On Mar 29, 2007 14:17 +0200, Vincent Caron wrote:
> I just noticed that 'tune2fs -l' did not returned a "lively" updated
> information regarding the free inodes count (looks like it's always
> correct after unmounting).
This is a bit of a defect in all 2.6 kernels. They never update the
on disk superblock free blocks/inodes information to avoid lock contention,
even if this info is available.
Can you please give the following patch a try? It fixes this issue,
and also makes statfs MUCH more efficient for large filesystems, because
the filesystem overhead is constant unless the filesystem size changes
and checking that for 16k groups is slow (hence hack to add cond_resched()
instead of fixing problem correctly). It has not been tested much, but
is very straight forward.
Only the last part is strictly necessary to fix your particular problem
(setting of es->s_free_inodes_count and es->s_free_blocks_count). This
is lazy, in the sense that you need a "statfs" to update the count, and
then a truncate or unlink or rmdir in order to dirty the superblock to
flush it to disk. However, it will be correct in the buffer cache, and
it is a lot better than what we have now. We don't want a non-lazy version
anyways, because of performance.
Signed-off-by: Andreas Dilger <[email protected]>
======================= ext3-statfs-2.6.20.diff ==========================
Index: linux-stage/fs/ext3/super.c
===================================================================
--- linux-stage.orig/fs/ext3/super.c 2007-03-22 17:29:30.000000000 -0600
+++ linux-stage/fs/ext3/super.c 2007-03-23 01:48:41.000000000 -0600
@@ -2389,19 +2389,22 @@ restore_opts:
struct super_block *sb = dentry->d_sb;
struct ext3_sb_info *sbi = EXT3_SB(sb);
struct ext3_super_block *es = sbi->s_es;
- ext3_fsblk_t overhead;
- int i;
+ static ext3_fsblk_t overhead_last;
+ static __le32 blocks_last;
u64 fsid;
- if (test_opt (sb, MINIX_DF))
- overhead = 0;
- else {
- unsigned long ngroups;
- ngroups = EXT3_SB(sb)->s_groups_count;
+ if (test_opt (sb, MINIX_DF)) {
+ overhead_last = 0;
+ } else if (blocks_last != es->s_blocks_count) {
+ unsigned long ngroups = sbi->s_groups_count, group, metabg = ~0;
+ unsigned three = 1, five = 5, seven = 7;
+ ext3_fsblk_t overhead = 0;
smp_rmb();
/*
- * Compute the overhead (FS structures)
+ * Compute the overhead (FS structures). This is constant
+ * for a given filesystem unless the number of block groups
+ * changes so we cache the previous value until it does.
*/
/*
@@ -2419,28 +2422,43 @@ static int ext3_statfs (struct super_blo
* block group descriptors. If the sparse superblocks
* feature is turned on, then not all groups have this.
*/
- for (i = 0; i < ngroups; i++) {
- overhead += ext3_bg_has_super(sb, i) +
- ext3_bg_num_gdb(sb, i);
- cond_resched();
- }
+ overhead += 1 + sbi->s_gdb_count +
+ le16_to_cpu(es->s_reserved_gdt_blocks); /* group 0 */
+ if (EXT3_HAS_INCOMPAT_FEATURE(sb,
+ EXT3_FEATURE_INCOMPAT_META_BG)) {
+ metabg = le32_to_cpu(es->s_first_meta_bg) *
+ sbi->s_desc_per_block;
+ group = ngroups - metabg;
+ overhead += (group + 1) / sbi->s_desc_per_block * 3 +
+ ((group%sbi->s_desc_per_block)>= 2?2:(group%2));
+ }
+
+ while ((group = ext3_list_backups(sb, &three, &five, &seven)) <
+ ngroups) /* sb + group descriptors backups */
+ overhead += 1 +(group >= metabg ? 0 : sbi->s_gdb_count +
+ le16_to_cpu(es->s_reserved_gdt_blocks));
/*
* Every block group has an inode bitmap, a block
* bitmap, and an inode table.
*/
- overhead += (ngroups * (2 + EXT3_SB(sb)->s_itb_per_group));
+ overhead += ngroups * (2 + sbi->s_itb_per_group);
+ overhead_last = overhead;
+ smp_wmb();
+ blocks_last = es->s_blocks_count;
}
buf->f_type = EXT3_SUPER_MAGIC;
buf->f_bsize = sb->s_blocksize;
- buf->f_blocks = le32_to_cpu(es->s_blocks_count) - overhead;
+ buf->f_blocks = le32_to_cpu(es->s_blocks_count) - overhead_last;
buf->f_bfree = percpu_counter_sum(&sbi->s_freeblocks_counter);
+ es->s_free_blocks_count = cpu_to_le32(buf->f_bfree);
buf->f_bavail = buf->f_bfree - le32_to_cpu(es->s_r_blocks_count);
if (buf->f_bfree < le32_to_cpu(es->s_r_blocks_count))
buf->f_bavail = 0;
buf->f_files = le32_to_cpu(es->s_inodes_count);
buf->f_ffree = percpu_counter_sum(&sbi->s_freeinodes_counter);
+ es->s_free_inodes_count = cpu_to_le32(buf->f_ffree);
buf->f_namelen = EXT3_NAME_LEN;
fsid = le64_to_cpup((void *)es->s_uuid) ^
le64_to_cpup((void *)es->s_uuid + sizeof(u64));
Index: linux-stage/fs/ext3/resize.c
===================================================================
--- linux-stage.orig/fs/ext3/resize.c 2007-03-22 17:29:30.000000000 -0600
+++ linux-stage/fs/ext3/resize.c 2007-03-23 01:16:38.000000000 -0600
@@ -292,8 +292,8 @@ exit_journal:
* sequence of powers of 3, 5, and 7: 1, 3, 5, 7, 9, 25, 27, 49, 81, ...
* For a non-sparse filesystem it will be every group: 1, 2, 3, 4, ...
*/
-static unsigned ext3_list_backups(struct super_block *sb, unsigned *three,
- unsigned *five, unsigned *seven)
+unsigned ext3_list_backups(struct super_block *sb, unsigned *three,
+ unsigned *five, unsigned *seven)
{
unsigned *min = three;
int mult = 3;
Index: linux-stage/include/linux/ext3_fs.h
===================================================================
--- linux-stage.orig/include/linux/ext3_fs.h 2007-03-22 17:29:30.000000000 -0600
+++ linux-stage/include/linux/ext3_fs.h 2007-03-23 00:41:22.000000000 -0600
@@ -846,6 +846,8 @@ extern int ext3_group_add(struct super_b
extern int ext3_group_extend(struct super_block *sb,
struct ext3_super_block *es,
ext3_fsblk_t n_blocks_count);
+extern unsigned ext3_list_backups(struct super_block *sb, unsigned *three,
+ unsigned *five, unsigned *seven);
/* super.c */
extern void ext3_error (struct super_block *, const char *, const char *, ...)
Cheers, Andreas
--
Andreas Dilger
Principal Software Engineer
Cluster File Systems, Inc.
On jeu, 2007-03-29 at 13:59 -0600, Andreas Dilger wrote:
> On Mar 29, 2007 14:17 +0200, Vincent Caron wrote:
> > I just noticed that 'tune2fs -l' did not returned a "lively" updated
> > information regarding the free inodes count (looks like it's always
> > correct after unmounting).
>
> This is a bit of a defect in all 2.6 kernels. They never update the
> on disk superblock free blocks/inodes information to avoid lock contention,
> even if this info is available.
It turns out it is okay in my case since 'df -i' reports correct
numbers.
> Can you please give the following patch a try? It fixes this issue,
> and also makes statfs MUCH more efficient for large filesystems, because
> the filesystem overhead is constant unless the filesystem size changes
> and checking that for 16k groups is slow (hence hack to add cond_resched()
> instead of fixing problem correctly). It has not been tested much, but
> is very straight forward.
>
> Only the last part is strictly necessary to fix your particular problem
> (setting of es->s_free_inodes_count and es->s_free_blocks_count). This
> is lazy, in the sense that you need a "statfs" to update the count, and
> then a truncate or unlink or rmdir in order to dirty the superblock to
> flush it to disk. However, it will be correct in the buffer cache, and
> it is a lot better than what we have now. We don't want a non-lazy version
> anyways, because of performance.
Unfortunately the problem shows on a production machine and I don't
have a similar one to test properly (it's a heavy-loaded filer).
BTW, if ondisk superblocks are not updated until specific events occur
(umount, statfs), what is the consequence of a system crash ? Does
journalization come into play (superblock=metadata?), does fsck fixes
figures from other ondisk structures ? Just being curious...
On Apr 03, 2007 15:55 +0200, Vincent Caron wrote:
> BTW, if ondisk superblocks are not updated until specific events occur
> (umount, statfs), what is the consequence of a system crash ? Does
> journalization come into play (superblock=metadata?), does fsck fixes
> figures from other ondisk structures ? Just being curious...
The free blocks/inodes count in the superblock aren't really used for
anything these days. At mount time the kernel sums up the free inode
and block counts from the group descriptors to put into the percpu
counters, and only the group descriptors are kept uptodate.
Cheers, Andreas
--
Andreas Dilger
Principal Software Engineer
Cluster File Systems, Inc.