MDEV-36760, MDEV-36781, MDEV-36759: Fix performance regressions after MDEV-29445

VERSION

@@ -1,4 +1,4 @@
 MYSQL_VERSION_MAJOR=10
 MYSQL_VERSION_MINOR=11
-MYSQL_VERSION_PATCH=12
+MYSQL_VERSION_PATCH=13
 SERVER_MATURITY=stable

include/my_sys.h

@@ -177,7 +177,9 @@ extern my_bool my_use_large_pages;
 
 int my_init_large_pages(void);
 uchar *my_large_malloc(size_t *size, myf my_flags);
-#if defined _WIN32 || defined HAVE_MMAP
+#ifdef _WIN32
+/* On Windows, use my_virtual_mem_reserve() and my_virtual_mem_commit(). */
+#else
 char *my_large_virtual_alloc(size_t *size);
 #endif
 void my_large_free(void *ptr, size_t size);

include/my_virtual_mem.h

@@ -24,7 +24,9 @@
 extern "C" {
 #endif
 
+# ifdef _WIN32
 char *my_virtual_mem_reserve(size_t *size);
+# endif
 char *my_virtual_mem_commit(char *ptr, size_t size);
 void my_virtual_mem_decommit(char *ptr, size_t size);
 void my_virtual_mem_release(char *ptr, size_t size);

mysql-test/suite/innodb/r/innodb_buffer_pool_resize.result

@@ -12,12 +12,19 @@ select @@innodb_buffer_pool_size;
 10485760
 create table t1 (id int primary key, val int not null)
 ENGINE=InnoDB ROW_FORMAT=COMPRESSED;
+create table t2 (id int primary key, val int not null)
+ENGINE=InnoDB ROW_FORMAT=COMPRESSED KEY_BLOCK_SIZE=$kbs;
 SET STATEMENT foreign_key_checks=0, unique_checks=0 FOR
 INSERT INTO t1 SELECT seq*4,seq*4 FROM seq_1_to_262144;
+SET STATEMENT foreign_key_checks=0, unique_checks=0 FOR
+INSERT INTO t2 SELECT seq*4,seq*4 FROM seq_1_to_16384;
 set global innodb_buffer_pool_size = 7340032;
 select count(val) from t1;
 count(val)
 262144
+select count(val) from t2;
+count(val)
+16384
 set global innodb_adaptive_hash_index=OFF;
 set global innodb_buffer_pool_size = 24117248;
 set global innodb_buffer_pool_size = 26214400;
@@ -29,7 +36,10 @@ select @@innodb_buffer_pool_size;
 select count(val) from t1;
 count(val)
 262144
-drop table t1;
+select count(val) from t2;
+count(val)
+16384
+drop table t1,t2;
 SET GLOBAL innodb_max_purge_lag_wait = 0;
 SET @save_pct= @@GLOBAL.innodb_max_dirty_pages_pct;
 SET @save_pct_lwm= @@GLOBAL.innodb_max_dirty_pages_pct_lwm;

mysql-test/suite/innodb/t/innodb_buffer_pool_resize.test

@@ -21,6 +21,7 @@ set global innodb_buffer_pool_size = 9437184;
 set global innodb_buffer_pool_size = 10485760;
 
 select @@innodb_buffer_pool_size;
+let $kbs=`SELECT CAST(@@innodb_page_size / 1024 AS INT)`;
 
 # fill buffer pool
 --disable_query_log
@@ -29,9 +30,13 @@ SET GLOBAL innodb_read_only_compressed=OFF;
 --enable_query_log
 create table t1 (id int primary key, val int not null)
 ENGINE=InnoDB ROW_FORMAT=COMPRESSED;
+evalp create table t2 (id int primary key, val int not null)
+ENGINE=InnoDB ROW_FORMAT=COMPRESSED KEY_BLOCK_SIZE=$kbs;
 
 SET STATEMENT foreign_key_checks=0, unique_checks=0 FOR
 INSERT INTO t1 SELECT seq*4,seq*4 FROM seq_1_to_262144;
+SET STATEMENT foreign_key_checks=0, unique_checks=0 FOR
+INSERT INTO t2 SELECT seq*4,seq*4 FROM seq_1_to_16384;
 
 --disable_query_log
 SET GLOBAL innodb_read_only_compressed=@save_innodb_read_only_compressed;
@@ -42,6 +47,7 @@ SET GLOBAL innodb_read_only_compressed=@save_innodb_read_only_compressed;
 set global innodb_buffer_pool_size = 7340032;
 
 select count(val) from t1;
+select count(val) from t2;
 
 set global innodb_adaptive_hash_index=OFF;
 
@@ -52,8 +58,9 @@ set global innodb_buffer_pool_size = 26214400;
 select @@innodb_buffer_pool_size;
 
 select count(val) from t1;
+select count(val) from t2;
 
-drop table t1;
+drop table t1,t2;
 
 SET GLOBAL innodb_max_purge_lag_wait = 0;
 SET @save_pct= @@GLOBAL.innodb_max_dirty_pages_pct;
@@ -66,8 +73,6 @@ SELECT variable_value = 0
 FROM information_schema.global_status
 WHERE variable_name = 'INNODB_BUFFER_POOL_PAGES_DIRTY';
 --source include/wait_condition.inc
-# this may occasionally be aborted on a heavily loaded builder
---error 0,ER_WRONG_USAGE
 SET GLOBAL innodb_buffer_pool_size = @old_innodb_buffer_pool_size;
 SET GLOBAL innodb_adaptive_hash_index = @old_innodb_adaptive_hash_index;
 SET GLOBAL innodb_max_dirty_pages_pct = @save_pct;

mysql-test/suite/rpl/t/rpl_semi_sync_ssl_stop.test

@@ -49,8 +49,8 @@ SET @@GLOBAL.rpl_semi_sync_slave_enabled= 1;
 
 --connection master
 --echo # Verify Semi-Sync is active
---let $status_var= Rpl_semi_sync_master_status
---let $status_var_value= ON
+--let $status_var= Rpl_semi_sync_master_clients
+--let $status_var_value= 1
 --source include/wait_for_status_var.inc
 SHOW STATUS LIKE 'Rpl_semi_sync_master_clients';
 
@@ -67,7 +67,7 @@ STOP SLAVE;
 --echo # MDEV-36663: Verifying dump thread connection is killed..
 # Prior to MDEV-36663 fixes, this would time out and
 # Rpl_semi_sync_master_clients would remain 1.
---let $wait_condition= SELECT COUNT(*)=0 FROM information_schema.PROCESSLIST WHERE COMMAND = 'Binlog Dump'
+--let $wait_condition= SELECT COUNT(*)=0 FROM information_schema.PROCESSLIST WHERE USER = 'replssl'
 --source include/wait_condition.inc
 
 --let $n_master_clients= query_get_value(SHOW STATUS LIKE 'Rpl_semi_sync_master_clients', Value, 1)

mysys/my_largepage.c

@@ -423,7 +423,7 @@ uchar *my_large_malloc(size_t *size, myf my_flags)
   DBUG_RETURN(ptr);
 }
 
-#ifdef _WIN32
+#ifndef _WIN32
 /**
   Special large pages allocator, with possibility to commit to allocating
   more memory later.
@@ -434,37 +434,10 @@ char *my_large_virtual_alloc(size_t *size)
   char *ptr;
   DBUG_ENTER("my_large_virtual_alloc");
 
-  if (my_use_large_pages)
-  {
-    size_t s= *size;
-    s= MY_ALIGN(s, (size_t) my_large_page_size);
-    ptr= VirtualAlloc(NULL, s, MEM_COMMIT | MEM_RESERVE | MEM_LARGE_PAGES,
-                      PAGE_READWRITE);
-    if (ptr)
-    {
-      *size= s;
-      DBUG_RETURN(ptr);
-    }
-  }
-
-  DBUG_RETURN(VirtualAlloc(NULL, *size, MEM_RESERVE, PAGE_READWRITE));
-}
-#elif defined HAVE_MMAP
-/**
-  Special large pages allocator, with possibility to commit to allocating
-  more memory later.
-  Every implementation returns a zero filled buffer here.
-*/
-char *my_large_mmap(size_t *size, int prot)
-{
-  char *ptr;
-  DBUG_ENTER("my_large_virtual_alloc");
-
   if (my_use_large_pages)
   {
     size_t large_page_size;
     int page_i= 0;
-    prot= PROT_READ | PROT_WRITE;
 
     while ((large_page_size= my_next_large_page_size(*size, &page_i)) != 0)
     {
@@ -488,7 +461,7 @@ char *my_large_mmap(size_t *size, int prot)
         OS_MAP_ANON;
 
       size_t aligned_size= MY_ALIGN(*size, (size_t) large_page_size);
-      ptr= mmap(NULL, aligned_size, prot, mapflag, -1, 0);
+      ptr= mmap(NULL, aligned_size, PROT_READ | PROT_WRITE, mapflag, -1, 0);
       if (ptr == (void*) -1)
       {
         ptr= NULL;
@@ -511,10 +484,7 @@ char *my_large_mmap(size_t *size, int prot)
     }
   }
 
-  ptr= mmap(NULL, *size, prot,
-# ifdef MAP_NORESERVE
-            MAP_NORESERVE |
-# endif
+  ptr= mmap(NULL, *size, PROT_READ | PROT_WRITE,
             MAP_PRIVATE | OS_MAP_ANON, -1, 0);
   if (ptr == MAP_FAILED)
   {
@@ -524,16 +494,6 @@ char *my_large_mmap(size_t *size, int prot)
 
   DBUG_RETURN(ptr);
 }
-
-/**
-  Special large pages allocator, with possibility to commit to allocating
-  more memory later.
-  Every implementation returns a zero filled buffer here.
-*/
-char *my_large_virtual_alloc(size_t *size)
-{
-  return my_large_mmap(size, PROT_READ | PROT_WRITE);
-}
 #endif
 
 /**

mysys/my_virtual_mem.c

@@ -34,13 +34,9 @@
 
   We try to respect use_large_pages setting, on Windows and Linux
 */
-#ifndef _WIN32
-char *my_large_mmap(size_t *size, int prot);
-#endif
-
+#ifdef _WIN32
 char *my_virtual_mem_reserve(size_t *size)
 {
-#ifdef _WIN32
   DWORD flags= my_use_large_pages
     ? MEM_LARGE_PAGES | MEM_RESERVE | MEM_COMMIT
     : MEM_RESERVE;
@@ -53,10 +49,8 @@ char *my_virtual_mem_reserve(size_t *size)
       my_error(EE_OUTOFMEMORY, MYF(ME_BELL + ME_ERROR_LOG), *size);
   }
   return ptr;
-#else
-  return my_large_mmap(size, PROT_NONE);
-#endif
 }
+#endif
 
 #if defined _WIN32 && !defined DBUG_OFF
 static my_bool is_memory_committed(char *ptr, size_t size)
@@ -88,7 +82,7 @@ char *my_virtual_mem_commit(char *ptr, size_t size)
   }
 #else
   if (my_use_large_pages)
-    /* my_large_mmap() already created a read/write mapping. */;
+    /* my_large_virtual_alloc() already created a read/write mapping. */;
   else
   {
 # ifdef _AIX

storage/innobase/btr/btr0sea.cc

@@ -276,7 +276,7 @@ ATTRIBUTE_COLD void btr_search_enable(bool resize)
 	}
 
 	btr_search_x_lock_all();
-	ulint hash_size = buf_pool.curr_size() / sizeof(void *) / 64;
+	ulint hash_size = buf_pool.curr_pool_size() / sizeof(void *) / 64;
 
 	if (btr_search_sys.parts[0].heap) {
 		ut_ad(btr_search_enabled);

storage/innobase/buf/buf0buddy.cc

@@ -637,7 +637,7 @@ void buf_buddy_free_low(void* buf, ulint i) noexcept
 	buf_buddy_add_to_free(reinterpret_cast<buf_buddy_free_t*>(buf), i);
 }
 
-/** Reallocate a ROW_FORMAT=COMPRESSED page frame during buf_pool_t::resize().
+/** Reallocate a ROW_FORMAT=COMPRESSED page frame during buf_pool_t::shrink().
 @param bpage page descriptor covering a ROW_FORMAT=COMPRESSED page
 @param block uncompressed block for storage
 @return block
@@ -672,10 +672,9 @@ buf_block_t *buf_buddy_shrink(buf_page_t *bpage, buf_block_t *block) noexcept
   bpage->zip.data= static_cast<page_zip_t*>(dst);
   buf_pool.buddy_stat[i].relocated++;
 
-  for (;;)
+  while (i < BUF_BUDDY_SIZES)
   {
     MEM_UNDEFINED(src, BUF_BUDDY_LOW << i);
-    ut_ad(i < BUF_BUDDY_SIZES);
     /* Try to combine adjacent blocks. */
     buf_buddy_free_t *buddy= reinterpret_cast<buf_buddy_free_t*>
       (buf_buddy_get(static_cast<byte*>(src), BUF_BUDDY_LOW << i));
@@ -684,20 +683,16 @@ buf_block_t *buf_buddy_shrink(buf_page_t *bpage, buf_block_t *block) noexcept
     {
       ut_ad(!buf_pool.contains_zip(src, BUF_BUDDY_LOW_SHIFT + i));
       buf_buddy_add_to_free(static_cast<buf_buddy_free_t*>(src), i);
-      break;
+      return block;
     }
 
     /* The buddy is free: recombine */
     buf_buddy_remove_from_free(buddy, i);
     i++;
     src= ut_align_down(src, BUF_BUDDY_LOW << i);
-    if (i == BUF_BUDDY_SIZES)
-    {
-      buf_buddy_block_free(src);
-      break;
-    }
   }
 
+  buf_buddy_block_free(src);
   return block;
 }
 

storage/innobase/buf/buf0buf.cc

@@ -1336,7 +1336,11 @@ bool buf_pool_t::create() noexcept
  retry:
   {
     NUMA_MEMPOLICY_INTERLEAVE_IN_SCOPE;
+#ifdef _WIN32
     memory_unaligned= my_virtual_mem_reserve(&size);
+#else
+    memory_unaligned= my_large_virtual_alloc(&size);
+#endif
   }
 
   if (!memory_unaligned)
@@ -1370,13 +1374,17 @@ bool buf_pool_t::create() noexcept
 #ifdef UNIV_PFS_MEMORY
   PSI_MEMORY_CALL(memory_alloc)(mem_key_buf_buf_pool, actual_size, &owner);
 #endif
+#ifdef _WIN32
   if (!my_virtual_mem_commit(memory, actual_size))
   {
     my_virtual_mem_release(memory_unaligned, size_unaligned);
     memory= nullptr;
     memory_unaligned= nullptr;
     goto oom;
   }
+#else
+  update_malloc_size(actual_size, 0);
+#endif
 
 #ifdef HAVE_LIBNUMA
   if (srv_numa_interleave)
@@ -1788,6 +1796,9 @@ ATTRIBUTE_COLD buf_pool_t::shrink_status buf_pool_t::shrink(size_t size)
     goto next;
   }
 
+  if (UT_LIST_GET_LEN(free) + UT_LIST_GET_LEN(LRU) < usable_size() / 20)
+    return SHRINK_ABORT;
+
   mysql_mutex_lock(&flush_list_mutex);
 
   if (LRU_warned && !UT_LIST_GET_FIRST(free))

storage/innobase/buf/buf0lru.cc

@@ -276,7 +276,10 @@ static void buf_LRU_check_size_of_non_data_objects() noexcept
 
   auto s= UT_LIST_GET_LEN(buf_pool.free) + UT_LIST_GET_LEN(buf_pool.LRU);
 
-  if (s < curr_size / 20)
+  if (s >= curr_size / 20);
+  else if (buf_pool.is_shrinking())
+    buf_pool.LRU_warn();
+  else
   {
     sql_print_error("[FATAL] InnoDB: Over 95 percent of the buffer pool is"
                     " occupied by lock heaps"

storage/innobase/dict/dict0dict.cc

@@ -933,7 +933,7 @@ void dict_sys_t::create() noexcept
   UT_LIST_INIT(table_LRU, &dict_table_t::table_LRU);
   UT_LIST_INIT(table_non_LRU, &dict_table_t::table_LRU);
 
-  const ulint hash_size = buf_pool.curr_size()
+  const ulint hash_size = buf_pool.curr_pool_size()
     / (DICT_POOL_PER_TABLE_HASH * UNIV_WORD_SIZE);
 
   table_hash.create(hash_size);
@@ -4399,7 +4399,7 @@ void dict_sys_t::resize() noexcept
   table_id_hash.free();
   temp_id_hash.free();
 
-  const ulint hash_size = buf_pool.curr_size()
+  const ulint hash_size = buf_pool.curr_pool_size()
     / (DICT_POOL_PER_TABLE_HASH * UNIV_WORD_SIZE);
   table_hash.create(hash_size);
   table_id_hash.create(hash_size);

storage/innobase/handler/ha_innodb.cc

@@ -4075,7 +4075,6 @@ static int innodb_init_params()
 #else
   ut_ad(srv_file_flush_method <= SRV_O_DIRECT_NO_FSYNC);
 #endif
-  srv_lock_table_size = 5 * buf_pool.curr_size();
   DBUG_RETURN(0);
 }
 

storage/innobase/include/buf0buddy.h

@@ -76,7 +76,7 @@ inline void buf_buddy_free(void* buf, ulint size) noexcept
 }
 
 ATTRIBUTE_COLD MY_ATTRIBUTE((nonnull, warn_unused_result))
-/** Reallocate a ROW_FORMAT=COMPRESSED page frame during buf_pool_t::resize().
+/** Reallocate a ROW_FORMAT=COMPRESSED page frame during buf_pool_t::shrink().
 @param bpage page descriptor covering a ROW_FORMAT=COMPRESSED page
 @param block uncompressed block for storage
 @return block