[Cluster-devel] cluster/gfs-kernel/src/gfs daemon.c gfs_ondisk ...
wcheng at sourceware.org
wcheng at sourceware.org
Tue Jun 19 21:26:13 UTC 2007
CVSROOT: /cvs/cluster
Module name: cluster
Changes by: wcheng at sourceware.org 2007-06-19 21:26:12
Modified files:
gfs-kernel/src/gfs: daemon.c gfs_ondisk.h incore.h ioctl.c
ops_fstype.c ops_super.c rgrp.c rgrp.h
super.c super.h
Log message:
Bugzilla 231904:
Port RHEL4 fast statfs (for commands such as "df") implementation over.
The "lvb" enhancement will be followed around RHEL 5.2 time frame.
Ballpark performance numbers:
dhcp145 (1 cpu HP): old df took 0.875 seconds, new df 0.008 second
dhcp146 (4 cpus DELL): old df took 0.808 seconds, new df 0.006 second.
Activated via "gfs_tool settune <mount point> statfs_fast 1" command.
Patches:
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/gfs-kernel/src/gfs/daemon.c.diff?cvsroot=cluster&r1=1.8&r2=1.9
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/gfs-kernel/src/gfs/gfs_ondisk.h.diff?cvsroot=cluster&r1=1.10&r2=1.11
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/gfs-kernel/src/gfs/incore.h.diff?cvsroot=cluster&r1=1.32&r2=1.33
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/gfs-kernel/src/gfs/ioctl.c.diff?cvsroot=cluster&r1=1.16&r2=1.17
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/gfs-kernel/src/gfs/ops_fstype.c.diff?cvsroot=cluster&r1=1.31&r2=1.32
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/gfs-kernel/src/gfs/ops_super.c.diff?cvsroot=cluster&r1=1.27&r2=1.28
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/gfs-kernel/src/gfs/rgrp.c.diff?cvsroot=cluster&r1=1.20&r2=1.21
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/gfs-kernel/src/gfs/rgrp.h.diff?cvsroot=cluster&r1=1.4&r2=1.5
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/gfs-kernel/src/gfs/super.c.diff?cvsroot=cluster&r1=1.23&r2=1.24
http://sourceware.org/cgi-bin/cvsweb.cgi/cluster/gfs-kernel/src/gfs/super.h.diff?cvsroot=cluster&r1=1.4&r2=1.5
--- cluster/gfs-kernel/src/gfs/daemon.c 2006/12/21 20:50:43 1.8
+++ cluster/gfs-kernel/src/gfs/daemon.c 2007/06/19 21:26:11 1.9
@@ -141,6 +141,18 @@
int error;
while (!kthread_should_stop()) {
+ /* Update statfs file */
+ if (gfs_tune_get(sdp, gt_statfs_fast) &&
+ time_after_eq(jiffies,
+ sdp->sd_statfs_sync_time +
+ gfs_tune_get(sdp, gt_statfs_fast) * HZ)) {
+ error = gfs_statfs_sync(sdp);
+ if (error && error != -EROFS &&
+ !test_bit(SDF_SHUTDOWN, &sdp->sd_flags))
+ printk("GFS: fsid=%s: statfs: error = %d\n",
+ sdp->sd_fsname, error);
+ sdp->sd_statfs_sync_time = jiffies;
+ }
/* Update quota file */
if (time_after_eq(jiffies,
sdp->sd_quota_sync_time +
--- cluster/gfs-kernel/src/gfs/gfs_ondisk.h 2006/07/17 21:38:13 1.10
+++ cluster/gfs-kernel/src/gfs/gfs_ondisk.h 2007/06/19 21:26:11 1.11
@@ -618,6 +618,24 @@
uint32_t ea_pad;
};
+/*
+ * Statfs change
+ * Describes an change to the pool of free and allocated
+ * blocks.
+ */
+
+struct gfs_statfs_change {
+ uint64_t sc_total;
+ uint64_t sc_free;
+ uint64_t sc_dinodes;
+};
+
+struct gfs_statfs_change_host {
+ int64_t sc_total;
+ int64_t sc_free;
+ int64_t sc_dinodes;
+};
+
/* Endian functions */
#define GFS_ENDIAN_BIG
--- cluster/gfs-kernel/src/gfs/incore.h 2007/06/17 05:34:53 1.32
+++ cluster/gfs-kernel/src/gfs/incore.h 2007/06/19 21:26:11 1.33
@@ -904,6 +904,7 @@
unsigned int gt_greedy_quantum;
unsigned int gt_greedy_max;
unsigned int gt_rgrp_try_threshold;
+ unsigned int gt_statfs_fast;
};
/*
@@ -966,6 +967,13 @@
struct gfs_tune sd_tune; /* Filesystem tuning structure */
+ /* statfs */
+ struct inode *sd_statfs_inode;
+ spinlock_t sd_statfs_spin;
+ struct gfs_statfs_change_host sd_statfs_master;
+ struct gfs_statfs_change_host sd_statfs_local;
+ unsigned long sd_statfs_sync_time;
+
/* Resource group stuff */
struct gfs_inode *sd_riinode; /* Resource Index (rindex) inode */
--- cluster/gfs-kernel/src/gfs/ioctl.c 2007/06/17 05:34:53 1.16
+++ cluster/gfs-kernel/src/gfs/ioctl.c 2007/06/19 21:26:11 1.17
@@ -483,6 +483,7 @@
gfs_printf("greedy_quantum %u\n", gt->gt_greedy_quantum);
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);
error = 0;
@@ -513,6 +514,7 @@
struct gfs_tune *gt = &sdp->sd_tune;
char param[ARG_SIZE], value[ARG_SIZE];
unsigned int x;
+ int error;
if (!capable(CAP_SYS_ADMIN))
return -EACCES;
@@ -752,6 +754,16 @@
return -EINVAL;
tune_set(gt_rgrp_try_threshold, x);
+ } else if (strcmp(param, "statfs_fast") == 0) {
+ if (sscanf(value, "%u", &x) != 1)
+ return -EINVAL;
+ error = gfs_statfs_init(sdp, x);
+ if (error)
+ return error;
+ else
+ tune_set(gt_statfs_fast, x);
+
+
} else
return -EINVAL;
--- cluster/gfs-kernel/src/gfs/ops_fstype.c 2007/06/19 14:23:29 1.31
+++ cluster/gfs-kernel/src/gfs/ops_fstype.c 2007/06/19 21:26:11 1.32
@@ -66,6 +66,8 @@
spin_lock_init(&sdp->sd_rg_recent_lock);
spin_lock_init(&sdp->sd_rg_forward_lock);
+ spin_lock_init(&sdp->sd_statfs_spin);
+
for (x = 0; x < GFS_GL_HASH_SIZE; x++) {
sdp->sd_gl_hash[x].hb_lock = RW_LOCK_UNLOCKED;
INIT_LIST_HEAD(&sdp->sd_gl_hash[x].hb_list);
@@ -367,6 +369,19 @@
goto fail_root_free;
}
+ /* Implement fast statfs on the unused license inode location.
+ * sb->sb_quota_di.no_formal_ino = jindex_dinode + 2;
+ * sb->sb_quota_di.no_addr = jindex_dinode + 2;
+ * sb->sb_license_di.no_formal_ino = jindex_dinode + 3;
+ * sb->sb_license_di.no_addr = jindex_dinode + 3;
+ */
+ error = gfs_get_linode(sdp);
+ if (error) {
+ printk("GFS: fsid=%s: can't get statfs file inode: %d\n",
+ sdp->sd_fsname, error);
+ goto fail_qi_free;
+ }
+
/* We're through with the superblock lock */
out:
gfs_glock_dq_uninit(&sb_gh);
@@ -377,6 +392,7 @@
dput(sb->s_root);
sb->s_root = NULL;
}
+fail_qi_free:
gfs_inode_put(sdp->sd_qinode);
fail_root_free:
gfs_inode_put(sdp->sd_rooti);
--- cluster/gfs-kernel/src/gfs/ops_super.c 2006/12/21 20:50:43 1.27
+++ cluster/gfs-kernel/src/gfs/ops_super.c 2007/06/19 21:26:11 1.28
@@ -144,6 +144,7 @@
gfs_inode_put(sdp->sd_jiinode);
gfs_inode_put(sdp->sd_rooti);
gfs_inode_put(sdp->sd_qinode);
+ gfs_inode_put(sdp->sd_linode);
gfs_glock_put(sdp->sd_trans_gl);
gfs_glock_put(sdp->sd_rename_gl);
@@ -267,6 +268,9 @@
atomic_inc(&sdp->sd_ops_super);
+ if (gfs_tune_get(sdp, gt_statfs_fast))
+ return(gfs_statfs_fast(sdp, (void *)buf));
+
error = gfs_stat_gfs(sdp, &sg, TRUE);
if (error)
return error;
--- cluster/gfs-kernel/src/gfs/rgrp.c 2007/02/02 21:01:04 1.20
+++ cluster/gfs-kernel/src/gfs/rgrp.c 2007/06/19 21:26:11 1.21
@@ -1501,7 +1501,7 @@
rgd = gfs_blk2rgrpd(sdp, bstart);
if (!rgd) {
if (gfs_consist(sdp))
- printk("GFS: fsid=%s: block = %"PRIu64"\n",
+ printk("GFS: fsid=%s: block = %llu\n",
sdp->sd_fsname, bstart);
return NULL;
}
@@ -1658,6 +1658,9 @@
al->al_alloced_data++;
gfs_trans_add_quota(sdp, +1, ip->i_di.di_uid, ip->i_di.di_gid);
+
+ /* total=0, free=-1, dinodes=0 */
+ gfs_statfs_modify(sdp, 0, -1, 0);
}
/**
@@ -1712,6 +1715,9 @@
gfs_trans_add_quota(sdp, +1, ip->i_di.di_uid, ip->i_di.di_gid);
+ /* total=0, free=-1, dinode=0 */
+ gfs_statfs_modify(sdp, 0, -1, 0);
+
return 0;
}
@@ -1727,6 +1733,7 @@
int
gfs_dialloc(struct gfs_inode *dip, uint64_t *block)
{
+ struct gfs_sbd *sdp = dip->i_sbd;
struct gfs_alloc *al = dip->i_alloc;
struct gfs_rgrpd *rgd = al->al_rgd;
uint32_t goal, blk;
@@ -1766,6 +1773,9 @@
al->al_alloced_di++;
al->al_alloced_meta++;
+ /* total=0, free=-1, dinodes=1 */
+ gfs_statfs_modify(sdp, 0, -1, +1);
+
return error;
}
@@ -1798,6 +1808,9 @@
gfs_trans_add_quota(sdp, -(int64_t)blen,
ip->i_di.di_uid,
ip->i_di.di_gid);
+
+ /* total=0, free=+blen, dinodes=0 */
+ gfs_statfs_modify(sdp, 0, blen, 0);
}
/**
@@ -1832,6 +1845,9 @@
gfs_trans_add_bh(rgd->rd_gl, rgd->rd_bh[0]);
gfs_rgrp_out(&rgd->rd_rg, rgd->rd_bh[0]->b_data);
+ /* total=0, free=blen, dinode=0 */
+ gfs_statfs_modify(sdp, 0, blen, 0);
+
gfs_trans_add_quota(sdp, -(int64_t)blen,
ip->i_di.di_uid,
ip->i_di.di_gid);
@@ -1866,6 +1882,9 @@
gfs_trans_add_bh(rgd->rd_gl, rgd->rd_bh[0]);
gfs_rgrp_out(&rgd->rd_rg, rgd->rd_bh[0]->b_data);
+
+ /* total=0, free=1, dinodes=-1 */
+ gfs_statfs_modify(rgd->rd_sbd, 0, +1, -1);
}
/**
@@ -1914,7 +1933,7 @@
rgd = gfs_blk2rgrpd(sdp, block);
if (!rgd) {
if (gfs_consist(sdp))
- printk("GFS: fsid=%s: block = %"PRIu64"\n",
+ printk("GFS: fsid=%s: block = %llu\n",
sdp->sd_fsname, block);
return;
}
--- cluster/gfs-kernel/src/gfs/rgrp.h 2006/07/10 23:22:34 1.4
+++ cluster/gfs-kernel/src/gfs/rgrp.h 2007/06/19 21:26:11 1.5
@@ -57,6 +57,11 @@
void gfs_difree_uninit(struct gfs_rgrpd *rgd, uint64_t addr);
void gfs_difree(struct gfs_rgrpd *rgd, struct gfs_inode *ip);
+extern void gfs_statfs_modify(struct gfs_sbd *sdp,
+ int64_t total,
+ int64_t free,
+ int64_t dinodes);
+
/*
* gfs_rgrp_list
*
--- cluster/gfs-kernel/src/gfs/super.c 2007/06/17 05:34:53 1.23
+++ cluster/gfs-kernel/src/gfs/super.c 2007/06/19 21:26:11 1.24
@@ -19,6 +19,7 @@
#include <linux/completion.h>
#include <linux/buffer_head.h>
#include <linux/vmalloc.h>
+#include <linux/statfs.h>
#include "gfs.h"
#include "dio.h"
@@ -33,6 +34,7 @@
#include "rgrp.h"
#include "super.h"
#include "unlinked.h"
+#include "trans.h"
/**
* gfs_tune_init - Fill a gfs_tune structure with default values
@@ -85,6 +87,7 @@
gt->gt_greedy_quantum = HZ / 40;
gt->gt_greedy_max = HZ / 4;
gt->gt_rgrp_try_threshold = 100;
+ gt->gt_statfs_fast = 0;
}
/**
@@ -652,6 +655,46 @@
}
/**
+ * gfs_get_linode - Read in the special (hidden) license inode
+ * @sdp: The GFS superblock
+ *
+ * If one is not on-disk already, create a new one.
+ * Does not read in file contents, just the dinode.
+ *
+ * Returns: errno
+ */
+
+int
+gfs_get_linode(struct gfs_sbd *sdp)
+{
+ struct gfs_holder i_gh;
+ int error;
+
+ /* Create, if not on-disk already */
+ if (!sdp->sd_sb.sb_license_di.no_formal_ino) {
+ error = gfs_alloc_linode(sdp);
+ if (error)
+ return error;
+ }
+
+ error = gfs_glock_nq_num(sdp,
+ sdp->sd_sb.sb_license_di.no_formal_ino,
+ &gfs_inode_glops,
+ LM_ST_SHARED, GL_LOCAL_EXCL,
+ &i_gh);
+ if (error)
+ return error;
+
+ /* iopen obtained in via gfs_glock_get(..gfs_iopen_glops) */
+ error = gfs_inode_get(i_gh.gh_gl, &sdp->sd_sb.sb_license_di,
+ CREATE, &sdp->sd_linode);
+
+ gfs_glock_dq_uninit(&i_gh);
+
+ return error;
+}
+
+/**
* gfs_make_fs_rw - Turn a Read-Only FS into a Read-Write one
* @sdp: the filesystem
*
@@ -731,6 +774,8 @@
!test_bit(SDF_SHUTDOWN, &sdp->sd_flags))
return error;
+ gfs_statfs_sync(sdp);
+
gfs_log_flush(sdp);
gfs_quota_sync(sdp);
gfs_quota_scan(sdp);
@@ -1044,3 +1089,199 @@
up(&sdp->sd_freeze_lock);
}
+
+/*
+ * Fast statfs implementation - mostly based on GFS2 implementation.
+ */
+
+void gfs_statfs_change_in(struct gfs_statfs_change_host *sc, const void *buf)
+{
+ const struct gfs_statfs_change *str = buf;
+
+ sc->sc_total = be64_to_cpu(str->sc_total);
+ sc->sc_free = be64_to_cpu(str->sc_free);
+ sc->sc_dinodes = be64_to_cpu(str->sc_dinodes);
+}
+
+void gfs_statfs_change_out(const struct gfs_statfs_change_host *sc, void *buf)
+{
+ struct gfs_statfs_change *str = buf;
+
+ str->sc_total = cpu_to_be64(sc->sc_total);
+ str->sc_free = cpu_to_be64(sc->sc_free);
+ str->sc_dinodes = cpu_to_be64(sc->sc_dinodes);
+}
+
+int gfs_statfs_start(struct gfs_sbd *sdp)
+{
+ struct gfs_stat_gfs sg;
+ struct gfs_inode *m_ip;
+ struct gfs_statfs_change_host *m_sc = &sdp->sd_statfs_master;
+ struct gfs_statfs_change_host *l_sc = &sdp->sd_statfs_local;
+ struct buffer_head *m_bh;
+ struct gfs_holder gh;
+ int error;
+
+ printk("GFS: fsid=%s: fast statfs start time = %lu\n",
+ sdp->sd_fsname, get_seconds());
+
+ /* created via gfs_get_linode() in fill_super(). */
+ /* gfs_inode_glops */
+ m_ip = sdp->sd_linode;
+
+ /* get real statistics */
+ error = gfs_stat_gfs(sdp, &sg, TRUE);
+ if (error)
+ return error;
+
+ /* make sure the page is refreshed via glock flushing */
+ error = gfs_glock_nq_init(m_ip->i_gl, LM_ST_EXCLUSIVE, GL_NOCACHE,
+ &gh);
+ if (error)
+ goto gfs_statfs_start_out;
+
+ error = gfs_get_inode_buffer(m_ip, &m_bh);
+ if (error)
+ goto gfs_statfs_start_unlock;
+
+ error = gfs_trans_begin(sdp, 1, 0);
+ if (error)
+ goto gfs_statfs_start_bh;
+
+ spin_lock(&sdp->sd_statfs_spin);
+ m_sc->sc_total = sg.sg_total_blocks;
+ m_sc->sc_free = sg.sg_free + sg.sg_free_dinode + sg.sg_free_meta;
+ m_sc->sc_dinodes = sg.sg_used_dinode;
+ memset(l_sc, 0, sizeof(struct gfs_statfs_change_host));
+ spin_unlock(&sdp->sd_statfs_spin);
+
+ gfs_trans_add_bh(m_ip->i_gl, m_bh);
+ gfs_statfs_change_out(m_sc, m_bh->b_data + sizeof(struct gfs_dinode));
+
+ gfs_trans_end(sdp);
+
+gfs_statfs_start_bh:
+ brelse(m_bh);
+
+gfs_statfs_start_unlock:
+ gfs_glock_dq_uninit(&gh);
+
+gfs_statfs_start_out:
+ return 0;
+}
+
+int gfs_statfs_init(struct gfs_sbd *sdp, int flag)
+{
+ int error;
+
+ /* if flag == 0, do we want to turn this off ? */
+ if (!flag)
+ return 0;
+
+ error = gfs_statfs_start(sdp);
+ if (error)
+ printk("GFS: fsid=%s: can't initialize statfs subsystem: %d\n",
+ sdp->sd_fsname, error);
+
+ return error;
+}
+
+void gfs_statfs_modify(struct gfs_sbd *sdp,
+ int64_t total,
+ int64_t free,
+ int64_t dinodes)
+{
+ struct gfs_statfs_change_host *l_sc = &sdp->sd_statfs_local;
+
+ spin_lock(&sdp->sd_statfs_spin);
+ l_sc->sc_total += total;
+ l_sc->sc_free += free;
+ l_sc->sc_dinodes += dinodes;
+ spin_unlock(&sdp->sd_statfs_spin);
+}
+
+int gfs_statfs_sync(struct gfs_sbd *sdp)
+{
+ struct gfs_inode *m_ip = sdp->sd_linode;
+ struct gfs_statfs_change_host *m_sc = &sdp->sd_statfs_master;
+ struct gfs_statfs_change_host *l_sc = &sdp->sd_statfs_local;
+ struct gfs_holder gh;
+ struct buffer_head *m_bh;
+ int error;
+
+ error = gfs_glock_nq_init(m_ip->i_gl, LM_ST_EXCLUSIVE, GL_NOCACHE,
+ &gh);
+ if (error)
+ return error;
+
+ error = gfs_get_inode_buffer(m_ip, &m_bh);
+ if (error)
+ goto gfs_statfs_sync_out;
+
+ /* if no change, simply return */
+ spin_lock(&sdp->sd_statfs_spin);
+ gfs_statfs_change_in(m_sc, m_bh->b_data +
+ sizeof(struct gfs_dinode));
+ if (!l_sc->sc_total && !l_sc->sc_free && !l_sc->sc_dinodes) {
+ spin_unlock(&sdp->sd_statfs_spin);
+ goto out_bh;
+ }
+ spin_unlock(&sdp->sd_statfs_spin);
+
+ error = gfs_trans_begin(sdp, 1, 0);
+ if (error)
+ goto out_bh;
+
+ spin_lock(&sdp->sd_statfs_spin);
+ m_sc->sc_total += l_sc->sc_total;
+ m_sc->sc_free += l_sc->sc_free;
+ m_sc->sc_dinodes += l_sc->sc_dinodes;
+ memset(l_sc, 0, sizeof(struct gfs_statfs_change_host));
+ spin_unlock(&sdp->sd_statfs_spin);
+
+ gfs_trans_add_bh(m_ip->i_gl, m_bh);
+ gfs_statfs_change_out(m_sc, m_bh->b_data + sizeof(struct gfs_dinode));
+
+ gfs_trans_end(sdp);
+
+out_bh:
+ brelse(m_bh);
+
+gfs_statfs_sync_out:
+ gfs_glock_dq_uninit(&gh);
+ return error;
+}
+
+int gfs_statfs_fast(struct gfs_sbd *sdp, void *b)
+{
+ struct kstatfs *buf = (struct kstatfs *)b;
+ struct gfs_statfs_change_host sc, *m_sc = &sdp->sd_statfs_master;
+ struct gfs_statfs_change_host *l_sc = &sdp->sd_statfs_local;
+
+ spin_lock(&sdp->sd_statfs_spin);
+
+ sc.sc_total = m_sc->sc_total + l_sc->sc_total;
+ sc.sc_free = m_sc->sc_free + l_sc->sc_free;
+ sc.sc_dinodes = m_sc->sc_dinodes + l_sc->sc_dinodes;
+ spin_unlock(&sdp->sd_statfs_spin);
+
+ if (sc.sc_free < 0)
+ sc.sc_free = 0;
+ if (sc.sc_free > sc.sc_total)
+ sc.sc_free = sc.sc_total;
+ if (sc.sc_dinodes < 0)
+ sc.sc_dinodes = 0;
+
+ /* fill in the statistics */
+ memset(buf, 0, sizeof(struct kstatfs));
+
+ buf->f_type = GFS_MAGIC; buf->f_bsize = sdp->sd_sb.sb_bsize;
+ buf->f_blocks = sc.sc_total;
+ buf->f_bfree = sc.sc_free;
+ buf->f_bavail = sc.sc_free;
+ buf->f_files = sc.sc_dinodes + sc.sc_free;
+ buf->f_ffree = sc.sc_free;
+ buf->f_namelen = GFS_FNAMESIZE;
+
+ return 0;
+}
--- cluster/gfs-kernel/src/gfs/super.h 2006/10/13 19:57:07 1.4
+++ cluster/gfs-kernel/src/gfs/super.h 2007/06/19 21:26:11 1.5
@@ -37,10 +37,15 @@
int gfs_get_riinode(struct gfs_sbd *sdp);
int gfs_get_rootinode(struct gfs_sbd *sdp);
int gfs_get_qinode(struct gfs_sbd *sdp);
+int gfs_get_linode(struct gfs_sbd *sdp);
int gfs_make_fs_rw(struct gfs_sbd *sdp);
int gfs_make_fs_ro(struct gfs_sbd *sdp);
+int gfs_statfs_init(struct gfs_sbd *sdp, int flag);
+int gfs_statfs_sync(struct gfs_sbd *sdp);
+int gfs_statfs_fast(struct gfs_sbd *sdp, void *buf);
+
struct gfs_stat_gfs {
uint64_t sg_total_blocks;
uint64_t sg_free;
More information about the Cluster-devel
mailing list