Add mmap support for read-only objects (a read-only object is an object

using a checkpoint other than the "live" system).  Basically, when a
read-only object reads pages from the backing file, it maps them into
memory rather than copying them into the buffers.

This change adds support for a new class of WT_ITEM buffer, one that has
WT_ITEM_MAPPED set.  The WT_ITEM_MAPPED flag gets set when the buffer's
object is mapped into memory by the underlying block manager read call.

Add configuration string "mmap", defaults to on, turns off mmap calls.

1 files changed, 21 insertions, 2 deletions

diff --git a/src/support/scratch.c b/src/support/scratch.c
index 6c4e77289a7..9cda695ec89 100644
--- a/src/support/scratch.c
+++ b/src/support/scratch.c
@@ -22,7 +22,11 @@ __wt_buf_clear(WT_ITEM *buf)
 	buf->mem = NULL;
 	buf->memsize = 0;
 
-	/* Note: don't clear the flags, the buffer remains marked in-use. */
+	/*
+	 * Note: don't clear the flags, the buffer remains marked for aligned
+	 * use as well as "in-use".
+	 */
+	F_CLR(buf, WT_ITEM_MAPPED);
 }
 
 /*
@@ -37,6 +41,10 @@ __wt_buf_grow(WT_SESSION_IMPL *session, WT_ITEM *buf, size_t size)
 
 	WT_ASSERT(session, size <= UINT32_MAX);
 
+	/* Clear buffers previously used for mapped returns. */
+	if (F_ISSET(buf, WT_ITEM_MAPPED))
+		__wt_buf_clear(buf);
+
 	if (size > buf->memsize) {
 		/*
 		 * Grow the buffer's memory: if the data reference is not set
@@ -132,6 +140,8 @@ __wt_buf_steal(WT_SESSION_IMPL *session, WT_ITEM *buf, uint32_t *sizep)
 {
 	void *retp;
 
+	WT_ASSERT(session, !F_ISSET(buf, WT_ITEM_MAPPED));
+
 	/*
 	 * Sometimes we steal a buffer for a different purpose, for example,
 	 * we've read in an overflow item, and now it's going to become a key
@@ -169,7 +179,8 @@ __wt_buf_steal(WT_SESSION_IMPL *session, WT_ITEM *buf, uint32_t *sizep)
 void
 __wt_buf_free(WT_SESSION_IMPL *session, WT_ITEM *buf)
 {
-	__wt_free(session, buf->mem);
+	if (!F_ISSET(buf, WT_ITEM_MAPPED))
+		__wt_free(session, buf->mem);
 	__wt_buf_clear(buf);
 }
 
@@ -184,6 +195,10 @@ __wt_buf_fmt(WT_SESSION_IMPL *session, WT_ITEM *buf, const char *fmt, ...)
 	va_list ap;
 	size_t len;
 
+	/* Clear buffers previously used for mapped returns. */
+	if (F_ISSET(buf, WT_ITEM_MAPPED))
+		__wt_buf_clear(buf);
+
 	for (;;) {
 		va_start(ap, fmt);
 		len = (size_t)vsnprintf(buf->mem, buf->memsize, fmt, ap);
@@ -217,6 +232,10 @@ __wt_buf_catfmt(WT_SESSION_IMPL *session, WT_ITEM *buf, const char *fmt, ...)
 	size_t len, space;
 	char *p;
 
+	/* Clear buffers previously used for mapped returns. */
+	if (F_ISSET(buf, WT_ITEM_MAPPED))
+		__wt_buf_clear(buf);
+
 	for (;;) {
 		va_start(ap, fmt);
 		p = (char *)((uint8_t *)buf->mem + buf->size);