From f851ef3784a4fbe9ce516947a85b6f9d39a908a2 Mon Sep 17 00:00:00 2001 From: Luke Chen Date: Wed, 7 Jul 2021 18:52:02 +1000 Subject: Import wiredtiger: f66db272184f1597839500e12b11a6eb9470b517 from branch mongodb-5.0 ref: dbe06a6778..f66db27218 for: 5.0.0-rc8 WT-7783 Fix RTS to restore tombstone when an on-disk update is out of order prepare update --- src/third_party/wiredtiger/import.data | 2 +- .../wiredtiger/src/txn/txn_rollback_to_stable.c | 13 +- .../test/suite/test_rollback_to_stable21.py | 145 ++++++++++++++++++++- 3 files changed, 154 insertions(+), 6 deletions(-) diff --git a/src/third_party/wiredtiger/import.data b/src/third_party/wiredtiger/import.data index e3eafa7d0ec..83a52630cf7 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-5.0", - "commit": "dbe06a6778c676506a2eddfda3e1fd5b69f61c57" + "commit": "f66db272184f1597839500e12b11a6eb9470b517" } diff --git a/src/third_party/wiredtiger/src/txn/txn_rollback_to_stable.c b/src/third_party/wiredtiger/src/txn/txn_rollback_to_stable.c index ff051d14564..bae70f90ac4 100644 --- a/src/third_party/wiredtiger/src/txn/txn_rollback_to_stable.c +++ b/src/third_party/wiredtiger/src/txn/txn_rollback_to_stable.c @@ -491,11 +491,16 @@ __rollback_ondisk_fixup_key(WT_SESSION_IMPL *session, WT_REF *ref, WT_PAGE *page WT_STAT_CONN_DATA_INCR(session, txn_rts_hs_restore_updates); /* - * We have a tombstone on the original update chain and it is behind the stable - * timestamp, we need to restore that as well. + * We have a tombstone on the original update chain and it is behind or equal to the + * stable timestamp, we need to restore that as well. */ - if (hs_stop_durable_ts <= rollback_timestamp && - hs_stop_durable_ts < newer_hs_durable_ts) { + if (hs_stop_durable_ts <= rollback_timestamp) { + /* + * The restoring tombstone timestamp must be less than previous update start + * timestamp or the on-disk update is an out of order prepared. + */ + WT_ASSERT(session, hs_stop_durable_ts < newer_hs_durable_ts || unpack->tw.prepare); + WT_ERR(__wt_upd_alloc_tombstone(session, &tombstone, NULL)); /* * Set the transaction id of updates to WT_TXN_NONE when called from recovery, diff --git a/src/third_party/wiredtiger/test/suite/test_rollback_to_stable21.py b/src/third_party/wiredtiger/test/suite/test_rollback_to_stable21.py index 1e898523c39..8404af43df9 100644 --- a/src/third_party/wiredtiger/test/suite/test_rollback_to_stable21.py +++ b/src/third_party/wiredtiger/test/suite/test_rollback_to_stable21.py @@ -26,7 +26,7 @@ # ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR # OTHER DEALINGS IN THE SOFTWARE. -from wiredtiger import stat +from wiredtiger import stat, WT_NOTFOUND from wtscenario import make_scenarios from helper import simulate_crash_restart from wtdataset import SimpleDataSet @@ -114,3 +114,146 @@ class test_rollback_to_stable21(test_rollback_to_stable_base): stat_cursor.close() self.assertGreater(hs_removed, 0) + + def test_rollback_to_stable_with_different_tombstone(self): + nrows = 1000 + + # Prepare transactions for column store table is not yet supported. + if self.key_format == 'r': + self.skipTest('Prepare transactions for column store table is not yet supported') + + # Create a table without logging. + uri = "table:rollback_to_stable21" + ds = SimpleDataSet( + self, uri, 0, key_format=self.key_format, value_format="S", config='log=(enabled=false)') + ds.populate() + + # Pin oldest and stable timestamps to 10. + self.conn.set_timestamp('oldest_timestamp=' + timestamp_str(10) + + ',stable_timestamp=' + timestamp_str(10)) + + valuea = 'a' * 400 + valueb = 'b' * 400 + + cursor = self.session.open_cursor(uri) + self.session.begin_transaction() + for i in range(1, nrows + 1): + cursor[i] = valuea + self.session.commit_transaction('commit_timestamp=' + timestamp_str(30)) + + self.session.begin_transaction() + for i in range(1, nrows + 1): + cursor.set_key(i) + cursor.remove() + self.session.commit_transaction('commit_timestamp=' + timestamp_str(40)) + + self.session.begin_transaction() + for i in range(1, nrows + 1): + cursor[i] = valueb + + cursor.reset() + cursor.close() + self.session.prepare_transaction('prepare_timestamp=' + timestamp_str(20)) + + s = self.conn.open_session() + s.begin_transaction('ignore_prepare = true, read_timestamp = ' + timestamp_str(30)) + # Configure debug behavior on a cursor to evict the page positioned on when the reset API is used. + evict_cursor = s.open_cursor(uri, None, "debug=(release_evict)") + + for i in range(1, nrows + 1): + evict_cursor.set_key(i) + self.assertEquals(evict_cursor.search(), 0) + self.assertEqual(evict_cursor.get_value(), valuea) + evict_cursor.reset() + + s.rollback_transaction() + self.conn.set_timestamp('stable_timestamp=' + timestamp_str(40)) + s.checkpoint() + + # Rollback the prepared transaction + self.session.rollback_transaction() + + # Simulate a server crash and restart. + self.pr("restart") + simulate_crash_restart(self, ".", "RESTART") + self.pr("restart complete") + + self.check(valuea, uri, nrows, 30) + self.check(valuea, uri, 0, 40) + + stat_cursor = self.session.open_cursor('statistics:', None, None) + hs_removed = stat_cursor[stat.conn.txn_rts_hs_removed][2] + hs_restored_tombstone = stat_cursor[stat.conn.txn_rts_hs_restore_tombstones][2] + stat_cursor.close() + + self.assertGreater(hs_removed, 0) + self.assertGreater(hs_restored_tombstone, 0) + + def test_rollback_to_stable_with_same_tombstone(self): + nrows = 1000 + + # Prepare transactions for column store table is not yet supported. + if self.key_format == 'r': + self.skipTest('Prepare transactions for column store table is not yet supported') + + # Create a table without logging. + uri = "table:rollback_to_stable21" + ds = SimpleDataSet( + self, uri, 0, key_format=self.key_format, value_format="S", config='log=(enabled=false)') + ds.populate() + + # Pin oldest and stable timestamps to 10. + self.conn.set_timestamp('oldest_timestamp=' + timestamp_str(10) + + ',stable_timestamp=' + timestamp_str(10)) + + valuea = 'a' * 400 + valueb = 'b' * 400 + + cursor = self.session.open_cursor(uri) + self.session.begin_transaction() + for i in range(1, nrows + 1): + cursor[i] = valuea + cursor.set_key(i) + cursor.remove() + + self.session.commit_transaction('commit_timestamp=' + timestamp_str(30)) + + self.session.begin_transaction() + for i in range(1, nrows + 1): + cursor[i] = valueb + + cursor.reset() + cursor.close() + self.session.prepare_transaction('prepare_timestamp=' + timestamp_str(20)) + + s = self.conn.open_session() + s.begin_transaction('ignore_prepare = true') + # Configure debug behavior on a cursor to evict the page positioned on when the reset API is used. + evict_cursor = s.open_cursor(uri, None, "debug=(release_evict)") + + for i in range(1, nrows + 1): + evict_cursor.set_key(i) + self.assertEquals(evict_cursor.search(), WT_NOTFOUND) + evict_cursor.reset() + + s.rollback_transaction() + self.conn.set_timestamp('stable_timestamp=' + timestamp_str(40)) + s.checkpoint() + + # Rollback the prepared transaction + self.session.rollback_transaction() + + # Simulate a server crash and restart. + self.pr("restart") + simulate_crash_restart(self, ".", "RESTART") + self.pr("restart complete") + + self.check(valuea, uri, 0, 40) + + stat_cursor = self.session.open_cursor('statistics:', None, None) + hs_removed = stat_cursor[stat.conn.txn_rts_hs_removed][2] + hs_restored_tombstone = stat_cursor[stat.conn.txn_rts_hs_restore_tombstones][2] + stat_cursor.close() + + self.assertGreater(hs_removed, 0) + self.assertGreater(hs_restored_tombstone, 0) -- cgit v1.2.1