MDEV-24167: Use lightweight srw_lock for btr_search_latch

Many InnoDB rw-locks unnecessarily depend on the complex
InnoDB rw_lock_t implementation that support the SX lock mode
as well as recursive acquisition of X or SX locks.
One of them is the bunch of adaptive hash index search latches,
instrumented as btr_search_latch in PERFORMANCE_SCHEMA.
Let us introduce a simpler lock for those in order to
reduce overhead.

srw_lock: A simple read-write lock that does not support recursion.
On Microsoft Windows, this wraps SRWLOCK, only adding
runtime overhead if PERFORMANCE_SCHEMA is enabled.
On Linux (all architectures), this is implemented with
std::atomic<uint32_t> and the futex system call.
On other platforms, we will wrap mysql_rwlock_t with
zero runtime overhead.

The PERFORMANCE_SCHEMA instrumentation differs
from InnoDB rw_lock_t in that we will only invoke
PSI_RWLOCK_CALL(start_rwlock_wrwait) or
PSI_RWLOCK_CALL(start_rwlock_rdwait)
if there is an actual conflict.

mysql-test/suite/perfschema/r/sxlock_func.result

@@ -7,7 +7,6 @@ TRUNCATE TABLE performance_schema.events_waits_current;
 select name from performance_schema.setup_instruments
 where name like "wait/synch/sxlock/%" order by name;
 name
-wait/synch/sxlock/innodb/btr_search_latch
 wait/synch/sxlock/innodb/dict_operation_lock
 wait/synch/sxlock/innodb/fil_space_latch
 wait/synch/sxlock/innodb/fts_cache_init_rw_lock
@@ -26,14 +25,6 @@ where name in
   )
 order by name;
 name
-wait/synch/sxlock/innodb/btr_search_latch
-wait/synch/sxlock/innodb/btr_search_latch
-wait/synch/sxlock/innodb/btr_search_latch
-wait/synch/sxlock/innodb/btr_search_latch
-wait/synch/sxlock/innodb/btr_search_latch
-wait/synch/sxlock/innodb/btr_search_latch
-wait/synch/sxlock/innodb/btr_search_latch
-wait/synch/sxlock/innodb/btr_search_latch
 wait/synch/sxlock/innodb/dict_operation_lock
 wait/synch/sxlock/innodb/trx_i_s_cache_lock
 wait/synch/sxlock/innodb/trx_purge_latch

storage/innobase/CMakeLists.txt

@@ -350,6 +350,12 @@ SET(INNOBASE_SOURCES
 	ut/ut0vec.cc
 	ut/ut0wqueue.cc)
 
+IF (UNIX)
+  IF(CMAKE_SYSTEM_NAME STREQUAL "Linux")
+    SET(INNOBASE_SOURCES ${INNOBASE_SOURCES} "sync/srw_lock_futex.cc")
+  ENDIF()
+ENDIF()
+
 MYSQL_ADD_PLUGIN(innobase ${INNOBASE_SOURCES} STORAGE_ENGINE
   MODULE_OUTPUT_NAME ha_innodb
   DEFAULT RECOMPILE_FOR_EMBEDDED

storage/innobase/btr/btr0cur.cc

@@ -1245,9 +1245,8 @@ btr_cur_search_to_nth_level_func(
 	btr_cur_t*	cursor, /*!< in/out: tree cursor; the cursor page is
 				s- or x-latched, but see also above! */
 #ifdef BTR_CUR_HASH_ADAPT
-	rw_lock_t*	ahi_latch,
-				/*!< in: currently held btr_search_latch
-				(in RW_S_LATCH mode), or NULL */
+	srw_lock*	ahi_latch,
+				/*!< in: currently held AHI rdlock, or NULL */
 #endif /* BTR_CUR_HASH_ADAPT */
 	const char*	file,	/*!< in: file name */
 	unsigned	line,	/*!< in: line where called */
@@ -1449,7 +1448,7 @@ btr_cur_search_to_nth_level_func(
 #ifdef BTR_CUR_HASH_ADAPT
 	if (ahi_latch) {
 		/* Release possible search latch to obey latching order */
-		rw_lock_s_unlock(ahi_latch);
+		ahi_latch->rd_unlock();
 	}
 #endif /* BTR_CUR_HASH_ADAPT */
 
@@ -2502,7 +2501,7 @@ btr_cur_search_to_nth_level_func(
 
 #ifdef BTR_CUR_HASH_ADAPT
 	if (ahi_latch) {
-		rw_lock_s_lock(ahi_latch);
+		ahi_latch->rd_lock();
 	}
 #endif /* BTR_CUR_HASH_ADAPT */
 
@@ -3584,7 +3583,7 @@ btr_cur_optimistic_insert(
 		ut_ad(index->is_instant());
 		ut_ad(flags == BTR_NO_LOCKING_FLAG);
 	} else {
-		rw_lock_t* ahi_latch = btr_search_sys.get_latch(*index);
+		srw_lock* ahi_latch = btr_search_sys.get_latch(*index);
 		if (!reorg && cursor->flag == BTR_CUR_HASH) {
 			btr_search_update_hash_node_on_insert(
 				cursor, ahi_latch);
@@ -4296,7 +4295,7 @@ btr_cur_update_in_place(
 
 #ifdef BTR_CUR_HASH_ADAPT
 	{
-		rw_lock_t* ahi_latch = block->index
+		srw_lock* ahi_latch = block->index
 			? btr_search_sys.get_latch(*index) : NULL;
 		if (ahi_latch) {
 			/* TO DO: Can we skip this if none of the fields
@@ -4316,7 +4315,7 @@ btr_cur_update_in_place(
 				btr_search_update_hash_on_delete(cursor);
 			}
 
-			rw_lock_x_lock(ahi_latch);
+			ahi_latch->wr_lock();
 		}
 
 		assert_block_ahi_valid(block);
@@ -4327,7 +4326,7 @@ btr_cur_update_in_place(
 
 #ifdef BTR_CUR_HASH_ADAPT
 		if (ahi_latch) {
-			rw_lock_x_unlock(ahi_latch);
+			ahi_latch->wr_unlock();
 		}
 	}
 #endif /* BTR_CUR_HASH_ADAPT */

storage/innobase/buf/buf0buf.cc

@@ -1980,16 +1980,11 @@ inline void buf_pool_t::resize()
 
 #ifdef BTR_CUR_HASH_ADAPT
 	/* disable AHI if needed */
-	const bool btr_search_disabled = btr_search_enabled;
-
 	buf_resize_status("Disabling adaptive hash index.");
 
 	btr_search_s_lock_all();
-	if (btr_search_disabled) {
-		btr_search_s_unlock_all();
-	} else {
+	const bool btr_search_disabled = btr_search_enabled;
 	btr_search_s_unlock_all();
-	}
 
 	btr_search_disable();
 

storage/innobase/handler/ha_innodb.cc

@@ -561,7 +561,9 @@ static PSI_mutex_info all_innodb_mutexes[] = {
 performance schema instrumented if "UNIV_PFS_RWLOCK"
 is defined */
 static PSI_rwlock_info all_innodb_rwlocks[] = {
-	PSI_RWLOCK_KEY(btr_search_latch),
+#  ifdef BTR_CUR_HASH_ADAPT
+  { &btr_search_latch_key, "btr_search_latch", 0 },
+#  endif
 	PSI_RWLOCK_KEY(dict_operation_lock),
 	PSI_RWLOCK_KEY(fil_space_latch),
 	PSI_RWLOCK_KEY(fts_cache_rw_lock),

storage/innobase/include/btr0cur.h

@@ -33,6 +33,9 @@ Created 10/16/1994 Heikki Tuuri
 #include "rem0types.h"
 #include "gis0type.h"
 #include "my_base.h"
+#ifdef BTR_CUR_HASH_ADAPT
+# include "srw_lock.h"
+#endif
 
 /** Mode flags for btr_cur operations; these can be ORed */
 enum {
@@ -202,9 +205,8 @@ btr_cur_search_to_nth_level_func(
 	btr_cur_t*	cursor, /*!< in/out: tree cursor; the cursor page is
 				s- or x-latched, but see also above! */
 #ifdef BTR_CUR_HASH_ADAPT
-	rw_lock_t*	ahi_latch,
-				/*!< in: currently held btr_search_latch
-				(in RW_S_LATCH mode), or NULL */
+	srw_lock*	ahi_latch,
+				/*!< in: currently held AHI rdlock, or NULL */
 #endif /* BTR_CUR_HASH_ADAPT */
 	const char*	file,	/*!< in: file name */
 	unsigned	line,	/*!< in: line where called */

storage/innobase/include/btr0pcur.h

@@ -137,9 +137,8 @@ btr_pcur_open_with_no_init_func(
 				that the ahi_latch protects the record! */
 	btr_pcur_t*	cursor, /*!< in: memory buffer for persistent cursor */
 #ifdef BTR_CUR_HASH_ADAPT
-	rw_lock_t*	ahi_latch,
-				/*!< in: adaptive hash index latch held
-				by the caller, or NULL if none */
+	srw_lock*	ahi_latch,
+				/*!< in: currently held AHI rdlock, or NULL */
 #endif /* BTR_CUR_HASH_ADAPT */
 	const char*	file,	/*!< in: file name */
 	unsigned	line,	/*!< in: line where called */

storage/innobase/include/btr0pcur.ic

@@ -483,9 +483,8 @@ btr_pcur_open_with_no_init_func(
 				that the ahi_latch protects the record! */
 	btr_pcur_t*	cursor, /*!< in: memory buffer for persistent cursor */
 #ifdef BTR_CUR_HASH_ADAPT
-	rw_lock_t*	ahi_latch,
-				/*!< in: adaptive hash index latch held
-				by the caller, or NULL if none */
+	srw_lock*	ahi_latch,
+				/*!< in: currently held AHI rdlock, or NULL */
 #endif /* BTR_CUR_HASH_ADAPT */
 	const char*	file,	/*!< in: file name */
 	unsigned	line,	/*!< in: line where called */

storage/innobase/include/btr0sea.h

@@ -30,7 +30,11 @@ Created 2/17/1996 Heikki Tuuri
 #include "dict0dict.h"
 #ifdef BTR_CUR_HASH_ADAPT
 #include "ha0ha.h"
-#include "sync0sync.h"
+#include "srw_lock.h"
+
+#ifdef UNIV_PFS_RWLOCK
+extern mysql_pfs_key_t btr_search_latch_key;
+#endif /* UNIV_PFS_RWLOCK */
 
 #define btr_search_sys_create() btr_search_sys.create()
 #define btr_search_sys_free() btr_search_sys.free()
@@ -77,7 +81,7 @@ btr_search_guess_on_hash(
 	ulint		mode,
 	ulint		latch_mode,
 	btr_cur_t*	cursor,
-	rw_lock_t*	ahi_latch,
+	srw_lock*	ahi_latch,
 	mtr_t*		mtr);
 
 /** Move or delete hash entries for moved records, usually in a page split.
@@ -109,8 +113,8 @@ void btr_search_drop_page_hash_when_freed(const page_id_t page_id);
 			using btr_cur_search_, and the new record has been
 			inserted next to the cursor.
 @param[in]	ahi_latch	the adaptive hash index latch */
-void
-btr_search_update_hash_node_on_insert(btr_cur_t* cursor, rw_lock_t* ahi_latch);
+void btr_search_update_hash_node_on_insert(btr_cur_t *cursor,
+                                           srw_lock *ahi_latch);
 
 /** Updates the page hash index when a single record is inserted on a page.
 @param[in,out]	cursor		cursor which was positioned to the
@@ -118,8 +122,8 @@ btr_search_update_hash_node_on_insert(btr_cur_t* cursor, rw_lock_t* ahi_latch);
 				and the new record has been inserted next
 				to the cursor
 @param[in]	ahi_latch	the adaptive hash index latch */
-void
-btr_search_update_hash_on_insert(btr_cur_t* cursor, rw_lock_t* ahi_latch);
+void btr_search_update_hash_on_insert(btr_cur_t *cursor,
+                                      srw_lock *ahi_latch);
 
 /** Updates the page hash index when a single record is deleted from a page.
 @param[in]	cursor	cursor which was positioned on the record to delete
@@ -139,23 +143,6 @@ static inline void btr_search_x_unlock_all();
 /** Lock all search latches in shared mode. */
 static inline void btr_search_s_lock_all();
 
-#ifdef UNIV_DEBUG
-/** Check if thread owns all the search latches.
-@param[in]	mode	lock mode check
-@retval true if owns all of them
-@retval false if does not own some of them */
-static inline bool btr_search_own_all(ulint mode);
-
-/** Check if thread owns any of the search latches.
-@param[in]	mode	lock mode check
-@retval true if owns any of them
-@retval false if owns no search latch */
-static inline bool btr_search_own_any(ulint mode);
-
-/** @return whether this thread holds any of the search latches */
-static inline bool btr_search_own_any();
-#endif /* UNIV_DEBUG */
-
 /** Unlock all search latches from shared mode. */
 static inline void btr_search_s_unlock_all();
 
@@ -250,20 +237,20 @@ struct btr_search_sys_t
   struct partition
   {
     /** latches protecting hash_table */
-    rw_lock_t latch;
+    srw_lock latch;
     /** mapping of dtuple_fold() to rec_t* in buf_block_t::frame */
     hash_table_t table;
     /** memory heap for table */
     mem_heap_t *heap;
 
-    char pad[(CPU_LEVEL1_DCACHE_LINESIZE - sizeof(rw_lock_t) -
+    char pad[(CPU_LEVEL1_DCACHE_LINESIZE - sizeof(srw_lock) -
               sizeof(hash_table_t) - sizeof(mem_heap_t)) &
              (CPU_LEVEL1_DCACHE_LINESIZE - 1)];
 
     void init()
     {
       memset((void*) this, 0, sizeof *this);
-      rw_lock_create(btr_search_latch_key, &latch, SYNC_SEARCH_SYS);
+      latch.init(btr_search_latch_key);
     }
 
     void alloc(ulint hash_size)
@@ -285,7 +272,7 @@ struct btr_search_sys_t
 
     void free()
     {
-      rw_lock_free(&latch);
+      latch.destroy();
       if (heap)
         clear();
     }
@@ -309,7 +296,7 @@ struct btr_search_sys_t
   }
 
   /** Get the search latch for the adaptive hash index partition */
-  rw_lock_t *get_latch(const dict_index_t &index) const
+  srw_lock *get_latch(const dict_index_t &index) const
   { return &get_part(index)->latch; }
 
   /** Create and initialize at startup */
@@ -354,10 +341,10 @@ inline ulint dict_index_t::n_ahi_pages() const
 {
   if (!btr_search_enabled)
     return 0;
-  rw_lock_t *latch = &btr_search_sys.get_part(*this)->latch;
-  rw_lock_s_lock(latch);
+  srw_lock *latch= &btr_search_sys.get_part(*this)->latch;
+  latch->rd_lock();
   ulint ref_count= search_info->ref_count;
-  rw_lock_s_unlock(latch);
+  latch->rd_unlock();
   return ref_count;
 }
 

storage/innobase/include/btr0sea.ic

@@ -59,9 +59,6 @@ btr_search_info_update(
 	dict_index_t*	index,	/*!< in: index of the cursor */
 	btr_cur_t*	cursor)	/*!< in: cursor which was just positioned */
 {
-	ut_ad(!btr_search_own_any(RW_LOCK_S));
-	ut_ad(!btr_search_own_any(RW_LOCK_X));
-
 	if (dict_index_is_spatial(index) || !btr_search_enabled) {
 		return;
 	}
@@ -88,7 +85,7 @@ btr_search_info_update(
 static inline void btr_search_x_lock_all()
 {
 	for (ulint i = 0; i < btr_ahi_parts; ++i) {
-		rw_lock_x_lock(&btr_search_sys.parts[i].latch);
+		btr_search_sys.parts[i].latch.wr_lock();
 	}
 }
 
@@ -96,7 +93,7 @@ static inline void btr_search_x_lock_all()
 static inline void btr_search_x_unlock_all()
 {
 	for (ulint i = 0; i < btr_ahi_parts; ++i) {
-		rw_lock_x_unlock(&btr_search_sys.parts[i].latch);
+		btr_search_sys.parts[i].latch.wr_unlock();
 	}
 }
 
@@ -104,7 +101,7 @@ static inline void btr_search_x_unlock_all()
 static inline void btr_search_s_lock_all()
 {
 	for (ulint i = 0; i < btr_ahi_parts; ++i) {
-		rw_lock_s_lock(&btr_search_sys.parts[i].latch);
+		btr_search_sys.parts[i].latch.rd_lock();
 	}
 }
 
@@ -112,49 +109,7 @@ static inline void btr_search_s_lock_all()
 static inline void btr_search_s_unlock_all()
 {
 	for (ulint i = 0; i < btr_ahi_parts; ++i) {
-		rw_lock_s_unlock(&btr_search_sys.parts[i].latch);
-	}
-}
-
-#ifdef UNIV_DEBUG
-/** Check if thread owns all the search latches.
-@param[in]	mode	lock mode check
-@retval true if owns all of them
-@retval false if does not own some of them */
-static inline bool btr_search_own_all(ulint mode)
-{
-	for (ulint i = 0; i < btr_ahi_parts; ++i) {
-		if (!rw_lock_own(&btr_search_sys.parts[i].latch, mode)) {
-			return(false);
-		}
-	}
-	return(true);
-}
-
-/** Check if thread owns any of the search latches.
-@param[in]	mode	lock mode check
-@retval true if owns any of them
-@retval false if owns no search latch */
-static inline bool btr_search_own_any(ulint mode)
-{
-	for (ulint i = 0; i < btr_ahi_parts; ++i) {
-		if (rw_lock_own(&btr_search_sys.parts[i].latch, mode)) {
-			return(true);
-		}
-	}
-	return(false);
-}
-
-/** @return whether this thread holds any of the search latches */
-static inline bool btr_search_own_any()
-{
-	for (ulint i = btr_ahi_parts; i--; ) {
-		if (rw_lock_own_flagged(&btr_search_sys.parts[i].latch,
-					RW_LOCK_FLAG_X | RW_LOCK_FLAG_S)) {
-			return true;
-		}
+		btr_search_sys.parts[i].latch.rd_unlock();
 	}
-	return false;
 }
-#endif /* UNIV_DEBUG */
 #endif /* BTR_CUR_HASH_ADAPT */

storage/innobase/include/rw_lock.h

@@ -36,9 +36,19 @@ class rw_lock
   /** Flag to indicate that write_lock() or write_lock_wait() is pending */
   static constexpr uint32_t WRITER_PENDING= WRITER | WRITER_WAITING;
 
-  /** Start waiting for an exclusive lock. */
-  void write_lock_wait_start()
-  { lock.fetch_or(WRITER_WAITING, std::memory_order_relaxed); }
+  /** Start waiting for an exclusive lock.
+  @return current value of the lock word */
+  uint32_t write_lock_wait_start()
+  { return lock.fetch_or(WRITER_WAITING, std::memory_order_relaxed); }
+  /** Wait for an exclusive lock.
+  @param l the value of the lock word
+  @return whether the exclusive lock was acquired */
+  bool write_lock_wait_try(uint32_t &l)
+  {
+    l= WRITER_WAITING;
+    return lock.compare_exchange_strong(l, WRITER, std::memory_order_acquire,
+                                        std::memory_order_relaxed);
+  }
   /** Try to acquire a shared lock.
   @param l the value of the lock word
   @return whether the lock was acquired */
@@ -58,9 +68,8 @@ class rw_lock
   @return whether the exclusive lock was acquired */
   bool write_lock_poll()
   {
-    auto l= WRITER_WAITING;
-    if (lock.compare_exchange_strong(l, WRITER, std::memory_order_acquire,
-                                     std::memory_order_relaxed))
+    uint32_t l;
+    if (write_lock_wait_try(l))
       return true;
     if (!(l & WRITER_WAITING))
       /* write_lock() must have succeeded for another thread */
@@ -72,12 +81,14 @@ class rw_lock
   /** Default constructor */
   rw_lock() : lock(UNLOCKED) {}
 
-  /** Release a shared lock */
-  void read_unlock()
+  /** Release a shared lock.
+  @return whether any writers may have to be woken up */
+  bool read_unlock()
   {
-    IF_DBUG_ASSERT(auto l=,) lock.fetch_sub(1, std::memory_order_release);
-    DBUG_ASSERT(l & ~WRITER_PENDING); /* at least one read lock */
+    auto l= lock.fetch_sub(1, std::memory_order_release);
+    DBUG_ASSERT(~WRITER_PENDING & l); /* at least one read lock */
     DBUG_ASSERT(!(l & WRITER)); /* no write lock must have existed */
+    return (~WRITER_PENDING & l) == 1;
   }
   /** Release an exclusive lock */
   void write_unlock()
@@ -106,6 +117,9 @@ class rw_lock
     auto l= lock.load(std::memory_order_relaxed);
     return (l & ~WRITER_PENDING) && !(l & WRITER);
   }
+  /** @return whether any lock is being held or waited for by any thread */
+  bool is_locked_or_waiting() const
+  { return lock.load(std::memory_order_relaxed) != 0; }
   /** @return whether any lock is being held by any thread */
   bool is_locked() const
   { return (lock.load(std::memory_order_relaxed) & ~WRITER_WAITING) != 0; }

storage/innobase/include/srw_lock.h

+/*****************************************************************************
+
+Copyright (c) 2020, MariaDB Corporation.
+
+This program is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free Software
+Foundation; version 2 of the License.
+
+This program is distributed in the hope that it will be useful, but WITHOUT
+ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along with
+this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1335 USA
+
+*****************************************************************************/
+
+#pragma once
+#include "univ.i"
+
+#if 0 // defined SAFE_MUTEX
+# define SRW_LOCK_DUMMY /* Use mysql_rwlock_t for debugging purposes */
+#endif
+
+#if defined SRW_LOCK_DUMMY || (!defined _WIN32 && !defined __linux__)
+#else
+# ifdef _WIN32
+#  include <windows.h>
+# else
+#  include "rw_lock.h"
+# endif
+#endif
+
+class srw_lock final
+#if defined __linux__ && !defined SRW_LOCK_DUMMY
+  : protected rw_lock
+#endif
+{
+#if defined SRW_LOCK_DUMMY || (!defined _WIN32 && !defined __linux__)
+  mysql_rwlock_t lock;
+public:
+  void init(mysql_pfs_key_t key) { mysql_rwlock_init(key, &lock); }
+  void destroy() { mysql_rwlock_destroy(&lock); }
+  void rd_lock() { mysql_rwlock_rdlock(&lock); }
+  void rd_unlock() { mysql_rwlock_unlock(&lock); }
+  void wr_lock() { mysql_rwlock_wrlock(&lock); }
+  void wr_unlock() { mysql_rwlock_unlock(&lock); }
+#else
+# ifdef UNIV_PFS_RWLOCK
+  PSI_rwlock *pfs_psi;
+# endif
+# ifdef _WIN32
+  SRWLOCK lock;
+  bool read_trylock() { return TryAcquireSRWLockShared(&lock); }
+  bool write_trylock() { return TryAcquireSRWLockExclusive(&lock); }
+  void read_lock() { AcquireSRWLockShared(&lock); }
+  void write_lock() { AcquireSRWLockExclusive(&lock); }
+# else
+  /** @return pointer to the lock word */
+  rw_lock *word() { return static_cast<rw_lock*>(this); }
+  /** Wait for a read lock.
+  @param l lock word from a failed read_trylock() */
+  void read_lock(uint32_t l);
+  /** Wait for a write lock after a failed write_trylock() */
+  void write_lock();
+# endif
+
+public:
+  void init(mysql_pfs_key_t key)
+  {
+# ifdef UNIV_PFS_RWLOCK
+    pfs_psi= PSI_RWLOCK_CALL(init_rwlock)(key, this);
+# endif
+    IF_WIN(lock= SRWLOCK_INIT, static_assert(4 == sizeof(rw_lock), "ABI"));
+  }
+  void destroy()
+  {
+# ifdef UNIV_PFS_RWLOCK
+    if (pfs_psi)
+    {
+      PSI_RWLOCK_CALL(destroy_rwlock)(pfs_psi);
+      pfs_psi= nullptr;
+    }
+# endif
+    DBUG_ASSERT(!is_locked_or_waiting());
+  }
+  void rd_lock()
+  {
+    IF_WIN(, uint32_t l);
+# ifdef UNIV_PFS_RWLOCK
+    if (read_trylock(IF_WIN(, l)))
+      return;
+    if (pfs_psi)
+    {
+      PSI_rwlock_locker_state state;
+      PSI_rwlock_locker *locker= PSI_RWLOCK_CALL(start_rwlock_rdwait)
+        (&state, pfs_psi, PSI_RWLOCK_READLOCK, __FILE__, __LINE__);
+      read_lock(IF_WIN(, l));
+      if (locker)
+        PSI_RWLOCK_CALL(end_rwlock_rdwait)(locker, 0);
+      return;
+    }
+# endif /* UNIV_PFS_RWLOCK */
+    IF_WIN(read_lock(), if (!read_trylock(l)) read_lock(l));
+  }
+  void wr_lock()
+  {
+# ifdef UNIV_PFS_RWLOCK
+    if (write_trylock())
+      return;
+    if (pfs_psi)
+    {
+      PSI_rwlock_locker_state state;
+      PSI_rwlock_locker *locker= PSI_RWLOCK_CALL(start_rwlock_wrwait)
+        (&state, pfs_psi, PSI_RWLOCK_WRITELOCK, __FILE__, __LINE__);
+      write_lock();
+      if (locker)
+        PSI_RWLOCK_CALL(end_rwlock_rdwait)(locker, 0);
+      return;
+    }
+# endif /* UNIV_PFS_RWLOCK */
+    IF_WIN(, if (!write_trylock())) write_lock();
+  }
+#ifdef _WIN32
+  void rd_unlock()
+  {
+#ifdef UNIV_PFS_RWLOCK
+    if (pfs_psi) PSI_RWLOCK_CALL(unlock_rwlock)(pfs_psi);
+#endif
+    ReleaseSRWLockShared(&lock);
+  }
+  void wr_unlock()
+  {
+#ifdef UNIV_PFS_RWLOCK
+    if (pfs_psi) PSI_RWLOCK_CALL(unlock_rwlock)(pfs_psi);
+#endif
+    ReleaseSRWLockExclusive(&lock);
+  }
+#else
+  void rd_unlock();
+  void wr_unlock();
+#endif
+#endif
+};

storage/innobase/include/sync0sync.h

@@ -87,7 +87,6 @@ extern mysql_pfs_key_t	read_view_mutex_key;
 #ifdef UNIV_PFS_RWLOCK
 /* Following are rwlock keys used to register with MySQL
 performance schema */
-extern	mysql_pfs_key_t btr_search_latch_key;
 extern	mysql_pfs_key_t	dict_operation_lock_key;
 extern	mysql_pfs_key_t	fil_space_latch_key;
 extern	mysql_pfs_key_t	fts_cache_rw_lock_key;

storage/innobase/include/sync0types.h

@@ -191,8 +191,6 @@ enum latch_level_t {
 	SYNC_POOL,
 	SYNC_POOL_MANAGER,
 
-	SYNC_SEARCH_SYS,
-
 	SYNC_WORK_QUEUE,
 
 	SYNC_FTS_TOKENIZE,
@@ -294,7 +292,6 @@ enum latch_id_t {
 	LATCH_ID_ROW_DROP_LIST,
 	LATCH_ID_INDEX_ONLINE_LOG,
 	LATCH_ID_WORK_QUEUE,
-	LATCH_ID_BTR_SEARCH,
 	LATCH_ID_BUF_BLOCK_LOCK,
 	LATCH_ID_BUF_BLOCK_DEBUG,
 	LATCH_ID_DICT_OPERATION,

storage/innobase/row/row0sel.cc

@@ -3857,15 +3857,15 @@ row_sel_try_search_shortcut_for_mysql(
 	ut_ad(dict_index_is_clust(index));
 	ut_ad(!prebuilt->templ_contains_blob);
 
-	rw_lock_t* ahi_latch = btr_search_sys.get_latch(*index);
-	rw_lock_s_lock(ahi_latch);
+	srw_lock* ahi_latch = btr_search_sys.get_latch(*index);
+	ahi_latch->rd_lock();
 	btr_pcur_open_with_no_init(index, search_tuple, PAGE_CUR_GE,
 				   BTR_SEARCH_LEAF, pcur, ahi_latch, mtr);
 	rec = btr_pcur_get_rec(pcur);
 
 	if (!page_rec_is_user_rec(rec) || rec_is_metadata(rec, *index)) {
 retry:
-		rw_lock_s_unlock(ahi_latch);
+		ahi_latch->rd_unlock();
 		return(SEL_RETRY);
 	}
 
@@ -3875,7 +3875,7 @@ row_sel_try_search_shortcut_for_mysql(
 
 	if (btr_pcur_get_up_match(pcur) < dtuple_get_n_fields(search_tuple)) {
 exhausted:
-		rw_lock_s_unlock(ahi_latch);
+		ahi_latch->rd_unlock();
 		return(SEL_EXHAUSTED);
 	}
 
@@ -3899,7 +3899,7 @@ row_sel_try_search_shortcut_for_mysql(
 
 	*out_rec = rec;
 
-	rw_lock_s_unlock(ahi_latch);
+	ahi_latch->rd_unlock();
 	return(SEL_FOUND);
 }
 #endif /* BTR_CUR_HASH_ADAPT */

storage/innobase/srv/srv0srv.cc

@@ -924,13 +924,13 @@ srv_printf_innodb_monitor(
 #ifdef BTR_CUR_HASH_ADAPT
 	for (ulint i = 0; i < btr_ahi_parts && btr_search_enabled; ++i) {
 		const auto part= &btr_search_sys.parts[i];
-		rw_lock_s_lock(&part->latch);
+		part->latch.rd_lock();
 		ut_ad(part->heap->type == MEM_HEAP_FOR_BTR_SEARCH);
 		fprintf(file, "Hash table size " ULINTPF
 			", node heap has " ULINTPF " buffer(s)\n",
 			part->table.n_cells,
 			part->heap->base.count - !part->heap->free_block);
-		rw_lock_s_unlock(&part->latch);
+		part->latch.rd_unlock();
 	}
 
 	fprintf(file,
@@ -1065,14 +1065,14 @@ srv_export_innodb_status(void)
 	ulint mem_adaptive_hash = 0;
 	for (ulong i = 0; i < btr_ahi_parts; i++) {
 		const auto part= &btr_search_sys.parts[i];
-		rw_lock_s_lock(&part->latch);
+		part->latch.rd_lock();
 		if (part->heap) {
 			ut_ad(part->heap->type == MEM_HEAP_FOR_BTR_SEARCH);
 
 			mem_adaptive_hash += mem_heap_get_size(part->heap)
 				+ part->table.n_cells * sizeof(hash_cell_t);
 		}
-		rw_lock_s_unlock(&part->latch);
+		part->latch.rd_unlock();
 	}
 	export_vars.innodb_mem_adaptive_hash = mem_adaptive_hash;
 #endif

storage/innobase/sync/srw_lock_futex.cc

+/*****************************************************************************
+
+Copyright (c) 2020, MariaDB Corporation.
+
+This program is free software; you can redistribute it and/or modify it under
+the terms of the GNU General Public License as published by the Free Software
+Foundation; version 2 of the License.
+
+This program is distributed in the hope that it will be useful, but WITHOUT
+ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
+FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
+
+You should have received a copy of the GNU General Public License along with
+this program; if not, write to the Free Software Foundation, Inc.,
+51 Franklin Street, Fifth Floor, Boston, MA 02110-1335 USA
+
+*****************************************************************************/
+
+#ifndef __linux__
+# error "This file is for Linux only"
+#endif
+
+#include "srw_lock.h"
+
+#ifdef SRW_LOCK_DUMMY
+/* Work around a potential build failure by preventing an empty .o file */
+int srw_lock_dummy_function() { return 0; }
+#else
+# include <linux/futex.h>
+# include <sys/syscall.h>
+
+# include "srv0srv.h"
+
+/** Wait for a read lock.
+@param lock word value from a failed read_trylock() */
+void srw_lock::read_lock(uint32_t l)
+{
+  do
+  {
+    if (l == WRITER_WAITING)
+    wake_writer:
+      syscall(SYS_futex, word(), FUTEX_WAKE_PRIVATE, 1, nullptr, nullptr, 0);
+    else
+      for (auto spin= srv_n_spin_wait_rounds; spin; spin--)
+      {
+        ut_delay(srv_spin_wait_delay);
+        if (read_trylock(l))
+          return;
+        else if (l == WRITER_WAITING)
+          goto wake_writer;
+      }
+
+    syscall(SYS_futex, word(), FUTEX_WAIT_PRIVATE, l, nullptr, nullptr, 0);
+  }
+  while (!read_trylock(l));
+}
+
+/** Wait for a write lock after a failed write_trylock() */
+void srw_lock::write_lock()
+{
+  for (;;)
+  {
+    uint32_t l= write_lock_wait_start();
+    /* We are the first writer to be granted the lock. Spin for a while. */
+    for (auto spin= srv_n_spin_wait_rounds; spin; spin--)
+    {
+      if (write_lock_wait_try(l))
+        return;
+      if (!(l & WRITER_WAITING))
+        l= write_lock_wait_start();
+      ut_delay(srv_spin_wait_delay);
+    }
+
+    if (write_lock_wait_try(l))
+      return;
+
+    if (!(l & WRITER_WAITING))
+    {
+      if (l == UNLOCKED && write_trylock())
+        return;
+      l= write_lock_wait_start() | WRITER_WAITING;
+    }
+    else
+      DBUG_ASSERT(~WRITER_WAITING & l);
+
+    syscall(SYS_futex, word(), FUTEX_WAIT_PRIVATE, l, nullptr, nullptr, 0);
+  }
+}
+
+void srw_lock::rd_unlock()
+{
+#ifdef UNIV_PFS_RWLOCK
+  if (pfs_psi) PSI_RWLOCK_CALL(unlock_rwlock)(pfs_psi);
+#endif
+  if (read_unlock())
+    syscall(SYS_futex, word(), FUTEX_WAKE_PRIVATE, 1, nullptr, nullptr, 0);
+}
+
+void srw_lock::wr_unlock()
+{
+#ifdef UNIV_PFS_RWLOCK
+  if (pfs_psi) PSI_RWLOCK_CALL(unlock_rwlock)(pfs_psi);
+#endif
+  write_unlock();
+  syscall(SYS_futex, word(), FUTEX_WAKE_PRIVATE, INT_MAX, nullptr, nullptr, 0);
+}
+#endif

storage/innobase/sync/sync0debug.cc

@@ -455,7 +455,6 @@ LatchDebug::LatchDebug()
 	LEVEL_MAP_INSERT(SYNC_ANY_LATCH);
 	LEVEL_MAP_INSERT(SYNC_POOL);
 	LEVEL_MAP_INSERT(SYNC_POOL_MANAGER);
-	LEVEL_MAP_INSERT(SYNC_SEARCH_SYS);
 	LEVEL_MAP_INSERT(SYNC_WORK_QUEUE);
 	LEVEL_MAP_INSERT(SYNC_FTS_TOKENIZE);
 	LEVEL_MAP_INSERT(SYNC_FTS_OPTIMIZE);
@@ -732,7 +731,6 @@ LatchDebug::check_order(
 	case SYNC_FTS_OPTIMIZE:
 	case SYNC_FTS_CACHE:
 	case SYNC_FTS_CACHE_INIT:
-	case SYNC_SEARCH_SYS:
 	case SYNC_LOCK_SYS:
 	case SYNC_LOCK_WAIT_SYS:
 	case SYNC_RW_TRX_HASH_ELEMENT:
@@ -1304,8 +1302,6 @@ sync_latch_meta_init()
 	LATCH_ADD_MUTEX(WORK_QUEUE, SYNC_WORK_QUEUE, PFS_NOT_INSTRUMENTED);
 
 	// Add the RW locks
-	LATCH_ADD_RWLOCK(BTR_SEARCH, SYNC_SEARCH_SYS, btr_search_latch_key);
-
 	LATCH_ADD_RWLOCK(BUF_BLOCK_LOCK, SYNC_LEVEL_VARYING,
 			 PFS_NOT_INSTRUMENTED);
 

storage/innobase/sync/sync0sync.cc

@@ -81,7 +81,6 @@ mysql_pfs_key_t	rw_trx_hash_element_mutex_key;
 mysql_pfs_key_t	read_view_mutex_key;
 #endif /* UNIV_PFS_MUTEX */
 #ifdef UNIV_PFS_RWLOCK
-mysql_pfs_key_t	btr_search_latch_key;
 mysql_pfs_key_t	dict_operation_lock_key;
 mysql_pfs_key_t	index_tree_rw_lock_key;
 mysql_pfs_key_t	index_online_log_key;