1 /* Copyright (c) 2018, 2020, Oracle and/or its affiliates.
2 
3    This program is free software; you can redistribute it and/or modify
4    it under the terms of the GNU General Public License, version 2.0,
5    as published by the Free Software Foundation.
6 
7    This program is also distributed with certain software (including
8    but not limited to OpenSSL) that is licensed under separate terms,
9    as designated in a particular file or component or in included license
10    documentation.  The authors of MySQL hereby grant you an additional
11    permission to link the program and your derivative works with the
12    separately licensed software that they have included with MySQL.
13 
14    This program is distributed in the hope that it will be useful,
15    but WITHOUT ANY WARRANTY; without even the implied warranty of
16    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
17    GNU General Public License, version 2.0, for more details.
18 
19    You should have received a copy of the GNU General Public License
20    along with this program; if not, write to the Free Software Foundation,
21    51 Franklin Street, Suite 500, Boston, MA 02110-1335 USA */
22 
23 #include "rpl_trx_tracking.h"
24 
25 #include "mysqld.h"
26 #include "binlog.h"
27 
28 #include "debug_sync.h"
29 
Logical_clock()30 Logical_clock::Logical_clock()
31   : state(SEQ_UNINIT), offset(0)
32 {}
33 
34 /**
35   Atomically fetch the current state.
36   @return  not subtracted "absolute" value.
37  */
get_timestamp()38 inline int64 Logical_clock::get_timestamp()
39 {
40   int64 retval= 0;
41   DBUG_ENTER("Logical_clock::get_timestamp");
42   retval= my_atomic_load64(&state);
43   DBUG_RETURN(retval);
44 }
45 
46 /**
47   Steps the absolute value of the clock (state) to return
48   an updated value.
49   The caller must be sure to call the method in no concurrent
50   execution context so either offset and state can't change.
51 
52   @return  incremented "absolute" value
53  */
step()54 inline int64 Logical_clock::step()
55 {
56   DBUG_ASSERT(SEQ_UNINIT == 0);
57   DBUG_EXECUTE_IF("logical_clock_step_2", ++state;);
58   return ++state;
59 }
60 
61 /**
62   To try setting the clock *forward*.
63   The clock does not change when the new value is in the past
64   which is reflected by the new value and by offset.
65   In other words the function main effects is described as
66     state= max(state, new_value).
67   Offset that exceeds the new value indicates the binary log rotation
68   to render such new value useless.
69 
70   @param  new_val  a new value (offset included)
71   @return a (new) value of state member regardless whether it's changed or not.
72  */
set_if_greater(int64 new_val)73 inline int64 Logical_clock::set_if_greater(int64 new_val)
74 {
75   longlong old_val= new_val - 1;
76   bool cas_rc;
77 
78   DBUG_ENTER("Logical_clock::set_if_greater");
79 
80   DBUG_ASSERT(new_val > 0);
81 
82   if (new_val <= offset)
83   {
84     /*
85       This function's invocation can be separated from the
86       transaction's flushing by few rotations. A late to log
87       transaction does not change the clock, similarly to how
88       its timestamps are handled at flushing.
89     */
90     DBUG_RETURN(SEQ_UNINIT);
91   }
92 
93   DBUG_ASSERT(new_val > 0);
94 
95   while (!(cas_rc= my_atomic_cas64(&state, &old_val, new_val)) &&
96          old_val < new_val)
97   {}
98 
99   DBUG_ASSERT(state >= new_val); // setting can't be done to past
100 
101   DBUG_ASSERT(cas_rc || old_val >= new_val);
102 
103   DBUG_RETURN(cas_rc ? new_val : old_val);
104 }
105 
106 
107 /**
108   Get the sequence_number for a transaction, and get the last_commit based
109   on parallel committing transactions.
110 
111   @param[in]     thd             Current THD from which to extract trx context.
112   @param[in,out] sequence_number Sequence number of current transaction.
113   @param[in,out] commit_parent   Commit_parent of current transaction,
114                                  pre-filled with the commit_parent calculated
115                                  by the logical clock logic.
116 */
117 void
get_dependency(THD * thd,int64 & sequence_number,int64 & commit_parent)118 Commit_order_trx_dependency_tracker::get_dependency(THD *thd,
119                                                     int64 &sequence_number,
120                                                     int64 &commit_parent)
121 {
122   Transaction_ctx *trn_ctx= thd->get_transaction();
123 
124   DBUG_ASSERT(trn_ctx->sequence_number
125               > m_max_committed_transaction.get_offset());
126   /*
127     Prepare sequence_number and commit_parent relative to the current
128     binlog.  This is done by subtracting the binlog's clock offset
129     from the values.
130 
131     A transaction that commits after the binlog is rotated, can have a
132     commit parent in the previous binlog. In this case, subtracting
133     the offset from the sequence number results in a negative
134     number. The commit parent dependency gets lost in such
135     case. Therefore, we log the value SEQ_UNINIT in this case.
136   */
137   sequence_number=
138     trn_ctx->sequence_number - m_max_committed_transaction.get_offset();
139 
140   commit_parent=
141     trn_ctx->last_committed <= m_max_committed_transaction.get_offset()
142            ? SEQ_UNINIT
143            : trn_ctx->last_committed - m_max_committed_transaction.get_offset();
144 }
145 
146 int64
step()147 Commit_order_trx_dependency_tracker::step()
148 {
149   return m_transaction_counter.step();
150 }
151 
152 void
rotate()153 Commit_order_trx_dependency_tracker::rotate()
154 {
155   m_max_committed_transaction.
156     update_offset(m_transaction_counter.get_timestamp());
157 
158   m_transaction_counter.
159     update_offset(m_transaction_counter.get_timestamp());
160 }
161 
162 void
update_max_committed(int64 sequence_number)163 Commit_order_trx_dependency_tracker::update_max_committed(int64 sequence_number)
164 {
165   mysql_mutex_assert_owner(&LOCK_slave_trans_dep_tracker);
166   m_max_committed_transaction.set_if_greater(sequence_number);
167 }
168 
169 /**
170   Get the writeset dependencies of a transaction.
171   This takes the commit_parent that must be previously set using
172   Commit_order_trx_dependency_tracker and tries to make the commit_parent as
173   low as possible, using the writesets of each transaction.
174   The commit_parent returned depends on how many row hashes are stored in the
175   writeset_history, which is cleared once it reaches the user-defined maximum.
176 
177   @param[in]     thd             Current THD from which to extract trx context.
178   @param[in,out] sequence_number Sequence number of current transaction.
179   @param[in,out] commit_parent   Commit_parent of current transaction,
180                                  pre-filled with the commit_parent calculated by
181                                  Commit_order_trx_dependency_tracker to use when
182                                  the writeset commit_parent is not valid.
183 */
184 void
get_dependency(THD * thd,int64 & sequence_number,int64 & commit_parent)185 Writeset_trx_dependency_tracker::get_dependency(THD *thd,
186                                                 int64 &sequence_number,
187                                                 int64 &commit_parent)
188 {
189   Rpl_transaction_write_set_ctx *write_set_ctx=
190     thd->get_transaction()->get_transaction_write_set_ctx();
191   std::set<uint64> *writeset= write_set_ctx->get_write_set();
192 
193 #ifndef DBUG_OFF
194   /* The writeset of an empty transaction must be empty. */
195   if (is_empty_transaction_in_binlog_cache(thd))
196     DBUG_ASSERT(writeset->size() == 0);
197 #endif
198 
199   /*
200     Check if this transaction has a writeset, if the writeset will overflow the
201     history size, if the transaction_write_set_extraction is consistent
202     between session and global or if changes in the tables referenced in this
203     transaction cascade to other tables. If that happens revert to using the
204     COMMIT_ORDER and clear the history to keep data consistent.
205   */
206   bool can_use_writesets=
207     // empty writeset implies DDL or similar, except if there are missing keys
208     (writeset->size() != 0 || write_set_ctx->get_has_missing_keys() ||
209      /*
210        The empty transactions do not need to clear the writeset history, since
211        they can be executed in parallel.
212      */
213      is_empty_transaction_in_binlog_cache(thd)) &&
214     // hashing algorithm for the session must be the same as used by other rows in history
215     (global_system_variables.transaction_write_set_extraction ==
216      thd->variables.transaction_write_set_extraction) &&
217     // must not use foreign keys
218     !write_set_ctx->get_has_related_foreign_keys();
219   bool exceeds_capacity= false;
220 
221   mysql_mutex_lock(&LOCK_slave_trans_dep_tracker);
222   if (can_use_writesets)
223   {
224     /*
225      Check if adding this transaction exceeds the capacity of the writeset
226      history. If that happens, m_writeset_history will be cleared only after
227      using its information for current transaction.
228     */
229     exceeds_capacity=
230       m_writeset_history.size() + writeset->size() > m_opt_max_history_size;
231 
232     /*
233      Compute the greatest sequence_number among all conflicts and add the
234      transaction's row hashes to the history.
235     */
236     DEBUG_SYNC(thd, "wait_in_get_dependency");
237     int64 last_parent= m_writeset_history_start;
238     for (std::set<uint64>::iterator it= writeset->begin();
239          it != writeset->end(); ++it)
240     {
241       Writeset_history::iterator hst= m_writeset_history.find(*it);
242       if (hst != m_writeset_history.end())
243       {
244         if (hst->second > last_parent && hst->second < sequence_number)
245           last_parent= hst->second;
246 
247         hst->second= sequence_number;
248       }
249       else
250       {
251         if (!exceeds_capacity)
252           m_writeset_history.insert(std::pair<uint64, int64>(*it, sequence_number));
253       }
254     }
255 
256     /*
257       If the transaction references tables with missing primary keys revert to
258       COMMIT_ORDER, update and not reset history, as it is unnecessary because
259       any transaction that refers this table will also revert to COMMIT_ORDER.
260     */
261     if (!write_set_ctx->get_has_missing_keys())
262     {
263       /*
264        The WRITESET commit_parent then becomes the minimum of largest parent
265        found using the hashes of the row touched by the transaction and the
266        commit parent calculated with COMMIT_ORDER.
267       */
268       commit_parent= std::min(last_parent, commit_parent);
269     }
270   }
271 
272   if (exceeds_capacity || !can_use_writesets)
273   {
274     m_writeset_history_start= sequence_number;
275     m_writeset_history.clear();
276   }
277   mysql_mutex_unlock(&LOCK_slave_trans_dep_tracker);
278 }
279 
280 void
rotate(int64 start)281 Writeset_trx_dependency_tracker::rotate(int64 start)
282 {
283   m_writeset_history_start= start;
284   m_writeset_history.clear();
285 }
286 
287 /**
288   Get the writeset commit parent of transactions using the session dependencies.
289 
290   @param[in]     thd             Current THD from which to extract trx context.
291   @param[in,out] sequence_number Sequence number of current transaction.
292   @param[in,out] commit_parent   Commit_parent of current transaction,
293                                  pre-filled with the commit_parent calculated
294                                  by the Write_set_trx_dependency_tracker as a
295                                  fall-back.
296 */
297 void
get_dependency(THD * thd,int64 & sequence_number,int64 & commit_parent)298 Writeset_session_trx_dependency_tracker::get_dependency(THD *thd,
299                                                         int64 &sequence_number,
300                                                         int64 &commit_parent)
301 {
302   int64 session_parent= thd->rpl_thd_ctx.dependency_tracker_ctx().
303                         get_last_session_sequence_number();
304 
305   if (session_parent != 0 && session_parent < sequence_number)
306     commit_parent= std::max(commit_parent, session_parent);
307 
308   thd->rpl_thd_ctx.dependency_tracker_ctx().
309     set_last_session_sequence_number(sequence_number);
310 }
311 
312 /**
313   Get the dependencies in a transaction, the main entry point for the
314   dependency tracking work.
315 */
316 void
get_dependency(THD * thd,int64 & sequence_number,int64 & commit_parent)317 Transaction_dependency_tracker::get_dependency(THD *thd,
318                                                int64 &sequence_number,
319                                                int64 &commit_parent)
320 {
321   sequence_number= commit_parent= 0;
322 
323   switch(my_atomic_load64(&m_opt_tracking_mode))
324   {
325     case DEPENDENCY_TRACKING_COMMIT_ORDER:
326       m_commit_order.get_dependency(thd, sequence_number, commit_parent);
327       break;
328     case DEPENDENCY_TRACKING_WRITESET:
329       m_commit_order.get_dependency(thd, sequence_number, commit_parent);
330       m_writeset.get_dependency(thd, sequence_number, commit_parent);
331       break;
332     case DEPENDENCY_TRACKING_WRITESET_SESSION:
333       m_commit_order.get_dependency(thd, sequence_number, commit_parent);
334       m_writeset.get_dependency(thd, sequence_number, commit_parent);
335       m_writeset_session.get_dependency(thd, sequence_number, commit_parent);
336       break;
337     default:
338       DBUG_ASSERT(0);  // blow up on debug
339       /*
340         Fallback to commit order on production builds.
341        */
342       m_commit_order.get_dependency(thd, sequence_number, commit_parent);
343   }
344 }
345 
346 void
tracking_mode_changed()347 Transaction_dependency_tracker::tracking_mode_changed()
348 {
349   Logical_clock max_committed_transaction=
350     m_commit_order.get_max_committed_transaction();
351   int64 timestamp= max_committed_transaction.get_timestamp()
352                    - max_committed_transaction.get_offset();
353 
354   m_writeset.rotate(timestamp);
355 }
356 
357 /**
358   The method is to be executed right before committing time.
359   It must be invoked even if the transaction does not commit
360   to engine being merely logged into the binary log.
361   max_committed_transaction is updated with a greater timestamp
362   value.
363   As a side effect, the transaction context's sequence_number
364   is reset.
365 
366   @param thd a pointer to THD instance
367 */
368 void
update_max_committed(THD * thd)369 Transaction_dependency_tracker::update_max_committed(THD *thd)
370 {
371   Transaction_ctx *trn_ctx= thd->get_transaction();
372   m_commit_order.update_max_committed(trn_ctx->sequence_number);
373   /*
374     sequence_number timestamp isn't needed anymore, so it's cleared off.
375   */
376   trn_ctx->sequence_number= SEQ_UNINIT;
377 
378   DBUG_ASSERT(trn_ctx->last_committed == SEQ_UNINIT ||
379               thd->commit_error == THD::CE_FLUSH_ERROR);
380 }
381 
382 int64
step()383 Transaction_dependency_tracker::step()
384 {
385   return m_commit_order.step();
386 }
387 
388 void
rotate()389 Transaction_dependency_tracker::rotate()
390 {
391   m_commit_order.rotate();
392   /*
393     To make slave appliers be able to execute transactions in parallel
394     after rotation, set the minimum commit_parent to 1 after rotation.
395   */
396   m_writeset.rotate(1);
397   if (current_thd)
398     current_thd->get_transaction()->sequence_number= 2;
399 }
400 
get_max_committed_timestamp()401 int64 Transaction_dependency_tracker::get_max_committed_timestamp()
402 {
403   return m_commit_order.get_max_committed_transaction().get_timestamp();
404 }
405 
406 
407