Second set of snapshot changes, ref #167:

Start using the alloc and discard lists instead of running everything
through the single avail list.  This primarily affects what happens when
a snapshot is deleted.  A snapshot's avail list is discarded; alloc and
discard lists are rolled-forward into the live system's corresponding
lists, and then the alloc and discard lists are checked for overlaps,
the overlaps are the blocks that are newly available for re-use.

Change extent writes to always append to the file (the problem is that
we can't allocate the avail list blocks from the avail list, and it's
simpler to always extend the file for extent lists rather than try to
figure out which extent lists writes are extending the file and which
aren't.)  This implies that we can tell the allocation code to always
extend the file, and the free code to simply put the free'd blocks on
the avail list, because extent blocks don't appear in the allocation
list.

Change page reconciliation to release all pages in the page-tracking
information before writing a snapshot, the underlying system has to be
fully consistent.  We can no longer retain currently unused overflow
pages to re-use, across reconciliation calls: if a reconciliation call
results in a snapshot, any overflow page not used as part of the page
reconciliation must be discarded, otherwise the underlying snapshot
won't be completely consistent.  This adds some (very!) few cases where
we don't re-use overflow items we would have re-used in earlier versions
of the system, but it shouldn't be common.  This required an extensive
re-work of the block/overflow page tracking code.

Changes for salvage: load and unload snapshots during salvage, blocks
free'd during salvage do not appear on the allocation list, so must be
free'd explicitly to the system's avail list.

Rename block-manager page type WT_PAGE_FREELIST to WT_PAGE_BLOCK_MANAGER.
We still handle block-manager pages in the btree layer during salvage,
but we could also handle them in the salvage code if we're willing to
look at the page type.

Set the snapshot field names when cracking the buffer cookie.

Don't display the by-size extent lists by default, it's not very useful.

1 files changed, 13 insertions, 7 deletions

diff --git a/src/block/block_slvg.c b/src/block/block_slvg.c
index 8b23547e7b6..baca1bb06a2 100644
--- a/src/block/block_slvg.c
+++ b/src/block/block_slvg.c
@@ -17,6 +17,9 @@ __wt_block_salvage_start(WT_SESSION_IMPL *session, WT_BLOCK *block)
 	off_t len;
 	uint32_t allocsize;
 
+	/* Reset the live snapshot information. */
+	WT_RET(__wt_block_snap_init(session, block, &block->live, 1));
+
 	/*
 	 * Truncate the file to an initial sector plus N allocation size
 	 * units (bytes trailing the last multiple of an allocation size
@@ -38,10 +41,11 @@ __wt_block_salvage_start(WT_SESSION_IMPL *session, WT_BLOCK *block)
 	block->slvg_off = WT_BLOCK_DESC_SECTOR;
 
 	/*
-	 * We don't currently need to do anything about the freelist because
-	 * we don't read it for salvage operations.
+	 * We don't currently need to do anything about the snapshot extents
+	 * because we don't read them for salvage operations.
 	 */
 
+	block->live_load = block->slvg = 1;
 	return (0);
 }
 
@@ -50,11 +54,13 @@ __wt_block_salvage_start(WT_SESSION_IMPL *session, WT_BLOCK *block)
  *	End a file salvage.
  */
 int
-__wt_block_salvage_end(WT_SESSION_IMPL *session, WT_BLOCK *block, int success)
+__wt_block_salvage_end(WT_SESSION_IMPL *session, WT_BLOCK *block)
 {
-	/* If not successful, discard the live snapshot we've created. */
-	if (!success)
-		(void)__wt_block_snap_unload(session, block);
+	/* Discard the live snapshot. */
+	WT_RET(__wt_block_snap_unload(session, block));
+
+	block->slvg = 0;
+
 	return (0);
 }
 
@@ -122,7 +128,7 @@ skip:			WT_VERBOSE(session, salvage,
 			 * than once.
 			 */
 			WT_RET(__wt_block_free(
-			    session, block, offset, (off_t)allocsize));
+			    session, block, offset, (off_t)allocsize, 0));
 			block->slvg_off = offset += allocsize;
 			continue;
 		}