Skip to contentSkip to conversation

VFS: File System Mount Wide O_DIRECT Support

RFC In review Li Wang 5 participants 5 messages Sep 4, 2012 VFS

lore
Share & tools
lore

Li Wang

author
#0 lore raw

RFC 1/1 VFS: File System Mount Wide O_DIRECT Support

For file system created on file-backed loop device, there will be two-levels of page cache present, which typically doubles the memory consumption.
In many cases, it is beneficial to turn on the O_DIRECT option while performing the upper file system file IO, to bypass the upper page cache, which not only reduces half of the memory consumption, but also improves the performance due to shorter copy path.

For example, the following iozone REREAD test with O_DIRECT turned on over the one without enjoys 10x speedup due to redundant cache elimination, consequently, avoiding page cache thrashing on a 2GB memory machine running 3.2.9 kernel.

losetup /dev/loop0 dummy // dummy is a ext4 file with a size of 1.1GB mkfs -t ext2 /dev/loop0
mount /dev/loop0 /dsk
cd /dsk
iozone -t 1 -s 1G -r 4M -i 0 -+n -w // produce a 1GB test file iozone -t 1 -s 1G -r 4M -i 1 -w // REREAD test without O_DIRECT echo 1 > /proc/sys/vm/drop_caches // cleanup the page cache iozone -t 1 -s 1G -r 4M -i 1 -w -I // REREAD test with O_DIRECT

This feature is also expected to be useful for virtualization situation, the file systems inside the guest operation system will use much less of guest memory, which, potencially results in less of host memory use. Especially, it may be more useful if multiple guests are running based on a same disk image file.

The idea is simple, leave the desicion for the file system user to enable file system mount wide O_DIRECT support with a new mount option, for example,

losetup /dev/loop0 dummy
mount /dev/loop0 -o MS_DIRECT /dsk

Below is the preliminary patch,

---
 fs/open.c          |    5 +++++
 fs/super.c         |    2 ++
 include/linux/fs.h |    1 +
 3 files changed, 8 insertions(+), 0 deletions(-)
fs/open.c +5 -0
index e1f2cdb..dacac30 100644
@@ -958,6 +958,11 @@ long do_sys_open(int dfd, const char __user *filename, int flags, umode_t mode)
 			} else {
 				fsnotify_open(f);
 				fd_install(fd, f);
+				if (f->f_vfsmnt->mnt_sb && f->f_vfsmnt->mnt_sb->s_flags & MS_DIRECT) {
+					if (S_ISREG(f->f_dentry->d_inode->i_mode)) {
+	if (!f->f_mapping->a_ops || ((!f->f_mapping->a_ops->direct_IO) && (!f->f_mapping->a_ops->get_xip_mem)))
+		f->f_flags |= O_DIRECT;
+				}
 			}
 		}
 		putname(tmp);
fs/super.c +2 -0
index 0902cfa..ab5c4a5 100644
@@ -1147,6 +1147,8 @@ mount_fs(struct file_system_type *type, int flags, const char *name, void *data)
 	WARN_ON(!sb->s_bdi);
 	WARN_ON(sb->s_bdi == &default_backing_dev_info);
 	sb->s_flags |= MS_BORN;
+	if (flags & MS_DIRECT)
+		sb->s_flags |= MS_DIRECT;
 
 	error = security_sb_kern_mount(sb, flags, secdata);
 	if (error)
include/linux/fs.h +1 -0
index aa11047..127cc85 100644
@@ -225,6 +225,7 @@ struct inodes_stat_t {
 #define MS_KERNMOUNT	(1<<22) /* this is a kern_mount call */
 #define MS_I_VERSION	(1<<23) /* Update inode I_version field */
 #define MS_STRICTATIME	(1<<24) /* Always perform atime updates */
+#define MS_DIRECT	(1<<27)
 #define MS_NOSEC	(1<<28)
 #define MS_BORN		(1<<29)
 #define MS_ACTIVE	(1<<30)
Signature · 1 line
1.7.6.5
#1 lore raw
Quoted 2 lines · Li Wang

For file system created on file-backed loop device, there will be two-levels of page cache present, which typically doubles the memory consumption.

And the right fix is to not use buffer I/O on the backing file instead of hacks like this.

#2 lore raw
Quoted 6 lines · Christoph Hellwig
Quoted 2 lines · Li Wang

For file system created on file-backed loop device, there will be two-levels of page cache present, which typically doubles the memory consumption.

And the right fix is to not use buffer I/O on the backing file instead of hacks like this.

That was my initial reaction too, but for the case of two VMs operating on the same device, it's better for it to be cached once in the hype-rvisor than twice in the VMs. Is that a common case worth optimising for? Probably not ...

Signature · 4 lines
Matthew Wilcox				Intel Open Source Technology Centre
"Bill, look, we understand that you're interested in selling us this
operating system, but compare it to ours.  We can't possibly take such
a retrograde step."
reviewer
#3 lore raw
In reply to Li Wang
Quoted 19 lines · Li Wang

For file system created on file-backed loop device, there will be two-levels of page cache present, which typically doubles the memory consumption.
In many cases, it is beneficial to turn on the O_DIRECT option while performing the upper file system file IO, to bypass the upper page cache, which not only reduces half of the memory consumption, but also improves the performance due to shorter copy path.

For example, the following iozone REREAD test with O_DIRECT turned on over the one without enjoys 10x speedup due to redundant cache elimination, consequently, avoiding page cache thrashing on a 2GB memory machine running 3.2.9 kernel.

losetup /dev/loop0 dummy // dummy is a ext4 file with a size of 1.1GB mkfs -t ext2 /dev/loop0
mount /dev/loop0 /dsk
cd /dsk
iozone -t 1 -s 1G -r 4M -i 0 -+n -w // produce a 1GB test file iozone -t 1 -s 1G -r 4M -i 1 -w // REREAD test without O_DIRECT echo 1 > /proc/sys/vm/drop_caches // cleanup the page cache iozone -t 1 -s 1G -r 4M -i 1 -w -I // REREAD test with O_DIRECT

Quoted 7 lines

This feature is also expected to be useful for virtualization situation, the file systems inside the guest operation system will use much less of guest memory, which, potencially results in less of host memory use. Especially, it may be more useful if multiple guests are running based on a same disk image file.

The idea is simple, leave the desicion for the file system user to enable file system mount wide O_DIRECT support with a new mount option, for example,

I believe a better approach to your problem is actually to enable
loopback device driver to use direct IO. Someone was actually working on this but I'm not sure where this ended up.

Honza

Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR
#4 lore raw
Quoted 6 lines
Quoted 2 lines

The idea is simple, leave the desicion for the file system user to enable file system mount wide O_DIRECT support with a new mount option, for example,

I believe a better approach to your problem is actually to enable
loopback device driver to use direct IO. Someone was actually working on this but I'm not sure where this ended up.

Dave's been working on getting those patches merged. I'm also not sure where the work currently is, but here's an older posting:

http://lwn.net/Articles/489647/

- z