WT-2737 Page scrubbing: more fixes. (#2899)

* WT-2737 Page scrubbing: more fixes.

* Swap endianness of the page header back after writing (for zSeries).
* Deal with restoring updates when the WT_EVICT_SCRUB flag is not set:
  we want to save disk images for only those pages that have updates to
  restore.  Unfortunately, that isn't easy to figure out down in the
  guts of raw compression.

* typo

* Move __wt_page_header_byteswap() calls to immediately before/after the write
call, so error paths don't leave the caller's content in an unexpected format.

* For now, WT_EVICT_UPDATE_RESTORE implies WT_EVICT_SCRUB; we may be able to
relax this in the future.

* whitespace

* If raw compression gives up, the split boundary's already-compressed
flag won't be true, but we will have a saved disk image. We can change
raw compression to only create disk images if the already-compressed
flag will be true, or always create a disk image in raw-compression,
regardless of whether or not we actually compressed data.  I'm going
with the latter, it's simpler.

* Fix a comment.

3 files changed, 26 insertions, 22 deletions

diff --git a/src/block/block_write.c b/src/block/block_write.c
index e6e992436a3..8a77641e1f3 100644
--- a/src/block/block_write.c
+++ b/src/block/block_write.c
@@ -254,12 +254,6 @@ __wt_block_write_off(WT_SESSION_IMPL *session, WT_BLOCK *block,
 	blk = WT_BLOCK_HEADER_REF(buf->mem);
 	memset(blk, 0, sizeof(*blk));
 
-	/*
-	 * Swap the page-header as needed; this doesn't belong here, but it's
-	 * the best place to catch all callers.
-	 */
-	__wt_page_header_byteswap(buf->mem);
-
 	/* Buffers should be aligned for writing. */
 	if (!F_ISSET(buf, WT_ITEM_ALIGNED)) {
 		WT_ASSERT(session, F_ISSET(buf, WT_ITEM_ALIGNED));
@@ -345,9 +339,16 @@ __wt_block_write_off(WT_SESSION_IMPL *session, WT_BLOCK *block,
 		__wt_spin_unlock(session, &block->live_lock);
 	WT_RET(ret);
 
-	/* Write the block. */
-	if ((ret =
-	    __wt_write(session, fh, offset, align_size, buf->mem)) != 0) {
+	/*
+	 * Ensure the page header is in little endian order; this doesn't belong
+	 * here, but it's the best place to catch all callers. After the write,
+	 * swap values back to native order so callers never see anything other
+	 * than their original content.
+	 */
+	__wt_page_header_byteswap(buf->mem);
+	ret = __wt_write(session, fh, offset, align_size, buf->mem);
+	__wt_page_header_byteswap(buf->mem);
+	if (ret != 0) {
 		if (!caller_locked)
 			__wt_spin_lock(session, &block->live_lock);
 		WT_TRET(__wt_block_off_free(
diff --git a/src/evict/evict_page.c b/src/evict/evict_page.c
index 5ef6a6370b0..6f5d04cc678 100644
--- a/src/evict/evict_page.c
+++ b/src/evict/evict_page.c
@@ -504,8 +504,7 @@ __evict_review(
 		LF_SET(WT_VISIBILITY_ERR);
 	else if (!WT_PAGE_IS_INTERNAL(page)) {
 		if (F_ISSET(S2C(session), WT_CONN_IN_MEMORY))
-			LF_SET(WT_EVICT_IN_MEMORY |
-			    WT_EVICT_UPDATE_RESTORE | WT_EVICT_SCRUB);
+			LF_SET(WT_EVICT_IN_MEMORY);
 		else if (page->read_gen == WT_READGEN_OLDEST ||
 		    page->memory_footprint > S2BT(session)->splitmempage)
 			LF_SET(WT_EVICT_UPDATE_RESTORE);
diff --git a/src/reconcile/rec_write.c b/src/reconcile/rec_write.c
index f08c6c5a825..a6a7c07ea59 100644
--- a/src/reconcile/rec_write.c
+++ b/src/reconcile/rec_write.c
@@ -382,6 +382,15 @@ __wt_reconcile(WT_SESSION_IMPL *session,
 	WT_ASSERT(session, WT_TXNID_LE(mod->last_oldest_id, oldest_id));
 	mod->last_oldest_id = oldest_id;
 
+	/*
+	 * Evicting in-memory uses the update/restore mechanisms.
+	 * The update/restore mechanisms use disk images.
+	 */
+	if (LF_ISSET(WT_EVICT_IN_MEMORY))
+		LF_SET(WT_EVICT_UPDATE_RESTORE);
+	if (LF_ISSET(WT_EVICT_UPDATE_RESTORE))
+		LF_SET(WT_EVICT_SCRUB);
+
 	/* Initialize the reconciliation structure for each new run. */
 	if ((ret = __rec_write_init(
 	    session, ref, flags, salvage, &session->reconcile)) != 0) {
@@ -2783,7 +2792,7 @@ no_slots:
 
 		/*
 		 * Optionally keep the disk image in cache. Update the initial
-		 * fields to reflect the actual disk image that was compressed.
+		 * page-header fields to reflect the actual data being written.
 		 */
 		if (F_ISSET(r, WT_EVICT_SCRUB)) {
 			WT_RET(__wt_strndup(session, dsk,
@@ -3327,17 +3336,12 @@ supd_check_complete:
 
 copy_image:
 	/*
-	 * Optionally keep the disk image in cache (raw compression has already
-	 * made a copy).
+	 * Optionally keep the disk image in cache (raw compression may have
+	 * already made a copy).
 	 */
-	if (F_ISSET(r, WT_EVICT_SCRUB)) {
-		WT_ASSERT(session,
-		    (bnd->already_compressed && bnd->disk_image != NULL) ||
-		    (!bnd->already_compressed && bnd->disk_image == NULL));
-		if (bnd->disk_image == NULL)
-			WT_ERR(__wt_strndup(
-			    session, buf->data, buf->size, &bnd->disk_image));
-	}
+	if (F_ISSET(r, WT_EVICT_SCRUB) && bnd->disk_image == NULL)
+		WT_ERR(__wt_strndup(
+		    session, buf->data, buf->size, &bnd->disk_image));
 
 err:	__wt_scr_free(session, &key);
 	return (ret);