Merge pull request #847 from wiredtiger/checkpoint-directio

Do direct I/O reads on checkpoints to reduce the VM pressure of LSM workloads

12 files changed, 40 insertions, 17 deletions

diff --git a/dist/api_data.py b/dist/api_data.py
index 5509032d148..412f0603efd 100644
--- a/dist/api_data.py
+++ b/dist/api_data.py
@@ -613,8 +613,13 @@ methods = {
 	    Use \c O_DIRECT to access files.  Options are given as a list,
 	    such as <code>"direct_io=[data]"</code>.  Configuring
 	    \c direct_io requires care, see @ref
-	    tuning_system_buffer_cache_direct_io for important warnings''',
-	    type='list', choices=['data', 'log']),
+	    tuning_system_buffer_cache_direct_io for important warnings.
+		Including \c "data" will cause WiredTiger data files to use
+		\c O_DIRECT, including \c "log" will cause WiredTiger log files
+		to use \c O_DIRECT, and including \c "checkpoint" will cause
+		WiredTiger data files opened at a checkpoint (i.e: read only) to
+		use \c O_DIRECT''',
+	    type='list', choices=['checkpoint', 'data', 'log']),
 	Config('extensions', '', r'''
 	    list of shared library extensions to load (using dlopen).
 	    Any values specified to an library extension are passed to
diff --git a/dist/flags.py b/dist/flags.py
index 7ca1ee3144d..2a6adfccf43 100644
--- a/dist/flags.py
+++ b/dist/flags.py
@@ -15,6 +15,7 @@ flags = {
 		'SYNC_WRITE_LEAVES',
 	],
 	'file_types' : [
+		'FILE_TYPE_CHECKPOINT',
 		'FILE_TYPE_DATA',
 		'FILE_TYPE_LOG'
 	],
diff --git a/src/block/block_mgr.c b/src/block/block_mgr.c
index 04c24a6a3b6..35201008622 100644
--- a/src/block/block_mgr.c
+++ b/src/block/block_mgr.c
@@ -405,8 +405,9 @@ __bm_method_set(WT_BM *bm, int readonly)
  *	Open a file.
  */
 int
-__wt_block_manager_open(WT_SESSION_IMPL *session, const char *filename,
-    const char *cfg[], int forced_salvage, uint32_t allocsize, WT_BM **bmp)
+__wt_block_manager_open(WT_SESSION_IMPL *session,
+    const char *filename, const char *cfg[],
+    int forced_salvage, int readonly, uint32_t allocsize, WT_BM **bmp)
 {
 	WT_BM *bm;
 	WT_DECL_RET;
@@ -416,8 +417,8 @@ __wt_block_manager_open(WT_SESSION_IMPL *session, const char *filename,
 	WT_RET(__wt_calloc_def(session, 1, &bm));
 	__bm_method_set(bm, 0);
 
-	WT_ERR(__wt_block_open(
-	    session, filename, cfg, forced_salvage, allocsize, &bm->block));
+	WT_ERR(__wt_block_open(session, filename, cfg,
+	    forced_salvage, readonly, allocsize, &bm->block));
 
 	*bmp = bm;
 	return (0);
diff --git a/src/block/block_open.c b/src/block/block_open.c
index 1132cb85a6c..b172d0bb009 100644
--- a/src/block/block_open.c
+++ b/src/block/block_open.c
@@ -95,7 +95,7 @@ __block_destroy(WT_SESSION_IMPL *session, WT_BLOCK *block)
 int
 __wt_block_open(WT_SESSION_IMPL *session,
     const char *filename, const char *cfg[],
-    int forced_salvage, uint32_t allocsize, WT_BLOCK **blockp)
+    int forced_salvage, int readonly, uint32_t allocsize, WT_BLOCK **blockp)
 {
 	WT_BLOCK *block;
 	WT_CONFIG_ITEM cval;
@@ -159,7 +159,9 @@ __wt_block_open(WT_SESSION_IMPL *session,
 
 	/* Open the underlying file handle. */
 	WT_ERR(__wt_open(
-	    session, filename, 0, 0, WT_FILE_TYPE_DATA, &block->fh));
+	    session, filename, 0, 0,
+	    readonly ? WT_FILE_TYPE_CHECKPOINT : WT_FILE_TYPE_DATA,
+	    &block->fh));
 
 	/* Initialize the live checkpoint's lock. */
 	WT_ERR(__wt_spin_init(session, &block->live_lock, "block manager"));
diff --git a/src/btree/bt_handle.c b/src/btree/bt_handle.c
index 2da886b4423..f6cc4cc6fb3 100644
--- a/src/btree/bt_handle.c
+++ b/src/btree/bt_handle.c
@@ -69,8 +69,8 @@ __wt_btree_open(WT_SESSION_IMPL *session, const char *op_cfg[])
 	if (!WT_PREFIX_SKIP(filename, "file:"))
 		WT_ERR_MSG(session, EINVAL, "expected a 'file:' URI");
 
-	WT_ERR(__wt_block_manager_open(session, filename,
-	    dhandle->cfg, forced_salvage, btree->allocsize, &btree->bm));
+	WT_ERR(__wt_block_manager_open(session, filename, dhandle->cfg,
+	    forced_salvage, readonly, btree->allocsize, &btree->bm));
 	bm = btree->bm;
 
 	/*
diff --git a/src/config/config_def.c b/src/config/config_def.c
index abe6713696c..2c01cac1a85 100644
--- a/src/config/config_def.c
+++ b/src/config/config_def.c
@@ -250,7 +250,9 @@ static const WT_CONFIG_CHECK confchk_wiredtiger_open[] = {
 	    ,
 	{ "checkpoint_sync", "boolean", NULL, NULL},
 	{ "create", "boolean", NULL, NULL},
-	{ "direct_io", "list", "choices=[\"data\",\"log\"]", NULL},
+	{ "direct_io", "list",
+	    "choices=[\"checkpoint\",\"data\",\"log\"]",
+	    NULL},
 	{ "error_prefix", "string", NULL, NULL},
 	{ "eviction_dirty_target", "int", "min=10,max=99", NULL},
 	{ "eviction_target", "int", "min=10,max=99", NULL},
diff --git a/src/conn/conn_api.c b/src/conn/conn_api.c
index bcbb8e3ceb8..b1992793827 100644
--- a/src/conn/conn_api.c
+++ b/src/conn/conn_api.c
@@ -1024,6 +1024,7 @@ wiredtiger_open(const char *home, WT_EVENT_HANDLER *event_handler,
 		const char *name;
 		uint32_t flag;
 	} *ft, file_types[] = {
+		{ "checkpoint",	WT_FILE_TYPE_CHECKPOINT },
 		{ "data",	WT_FILE_TYPE_DATA },
 		{ "log",	WT_FILE_TYPE_LOG },
 		{ NULL, 0 }
diff --git a/src/include/extern.h b/src/include/extern.h
index bd24cb02d61..ba026253a5b 100644
--- a/src/include/extern.h
+++ b/src/include/extern.h
@@ -126,6 +126,7 @@ extern int __wt_block_manager_open(WT_SESSION_IMPL *session,
     const char *filename,
     const char *cfg[],
     int forced_salvage,
+    int readonly,
     uint32_t allocsize,
     WT_BM **bmp);
 extern int __wt_block_manager_truncate( WT_SESSION_IMPL *session,
@@ -138,6 +139,7 @@ extern int __wt_block_open(WT_SESSION_IMPL *session,
     const char *filename,
     const char *cfg[],
     int forced_salvage,
+    int readonly,
     uint32_t allocsize,
     WT_BLOCK **blockp);
 extern int __wt_block_close(WT_SESSION_IMPL *session, WT_BLOCK *block);
diff --git a/src/include/flags.h b/src/include/flags.h
index 31eed83e351..89f2450f3af 100644
--- a/src/include/flags.h
+++ b/src/include/flags.h
@@ -10,6 +10,7 @@
 #define	WT_CONN_PANIC					0x00000002
 #define	WT_CONN_SERVER_RUN				0x00000001
 #define	WT_EVICTION_SERVER_LOCKED			0x00000004
+#define	WT_FILE_TYPE_CHECKPOINT				0x00000004
 #define	WT_FILE_TYPE_DATA				0x00000002
 #define	WT_FILE_TYPE_LOG				0x00000001
 #define	WT_LOGSCAN_FIRST				0x00000008
diff --git a/src/include/wiredtiger.in b/src/include/wiredtiger.in
index 8bcdc431bd9..6334bca2061 100644
--- a/src/include/wiredtiger.in
+++ b/src/include/wiredtiger.in
@@ -1493,8 +1493,12 @@ struct __wt_connection {
  * @config{direct_io, Use \c O_DIRECT to access files.  Options are given as a
  * list\, such as <code>"direct_io=[data]"</code>. Configuring \c direct_io
  * requires care\, see @ref tuning_system_buffer_cache_direct_io for important
- * warnings., a list\, with values chosen from the following options: \c
- * "data"\, \c "log"; default empty.}
+ * warnings.  Including \c "data" will cause WiredTiger data files to use \c
+ * O_DIRECT\, including \c "log" will cause WiredTiger log files to use \c
+ * O_DIRECT\, and including \c "checkpoint" will cause WiredTiger data files
+ * opened at a checkpoint (i.e: read only) to use \c O_DIRECT., a list\, with
+ * values chosen from the following options: \c "checkpoint"\, \c "data"\, \c
+ * "log"; default empty.}
  * @config{error_prefix, prefix string for error messages., a string; default
  * empty.}
  * @config{eviction_dirty_target, continue evicting until the cache has less
diff --git a/src/os_posix/os_open.c b/src/os_posix/os_open.c
index 6ef4caadd0a..c6938dad9fd 100644
--- a/src/os_posix/os_open.c
+++ b/src/os_posix/os_open.c
@@ -107,7 +107,8 @@ __wt_open(WT_SESSION_IMPL *session,
 #endif
 #ifdef O_NOATIME
 	/* Avoid updating metadata for read-only workloads. */
-	if (dio_type == WT_FILE_TYPE_DATA)
+	if (dio_type == WT_FILE_TYPE_DATA ||
+	    dio_type == WT_FILE_TYPE_CHECKPOINT)
 		f |= O_NOATIME;
 #endif
 
@@ -157,7 +158,8 @@ __wt_open(WT_SESSION_IMPL *session,
 
 #if defined(HAVE_POSIX_FADVISE)
 	/* Disable read-ahead on trees: it slows down random read workloads. */
-	if (dio_type == WT_FILE_TYPE_DATA)
+	if (dio_type == WT_FILE_TYPE_DATA ||
+	    dio_type == WT_FILE_TYPE_CHECKPOINT)
 		WT_ERR(posix_fadvise(fd, 0, 0, POSIX_FADV_RANDOM));
 #endif
 
@@ -174,7 +176,8 @@ __wt_open(WT_SESSION_IMPL *session,
 	WT_ERR(__wt_filesize(session, fh, &fh->size));
 
 	/* Configure file extension. */
-	if (dio_type == WT_FILE_TYPE_DATA)
+	if (dio_type == WT_FILE_TYPE_DATA ||
+	    dio_type == WT_FILE_TYPE_CHECKPOINT)
 		fh->extend_len = conn->data_extend_len;
 
 	/*
diff --git a/src/schema/schema_create.c b/src/schema/schema_create.c
index 8d620198c09..8bd78ce0ac4 100644
--- a/src/schema/schema_create.c
+++ b/src/schema/schema_create.c
@@ -33,7 +33,8 @@ __wt_direct_io_size_check(WT_SESSION_IMPL *session,
 	 * if you configure direct I/O and then don't do I/O in alignments and
 	 * units of its happy place.
 	 */
-	if (FLD_ISSET(conn->direct_io, WT_FILE_TYPE_DATA)) {
+	if (FLD_ISSET(conn->direct_io,
+	   WT_FILE_TYPE_CHECKPOINT | WT_FILE_TYPE_DATA)) {
 		align = (int64_t)conn->buffer_alignment;
 		if (align != 0 && (cval.val < align || cval.val % align != 0))
 			WT_RET_MSG(session, EINVAL,