Import wiredtiger: 38e948b9eb4ab532372ea62d0bc18b927007b079 from branch mongodb-4.4

ref: 885403077c..38e948b9eb
for: 4.4.6

WT-7426       Set write generation number when the page image gets created

4 files changed, 33 insertions, 17 deletions

diff --git a/src/third_party/wiredtiger/import.data b/src/third_party/wiredtiger/import.data
index 1c581648cb2..d952484a5f3 100644
--- a/src/third_party/wiredtiger/import.data
+++ b/src/third_party/wiredtiger/import.data
@@ -2,5 +2,5 @@
     "vendor": "wiredtiger",
     "github": "wiredtiger/wiredtiger.git",
     "branch": "mongodb-4.4",
-    "commit": "885403077cfdbabc8107b9522af57513952e6f85"
+    "commit": "38e948b9eb4ab532372ea62d0bc18b927007b079"
 }
diff --git a/src/third_party/wiredtiger/src/btree/bt_io.c b/src/third_party/wiredtiger/src/btree/bt_io.c
index 611b4d3dc64..373cc7b71f1 100644
--- a/src/third_party/wiredtiger/src/btree/bt_io.c
+++ b/src/third_party/wiredtiger/src/btree/bt_io.c
@@ -303,20 +303,7 @@ __wt_bt_write(WT_SESSION_IMPL *session, WT_ITEM *buf, uint8_t *addr, size_t *add
     if (encrypted)
         F_SET(dsk, WT_PAGE_ENCRYPTED);
 
-    /*
-     * We increment the block's write generation so it's easy to identify newer versions of blocks
-     * during salvage. (It's common in WiredTiger, at least for the default block manager, for
-     * multiple blocks to be internally consistent with identical first and last keys, so we need a
-     * way to know the most recent state of the block. We could check which leaf is referenced by a
-     * valid internal page, but that implies salvaging internal pages, which I don't want to do, and
-     * it's not as good anyway, because the internal page may not have been written after the leaf
-     * page was updated. So, write generations it is.
-     *
-     * Nothing is locked at this point but two versions of a page with the same generation is pretty
-     * unlikely, and if we did, they're going to be roughly identical for the purposes of salvage,
-     * anyway.
-     */
-    dsk->write_gen = ++btree->write_gen;
+    WT_ASSERT(session, (dsk->write_gen != 0 && dsk->write_gen > btree->base_write_gen));
 
     /*
      * Checksum the data if the buffer isn't compressed or checksums are configured.
diff --git a/src/third_party/wiredtiger/src/include/cell_inline.h b/src/third_party/wiredtiger/src/include/cell_inline.h
index 70233e3f9cd..d3e0f85359d 100644
--- a/src/third_party/wiredtiger/src/include/cell_inline.h
+++ b/src/third_party/wiredtiger/src/include/cell_inline.h
@@ -970,7 +970,8 @@ __cell_unpack_window_cleanup(WT_SESSION_IMPL *session, const WT_PAGE_HEADER *dsk
      * No delete              txnid=MAX, ts=MAX,            txnid=MAX, ts=MAX,
      *                        durable_ts=NONE               durable_ts=NONE
      */
-    if (dsk->write_gen == 0 || dsk->write_gen > S2BT(session)->base_write_gen)
+    WT_ASSERT(session, dsk->write_gen != 0);
+    if (dsk->write_gen > S2BT(session)->base_write_gen)
         return;
 
     /* Tell reconciliation we cleared the transaction ids and the cell needs to be rebuilt. */
diff --git a/src/third_party/wiredtiger/src/reconcile/rec_write.c b/src/third_party/wiredtiger/src/reconcile/rec_write.c
index 3e21b7f3ca0..97743d7c3ee 100644
--- a/src/third_party/wiredtiger/src/reconcile/rec_write.c
+++ b/src/third_party/wiredtiger/src/reconcile/rec_write.c
@@ -1514,6 +1514,32 @@ err:
 }
 
 /*
+ * __rec_set_page_write_gen --
+ *     Initialize the page write generation number.
+ */
+static void
+__rec_set_page_write_gen(WT_PAGE_HEADER *dsk, WT_BTREE *btree)
+{
+    /*
+     * We increment the block's write generation so it's easy to identify newer versions of blocks
+     * during salvage. (It's common in WiredTiger, at least for the default block manager, for
+     * multiple blocks to be internally consistent with identical first and last keys, so we need a
+     * way to know the most recent state of the block. We could check which leaf is referenced by a
+     * valid internal page, but that implies salvaging internal pages, which I don't want to do, and
+     * it's not as good anyway, because the internal page may not have been written after the leaf
+     * page was updated. So, write generations it is.
+     *
+     * Nothing is locked at this point but two versions of a page with the same generation is pretty
+     * unlikely, and if we did, they're going to be roughly identical for the purposes of salvage,
+     * anyway.
+     *
+     * Other than salvage, the write generation number is used to reset the stale transaction id's
+     * present on the page upon server restart.
+     */
+    dsk->write_gen = ++btree->write_gen;
+}
+
+/*
  * __rec_split_write_header --
  *     Initialize a disk page's header.
  */
@@ -1528,7 +1554,8 @@ __rec_split_write_header(WT_SESSION_IMPL *session, WT_RECONCILE *r, WT_REC_CHUNK
     page = r->page;
 
     dsk->recno = btree->type == BTREE_ROW ? WT_RECNO_OOB : multi->key.recno;
-    dsk->write_gen = 0;
+
+    __rec_set_page_write_gen(dsk, btree);
     dsk->mem_size = multi->size;
     dsk->u.entries = chunk->entries;
     dsk->type = page->type;
@@ -2341,6 +2368,7 @@ __wt_rec_cell_build_ovfl(WT_SESSION_IMPL *session, WT_RECONCILE *r, WT_REC_KV *k
         dsk = tmp->mem;
         memset(dsk, 0, WT_PAGE_HEADER_SIZE);
         dsk->type = WT_PAGE_OVFL;
+        __rec_set_page_write_gen(dsk, btree);
         dsk->u.datalen = (uint32_t)kv->buf.size;
         memcpy(WT_PAGE_HEADER_BYTE(btree, dsk), kv->buf.data, kv->buf.size);
         dsk->mem_size = WT_PAGE_HEADER_BYTE_SIZE(btree) + (uint32_t)kv->buf.size;