MDEV-32050: Do not copy undo records in purge

Also, default to innodb_purge_batch_size=1000,
replacing the old default value of processing 300 undo log pages
in a batch. Axel Schwenke found this value to help reduce purge lag
without having a significant impact on workload throughput.

In purge, we can simply acquire a shared latch on the undo log page
(to avoid a race condition like the one that was fixed in
commit b102872a) and retain a buffer-fix
after releasing the latch. The buffer-fix will prevent the undo log
page from being evicted from the buffer pool. Concurrent modification
is prevented by design. Only the purge_coordinator_task
(or its accomplice purge_truncation_task) may free the undo log pages,
after any purge_worker_task have completed execution. Hence, we do not
have to worry about any overwriting or reuse of the undo log records.

trx_undo_rec_copy(): Remove. The only remaining caller would have been
trx_undo_get_undo_rec_low(), which is where the logic was merged.

purge_sys_t::m_initialized: Replaces heap.

purge_sys_t::pages: A cache of buffer-fixed pages that have been
looked up from buf_pool.page_hash.

purge_sys_t::get_page(): Return a buffer-fixed undo page, using the
pages cache.

trx_purge_t::batch_cleanup(): Renamed from clone_end_view().
Clear the pages cache and clone the end_view at the end of a batch.

purge_sys_t::n_pages_handled(): Return pages.size(). This determines
if innodb_purge_batch_size was exceeded.

purge_sys_t::rseg_get_next_history_log(): Replaces
trx_purge_rseg_get_next_history_log().

purge_sys_t::choose_next_log(): Replaces trx_purge_choose_next_log()
and trx_purge_read_undo_rec().

purge_sys_t::get_next_rec(): Replaces trx_purge_get_next_rec()
and trx_undo_get_next_rec().

purge_sys_t::fetch_next_rec(): Replaces trx_purge_fetch_next_rec()
and some use of trx_undo_get_first_rec().

trx_purge_attach_undo_recs(): Do not allow purge_sys.n_pages_handled()
exceed the innodb_purge_batch_size or ¾ of the buffer pool, whichever
is smaller.

Reviewed by: Vladislav Lesin
Tested by: Matthias Leich and Axel Schwenke

mysql-test/suite/sys_vars/r/innodb_purge_batch_size_basic.result

 SET @global_start_value = @@global.innodb_purge_batch_size;
 SELECT @global_start_value;
 @global_start_value
-300
+1000
 '#--------------------FN_DYNVARS_046_01------------------------#'
 SET @@global.innodb_purge_batch_size = 1;
 SET @@global.innodb_purge_batch_size = DEFAULT;
 SELECT @@global.innodb_purge_batch_size;
 @@global.innodb_purge_batch_size
-300
+1000
 '#---------------------FN_DYNVARS_046_02-------------------------#'
 SET innodb_purge_batch_size = 1;
 ERROR HY000: Variable 'innodb_purge_batch_size' is a GLOBAL variable and should be set with SET GLOBAL
 SELECT @@innodb_purge_batch_size;
 @@innodb_purge_batch_size
-300
+1000
 SELECT local.innodb_purge_batch_size;
 ERROR 42S02: Unknown table 'local' in field list
 SET global innodb_purge_batch_size = 1;
@@ -112,4 +112,4 @@ SELECT @@global.innodb_purge_batch_size;
 SET @@global.innodb_purge_batch_size = @global_start_value;
 SELECT @@global.innodb_purge_batch_size;
 @@global.innodb_purge_batch_size
-300
+1000

mysql-test/suite/sys_vars/r/sysvars_innodb,32bit.rdiff

@@ -307,7 +307,7 @@
  NUMERIC_MAX_VALUE	65536
 @@ -1345,7 +1345,7 @@
  SESSION_VALUE	NULL
- DEFAULT_VALUE	300
+ DEFAULT_VALUE	1000
  VARIABLE_SCOPE	GLOBAL
 -VARIABLE_TYPE	BIGINT UNSIGNED
 +VARIABLE_TYPE	INT UNSIGNED

mysql-test/suite/sys_vars/r/sysvars_innodb.result

@@ -1293,7 +1293,7 @@ READ_ONLY	NO
 COMMAND_LINE_ARGUMENT	OPTIONAL
 VARIABLE_NAME	INNODB_PURGE_BATCH_SIZE
 SESSION_VALUE	NULL
-DEFAULT_VALUE	300
+DEFAULT_VALUE	1000
 VARIABLE_SCOPE	GLOBAL
 VARIABLE_TYPE	BIGINT UNSIGNED
 VARIABLE_COMMENT	Number of UNDO log pages to purge in one batch from the history list.

storage/innobase/handler/ha_innodb.cc

@@ -18852,7 +18852,7 @@ static MYSQL_SYSVAR_ULONG(purge_batch_size, srv_purge_batch_size,
   PLUGIN_VAR_OPCMDARG,
   "Number of UNDO log pages to purge in one batch from the history list.",
   NULL, NULL,
-  300,			/* Default setting */
+  1000,			/* Default setting */
   1,			/* Minimum value */
   5000, 0);		/* Maximum value */
 

storage/innobase/include/row0purge.h

@@ -80,15 +80,6 @@ row_purge_step(
 	que_thr_t*	thr)	/*!< in: query thread */
 	MY_ATTRIBUTE((nonnull, warn_unused_result));
 
-/** Info required to purge a record */
-struct trx_purge_rec_t
-{
-  /** Record to purge */
-  const trx_undo_rec_t *undo_rec;
-  /** File pointer to undo record */
-  roll_ptr_t roll_ptr;
-};
-
 /** Purge worker context */
 struct purge_node_t
 {

storage/innobase/include/trx0purge.h

@@ -31,6 +31,7 @@ Created 3/26/1996 Heikki Tuuri
 #include "srw_lock.h"
 
 #include <queue>
+#include <unordered_map>
 
 /** Prepend the history list with an undo log.
 Remove the undo log segment from the rseg slot if it is too big for reuse.
@@ -127,6 +128,7 @@ struct TrxUndoRsegsIterator {
 /** The control structure used in the purge operation */
 class purge_sys_t
 {
+  friend TrxUndoRsegsIterator;
 public:
   /** latch protecting view, m_enabled */
   alignas(CPU_LEVEL1_DCACHE_LINESIZE) mutable srw_spin_lock latch;
@@ -134,6 +136,8 @@ class purge_sys_t
   /** Read view at the start of a purge batch. Any encountered index records
   that are older than view will be removed. */
   ReadViewBase view;
+  /** whether the subsystem has been initialized */
+  bool m_initialized{false};
   /** whether purge is enabled; protected by latch and std::atomic */
   std::atomic<bool> m_enabled{false};
 public:
@@ -152,7 +156,34 @@ class purge_sys_t
   /** Read view at the end of a purge batch (copied from view). Any undo pages
   containing records older than end_view may be freed. */
   ReadViewBase end_view;
+
+  struct hasher
+  {
+    size_t operator()(const page_id_t &id) const { return size_t(id.raw()); }
+  };
+
+  using unordered_map =
+    std::unordered_map<const page_id_t, buf_block_t*, hasher,
+#if defined __GNUC__ && __GNUC__ == 4 && __GNUC_MINOR__ >= 8
+                       std::equal_to<page_id_t>
+                       /* GCC 4.8.5 would fail to find a matching allocator */
+#else
+                       std::equal_to<page_id_t>,
+                       ut_allocator<std::pair<const page_id_t, buf_block_t*>>
+#endif
+                       >;
+  /** map of buffer-fixed undo log pages processed during a purge batch */
+  unordered_map pages;
 public:
+  /** @return the number of processed undo pages */
+  size_t n_pages_handled() const { return pages.size(); }
+
+  /** Look up an undo log page.
+  @param id    undo page identifier
+  @return undo page
+  @retval nullptr in case the page is corrupted */
+  buf_block_t *get_page(page_id_t id);
+
 	que_t*		query;		/*!< The query graph which will do the
 					parallelized purge operation */
 
@@ -188,6 +219,7 @@ class purge_sys_t
 					to purge */
 	trx_rseg_t*	rseg;		/*!< Rollback segment for the next undo
 					record to purge */
+private:
 	uint32_t	page_no;	/*!< Page number for the next undo
 					record to purge, page number of the
 					log header, if dummy record */
@@ -202,7 +234,7 @@ class purge_sys_t
 	TrxUndoRsegsIterator
 			rseg_iter;	/*!< Iterator to get the next rseg
 					to process */
-
+public:
 	purge_pq_t	purge_queue;	/*!< Binary min-heap, ordered on
 					TrxUndoRsegs::trx_no. It is protected
 					by the pq_mutex */
@@ -217,17 +249,6 @@ class purge_sys_t
 		fil_space_t*	last;
 	} truncate;
 
-	/** Heap for reading the undo log records */
-	mem_heap_t*	heap;
-  /**
-    Constructor.
-
-    Some members may require late initialisation, thus we just mark object as
-    uninitialised. Real initialisation happens in create().
-  */
-
-  purge_sys_t(): m_enabled(false), heap(nullptr) {}
-
   /** Create the instance */
   void create();
 
@@ -281,6 +302,32 @@ class purge_sys_t
   /** @return whether stop_SYS() is in effect */
   bool must_wait_FTS() const { return m_FTS_paused; }
 
+private:
+  /**
+  Get the next record to purge and update the info in the purge system.
+  @param roll_ptr           undo log pointer to the record
+  @return buffer-fixed reference to undo log record
+  @retval {nullptr,1} if the whole undo log can skipped in purge
+  @retval {nullptr,0} if nothing is left, or on corruption */
+  inline trx_purge_rec_t get_next_rec(roll_ptr_t roll_ptr);
+
+  /** Choose the next undo log to purge.
+  @return whether anything is to be purged */
+  bool choose_next_log();
+
+  /** Update the last not yet purged history log info in rseg when
+  we have purged a whole undo log. Advances also purge_trx_no
+  past the purged log. */
+  void rseg_get_next_history_log();
+
+public:
+  /**
+  Fetch the next undo log record from the history list to purge.
+  @return buffer-fixed reference to undo log record
+  @retval {nullptr,1} if the whole undo log can skipped in purge
+  @retval {nullptr,0} if nothing is left, or on corruption */
+  inline trx_purge_rec_t fetch_next_rec();
+
   /** Determine if the history of a transaction is purgeable.
   @param trx_id  transaction identifier
   @return whether the history is purgeable */
@@ -327,9 +374,10 @@ class purge_sys_t
   /** Wake up the purge threads if there is work to do. */
   void wake_if_not_active();
 
-  /** Update end_view at the end of a purge batch.
-  @param head   the new head of the purge queue */
-  inline void clone_end_view(const iterator &head);
+  /** Release undo pages and update end_view at the end of a purge batch.
+  @retval false when nothing is to be purged
+  @retval true  when purge_sys.rseg->latch was locked  */
+  inline void batch_cleanup(const iterator &head);
 
   struct view_guard
   {

storage/innobase/include/trx0rec.h

@@ -28,32 +28,9 @@ Created 3/26/1996 Heikki Tuuri
 
 #include "trx0types.h"
 #include "row0types.h"
-#include "mtr0mtr.h"
-#include "rem0types.h"
 #include "page0types.h"
-#include "row0log.h"
 #include "que0types.h"
 
-/***********************************************************************//**
-Copies the undo record to the heap.
-@param undo_rec   record in an undo log page
-@param heap       memory heap
-@return copy of undo_rec
-@retval nullptr if the undo log record is corrupted */
-inline trx_undo_rec_t* trx_undo_rec_copy(const trx_undo_rec_t *undo_rec,
-                                         mem_heap_t *heap)
-{
-  const size_t offset= ut_align_offset(undo_rec, srv_page_size);
-  const size_t end= mach_read_from_2(undo_rec);
-  if (end <= offset || end >= srv_page_size - FIL_PAGE_DATA_END)
-    return nullptr;
-  const size_t len= end - offset;
-  trx_undo_rec_t *rec= static_cast<trx_undo_rec_t*>
-    (mem_heap_dup(heap, undo_rec, len));
-  mach_write_to_2(rec, len);
-  return rec;
-}
-
 /**********************************************************************//**
 Reads the undo log record number.
 @return undo no */

storage/innobase/include/trx0types.h

@@ -107,6 +107,15 @@ typedef	byte	trx_undo_rec_t;
 
 /* @} */
 
+/** Info required to purge a record */
+struct trx_purge_rec_t
+{
+  /** Undo log record, or nullptr (roll_ptr!=0 if the log can be skipped) */
+  const trx_undo_rec_t *undo_rec;
+  /** File pointer to undo_rec */
+  roll_ptr_t roll_ptr;
+};
+
 typedef std::vector<trx_id_t, ut_allocator<trx_id_t> >	trx_ids_t;
 
 /** Number of std::unordered_map hash buckets expected to be needed

storage/innobase/include/trx0undo.h

@@ -116,31 +116,16 @@ trx_undo_page_get_next_rec(const buf_block_t *undo_page, uint16_t rec,
 trx_undo_rec_t*
 trx_undo_get_prev_rec(buf_block_t *&block, uint16_t rec, uint32_t page_no,
                       uint16_t offset, bool shared, mtr_t *mtr);
-/** Get the next record in an undo log.
-@param[in,out]  block   undo log page
-@param[in]      rec     undo record offset in the page
-@param[in]      page_no undo log header page number
-@param[in]      offset  undo log header offset on page
-@param[in,out]  mtr     mini-transaction
-@return undo log record, the page latched, NULL if none */
-trx_undo_rec_t*
-trx_undo_get_next_rec(const buf_block_t *&block, uint16_t rec,
-                      uint32_t page_no, uint16_t offset, mtr_t *mtr);
 
-/** Get the first record in an undo log.
-@param[in]      space   undo log header space
+/** Get the first undo log record on a page.
+@param[in]	block	undo log page
 @param[in]	page_no	undo log header page number
-@param[in]      offset  undo log header offset on page
-@param[in]      mode    latching mode: RW_S_LATCH or RW_X_LATCH
-@param[out]     block   undo log page
-@param[in,out]  mtr     mini-transaction
-@param[out]     err     error code
-@return undo log record, the page latched
-@retval nullptr if none */
+@param[in]	offset	undo log header page offset
+@return	pointer to first record
+@retval	nullptr	if none exists */
 trx_undo_rec_t*
-trx_undo_get_first_rec(const fil_space_t &space, uint32_t page_no,
-                       uint16_t offset, ulint mode, const buf_block_t*& block,
-                       mtr_t *mtr, dberr_t *err);
+trx_undo_page_get_first_rec(const buf_block_t *block, uint32_t page_no,
+                            uint16_t offset);
 
 /** Initialize an undo log page.
 NOTE: This corresponds to a redo log record and must not be changed!

storage/innobase/trx/trx0rec.cc

@@ -2061,12 +2061,23 @@ trx_undo_get_undo_rec_low(
 
   mtr.start();
 
-  const buf_block_t* undo_page=
-    buf_page_get(page_id_t(rseg->space->id, page_no), 0, RW_S_LATCH, &mtr);
-
-  trx_undo_rec_t *undo_rec= undo_page
-    ? trx_undo_rec_copy(undo_page->page.frame + offset, heap)
-    : nullptr;
+  trx_undo_rec_t *undo_rec= nullptr;
+  if (const buf_block_t* undo_page=
+      buf_page_get(page_id_t(rseg->space->id, page_no), 0, RW_S_LATCH, &mtr))
+  {
+    undo_rec= undo_page->page.frame + offset;
+    const size_t end= mach_read_from_2(undo_rec);
+    if (UNIV_UNLIKELY(end <= offset ||
+                      end >= srv_page_size - FIL_PAGE_DATA_END))
+      undo_rec= nullptr;
+    else
+    {
+      size_t len{end - offset};
+      undo_rec=
+        static_cast<trx_undo_rec_t*>(mem_heap_dup(heap, undo_rec, len));
+      mach_write_to_2(undo_rec, len);
+    }
+  }
 
   mtr.commit();
   return undo_rec;

storage/innobase/trx/trx0undo.cc

@@ -128,8 +128,8 @@ uint16_t trx_undo_page_get_start(const buf_block_t *block, uint32_t page_no,
 @param[in]	page_no	undo log header page number
 @param[in]	offset	undo log header page offset
 @return	pointer to first record
-@retval	NULL	if none exists */
-static trx_undo_rec_t*
+@retval	nullptr	if none exists */
+trx_undo_rec_t*
 trx_undo_page_get_first_rec(const buf_block_t *block, uint32_t page_no,
                             uint16_t offset)
 {
@@ -253,25 +253,6 @@ trx_undo_get_next_rec_from_next_page(const buf_block_t *&block,
   return block ? trx_undo_page_get_first_rec(block, page_no, offset) : nullptr;
 }
 
-/** Get the next record in an undo log.
-@param[in,out]  block   undo log page
-@param[in]      rec     undo record offset in the page
-@param[in]      page_no undo log header page number
-@param[in]      offset  undo log header offset on page
-@param[in,out]  mtr     mini-transaction
-@return undo log record, the page latched, NULL if none */
-trx_undo_rec_t*
-trx_undo_get_next_rec(const buf_block_t *&block, uint16_t rec,
-                      uint32_t page_no, uint16_t offset, mtr_t *mtr)
-{
-  if (trx_undo_rec_t *next= trx_undo_page_get_next_rec(block, rec, page_no,
-                                                       offset))
-    return next;
-
-  return trx_undo_get_next_rec_from_next_page(block, page_no, offset,
-                                              RW_S_LATCH, mtr);
-}
-
 /** Get the first record in an undo log.
 @param[in]      space   undo log header space
 @param[in]      page_no undo log header page number
@@ -282,7 +263,7 @@ trx_undo_get_next_rec(const buf_block_t *&block, uint16_t rec,
 @param[out]     err     error code
 @return undo log record, the page latched
 @retval nullptr if none */
-trx_undo_rec_t*
+static trx_undo_rec_t*
 trx_undo_get_first_rec(const fil_space_t &space, uint32_t page_no,
                        uint16_t offset, ulint mode, const buf_block_t*& block,
                        mtr_t *mtr, dberr_t *err)