InnoDB redo log IO methodsbb-10.5-svoj-MDEV-17084-redo

normal - conventional IO using read()/write()/fdatasync()
mmap - memory mapped IO, expected to be faster but less secure

5 files changed, 193 insertions, 27 deletions

diff --git a/mysql-test/suite/sys_vars/r/sysvars_innodb.result b/mysql-test/suite/sys_vars/r/sysvars_innodb.result
index 5f4b031ada0..2ddca40d4d4 100644
--- a/mysql-test/suite/sys_vars/r/sysvars_innodb.result
+++ b/mysql-test/suite/sys_vars/r/sysvars_innodb.result
@@ -1257,6 +1257,18 @@ NUMERIC_BLOCK_SIZE	NULL
 ENUM_VALUE_LIST	NULL
 READ_ONLY	YES
 COMMAND_LINE_ARGUMENT	REQUIRED
+VARIABLE_NAME	INNODB_LOG_IO_METHOD
+SESSION_VALUE	NULL
+DEFAULT_VALUE	normal
+VARIABLE_SCOPE	GLOBAL
+VARIABLE_TYPE	ENUM
+VARIABLE_COMMENT	InnoDB redo log IO method: normal (default), mmap.
+NUMERIC_MIN_VALUE	NULL
+NUMERIC_MAX_VALUE	NULL
+NUMERIC_BLOCK_SIZE	NULL
+ENUM_VALUE_LIST	normal,mmap
+READ_ONLY	YES
+COMMAND_LINE_ARGUMENT	REQUIRED
 VARIABLE_NAME	INNODB_LOG_OPTIMIZE_DDL
 SESSION_VALUE	NULL
 DEFAULT_VALUE	ON
diff --git a/storage/innobase/CMakeLists.txt b/storage/innobase/CMakeLists.txt
index 593567bb097..d21ea39c11a 100644
--- a/storage/innobase/CMakeLists.txt
+++ b/storage/innobase/CMakeLists.txt
@@ -155,7 +155,8 @@ MYSQL_ADD_PLUGIN(innobase ${INNOBASE_SOURCES} STORAGE_ENGINE
 	${CRC32_LIBRARY}
 	${NUMA_LIBRARY}
 	${LIBSYSTEMD}
-	${LINKER_SCRIPT})
+	${LINKER_SCRIPT}
+	${LIBPMEM})
 
 IF(NOT TARGET innobase)
   RETURN()
diff --git a/storage/innobase/handler/ha_innodb.cc b/storage/innobase/handler/ha_innodb.cc
index 56dc0087045..484de2f840e 100644
--- a/storage/innobase/handler/ha_innodb.cc
+++ b/storage/innobase/handler/ha_innodb.cc
@@ -406,6 +406,21 @@ static TYPELIB innodb_change_buffering_typelib = {
 	NULL
 };
 
+static const char *innodb_log_io_method_names[]=
+{
+  "normal",
+  "mmap",
+  NullS
+};
+
+static TYPELIB innodb_log_io_method_typelib=
+{
+  array_elements(innodb_log_io_method_names) - 1,
+  "innodb_log_io_method_typelib",
+  innodb_log_io_method_names,
+  NULL
+};
+
 /** Retrieve the FTS Relevance Ranking result for doc with doc_id
 of m_prebuilt->fts_doc_id
 @param[in,out]	fts_hdl	FTS handler
@@ -19413,6 +19428,12 @@ static MYSQL_SYSVAR_BOOL(log_optimize_ddl, innodb_log_optimize_ddl,
   " allows concurrent backup.",
   NULL, NULL, TRUE);
 
+static MYSQL_SYSVAR_ENUM(log_io_method, innodb_log_io_method,
+  PLUGIN_VAR_RQCMDARG | PLUGIN_VAR_READONLY,
+  "InnoDB redo log IO method: normal (default), mmap.",
+  NULL, NULL, 0,
+  &innodb_log_io_method_typelib);
+
 static MYSQL_SYSVAR_ULONG(autoextend_increment,
   sys_tablespace_auto_extend_increment,
   PLUGIN_VAR_RQCMDARG,
@@ -20303,6 +20324,7 @@ static struct st_mysql_sys_var* innobase_system_variables[]= {
   MYSQL_SYSVAR(log_group_home_dir),
   MYSQL_SYSVAR(log_compressed_pages),
   MYSQL_SYSVAR(log_optimize_ddl),
+  MYSQL_SYSVAR(log_io_method),
   MYSQL_SYSVAR(max_dirty_pages_pct),
   MYSQL_SYSVAR(max_dirty_pages_pct_lwm),
   MYSQL_SYSVAR(adaptive_flushing_lwm),
diff --git a/storage/innobase/include/log0log.h b/storage/innobase/include/log0log.h
index 4e619a203d9..cf19eaf297a 100644
--- a/storage/innobase/include/log0log.h
+++ b/storage/innobase/include/log0log.h
@@ -40,6 +40,7 @@ Created 12/9/1995 Heikki Tuuri
 #include "os0file.h"
 #include "span.h"
 #include <atomic>
+#include <memory>
 
 using st_::span;
 
@@ -449,6 +450,7 @@ typedef ib_mutex_t	LogSysMutex;
 typedef ib_mutex_t	FlushOrderMutex;
 
 extern my_bool	srv_read_only_mode;
+extern ulong innodb_log_io_method;
 
 /** Redo log buffer */
 struct log_t{
@@ -516,6 +518,22 @@ struct log_t{
 
   /** Log files. Protected by mutex or write_mutex. */
   struct files {
+    class file_io
+    {
+    protected:
+      bool durable_writes;
+    public:
+      virtual ~file_io() {}
+      virtual dberr_t open(const char *path)= 0;
+      virtual dberr_t close()= 0;
+      virtual dberr_t read(os_offset_t offset, span<byte> buf)= 0;
+      virtual dberr_t write(const char *path, os_offset_t offset,
+                            span<byte> buf)= 0;
+      virtual dberr_t flush_data_only()= 0;
+
+      bool writes_are_durable() const { return durable_writes; }
+    };
+
     /** number of files */
     ulint				n_files;
     /** format of the redo log: e.g., FORMAT_10_4 */
@@ -530,14 +548,14 @@ struct log_t{
     lsn_t				lsn;
     /** the byte offset of the above lsn */
     lsn_t				lsn_offset;
+    /** file descriptors for all log files */
+    std::vector<std::unique_ptr<file_io>> files;
 
   public:
     /** used only in recovery: recovery scan succeeded up to this
     lsn in this log group */
     lsn_t				scanned_lsn;
 
-    /** file descriptors for all log files */
-    std::vector<pfs_os_file_t> files;
     /** file names for all log files */
     std::vector<std::string> file_names;
 
@@ -553,6 +571,11 @@ struct log_t{
     @param[in]	total_offset	offset in log files treated as a single file
     @param[in]	buf		buffer from which to write */
     void write(size_t total_offset, span<byte> buf);
+    /** checks whether flush_data_only() is needed to make data persistend */
+    bool writes_are_durable() const
+    {
+      return files.front()->writes_are_durable();
+    }
     /** flushes OS page cache (excluding metadata!) for all log files */
     void flush_data_only();
     /** closes all log files */
diff --git a/storage/innobase/log/log0log.cc b/storage/innobase/log/log0log.cc
index d0b325c05aa..043de52d737 100644
--- a/storage/innobase/log/log0log.cc
+++ b/storage/innobase/log/log0log.cc
@@ -84,6 +84,8 @@ reduce the size of the log.
 /** Redo log system */
 log_t	log_sys;
 
+ulong innodb_log_io_method;
+
 /* Next log block number to do dummy record filling if no log records written
 for a while */
 static ulint		next_lbn_to_pad = 0;
@@ -587,6 +589,112 @@ void log_t::create()
   }
 }
 
+
+class file_os_io final: public log_t::files::file_io
+{
+  pfs_os_file_t fd;
+public:
+  dberr_t open(const char *path) final
+  {
+    bool success;
+    fd= os_file_create(innodb_log_file_key, path,
+                       OS_FILE_OPEN | OS_FILE_ON_ERROR_NO_EXIT,
+                       OS_FILE_NORMAL, OS_LOG_FILE,
+                       srv_read_only_mode, &success);
+    durable_writes= srv_file_flush_method == SRV_O_DSYNC;
+    return success ? DB_SUCCESS : DB_ERROR;
+  }
+  dberr_t close() final { return os_file_close(fd) ? DB_SUCCESS : DB_ERROR; }
+  dberr_t read(os_offset_t offset, span<byte> buf) final
+  {
+    return os_file_read(IORequestRead, fd, buf.data(), offset, buf.size());
+  }
+  dberr_t write(const char *path, os_offset_t offset, span<byte> buf) final
+  {
+    return os_file_write(IORequestWrite, path, fd, buf.data(), offset,
+                         buf.size());
+  }
+  dberr_t flush_data_only() final
+  {
+    return os_file_flush_data(fd) ? DB_SUCCESS : DB_ERROR;
+  }
+};
+
+
+#ifdef HAVE_PMEM
+#include <libpmem.h>
+#endif
+class file_mmap_io final: public log_t::files::file_io
+{
+  File fd;
+  void *addr;
+  size_t length;
+public:
+  dberr_t open(const char *path) final
+  {
+    fd= mysql_file_open(innodb_log_file_key, path,
+                        srv_read_only_mode ? O_RDONLY : O_RDWR, MYF(MY_WME));
+    if (fd >= 0)
+    {
+      MY_STAT sb;
+      if (!mysql_file_fstat(fd, &sb, MYF(0)))
+      {
+        int prot= srv_read_only_mode ? PROT_READ : PROT_READ | PROT_WRITE;
+        length= sb.st_size;
+        addr= my_mmap(0, length, prot, MAP_SHARED_VALIDATE | MAP_SYNC, fd, 0);
+        if (addr != MAP_FAILED)
+        {
+#ifdef HAVE_PMEM
+          durable_writes= true;
+          ib::info() << "The redo log file is located on a DAX storage. "
+                        "Writes are durable, sync disabled.";
+#else
+          durable_writes= false;
+          ib::info() << "The redo log file is located on a DAX storage, "
+                        "but persistent memory features were disabled "
+                        "(WITH_PMEM=OFF). Page cache is bypassed, sync is "
+                        "required to make writes durable.";
+#endif
+          return DB_SUCCESS;
+        }
+        addr= my_mmap(0, length, prot, MAP_SHARED, fd, 0);
+        if (addr != MAP_FAILED)
+        {
+          durable_writes= false;
+          return DB_SUCCESS;
+        }
+      }
+      mysql_file_close(fd, MYF(MY_WME));
+    }
+    return DB_ERROR;
+  }
+  dberr_t close() final
+  {
+    int err= my_munmap(addr, length);
+    return (!mysql_file_close(fd, MYF(MY_WME)) && !err) ? DB_SUCCESS : DB_ERROR;
+  }
+  dberr_t read(os_offset_t offset, span<byte> buf) final
+  {
+    memcpy(buf.data(), (char*) addr + offset, buf.size());
+    return DB_SUCCESS;
+  }
+  dberr_t write(const char *, os_offset_t offset, span<byte> buf) final
+  {
+#ifdef HAVE_PMEM
+    pmem_memcpy_persist((char*) addr + offset, buf.data(), buf.size());
+#else
+    memcpy((char*) addr + offset, buf.data(), buf.size());
+#endif
+    return DB_SUCCESS;
+  }
+  dberr_t flush_data_only() final
+  {
+    ut_ad(!durable_writes);
+    return my_msync(fd, addr, length, MS_SYNC) ? DB_ERROR : DB_SUCCESS;
+  }
+};
+
+
 void log_t::files::set_file_names(std::vector<std::string> names)
 {
   file_names= std::move(names);
@@ -598,15 +706,18 @@ void log_t::files::open_files()
   files.reserve(file_names.size());
   for (const auto &name : file_names)
   {
-    bool success;
-    files.push_back(os_file_create(innodb_log_file_key, name.c_str(),
-                                   OS_FILE_OPEN | OS_FILE_ON_ERROR_NO_EXIT,
-                                   OS_FILE_NORMAL, OS_LOG_FILE,
-                                   srv_read_only_mode, &success));
-    if (!success)
+    file_io *io;
+
+    switch (innodb_log_io_method)
     {
-      ib::fatal() << "os_file_create(" << name << ") failed";
+      case 1: io= new file_mmap_io; break;
+      default: io= new file_os_io;
     }
+    ut_a(io);
+
+    if (io->open(name.c_str()))
+      ib::fatal() << "open(" << name << ") failed";
+    files.emplace_back(io);
   }
 }
 
@@ -617,12 +728,8 @@ void log_t::files::read(size_t total_offset, span<byte> buf)
   const size_t file_idx= total_offset / static_cast<size_t>(file_size);
   const size_t offset= total_offset % static_cast<size_t>(file_size);
 
-  if (const dberr_t err= os_file_read(IORequestRead, files[file_idx],
-                                      buf.data(), offset, buf.size()))
-  {
-    ib::fatal() << "os_file_read(" << file_names[file_idx] << ") returned "
-                << err;
-  }
+  if (const dberr_t err= files[file_idx]->read(offset, buf))
+    ib::fatal() << "read(" << file_names[file_idx] << ") returned " << err;
 }
 
 void log_t::files::write(size_t total_offset, span<byte> buf)
@@ -632,13 +739,9 @@ void log_t::files::write(size_t total_offset, span<byte> buf)
   const size_t file_idx= total_offset / static_cast<size_t>(file_size);
   const size_t offset= total_offset % static_cast<size_t>(file_size);
 
-  if (const dberr_t err=
-          os_file_write(IORequestWrite, file_names[file_idx].c_str(),
-                        files[file_idx], buf.data(), offset, buf.size()))
-  {
-    ib::fatal() << "os_file_write(" << file_names[file_idx] << ") returned "
-                << err;
-  }
+  if (const dberr_t err= files[file_idx]->write(file_names[file_idx].c_str(),
+                                                offset, buf))
+    ib::fatal() << "write(" << file_names[file_idx] << ") returned " << err;
 }
 
 void log_t::files::flush_data_only()
@@ -648,10 +751,10 @@ void log_t::files::flush_data_only()
   log_sys.pending_flushes.fetch_add(1, std::memory_order_acquire);
   for (auto it= files.begin(), end= files.end(); it != end; ++it)
   {
-    if (!os_file_flush_data(*it))
+    if ((*it)->flush_data_only())
     {
       const auto idx= std::distance(files.begin(), it);
-      ib::fatal() << "os_file_flush_data(" << file_names[idx] << ") failed";
+      ib::fatal() << "flush_data_only(" << file_names[idx] << ") failed";
     }
   }
   log_sys.pending_flushes.fetch_sub(1, std::memory_order_release);
@@ -662,10 +765,10 @@ void log_t::files::close_files()
 {
   for (auto it= files.begin(), end= files.end(); it != end; ++it)
   {
-    if (!os_file_close(*it))
+    if ((*it)->close())
     {
       const auto idx= std::distance(files.begin(), it);
-      ib::fatal() << "os_file_close(" << file_names[idx] << ") failed";
+      ib::fatal() << "close(" << file_names[idx] << ") failed";
     }
   }
   files.clear();
@@ -934,6 +1037,11 @@ void log_write_up_to(lsn_t lsn, bool flush_to_disk, bool rotate_key)
 		return;
 	}
 
+	/* FIXME!!! */
+	if (log_sys.log.writes_are_durable()) {
+		flush_to_disk= false;
+	}
+
 loop:
 	ut_ad(++loop_count < 128);