[Cluster-devel] cluster/gfs-kernel/src/gfs diaper.c gfs.h inco ...

[wcheng@sourceware.org <wcheng@sourceware.org>]

CVSROOT:	/cvs/cluster
Module name:	cluster
Branch: 	RHEL4
Changes by:	wcheng at sourceware.org	2007-08-28 14:58:48

Modified files:
	gfs-kernel/src/gfs: diaper.c gfs.h incore.h ioctl.c ops_file.c 
	                    super.h 

Log message:
	Bugzilla 230803:
	
	Allow GFS to take advantage of VFS's "bd_inode_backing_dev_info" setting
	that overrides diaper device's generic readahead value. Various benchmarks
	show we can achieve 2x more performance with large sequential write.
	
	Option is controlled by a new GFS tunable (seq_readahead).

Patches:
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/gfs-kernel/src/gfs/diaper.c.diff?cvsroot=cluster&only_with_tag=RHEL4&r1=1.1.2.3&r2=1.1.2.4
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/gfs-kernel/src/gfs/gfs.h.diff?cvsroot=cluster&only_with_tag=RHEL4&r1=1.5&r2=1.5.2.1
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/gfs-kernel/src/gfs/incore.h.diff?cvsroot=cluster&only_with_tag=RHEL4&r1=1.17.2.9&r2=1.17.2.10
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/gfs-kernel/src/gfs/ioctl.c.diff?cvsroot=cluster&only_with_tag=RHEL4&r1=1.7.2.7&r2=1.7.2.8
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/gfs-kernel/src/gfs/ops_file.c.diff?cvsroot=cluster&only_with_tag=RHEL4&r1=1.16.2.15&r2=1.16.2.16
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/gfs-kernel/src/gfs/super.h.diff?cvsroot=cluster&only_with_tag=RHEL4&r1=1.2.2.1&r2=1.2.2.2

--- cluster/gfs-kernel/src/gfs/Attic/diaper.c	2005/05/05 23:19:39	1.1.2.3
+++ cluster/gfs-kernel/src/gfs/Attic/diaper.c	2007/08/28 14:58:48	1.1.2.4
@@ -371,6 +371,7 @@
 		diaper->bd_inode->i_data.backing_dev_info = &gd->queue->backing_dev_info;
 	} else
 		printk("GFS: diaper: reopening\n");
+
 	diaper->bd_openers++;
 	up(&diaper->bd_sem);
 
--- cluster/gfs-kernel/src/gfs/gfs.h	2005/01/04 10:07:11	1.5
+++ cluster/gfs-kernel/src/gfs/gfs.h	2007/08/28 14:58:48	1.5.2.1
@@ -21,6 +21,7 @@
 
 #include "fixed_div64.h"
 #include "lvb.h"
+#include <linux/backing-dev.h>
 #include "incore.h"
 #include "util.h"
 
--- cluster/gfs-kernel/src/gfs/incore.h	2007/03/13 21:21:21	1.17.2.9
+++ cluster/gfs-kernel/src/gfs/incore.h	2007/08/28 14:58:48	1.17.2.10
@@ -894,6 +894,7 @@
 	unsigned int gt_greedy_max;
 	unsigned int gt_rgrp_try_threshold;
 	unsigned int gt_statfs_fast;
+	unsigned int gt_seq_readahead;
 };
 
 /*
@@ -959,6 +960,10 @@
 
 	struct super_block *sd_vfs;     /* Linux VFS device independent sb */
 
+	/* readahead */
+	struct backing_dev_info sd_dev_info;
+	spinlock_t sd_dev_info_spin;
+
 	struct gfs_args sd_args;        /* Mount arguments */
 	unsigned long sd_flags;         /* SDF_... see above */
 
--- cluster/gfs-kernel/src/gfs/ioctl.c	2007/03/13 21:21:21	1.7.2.7
+++ cluster/gfs-kernel/src/gfs/ioctl.c	2007/08/28 14:58:48	1.7.2.8
@@ -484,6 +484,7 @@
         gfs_printf("greedy_max %u\n", gt->gt_greedy_max);
         gfs_printf("rgrp_try_threshold %u\n", gt->gt_rgrp_try_threshold);
         gfs_printf("statfs_fast %u\n", gt->gt_statfs_fast);
+        gfs_printf("seq_readahead %u\n", gt->gt_seq_readahead);
 
         error = 0;
 
@@ -763,6 +764,14 @@
 		else 
 			tune_set(gt_statfs_fast, x);
 
+	} else if (strcmp(param, "seq_readahead") == 0) {
+		if (sscanf(value, "%u", &x) != 1)
+			return -EINVAL;
+		error = gfs_reset_readahead(sdp, x);
+		if (error)
+			return error;
+		tune_set(gt_seq_readahead, x);
+
 	} else
 		return -EINVAL;
 
--- cluster/gfs-kernel/src/gfs/ops_file.c	2006/11/17 20:33:15	1.16.2.15
+++ cluster/gfs-kernel/src/gfs/ops_file.c	2007/08/28 14:58:48	1.16.2.16
@@ -1743,6 +1743,57 @@
 		return do_flock(file, cmd, fl);
 }
 
+/*
+ * VFS layer offers a "bd_inode_backing_dev_info" pointer to allow
+ * filesystem overriding its block device generic readahead value.
+ * We'll take this advantage to tune for large sequential write.
+ */
+struct backing_dev_info gfs_file_backing_dev_info = {
+        .ra_pages       = 0,    /* place holder */
+        .memory_backed  = 0,    /* place holder */
+        .unplug_io_fn   = default_unplug_io_fn,
+};
+
+int gfs_reset_readahead(struct gfs_sbd *sdp, unsigned int file_ra)
+{
+	unsigned int max_ra;
+	struct backing_dev_info *bdi;
+	struct super_block *vfs_sb=sdp->sd_vfs;
+
+	/* user wants to reset back to default */
+	if (file_ra == 0) {
+		vfs_sb->s_bdev->bd_inode_backing_dev_info = NULL;
+		return 0;
+	}
+
+	/* can't exceed gfs max readahead */
+	max_ra = gfs_tune_get(sdp, gt_max_readahead);
+	if (file_ra > max_ra) {
+		printk("GFS: einval - readahead exceeds max\n");
+		return -EINVAL;
+	}
+
+	bdi = sdp->sd_vfs->s_bdev->bd_inode_backing_dev_info;
+
+	/* sanity check */
+	if (gfs_tune_get(sdp, gt_seq_readahead) && !bdi) {
+		printk("GFS: error - gt_seq_readahead set but bdi is NULL\n");
+		return -EIO;
+	}
+		
+	/* add pointer */
+	if (!bdi) {
+		bdi = &sdp->sd_dev_info;
+		vfs_sb->s_bdev->bd_inode_backing_dev_info = bdi;
+		*bdi = *(vfs_sb->s_bdev->bd_inode->i_data.backing_dev_info);
+	}
+
+	/* now change the setting */
+	bdi->ra_pages = file_ra;
+
+	return 0;
+}
+
 struct file_operations gfs_file_fops = {
 	.llseek = gfs_llseek,
 	.read = gfs_read,
--- cluster/gfs-kernel/src/gfs/super.h	2007/03/13 21:21:21	1.2.2.1
+++ cluster/gfs-kernel/src/gfs/super.h	2007/08/28 14:58:48	1.2.2.2
@@ -63,4 +63,6 @@
 int gfs_freeze_fs(struct gfs_sbd *sdp);
 void gfs_unfreeze_fs(struct gfs_sbd *sdp);
 
+int gfs_reset_readahead(struct gfs_sbd *sdp, unsigned int file_ra);
+
 #endif /* __SUPER_DOT_H__ */