Import wiredtiger-wiredtiger-mongodb-2.8-rc5-9-g46211a1.tar.gz from wiredtiger branch mongodb-2.8r2.8.0-rc5

4 files changed, 63 insertions, 14 deletions

diff --git a/src/third_party/wiredtiger/src/btree/bt_split.c b/src/third_party/wiredtiger/src/btree/bt_split.c
index ca7f2266f2a..4041bbb71b0 100644
--- a/src/third_party/wiredtiger/src/btree/bt_split.c
+++ b/src/third_party/wiredtiger/src/btree/bt_split.c
@@ -719,6 +719,14 @@ __split_multi_inmem(
 	 */
 	page->modify->first_dirty_txn = WT_TXN_FIRST;
 
+	/*
+	 * XXX Don't allow this page to be evicted immediately.
+	 *
+	 * In some cases involving forced eviction during truncates, a reader
+	 * ends up looking at an evicted page.  This is a temporary workaround.
+	 */
+	page->modify->inmem_split_txn = __wt_txn_new_id(session);
+
 err:	/* Free any resources that may have been cached in the cursor. */
 	WT_TRET(__wt_btcur_close(&cbt));
 
diff --git a/src/third_party/wiredtiger/src/conn/conn_log.c b/src/third_party/wiredtiger/src/conn/conn_log.c
index 2799a58f327..d15b27f9bc9 100644
--- a/src/third_party/wiredtiger/src/conn/conn_log.c
+++ b/src/third_party/wiredtiger/src/conn/conn_log.c
@@ -126,11 +126,13 @@ __log_archive_once(WT_SESSION_IMPL *session, uint32_t backup_file)
 	/*
 	 * If we're coming from a backup cursor we want the smaller of
 	 * the last full log file copied in backup or the checkpoint LSN.
+	 * Otherwise we want the minimum of the last log file written to
+	 * disk and the checkpoint LSN.
 	 */
 	if (backup_file != 0)
 		min_lognum = WT_MIN(log->ckpt_lsn.file, backup_file);
 	else
-		min_lognum = log->ckpt_lsn.file;
+		min_lognum = WT_MIN(log->ckpt_lsn.file, log->sync_lsn.file);
 	WT_RET(__wt_verbose(session, WT_VERB_LOG,
 	    "log_archive: archive to log number %" PRIu32, min_lognum));
 
diff --git a/src/third_party/wiredtiger/src/log/log.c b/src/third_party/wiredtiger/src/log/log.c
index 944e748a6a8..181e77128a9 100644
--- a/src/third_party/wiredtiger/src/log/log.c
+++ b/src/third_party/wiredtiger/src/log/log.c
@@ -859,7 +859,7 @@ __log_release(WT_SESSION_IMPL *session, WT_LOGSLOT *slot)
 	WT_DECL_RET;
 	WT_FH *close_fh;
 	WT_LOG *log;
-	WT_LSN sync_lsn;
+	WT_LSN close_end_lsn, close_lsn, sync_lsn;
 	size_t write_size;
 	int locked;
 	WT_DECL_SPINLOCK_ID(id);			/* Must appear last */
@@ -873,8 +873,22 @@ __log_release(WT_SESSION_IMPL *session, WT_LOGSLOT *slot)
 	 * of the file handle structure.
 	 */
 	close_fh = NULL;
+	WT_INIT_LSN(&close_lsn);
+	WT_INIT_LSN(&close_end_lsn);
 	if (F_ISSET(slot, SLOT_CLOSEFH)) {
 		close_fh = log->log_close_fh;
+		/*
+		 * Set the close_end_lsn to the LSN immediately after ours.
+		 * That is, the beginning of the next log file.  We need to
+		 * know the LSN file number of our own close in case earlier
+		 * calls are still in progress and the next one to move the
+		 * sync_lsn into the next file for later syncs.
+		 */
+		WT_ERR(__wt_log_extract_lognum(session, close_fh->name,
+		    &close_lsn.file));
+		close_lsn.offset = 0;
+		close_end_lsn = close_lsn;
+		close_end_lsn.file++;
 		log->log_close_fh = NULL;
 		F_CLR(slot, SLOT_CLOSEFH);
 	}
@@ -896,12 +910,45 @@ __log_release(WT_SESSION_IMPL *session, WT_LOGSLOT *slot)
 	log->write_lsn = slot->slot_end_lsn;
 
 	/*
+	 * If we have a file to close, close it now.  First fsync so
+	 * that a later sync will be assured all earlier transactions
+	 * in earlier log files are also on disk.  We have to do this
+	 * before potentially syncing our own operation.
+	 */
+	if (close_fh) {
+		WT_ERR(__wt_fsync(session, close_fh));
+		/*
+		 * This loop guarantees sync_lsn is updated in file order
+		 * so that when sync_lsn is updated, we know all earlier files
+		 * have already been fully processed.
+		 */
+		while (log->sync_lsn.file < close_lsn.file ||
+		    __wt_spin_trylock(session, &log->log_sync_lock, &id) != 0) {
+			WT_ERR(__wt_cond_wait(
+			    session, log->log_sync_cond, 10000));
+			continue;
+		}
+		locked = 1;
+		WT_ERR(__wt_close(session, close_fh));
+		log->sync_lsn = close_end_lsn;
+		WT_ERR(__wt_cond_signal(session, log->log_sync_cond));
+		locked = 0;
+		__wt_spin_unlock(session, &log->log_sync_lock);
+	}
+
+	/*
 	 * Try to consolidate calls to fsync to wait less.  Acquire a spin lock
 	 * so that threads finishing writing to the log will wait while the
 	 * current fsync completes and advance log->sync_lsn.
 	 */
 	while (F_ISSET(slot, SLOT_SYNC | SLOT_SYNC_DIR)) {
-		if (__wt_spin_trylock(session, &log->log_sync_lock, &id) != 0) {
+		/*
+		 * We have to wait until earlier log files have finished their
+		 * sync operations.  The most recent one will set the LSN to the
+		 * beginning of our file.
+		 */
+		if (log->sync_lsn.file < slot->slot_end_lsn.file ||
+		    __wt_spin_trylock(session, &log->log_sync_lock, &id) != 0) {
 			WT_ERR(__wt_cond_wait(
 			    session, log->log_sync_cond, 10000));
 			continue;
@@ -956,16 +1003,6 @@ __log_release(WT_SESSION_IMPL *session, WT_LOGSLOT *slot)
 		WT_ERR(__wt_buf_grow(session,
 		    &slot->slot_buf, slot->slot_buf.memsize * 2));
 	}
-	/*
-	 * If we have a file to close, close it now.  First fsync so
-	 * that a later sync will be assured all earlier transactions
-	 * in earlier log files are also on disk.
-	 */
-	if (close_fh) {
-		WT_ERR(__wt_fsync(session, close_fh));
-		WT_ERR(__wt_close(session, close_fh));
-	}
-
 err:	if (locked)
 		__wt_spin_unlock(session, &log->log_sync_lock);
 	if (ret != 0 && slot->slot_error == 0)
diff --git a/src/third_party/wiredtiger/src/txn/txn_log.c b/src/third_party/wiredtiger/src/txn/txn_log.c
index f706efa8a70..789be2ceef4 100644
--- a/src/third_party/wiredtiger/src/txn/txn_log.c
+++ b/src/third_party/wiredtiger/src/txn/txn_log.c
@@ -337,7 +337,9 @@ __wt_txn_checkpoint_log(
 		    rectype, ckpt_lsn->file, ckpt_lsn->offset,
 		    txn->ckpt_nsnapshot, ckpt_snapshot));
 		logrec->size += (uint32_t)recsize;
-		WT_ERR(__wt_log_write(session, logrec, lsnp, 0));
+		WT_ERR(__wt_log_write(session, logrec, lsnp,
+		    F_ISSET(S2C(session), WT_CONN_CKPT_SYNC) ?
+		    WT_LOG_FSYNC : 0));
 
 		/*
 		 * If this full checkpoint completed successfully and there is