From: Rusty Russell <rusty@rustcorp.com.au>
Date: Tue, 20 Dec 2011 07:09:28 +0000 (+1030)
Subject: tdb: use same expansion factor logic when expanding for new recovery area.
X-Git-Url: http://git.ozlabs.org/?p=ccan;a=commitdiff_plain;h=08b460e9dc91af96b8a2628e275fd3495fbd56e0

tdb: use same expansion factor logic when expanding for new recovery area.

If we're expanding because the current recovery area is too small, we
expand only the amount we need.  This can quickly lead to exponential
growth when we have a slowly-expanding record (hence a
slowly-expanding transaction size).
---

diff --git a/ccan/tdb/io.c b/ccan/tdb/io.c
index 2262eabf..77f26adf 100644
--- a/ccan/tdb/io.c
+++ b/ccan/tdb/io.c
@@ -303,13 +303,39 @@ static int tdb_expand_file(struct tdb_context *tdb, tdb_off_t size, tdb_off_t ad
 	return 0;
 }
 
+/* You need 'size', this tells you how much you should expand by. */
+tdb_off_t tdb_expand_adjust(tdb_off_t map_size, tdb_off_t size, int page_size)
+{
+	tdb_off_t new_size, top_size;
+
+	/* limit size in order to avoid using up huge amounts of memory for
+	 * in memory tdbs if an oddball huge record creeps in */
+	if (size > 100 * 1024) {
+		top_size = map_size + size * 2;
+	} else {
+		top_size = map_size + size * 100;
+	}
+
+	/* always make room for at least top_size more records, and at
+	   least 25% more space. if the DB is smaller than 100MiB,
+	   otherwise grow it by 10% only. */
+	if (map_size > 100 * 1024 * 1024) {
+		new_size = map_size * 1.10;
+	} else {
+		new_size = map_size * 1.25;
+	}
+
+	/* Round the database up to a multiple of the page size */
+	new_size = MAX(top_size, new_size);
+	return TDB_ALIGN(new_size, page_size) - map_size;
+}
 
 /* expand the database at least size bytes by expanding the underlying
    file and doing the mmap again if necessary */
 int tdb_expand(struct tdb_context *tdb, tdb_off_t size)
 {
 	struct tdb_record rec;
-	tdb_off_t offset, new_size;	
+	tdb_off_t offset;
 
 	if (tdb_lock(tdb, -1, F_WRLCK) == -1) {
 		TDB_LOG((tdb, TDB_DEBUG_ERROR, "lock failed in tdb_expand\n"));
@@ -319,11 +345,7 @@ int tdb_expand(struct tdb_context *tdb, tdb_off_t size)
 	/* must know about any previous expansions by another process */
 	tdb->methods->tdb_oob(tdb, tdb->map_size + 1, 1);
 
-	/* always make room for at least 100 more records, and at
-           least 25% more space. Round the database up to a multiple
-           of the page size */
-	new_size = MAX(tdb->map_size + size*100, tdb->map_size * 1.25);
-	size = TDB_ALIGN(new_size, tdb->page_size) - tdb->map_size;
+	size = tdb_expand_adjust(tdb->map_size, size, tdb->page_size);
 
 	if (!(tdb->flags & TDB_INTERNAL))
 		tdb_munmap(tdb);
diff --git a/ccan/tdb/tdb_private.h b/ccan/tdb/tdb_private.h
index 4faf8989..f1c841c2 100644
--- a/ccan/tdb/tdb_private.h
+++ b/ccan/tdb/tdb_private.h
@@ -298,6 +298,7 @@ tdb_off_t tdb_find_lock_hash(struct tdb_context *tdb, TDB_DATA key, uint32_t has
 			   struct tdb_record *rec);
 void tdb_io_init(struct tdb_context *tdb);
 int tdb_expand(struct tdb_context *tdb, tdb_off_t size);
+tdb_off_t tdb_expand_adjust(tdb_off_t map_size, tdb_off_t size, int page_size);
 int tdb_rec_free_read(struct tdb_context *tdb, tdb_off_t off,
 		      struct tdb_record *rec);
 void tdb_header_hash(struct tdb_context *tdb,
diff --git a/ccan/tdb/test/run-transaction-expand.c b/ccan/tdb/test/run-transaction-expand.c
index 463cb106..a46669d6 100644
--- a/ccan/tdb/test/run-transaction-expand.c
+++ b/ccan/tdb/test/run-transaction-expand.c
@@ -41,13 +41,10 @@ static void write_record(struct tdb_context *tdb, size_t extra_len,
 	key.dsize = strlen("hi");
 	key.dptr = (void *)"hi";
 
-	data->dptr = realloc(data->dptr, data->dsize + extra_len);
-	memset(data->dptr + data->dsize, 'x', extra_len);
 	data->dsize += extra_len;
 	tdb_transaction_start(tdb);
 	tdb_store(tdb, key, *data, TDB_REPLACE);
 	tdb_transaction_commit(tdb);
-	diag("TDB size = %zu", (size_t)tdb->map_size);
 }
 
 int main(int argc, char *argv[])
@@ -55,6 +52,8 @@ int main(int argc, char *argv[])
 	struct tdb_context *tdb;
 	size_t i;
 	TDB_DATA data;
+	struct tdb_record rec;
+	tdb_off_t off;
 
 	plan_tests(2);
 	tdb = tdb_open_ex("run-transaction-expand.tdb",
@@ -63,14 +62,19 @@ int main(int argc, char *argv[])
 	ok1(tdb);
 
 	data.dsize = 0;
-	data.dptr = NULL;
+	data.dptr = calloc(1000, getpagesize());
 
 	/* Simulate a slowly growing record. */
 	for (i = 0; i < 1000; i++)
 		write_record(tdb, getpagesize(), &data);
 
-	/* We should only be about 3 times larger than largest record. */
-	ok1(tdb->map_size < 3 * i * getpagesize());
+	tdb_ofs_read(tdb, TDB_RECOVERY_HEAD, &off);
+	tdb_read(tdb, off, &rec, sizeof(rec), DOCONV());
+	diag("TDB size = %zu, recovery = %u-%u",
+	     (size_t)tdb->map_size, off, off + sizeof(rec) + rec.rec_len);
+
+	/* We should only be about 5 times larger than largest record. */
+	ok1(tdb->map_size < 6 * i * getpagesize());
 	tdb_close(tdb);
 	free(data.dptr);
 
diff --git a/ccan/tdb/transaction.c b/ccan/tdb/transaction.c
index 11194773..525e0642 100644
--- a/ccan/tdb/transaction.c
+++ b/ccan/tdb/transaction.c
@@ -699,7 +699,11 @@ static int tdb_recovery_allocate(struct tdb_context *tdb,
 	*recovery_size = tdb_recovery_size(tdb);
 
 	/* round up to a multiple of page size */
-	*recovery_max_size = TDB_ALIGN(sizeof(rec) + *recovery_size, tdb->page_size) - sizeof(rec);
+	*recovery_max_size = tdb_expand_adjust(tdb->map_size,
+					       *recovery_size,
+					       tdb->page_size)
+		- sizeof(rec);
+
 	*recovery_offset = tdb->map_size;
 	recovery_head = *recovery_offset;