MDEV-34705: Binlog in Engine: Searchability for GTID position

Every N bytes (hardcoded at 64k for now, to become a configurable setting),
write the binlog GTID state into the binlog tablespace. This allows to
quickly find a given GTID position by binary search to the prior GTID state
in the tablespace and then a small linear scan from that point.

The full binlog state is dumped at the start of the binlog file; remaining
states dumped are differential states containing only the changed
(domain_id, server_id) pairs, to save space if binlog space is large.

This commit only implements the writing of the binlog state to the
tablespace at regular intervals. The binary search to be implemented in a
subsequent commit.

Signed-off-by: Kristian Nielsen <knielsen@knielsen-hq.org>

sql/handler.h

@@ -60,6 +60,7 @@ class Field_blob;
 class Column_definition;
 class select_result;
 class handler_binlog_reader;
+struct rpl_gtid;
 
 // the following is for checking tables
 
@@ -1531,7 +1532,8 @@ struct handlerton
 
   /* Optional implementation of binlog in the engine. */
   bool (*binlog_init)(size_t binlog_size);
-  bool (*binlog_write_direct)(IO_CACHE *cache, size_t main_size);
+  bool (*binlog_write_direct)(IO_CACHE *cache, size_t main_size,
+                              const rpl_gtid *gtid);
   handler_binlog_reader * (*get_binlog_reader)();
 
    /*

sql/log.cc

@@ -1836,11 +1836,13 @@ binlog_flush_cache(THD *thd, binlog_cache_mngr *cache_mngr,
 
 extern "C"
 void
-binlog_get_cache(THD *thd, IO_CACHE **out_cache, size_t *out_main_size)
+binlog_get_cache(THD *thd, IO_CACHE **out_cache, size_t *out_main_size,
+                 const rpl_gtid **out_gtid)
 {
   IO_CACHE *cache= nullptr;
   size_t main_size= 0;
   binlog_cache_mngr *cache_mngr;
+  const rpl_gtid *gtid= nullptr;
   if (likely(!opt_bootstrap /* ToDo needed? */) &&
       opt_binlog_engine_hton &&
       (cache_mngr= thd->binlog_get_cache_mngr()))
@@ -1848,9 +1850,11 @@ binlog_get_cache(THD *thd, IO_CACHE **out_cache, size_t *out_main_size)
     main_size= cache_mngr->gtid_cache_offset;
     cache= !cache_mngr->trx_cache.empty() ?
       &cache_mngr->trx_cache.cache_log : &cache_mngr->stmt_cache.cache_log;
+    gtid= thd->get_last_commit_gtid();
   }
   *out_cache= cache;
   *out_main_size= main_size;
+  *out_gtid= gtid;
 }
 
 
@@ -7087,6 +7091,17 @@ MYSQL_BIN_LOG::get_most_recent_gtid_list(rpl_gtid **list, uint32 *size)
 }
 
 
+/* ToDo: Should this be a service? */
+bool
+load_global_binlog_state(rpl_binlog_state_base *state)
+{
+  mysql_mutex_lock(&rpl_global_gtid_binlog_state.LOCK_binlog_state);
+  bool err= state->load_nolock(&rpl_global_gtid_binlog_state);
+  mysql_mutex_unlock(&rpl_global_gtid_binlog_state.LOCK_binlog_state);
+  return err;
+}
+
+
 bool
 MYSQL_BIN_LOG::append_state_pos(String *str)
 {
@@ -7471,7 +7486,8 @@ err:
         mysql_mutex_unlock(&LOCK_log);
         mysql_mutex_lock(&LOCK_commit_ordered);
         mysql_mutex_unlock(&LOCK_after_binlog_sync);
-        if ((*opt_binlog_engine_hton->binlog_write_direct)(file, binlog_main_bytes))
+        if ((*opt_binlog_engine_hton->binlog_write_direct)
+            (file, binlog_main_bytes, thd->get_last_commit_gtid()))
           goto engine_fail;
         /* ToDo: Need to set last_commit_pos_offset here? */
         /* ToDo: Maybe binlog_write_direct() could return the coords. */
@@ -7576,14 +7592,14 @@ err:
 
 
 void
-MYSQL_BIN_LOG::update_gtid_index(uint32 offset, rpl_gtid gtid)
+MYSQL_BIN_LOG::update_gtid_index(uint32 offset, const rpl_gtid *gtid)
 {
   if (!unlikely(gtid_index))
     return;
 
   rpl_gtid *gtid_list;
   uint32 gtid_count;
-  int err= gtid_index->process_gtid_check_batch(offset, &gtid,
+  int err= gtid_index->process_gtid_check_batch(offset, gtid,
                                                 &gtid_list, &gtid_count);
   if (err)
     return;

sql/log.h

@@ -600,6 +600,7 @@ class binlog_cache_mngr;
 class binlog_cache_data;
 struct rpl_gtid;
 struct wait_for_commit;
+struct rpl_binlog_state_base;
 
 class MYSQL_BIN_LOG: public TC_LOG, private Event_log
 {
@@ -763,7 +764,7 @@ class MYSQL_BIN_LOG: public TC_LOG, private Event_log
   bool write_transaction_to_binlog_events(group_commit_entry *entry);
   void trx_group_commit_leader(group_commit_entry *leader);
   bool is_xidlist_idle_nolock();
-  void update_gtid_index(uint32 offset, rpl_gtid gtid);
+  void update_gtid_index(uint32 offset, const rpl_gtid *gtid);
 
 public:
   void purge(bool all);
@@ -1191,6 +1192,9 @@ public:
   char binlog_end_pos_file[FN_REFLEN];
 };
 
+extern bool load_global_binlog_state(rpl_binlog_state_base *state);
+
+
 class Log_event_handler
 {
 public:

sql/sql_class.h

@@ -5512,7 +5512,7 @@ private:
   rpl_gtid m_last_commit_gtid;
 
 public:
-  rpl_gtid get_last_commit_gtid() { return m_last_commit_gtid; }
+  const rpl_gtid *get_last_commit_gtid() { return &m_last_commit_gtid; }
   void set_last_commit_gtid(rpl_gtid &gtid);
 
 

sql/sys_vars.cc

@@ -2324,7 +2324,7 @@ Sys_var_last_gtid::session_value_ptr(THD *thd, const LEX_CSTRING *base) const
   bool first= true;
 
   str.length(0);
-  rpl_gtid gtid= thd->get_last_commit_gtid();
+  rpl_gtid gtid= *thd->get_last_commit_gtid();
   if ((gtid.seq_no > 0 &&
        rpl_slave_state_tostring_helper(&str, &gtid, &first)) ||
       !(p= thd->strmake(str.ptr(), str.length())))

storage/innobase/fsp/fsp0fsp.cc

@@ -51,6 +51,7 @@ Created 11/29/1995 Heikki Tuuri
 #include <unordered_set>
 #include "trx0undo.h"
 #include "trx0trx.h"
+#include "ut0compr_int.h"
 #include "rpl_gtid_base.h"
 
 /** Returns the first extent descriptor for a segment.
@@ -4277,6 +4278,24 @@ func_exit:
 }
 
 
+/*
+  Binlog implementation in InnoDB.
+  ToDo: Move this somewhere reasonable, its own file(s) etc.
+*/
+
+/*
+  How often (in terms of bytes written) to dump a (differential) binlog state
+  at the start of the page, to speed up finding the initial GTID position for
+  a connecting slave.
+
+  Must be a power-of-two multiple of the page size.
+
+  ToDo: An InnoDB setting for this - which needs to be persisted in the
+  tablespace header or maybe initial full binlog state record.
+  ToDo: Default should be like 1M or so, this 64k is for testing.
+*/
+#define INNODB_BINLOG_STATE_INTERVAL (64*1024)
+
 static uint32_t binlog_size_in_pages;
 buf_block_t *binlog_cur_block;
 uint32_t binlog_cur_page_no;
@@ -4309,6 +4328,12 @@ std::atomic<uint64_t> first_open_binlog_file_no;
 uint64_t last_created_binlog_file_no;
 fil_space_t *last_created_binlog_space;
 
+/*
+  Differential binlog state in the currently active binlog tablespace, relative
+  to the state at the start.
+*/
+static rpl_binlog_state_base binlog_diff_state;
+
 /*
   Point at which it is guaranteed that all data has been written out to the
   binlog file (on the OS level; not necessarily fsync()'ed yet).
@@ -4386,15 +4411,17 @@ fsp_binlog_tablespace_close(uint64_t file_no)
   /*
     Write out any remaining pages in the buffer pool to the binlog tablespace.
     Then flush the file to disk, and close the old tablespace.
+
+    ToDo: Will this turn into a busy-wait if some of the pages are still latched
+    in this tablespace, maybe because even though the tablespace has been
+    written full, the mtr that's ending in the next tablespace may still be
+    active? This will need fixing, no busy-wait should be done here.
   */
   while (buf_flush_list_space(space))
     ;
   os_aio_wait_until_no_pending_writes(false);
   space->flush<false>();
-  mysql_mutex_lock(&fil_system.mutex);
+  fil_space_free(space_id, false);
-  fil_system.detach(space, false);
-  mysql_mutex_unlock(&fil_system.mutex);
-
   res= DB_SUCCESS;
 end:
   return res;
@@ -4411,6 +4438,7 @@ fsp_binlog_init()
 {
   mysql_mutex_init(fsp_active_binlog_mutex_key, &active_binlog_mutex, nullptr);
   pthread_cond_init(&active_binlog_cond, nullptr);
+  binlog_diff_state.init();
 }
 
 
@@ -4849,6 +4877,7 @@ void fsp_binlog_close()
 ==3464576==    by 0x1558445: fsp_binlog_tablespace_create(unsigned long) (fsp0fsp.cc:3900)
 ==3464576==    by 0x1558C70: fsp_binlog_write_cache(st_io_cache*, unsigned long, mtr_t*) (fsp0fsp.cc:4013)
   */
+  binlog_diff_state.free();
   pthread_cond_destroy(&active_binlog_cond);
   mysql_mutex_destroy(&active_binlog_mutex);
 }
@@ -4987,54 +5016,127 @@ fsp_binlog_prealloc_thread()
 }
 
 
-void fsp_binlog_write_start(uint32_t page_no,
+__attribute__((noinline))
-                            const uchar *data, uint32_t len, mtr_t *mtr)
+static ssize_t
+serialize_gtid_state(rpl_binlog_state_base *state, byte *buf, size_t buf_size,
+                     bool is_first_page)
 {
-	buf_block_t *block= fsp_page_create(active_binlog_space, page_no, mtr);
+  unsigned char *p= (unsigned char *)buf;
-	mtr->memcpy<mtr_t::MAYBE_NOP>(*block, FIL_PAGE_DATA + block->page.frame,
+  ut_ad(buf_size >= 2*COMPR_INT_MAX32 + 2*COMPR_INT_MAX64);
-				      data, len);
+  if (is_first_page) {
-	binlog_cur_block= block;
+    /*
-}
+      In the first page where we put the full state, include the value of the
-
+      setting for the interval at which differential states are binlogged, so
-void fsp_binlog_write_offset(uint32_t page_no, uint32_t offset,
+      we know how to search them independent of how the setting changes.
-                             const uchar *data, uint32_t len, mtr_t *mtr)
+    */
-{
+    p= compr_int_write(p, INNODB_BINLOG_STATE_INTERVAL);
-	dberr_t err;
-        /* ToDo: Is RW_SX_LATCH appropriate here? */
-	buf_block_t *block= buf_page_get_gen(page_id_t{active_binlog_space->id, page_no},
-					     0, RW_SX_LATCH, binlog_cur_block,
-					     BUF_GET, mtr, &err);
-	ut_a(err == DB_SUCCESS);
-	mtr->memcpy<mtr_t::MAYBE_NOP>(*block,
-                                      offset + block->page.frame,
-                                      data, len);
-}
-
-void fsp_binlog_append(const uchar *data, uint32_t len, mtr_t *mtr)
-{
-  ut_ad(binlog_cur_page_offset <= srv_page_size - FIL_PAGE_DATA_END);
-  uint32_t remain= ((uint32_t)srv_page_size - FIL_PAGE_DATA_END) -
-    binlog_cur_page_offset;
-  // ToDo: Some kind of mutex to protect binlog access.
-  while (len > 0) {
-    if (remain < 4) {
-      binlog_cur_page_offset= FIL_PAGE_DATA;
-      remain= ((uint32_t)srv_page_size - FIL_PAGE_DATA_END) -
-        binlog_cur_page_offset;
-      ++binlog_cur_page_no;
-    }
-    uint32_t this_len= std::min<uint32_t>(len, remain);
-    if (binlog_cur_page_offset == FIL_PAGE_DATA)
-      fsp_binlog_write_start(binlog_cur_page_no, data, this_len, mtr);
-    else
-      fsp_binlog_write_offset(binlog_cur_page_no, binlog_cur_page_offset,
-                              data, this_len, mtr);
-    len-= this_len;
-    data+= this_len;
-    binlog_cur_page_offset+= this_len;
   }
+  unsigned char * const pmax=
+    p + (buf_size - (2*COMPR_INT_MAX32 + COMPR_INT_MAX64));
+
+  if (state->iterate(
+    [buf, buf_size, pmax, &p] (const rpl_gtid *gtid) {
+      if (UNIV_UNLIKELY(p > pmax))
+        return true;
+      p= compr_int_write(p, gtid->domain_id);
+      p= compr_int_write(p, gtid->server_id);
+      p= compr_int_write(p, gtid->seq_no);
+      return false;
+    }))
+    return -1;
+  else
+    return p - (unsigned char *)buf;
 }
 
+
+static bool
+binlog_gtid_state(rpl_binlog_state_base *state, mtr_t *mtr,
+                  buf_block_t * &block, uint32_t &page_no,
+                  uint32_t &page_offset, fil_space_t *space)
+{
+  /*
+    Use a small, efficient stack-allocated buffer by default, falling back to
+    malloc() if needed for large GTID state.
+  */
+  byte small_buf[192];
+  byte *buf, *alloced_buf;
+
+  ssize_t used_bytes= serialize_gtid_state(state, small_buf, sizeof(small_buf),
+                                           page_no==0);
+  if (used_bytes >= 0)
+  {
+    buf= small_buf;
+    alloced_buf= nullptr;
+  }
+  else
+  {
+    size_t buf_size=
+      state->count_nolock() * (2*COMPR_INT_MAX32 + COMPR_INT_MAX64);
+    /* ToDo: InnoDB alloc function? */
+    alloced_buf= (byte *)my_malloc(PSI_INSTRUMENT_ME, buf_size, MYF(MY_WME));
+    if (UNIV_UNLIKELY(!alloced_buf))
+      return true;
+    buf= alloced_buf;
+    used_bytes= serialize_gtid_state(state, buf, buf_size, page_no==0);
+    if (UNIV_UNLIKELY(used_bytes < 0))
+    {
+      ut_ad(0 /* Shouldn't happen, as we allocated maximum needed size. */);
+      my_free(alloced_buf);
+      return true;
+    }
+  }
+
+  const uint32_t page_size= (uint32_t)srv_page_size;
+  const uint32_t page_room= page_size - (FIL_PAGE_DATA + FIL_PAGE_DATA_END);
+  uint32_t needed_pages= (uint32_t)((used_bytes + page_room - 1) / page_room);
+
+  /* For now, GTID state always at the start of a page. */
+  ut_ad(page_offset == FIL_PAGE_DATA);
+
+  /*
+    Only write the GTID state record if there is room for actual event data
+    afterwards. There is no point in using space to allow fast search to a
+    point if there is no data to search for after that point.
+  */
+  if (page_no + needed_pages < space->size)
+  {
+    while (used_bytes > 0)
+    {
+      ut_ad(page_no < space->size);
+      block= fsp_page_create(space, page_no, mtr);
+      ut_a(block /* ToDo: error handling? */);
+      page_offset= FIL_PAGE_DATA;
+      byte *ptr= page_offset + block->page.frame;
+      ssize_t chunk= used_bytes;
+      if (chunk > page_room - 3) {
+        chunk= page_room - 3;
+        ++page_no;
+      }
+      ptr[0]= 0x02 /* ToDo: FSP_BINLOG_TYPE_GTID_STATE */ | ((chunk < used_bytes) << 7);
+      ptr[1] = (byte)chunk & 0xff;
+      ptr[2] = (byte)(chunk >> 8);
+      ut_ad(chunk <= 0xffff);
+      memcpy(ptr+3, buf, chunk);
+      mtr->memcpy(*block, page_offset, chunk+3);
+      page_offset+= (uint32_t)(chunk+3);
+      buf+= chunk;
+      used_bytes-= chunk;
+    }
+
+    if (page_offset == FIL_PAGE_DATA_END) {
+      block= nullptr;
+      page_offset= FIL_PAGE_DATA;
+      ++page_no;
+      block= fsp_page_create(space, page_no, mtr);
+      ut_a(block /* ToDo: error handling? */);
+    }
+  }
+
+  my_free(alloced_buf);
+  return false;  // No error
+}
+
+
 void fsp_binlog_write_cache(IO_CACHE *cache, size_t main_size, mtr_t *mtr)
 {
   uint32_t page_size= (uint32_t)srv_page_size;
@@ -5114,7 +5216,32 @@ void fsp_binlog_write_cache(IO_CACHE *cache, size_t main_size, mtr_t *mtr)
         mysql_mutex_unlock(&active_binlog_mutex);
         binlog_cur_page_no= page_no= 0;
       }
-      block= fsp_page_create(space, page_no, mtr);
+
+      /* Must be a power of two and larger than page size. */
+      ut_ad(INNODB_BINLOG_STATE_INTERVAL > page_size);
+      ut_ad(INNODB_BINLOG_STATE_INTERVAL ==
+            (uint64_t)1 << (63 - nlz((uint64_t)INNODB_BINLOG_STATE_INTERVAL)));
+
+      if (0 == (page_no &
+                ((INNODB_BINLOG_STATE_INTERVAL >> page_size_shift) - 1))) {
+        if (page_no == 0) {
+          rpl_binlog_state_base full_state;
+          full_state.init();
+          bool err= load_global_binlog_state(&full_state);
+          ut_a(!err /* ToDo error handling */);
+          err= binlog_gtid_state(&full_state, mtr, block, page_no,
+                                 page_offset, space);
+          ut_a(!err /* ToDo error handling */);
+          ut_ad(block);
+          full_state.free();
+          binlog_diff_state.reset_nolock();
+        } else {
+          bool err= binlog_gtid_state(&binlog_diff_state, mtr, block, page_no,
+                                      page_offset, space);
+          ut_a(!err /* ToDo error handling */);
+        }
+      } else
+        block= fsp_page_create(space, page_no, mtr);
     } else {
       dberr_t err;
       /* ToDo: Is RW_SX_LATCH appropriate here? */
@@ -5185,30 +5312,23 @@ void fsp_binlog_write_cache(IO_CACHE *cache, size_t main_size, mtr_t *mtr)
 }
 
 
-extern "C" void binlog_get_cache(THD *, IO_CACHE **, size_t *);
+extern "C" void binlog_get_cache(THD *, IO_CACHE **, size_t *,
+                                 const rpl_gtid **);
 
 void
 fsp_binlog_trx(trx_t *trx, mtr_t *mtr)
 {
   IO_CACHE *cache;
   size_t main_size;
+  const rpl_gtid *gtid;
 
   if (!trx->mysql_thd)
     return;
-  binlog_get_cache(trx->mysql_thd, &cache, &main_size);
+  binlog_get_cache(trx->mysql_thd, &cache, &main_size, &gtid);
-  if (main_size)
+  if (main_size) {
+    binlog_diff_state.update_nolock(gtid);
     fsp_binlog_write_cache(cache, main_size, mtr);
-}
+  }
-
-
-void fsp_binlog_test(const uchar *data, uint32_t len)
-{
-  mtr_t mtr;
-  mtr.start();
-  if (!active_binlog_space)
-    fsp_binlog_tablespace_create(0, &active_binlog_space);
-  fsp_binlog_append(data, len, &mtr);
-  mtr.commit();
 }
 
 
@@ -5276,7 +5396,9 @@ int ha_innodb_binlog_reader::read_binlog_data(uchar *buf, uint32_t len)
   uint64_t file_no= cur_file_no;
   uint64_t offset= cur_file_offset;
   uint64_t active_file_no= active_binlog_file_no.load(std::memory_order_acquire);
-  if (first_open_binlog_file_no.load(std::memory_order_relaxed) > file_no + /* Temporary hack to work-around the next line ToDo: comment */ (offset >= cur_file_length)) {
+  if (first_open_binlog_file_no.load(std::memory_order_relaxed) > file_no +
+      /* Temporary hack to work-around the next line ToDo: comment */
+      (cur_file != (File)-1 && offset >= cur_file_length)) {
     // ToDo: I think there is a bug here, if we're at the end of active_file_no-2, we will be reading directly from active_file_no-1 without checking properly if buffer pool is needed instead. */
     return read_from_file(~(uint64_t)0, buf, len);
   }
@@ -5296,6 +5418,10 @@ int ha_innodb_binlog_reader::read_binlog_data(uchar *buf, uint32_t len)
   } else {  /* file_no == active_file_no - 1 */
     if (offset >= end_offset) {  // ToDo what if this end_pos is stale? Need somehow an extra check afterwards if we are now active-2, and then do a simple file read, not EOF on the stale end_offset.
       /* Handle moving to the currently active file. */
+      if (!(cur_file < (File)0)) {
+        my_close(cur_file, MYF(0));
+        cur_file= (File)-1;
+      }
       cur_file_no= ++file_no;
       cur_file_offset= offset= 0;
       idx= file_no & 1;
@@ -5372,11 +5498,14 @@ ha_innodb_binlog_reader::read_from_file(uint64_t end_offset,
   uint64_t offset= cur_file_offset;
   uint64_t page_start_offset;
 
-  if (cur_file < (File)0 || cur_file_offset >= cur_file_length) {
+  if (cur_file >= (File)0 && cur_file_offset >= cur_file_length) {
-    if (!(cur_file < (File)0)) {
+    /* At the end of currently open file, move to the next file. */
-      my_close(cur_file, MYF(0));
+    my_close(cur_file, MYF(0));
-      ++cur_file_no;
+    cur_file= (File)-1;
-    }
+    ++cur_file_no;
+    cur_file_offset= offset= 0;
+  }
+  if (cur_file < (File)0) {
     char filename[BINLOG_NAME_LEN];
     binlog_name_make(filename, cur_file_no);
     if ((cur_file= my_open(filename, O_RDONLY | O_BINARY, MYF(MY_WME))) < (File)0)
@@ -5390,7 +5519,6 @@ ha_innodb_binlog_reader::read_from_file(uint64_t end_offset,
       return -1;
     }
     cur_file_length= stat_buf.st_size;
-    cur_file_offset= offset= 0;
   }
 
   page_start_offset= offset & ~mask;
@@ -5501,9 +5629,12 @@ innodb_get_binlog_reader()
 
 
 bool
-innobase_binlog_write_direct(IO_CACHE *cache, size_t main_size)
+innobase_binlog_write_direct(IO_CACHE *cache, size_t main_size,
+                             const rpl_gtid *gtid)
 {
   mtr_t mtr;
+  if (gtid)
+    binlog_diff_state.update_nolock(gtid);
   mtr.start();
   fsp_binlog_write_cache(cache, main_size, &mtr);
   mtr.commit();

storage/innobase/include/fsp0fsp.h

@@ -556,10 +556,10 @@ void fsp_system_tablespace_truncate(bool shutdown);
 /** Truncate the temporary tablespace */
 void fsp_shrink_temp_space();
 
-extern void fsp_binlog_test(const uchar *data, uint32_t len);
 extern void fsp_binlog_trx(trx_t *trx, mtr_t *mtr);
 class handler_binlog_reader;
-extern bool innobase_binlog_write_direct(IO_CACHE *cache, size_t main_size);
+extern bool innobase_binlog_write_direct(IO_CACHE *cache, size_t main_size,
+                                         const struct rpl_gtid *gtid);
 extern handler_binlog_reader *innodb_get_binlog_reader();
 extern void fsp_binlog_init();
 extern bool innodb_binlog_init(size_t binlog_size);