Add support for direct I/O, with the config "direct_io=(data,log)".

--HG--
extra : rebase_source : c9ece039e84f16dcb4cf4fe8419e6de04f970a13

8 files changed, 66 insertions, 32 deletions

diff --git a/dist/api_data.py b/dist/api_data.py
index c3fb21366ba..e689b28446b 100644
--- a/dist/api_data.py
+++ b/dist/api_data.py
@@ -328,6 +328,10 @@ methods = {
 	Config('create', 'false', r'''
 		create the database if it does not exist''',
 		type='boolean'),
+	Config('direct_io', '', r'''
+		Use \c O_DIRECT to access files.  Options are given as a
+		list, such as <code>"direct_io=[data]"</code>''',
+		type='list', choices=['data', 'log']),
 	Config('home_environment', 'false', r'''
 		use the \c WIREDTIGER_HOME environment variable for naming
 		unless the process is running with special privileges.
@@ -394,6 +398,7 @@ flags = {
 ###################################################
 # Internal routine flag declarations
 ###################################################
+	'direct_io' : [ 'DIRECTIO_DATA', 'DIRECTIO_LOG' ],
 	'page_free' : [ 'PAGE_FREE_IGNORE_DISK' ],
 	'rec_evict' : [ 'REC_SINGLE' ],
 	'verbose' : [
diff --git a/src/block/block_open.c b/src/block/block_open.c
index bdab374ce78..eba5de0c1a7 100644
--- a/src/block/block_open.c
+++ b/src/block/block_open.c
@@ -21,7 +21,7 @@ __wt_block_truncate(WT_SESSION_IMPL *session, const char *filename)
 	int ret;
 
 	/* Open the underlying file handle. */
-	WT_RET(__wt_open(session, filename, 0, &fh));
+	WT_RET(__wt_open(session, filename, 0, 1, &fh));
 
 	/* Truncate the file. */
 	WT_ERR(__wt_ftruncate(session, fh, (off_t)0));
@@ -54,7 +54,7 @@ __wt_block_create(WT_SESSION_IMPL *session, const char *filename)
 		    filename);
 
 	/* Open the underlying file handle. */
-	WT_RET(__wt_open(session, filename, 1, &fh));
+	WT_RET(__wt_open(session, filename, 1, 1, &fh));
 
 	/* Write out the file's meta-data. */
 	ret = __wt_desc_init(session, fh);
@@ -96,7 +96,7 @@ __wt_block_open(WT_SESSION_IMPL *session, const char *filename,
 	__wt_block_freelist_open(session, block);
 
 	/* Open the underlying file handle. */
-	WT_ERR(__wt_open(session, filename, 1, &block->fh));
+	WT_ERR(__wt_open(session, filename, 1, 1, &block->fh));
 
 	/* Get the allocation size. */
 	WT_ERR(__wt_config_getones(session, config, "allocation_size", &cval));
diff --git a/src/config/config_def.c b/src/config/config_def.c
index 416c4be9797..7d5acc492df 100644
--- a/src/config/config_def.c
+++ b/src/config/config_def.c
@@ -266,19 +266,20 @@ __wt_confchk_table_meta =
 
 const char *
 __wt_confdfl_wiredtiger_open =
-    "cache_size=100MB,create=false,error_prefix="",eviction_target=80,"
-    "eviction_trigger=95,extensions=(),hazard_max=30,home_environment=false,"
-    "home_environment_priv=false,logging=false,multiprocess=false,"
-    "session_max=50,transactional=false,verbose=()";
+    "cache_size=100MB,create=false,direct_io=(),error_prefix="","
+    "eviction_target=80,eviction_trigger=95,extensions=(),hazard_max=30,"
+    "home_environment=false,home_environment_priv=false,logging=false,"
+    "multiprocess=false,session_max=50,transactional=false,verbose=()";
 
 const char *
 __wt_confchk_wiredtiger_open =
     "cache_size=(type=int,min=1MB,max=10TB),create=(type=boolean),"
-    "error_prefix=(),eviction_target=(type=int,min=10,max=99),"
-    "eviction_trigger=(type=int,min=10,max=99),extensions=(type=list),"
-    "hazard_max=(type=int,min=15),home_environment=(type=boolean),"
-    "home_environment_priv=(type=boolean),logging=(type=boolean),"
-    "multiprocess=(type=boolean),session_max=(type=int,min=1),"
-    "transactional=(type=boolean),verbose=(type=list,choices=[\"block\","
-    "\"evict\",\"evictserver\",\"fileops\",\"hazard\",\"mutex\",\"read\","
-    "\"readserver\",\"reconcile\",\"salvage\",\"verify\",\"write\"])";
+    "direct_io=(type=list,choices=[\"data\",\"log\"]),error_prefix=(),"
+    "eviction_target=(type=int,min=10,max=99),eviction_trigger=(type=int,"
+    "min=10,max=99),extensions=(type=list),hazard_max=(type=int,min=15),"
+    "home_environment=(type=boolean),home_environment_priv=(type=boolean),"
+    "logging=(type=boolean),multiprocess=(type=boolean),session_max=(type=int"
+    ",min=1),transactional=(type=boolean),verbose=(type=list,"
+    "choices=[\"block\",\"evict\",\"evictserver\",\"fileops\",\"hazard\","
+    "\"mutex\",\"read\",\"readserver\",\"reconcile\",\"salvage\",\"verify\","
+    "\"write\"])";
diff --git a/src/conn/conn_api.c b/src/conn/conn_api.c
index 47d3f71f2d3..00ff16cbe08 100644
--- a/src/conn/conn_api.c
+++ b/src/conn/conn_api.c
@@ -346,9 +346,9 @@ wiredtiger_open(const char *home, WT_EVENT_HANDLER *event_handler,
 		__conn_open_session
 	};
 	static struct {
-		const char *vname;
-		uint32_t vflag;
-	} *vt, verbtypes[] = {
+		const char *name;
+		uint32_t flag;
+	} *ft, verbtypes[] = {
 		{ "block",	WT_VERB_block },
 		{ "evict",	WT_VERB_evict },
 		{ "evictserver",WT_VERB_evictserver },
@@ -362,6 +362,10 @@ wiredtiger_open(const char *home, WT_EVENT_HANDLER *event_handler,
 		{ "verify",	WT_VERB_verify },
 		{ "write",	WT_VERB_write },
 		{ NULL, 0 }
+	}, directio_types[] = {
+		{ "data",	WT_DIRECTIO_DATA },
+		{ "log",	WT_DIRECTIO_LOG },
+		{ NULL, 0 }
 	};
 	WT_CONFIG subconfig;
 	WT_CONFIG_ITEM cval, skey, sval;
@@ -444,21 +448,30 @@ wiredtiger_open(const char *home, WT_EVENT_HANDLER *event_handler,
 	conn->verbose = 0;
 #ifdef HAVE_VERBOSE
 	WT_ERR(__wt_config_gets(session, cfg, "verbose", &cval));
-	for (vt = verbtypes; vt->vname != NULL; vt++) {
-		WT_ERR(__wt_config_subinit(session, &subconfig, &cval));
-		skey.str = vt->vname;
-		skey.len = strlen(vt->vname);
-		ret = __wt_config_getraw(&subconfig, &skey, &sval);
-		if (ret == 0 && sval.val)
-			FLD_SET(conn->verbose, vt->vflag);
-		else if (ret != WT_NOTFOUND)
+	for (ft = verbtypes; ft->name != NULL; ft++) {
+		ret = __wt_config_subgets(session, &cval, ft->name, &sval);
+		if (ret == 0) {
+			if (sval.val)
+				FLD_SET(conn->verbose, ft->flag);
+		} else if (ret != WT_NOTFOUND)
 			goto err;
 	}
 #endif
 
 	WT_ERR(__wt_config_gets(session, cfg, "logging", &cval));
 	if (cval.val != 0)
-		WT_ERR(__wt_open(session, WT_LOG_FILENAME, 1, &conn->log_fh));
+		WT_ERR(__wt_open(session, WT_LOG_FILENAME, 1, 0, &conn->log_fh));
+
+	/* Configure direct I/O flags. */
+	WT_ERR(__wt_config_gets(session, cfg, "direct_io", &cval));
+	for (ft = directio_types; ft->name != NULL; ft++) {
+		ret = __wt_config_subgets(session, &cval, ft->name, &sval);
+		if (ret == 0) {
+			if (sval.val)
+				FLD_SET(conn->direct_io, ft->flag);
+		} else if (ret != WT_NOTFOUND)
+			goto err;
+	}
 
 	/* Load any extensions referenced in the config. */
 	WT_ERR(__wt_config_gets(session, cfg, "extensions", &cval));
@@ -578,7 +591,7 @@ __conn_single(WT_CONNECTION_IMPL *conn, const char **cfg)
 	 */
 	WT_RET(__wt_config_gets(session, cfg, "create", &cval));
 	WT_RET(__wt_open(session,
-	    WT_FLAGFILE, cval.val == 0 ? 0 : 1, &conn->lock_fh));
+	    WT_FLAGFILE, cval.val == 0 ? 0 : 1, 0, &conn->lock_fh));
 
 	/*
 	 * Lock a byte of the file: if we don't get the lock, some other process
@@ -670,7 +683,7 @@ __conn_config(WT_CONNECTION_IMPL *conn, const char **cfg, WT_ITEM **cbufp)
 		return (0);
 
 	/* Open the configuration file. */
-	WT_RET(__wt_open(session, WT_CONFIGFILE, 0, &fh));
+	WT_RET(__wt_open(session, WT_CONFIGFILE, 0, 0, &fh));
 	WT_ERR(__wt_filesize(session, fh, &size));
 	if (size == 0)
 		goto err;
diff --git a/src/include/api.h b/src/include/api.h
index 571c40eca43..683994911b0 100644
--- a/src/include/api.h
+++ b/src/include/api.h
@@ -202,6 +202,7 @@ struct __wt_connection_impl {
 	FILE *msgfile;
 	void (*msgcall)(const WT_CONNECTION_IMPL *, const char *);
 
+	uint32_t direct_io;
 	uint32_t verbose;
 
 	uint32_t flags;
@@ -266,6 +267,8 @@ extern WT_PROCESS __wt_process;
  * DO NOT EDIT: automatically built by dist/api_flags.py.
  * API flags section: BEGIN
  */
+#define	WT_DIRECTIO_DATA				0x00000002
+#define	WT_DIRECTIO_LOG					0x00000001
 #define	WT_PAGE_FREE_IGNORE_DISK			0x00000001
 #define	WT_REC_SINGLE					0x00000001
 #define	WT_SERVER_RUN					0x00000001
diff --git a/src/include/extern.h b/src/include/extern.h
index c0b7bb641ee..308c62ac56c 100644
--- a/src/include/extern.h
+++ b/src/include/extern.h
@@ -551,7 +551,8 @@ extern void __wt_rwunlock(WT_SESSION_IMPL *session, WT_RWLOCK *rwlock);
 extern int __wt_rwlock_destroy(WT_SESSION_IMPL *session, WT_RWLOCK *rwlock);
 extern int __wt_open(WT_SESSION_IMPL *session,
     const char *name,
-    int create,
+    int ok_create,
+    int is_tree,
     WT_FH **fhp);
 extern int __wt_close(WT_SESSION_IMPL *session, WT_FH *fh);
 extern int __wt_has_priv(void);
diff --git a/src/include/wiredtiger.in b/src/include/wiredtiger.in
index 757c6fceb38..960ed49a3c0 100644
--- a/src/include/wiredtiger.in
+++ b/src/include/wiredtiger.in
@@ -1039,6 +1039,9 @@ struct __wt_connection {
  * between 1MB and 10TB; default \c 100MB.}
  * @config{create, create the database if it does not exist.,a boolean flag;
  * default \c false.}
+ * @config{direct_io, Use \c O_DIRECT to access files.  Options are given as a
+ * list\, such as <code>"direct_io=[data]"</code>.,a list\, with values chosen
+ * from the following options: \c "data"\, \c "log"; default empty.}
  * @config{error_prefix, prefix string for error messages.,a string; default
  * empty.}
  * @config{eviction_target, continue evicting until the cache becomes less full
diff --git a/src/os_posix/os_open.c b/src/os_posix/os_open.c
index b82d88987a0..281db425917 100644
--- a/src/os_posix/os_open.c
+++ b/src/os_posix/os_open.c
@@ -12,7 +12,8 @@
  *	Open a file handle.
  */
 int
-__wt_open(WT_SESSION_IMPL *session, const char *name, int create, WT_FH **fhp)
+__wt_open(WT_SESSION_IMPL *session,
+    const char *name, int ok_create, int is_tree, WT_FH **fhp)
 {
 	const char *path;
 	WT_CONNECTION_IMPL *conn;
@@ -49,12 +50,19 @@ __wt_open(WT_SESSION_IMPL *session, const char *name, int create, WT_FH **fhp)
 	/* Windows clones: we always want to treat the file as a binary. */
 	f |= O_BINARY;
 #endif
-	if (create) {
+	if (ok_create) {
 		f |= O_CREAT;
 		mode = 0666;
 	} else
 		mode = 0;
 
+#ifdef O_DIRECT
+	if (is_tree && FLD_ISSET(conn->direct_io, WT_DIRECTIO_DATA))
+		f |= O_DIRECT;
+#else
+	WT_UNUSED(is_tree);
+#endif
+
 	WT_SYSCALL_RETRY(((fd = open(path, f, mode)) == -1 ? 1 : 0), ret);
 	if (ret != 0)
 		WT_ERR_MSG(session, ret, "%s", name);