb_cursor_sibling(MDB_cursor *mc, int move_right)
 	}
 	mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
 
+	MDB_PAGE_UNREF(mc->mc_txn, op);
+
 	indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
 	if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
 		/* mc will be inconsistent if caller does mc_snum++ as above */
@@ -5851,6 +6391,13 @@ mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
 					return rc;
 				}
 			}
+#ifdef MDB_VL32
+			else {
+				if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
+					mdb_cursor_unref(&mc->mc_xcursor->mx_cursor);
+				}
+			}
+#endif
 		} else {
 			mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
 			if (op == MDB_NEXT_DUP)
@@ -5930,6 +6477,13 @@ mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
 					return rc;
 				}
 			}
+#ifdef MDB_VL32
+			else {
+				if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
+					mdb_cursor_unref(&mc->mc_xcursor->mx_cursor);
+				}
+			}
+#endif
 		} else {
 			mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
 			if (op == MDB_PREV_DUP)
@@ -6129,6 +6683,11 @@ set1:
 		return MDB_SUCCESS;
 	}
 
+#ifdef MDB_VL32
+	if (mc->mc_xcursor && mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
+		mdb_cursor_unref(&mc->mc_xcursor->mx_cursor);
+	}
+#endif
 	if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
 		mdb_xcursor_init1(mc, leaf);
 	}
@@ -6189,8 +6748,14 @@ mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
 	int		 rc;
 	MDB_node	*leaf;
 
-	if (mc->mc_xcursor)
+	if (mc->mc_xcursor) {
+#ifdef MDB_VL32
+		if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
+			mdb_cursor_unref(&mc->mc_xcursor->mx_cursor);
+		}
+#endif
 		mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
+	}
 
 	if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
 		rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
@@ -6233,8 +6798,14 @@ mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
 	int		 rc;
 	MDB_node	*leaf;
 
-	if (mc->mc_xcursor)
+	if (mc->mc_xcursor) {
+#ifdef MDB_VL32
+		if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
+			mdb_cursor_unref(&mc->mc_xcursor->mx_cursor);
+		}
+#endif
 		mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
+	}
 
 	if (!(mc->mc_flags & C_EOF)) {
 
@@ -9854,6 +10425,11 @@ mdb_drop0(MDB_cursor *mc, int subs)
 			mdb_cursor_pop(mc);
 
 		mdb_cursor_copy(mc, &mx);
+#ifdef MDB_VL32
+		/* bump refcount for mx's pages */
+		for (i=0; i<mc->mc_snum; i++)
+			mdb_page_get(txn, mc->mc_pg[i]->mp_pgno, &mx.mc_pg[i], NULL);
+#endif
 		while (mc->mc_snum > 0) {
 			MDB_page *mp = mc->mc_pg[mc->mc_top];
 			unsigned n = NUMKEYS(mp);
@@ -9913,6 +10489,10 @@ mdb_drop0(MDB_cursor *mc, int subs)
 done:
 		if (rc)
 			txn->mt_flags |= MDB_TXN_ERROR;
+#ifdef MDB_VL32
+		/* drop refcount for mx's pages */
+		mdb_cursor_unref(&mx);
+#endif
 	} else if (rc == MDB_NOTFOUND) {
 		rc = MDB_SUCCESS;
 	}

libraries/liblmdb/midl.c

@@ -354,5 +354,67 @@ int mdb_mid2l_append( MDB_ID2L ids, MDB_ID2 *id )
 	return 0;
 }
 
+#ifdef MDB_VL32
+unsigned mdb_mid3l_search( MDB_ID3L ids, MDB_ID id )
+{
+	/*
+	 * binary search of id in ids
+	 * if found, returns position of id
+	 * if not found, returns first position greater than id
+	 */
+	unsigned base = 0;
+	unsigned cursor = 1;
+	int val = 0;
+	unsigned n = (unsigned)ids[0].mid;
+
+	while( 0 < n ) {
+		unsigned pivot = n >> 1;
+		cursor = base + pivot + 1;
+		val = CMP( id, ids[cursor].mid );
+
+		if( val < 0 ) {
+			n = pivot;
+
+		} else if ( val > 0 ) {
+			base = cursor;
+			n -= pivot + 1;
+
+		} else {
+			return cursor;
+		}
+	}
+
+	if( val > 0 ) {
+		++cursor;
+	}
+	return cursor;
+}
+
+int mdb_mid3l_insert( MDB_ID3L ids, MDB_ID3 *id )
+{
+	unsigned x, i;
+
+	x = mdb_mid3l_search( ids, id->mid );
+
+	if( x < 1 ) {
+		/* internal error */
+		return -2;
+	}
+
+	if ( x <= ids[0].mid && ids[x].mid == id->mid ) {
+		/* duplicate */
+		return -1;
+	}
+
+	/* insert id */
+	ids[0].mid++;
+	for (i=(unsigned)ids[0].mid; i>x; i--)
+		ids[i] = ids[i-1];
+	ids[x] = *id;
+
+	return 0;
+}
+#endif /* MDB_VL32 */
+
 /** @} */
 /** @} */

libraries/liblmdb/midl.h

@@ -186,6 +186,20 @@ int mdb_mid2l_insert( MDB_ID2L ids, MDB_ID2 *id );
 	 */
 int mdb_mid2l_append( MDB_ID2L ids, MDB_ID2 *id );
 
+#ifdef MDB_VL32
+typedef struct MDB_ID3 {
+	MDB_ID mid;		/**< The ID */
+	void *mptr;		/**< The pointer */
+	unsigned int mcnt;		/**< Number of pages */
+	unsigned int mref;		/**< Refcounter */
+} MDB_ID3;
+
+typedef MDB_ID3 *MDB_ID3L;
+
+unsigned mdb_mid3l_search( MDB_ID3L ids, MDB_ID id );
+int mdb_mid3l_insert( MDB_ID3L ids, MDB_ID3 *id );
+
+#endif /* MDB_VL32 */
 /** @} */
 /** @} */
 #ifdef __cplusplus