Patch to fix OCFS bug 73 and 74.

Hi,

After investigate in the JBD code in kernel 2.6.6 and OCFS2 svn
929 on IPF, I think I find the cause of dead lock bug 73 and 74.

In kernel 2.6.6, the bit 12 to 19 of variable b_state in
structure buffer_head is occupied by the linux journaling code as follows.

enum jbd_state_bits {
BH_JBD /* Has an attached ext3 journal_head */
= BH_PrivateStart,
BH_JWrite, /* Being written to log (@@@ DEBUGGING)
*/

        BH_Freed,               /* Has been freed (truncated) */
        BH_Revoked,             /* Has been revoked from the log */
        BH_RevokeValid,         /* Revoked flag is valid */
        BH_JBDDirty,            /* Is dirty but journaled */
        BH_State,               /* Pins most journal_head state */
        BH_JournalHead,         /* Pins bh->b_private and jh->b_bh */

};

But, in ocfs2 head io.h, variable BH_Modified and the inode
clean sequence number also occupy some bits in this frame as follows.

#define BH_Modified 18
#define STATE_BIT_MASK ((~0UL)<<19)

The bit 18 and bit 19 conflict with Linux journaling code in
kernel 2.6.6. When copy a file to the OCFS2 volume, the journaling code will try to lock bit BH_JournalHead and BH_State of variable b_state in buffer_head, which have already been set to 1 in OCFS2 routine ocfs_read_dirnode() and ocfs_insert_dir_node(). So, the dead locks occur at journaling routine jbd_lock_bh_journal_head() and jbd_lock_bh_state().

My patch fixes these 2 bugs by using the empty bits 20 ~ 31. It
is against svn 932. Please refer to the attachment.

Any comment?

------------------------------------------------------------------------
-----------

@@ -75,7 +75,7 @@
 
 #define OCFS_NONCACHED(osb,off)  ((off) <

(osb)->vol_layout.data_start_off)

-#define BH_Modified 18
+#define BH_Modified 20

#if LINUX_VERSION_CODE >= KERNEL_VERSION(2,6,0)
/*
@@ -214,10 +214,10 @@
ocfs_bh_sem_unlock(bh);
}

-
+#define JBD_STATE_BIT_MAX 21
#define STATE_BIT_MAX (1 << 12)
#define STATE_BIT_MAX_MASK ((1 << 12)-1)
-#define STATE_BIT_MASK ((~0UL)<<19)
+#define STATE_BIT_MASK ((~0UL)<<JBD_STATE_BIT_MAX)

static inline void CLEAR_BH_SEQNUM(struct buffer_head *bh)
@@ -229,7 +229,7 @@
static inline void SET_BH_SEQNUM(struct inode *inode, struct
buffer_head *bh)
{
unsigned int seq = (atomic_read(GET_INODE_CLEAN_SEQ(inode)) &
- STATE_BIT_MAX_MASK) << 19;
+ STATE_BIT_MAX_MASK) << JBD_STATE_BIT_MAX;

CLEAR_BH_SEQNUM(bh);
bh->b_state |= seq;
@@ -238,7 +238,7 @@
static inline int TEST_BH_SEQNUM(struct inode *inode, struct
buffer_head *bh)
{
int ret;
- unsigned int seq = (bh->b_state & STATE_BIT_MASK) >> 19;
+ unsigned int seq = (bh->b_state & STATE_BIT_MASK) >>
JBD_STATE_BIT_MAX;

ret = (seq == atomic_read(GET_INODE_CLEAN_SEQ(inode))); return ret;

*********************************************
Sonic Zhang
Software Engineer
Intel China Software Lab
Tel: (086)021-52574545-1667
iNet: 752-1667
*********************************************

-------------- next part --------------
A non-text attachment was scrubbed...
Name: ocfs2-jbd-statebit.patch
Type: application/octet-stream
Size: 1274 bytes
Desc: ocfs2-jbd-statebit.patch
Url : http://oss.oracle.com/pipermail/ocfs2-devel/attachments/20040525/895dece7/ocfs2-jbd-statebit.obj

Yeah, we saw that already, and I've managed to fail to commit
the change. Anyway, this is something that cannot depend on JBD numbers or anything, we just have to be vigilant wrt BH_flags. I'll commit.

Joel

"Win95 file and print sharing are for relatively friendly nets."
	- Paul Leach, Microsoft

Joel Becker
Senior Member of Technical Staff
Oracle Corporation
E-mail: [email protected]
Phone: (650) 506-8127