I noticed when filling a 1T filesystem with 4 threads using the
fs_mark benchmark:
fs_mark -d /mnt/test -D 256 -n 100000 -t 4 -s 20480 -F -S 0
that I occasionally got checksum mismatch errors:
EXT4-fs error (device sdb): ext4_init_inode_bitmap: Checksum bad for group 6935
etc. I'd reliably get 4-5 of them during the run.
It appears that the problem is likely a race to init the bg's
when the uninit_bg feature is enabled.
With the patch below I was able to complete 2 runs with no errors
or warnings. However, I did hit a hang on one run that I can't yet
explain, so maybe this bears more inspection or testing.
Signed-off-by: Eric Sandeen <[email protected]>
---
Index: linux-2.6/fs/ext4/balloc.c
===================================================================
--- linux-2.6.orig/fs/ext4/balloc.c 2008-07-14 16:50:42.252353479 -0500
+++ linux-2.6/fs/ext4/balloc.c 2008-07-15 22:08:29.944291399 -0500
@@ -321,12 +321,15 @@ ext4_read_block_bitmap(struct super_bloc
if (bh_uptodate_or_lock(bh))
return bh;
+ spin_lock(sb_bgl_lock(EXT4_SB(sb), block_group));
if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
ext4_init_block_bitmap(sb, bh, block_group, desc);
set_buffer_uptodate(bh);
unlock_buffer(bh);
+ spin_unlock(sb_bgl_lock(EXT4_SB(sb), block_group));
return bh;
}
+ spin_unlock(sb_bgl_lock(EXT4_SB(sb), block_group));
if (bh_submit_read(bh) < 0) {
put_bh(bh);
ext4_error(sb, __func__,
Index: linux-2.6/fs/ext4/ialloc.c
===================================================================
--- linux-2.6.orig/fs/ext4/ialloc.c 2008-07-14 16:50:41.750354227 -0500
+++ linux-2.6/fs/ext4/ialloc.c 2008-07-15 22:09:21.682353674 -0500
@@ -105,6 +105,7 @@ read_inode_bitmap(struct super_block *sb
desc = ext4_get_group_desc(sb, block_group, NULL);
if (!desc)
goto error_out;
+ spin_lock(sb_bgl_lock(EXT4_SB(sb), block_group));
if (desc->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)) {
bh = sb_getblk(sb, ext4_inode_bitmap(sb, desc));
if (!buffer_uptodate(bh)) {
@@ -119,6 +120,7 @@ read_inode_bitmap(struct super_block *sb
} else {
bh = sb_bread(sb, ext4_inode_bitmap(sb, desc));
}
+ spin_unlock(sb_bgl_lock(EXT4_SB(sb), block_group));
if (!bh)
ext4_error(sb, "read_inode_bitmap",
"Cannot read inode bitmap - "
@@ -728,7 +730,7 @@ got:
/* When marking the block group with
* ~EXT4_BG_INODE_UNINIT we don't want to depend
- * on the value of bg_itable_unsed even though
+ * on the value of bg_itable_unused even though
* mke2fs could have initialized the same for us.
* Instead we calculated the value below
*/
Index: linux-2.6/fs/ext4/mballoc.c
===================================================================
--- linux-2.6.orig/fs/ext4/mballoc.c 2008-07-14 16:50:42.326353353 -0500
+++ linux-2.6/fs/ext4/mballoc.c 2008-07-15 22:10:06.249291399 -0500
@@ -787,13 +787,16 @@ static int ext4_mb_init_cache(struct pag
if (bh_uptodate_or_lock(bh[i]))
continue;
+ spin_lock(sb_bgl_lock(EXT4_SB(sb), first_group + i));
if (desc->bg_flags & cpu_to_le16(EXT4_BG_BLOCK_UNINIT)) {
ext4_init_block_bitmap(sb, bh[i],
first_group + i, desc);
set_buffer_uptodate(bh[i]);
unlock_buffer(bh[i]);
+ spin_unlock(sb_bgl_lock(EXT4_SB(sb), first_group + i));
continue;
}
+ spin_unlock(sb_bgl_lock(EXT4_SB(sb), first_group + i));
get_bh(bh[i]);
bh[i]->b_end_io = end_buffer_read_sync;
submit_bh(READ, bh[i]);
Index: linux-2.6/fs/ext4/super.c
===================================================================
--- linux-2.6.orig/fs/ext4/super.c 2008-07-14 16:50:41.775353495 -0500
+++ linux-2.6/fs/ext4/super.c 2008-07-15 22:19:02.395291074 -0500
@@ -1621,6 +1621,7 @@ static int ext4_check_descriptors(struct
"(block %llu)!", i, inode_table);
return 0;
}
+ spin_lock(sb_bgl_lock(sbi, i));
if (!ext4_group_desc_csum_verify(sbi, i, gdp)) {
printk(KERN_ERR "EXT4-fs: ext4_check_descriptors: "
"Checksum for group %lu failed (%u!=%u)\n",
@@ -1628,6 +1629,7 @@ static int ext4_check_descriptors(struct
gdp)), le16_to_cpu(gdp->bg_checksum));
return 0;
}
+ spin_unlock(sb_bgl_lock(sbi, i));
if (!flexbg_flag)
first_block += EXT4_BLOCKS_PER_GROUP(sb);
}
Eric Sandeen wrote:
> I noticed when filling a 1T filesystem with 4 threads using the
> fs_mark benchmark:
>
> fs_mark -d /mnt/test -D 256 -n 100000 -t 4 -s 20480 -F -S 0
>
> that I occasionally got checksum mismatch errors:
>
> EXT4-fs error (device sdb): ext4_init_inode_bitmap: Checksum bad for group 6935
>
> etc. I'd reliably get 4-5 of them during the run.
>
> It appears that the problem is likely a race to init the bg's
> when the uninit_bg feature is enabled.
>
> With the patch below I was able to complete 2 runs with no errors
> or warnings. However, I did hit a hang on one run that I can't yet
> explain, so maybe this bears more inspection or testing.
Crud hit it again, looks like it's my fault. So hold off on this one :)
-Eric
On Jul 16, 2008 00:26 -0500, Eric Sandeen wrote:
> @@ -105,6 +105,7 @@ read_inode_bitmap(struct super_block *sb
> desc = ext4_get_group_desc(sb, block_group, NULL);
> if (!desc)
> goto error_out;
> + spin_lock(sb_bgl_lock(EXT4_SB(sb), block_group));
> if (desc->bg_flags & cpu_to_le16(EXT4_BG_INODE_UNINIT)) {
> bh = sb_getblk(sb, ext4_inode_bitmap(sb, desc));
> if (!buffer_uptodate(bh)) {
sb_getblk() calls __getblk(), which is might_sleep() so is a no-no.
> @@ -119,6 +120,7 @@ read_inode_bitmap(struct super_block *sb
> } else {
> bh = sb_bread(sb, ext4_inode_bitmap(sb, desc));
> }
> + spin_unlock(sb_bgl_lock(EXT4_SB(sb), block_group));
Likewise "sb_bread" is doing disk access. I guess you don't have
CONFIG_DEBUG_SPINLOCK_SLEEP enabled in your kernel.
Cheers, Andreas
--
Andreas Dilger
Sr. Staff Engineer, Lustre Group
Sun Microsystems of Canada, Inc.