[PATCH]JBD2/JBD: race condition while writing updates to journal

[dingdinghua <dingdinghua85@gmail.com>]

At committing phase, we call jbd2_journal_write_metadata_buffer to
prepare log block's buffer_head, in this function, new_bh->b_data is set
to b_frozen_data or bh_in->b_data. We call "jbd_unlock_bh_state(bh_in)"
too early, since at this point , we haven't file bh_in to BJ_shadow list,
and we may set new_bh->b_data to bh_in->b_data, at this time, another
thread may call get write access of bh_in, modify bh_in->b_data and
dirty it. So , if new_bh->b_data is set to bh_in->b_data, the committing
transaction may flush the newly modified buffer content to disk,
preserve work done in jbd2_journal_get_write_access is useless. jbd also
has this problem.

here is the patch based on kernel version 2.6.30:

Signed-off-by: dingdinghua <dingdinghua85@gmail.com>
Acked-by: Jan Kara <jack@suse.cz>

---

--- fs/jbd2/journal.c.old 2009-06-21 16:18:18.000000000 +0800
+++ fs/jbd2/journal.c 2009-06-21 16:38:53.000000000 +0800
@@ -297,6 +297,8 @@ int jbd2_journal_write_metadata_buffer(t
unsigned int new_offset;
struct buffer_head *bh_in = jh2bh(jh_in);
struct jbd2_buffer_trigger_type *triggers;
+ journal_t *journal = transaction->t_journal;
+

/*
* The buffer really shouldn't be locked: only the current committing
@@ -310,6 +312,11 @@ int jbd2_journal_write_metadata_buffer(t
J_ASSERT_BH(bh_in, buffer_jbddirty(bh_in));

new_bh = alloc_buffer_head(GFP_NOFS|__GFP_NOFAIL);
+ /* keep subsequent assertions sane */
+ new_bh->b_state = 0;
+ init_buffer(new_bh, NULL, NULL);
+ atomic_set(&new_bh->b_count, 1);
+ new_jh = jbd2_journal_add_journal_head(new_bh); /* This sleeps */

/*
* If a new transaction has already done a buffer copy-out, then
@@ -388,14 +395,6 @@ repeat:
kunmap_atomic(mapped_data, KM_USER0);
}

- /* keep subsequent assertions sane */
- new_bh->b_state = 0;
- init_buffer(new_bh, NULL, NULL);
- atomic_set(&new_bh->b_count, 1);
- jbd_unlock_bh_state(bh_in);
-
- new_jh = jbd2_journal_add_journal_head(new_bh); /* This sleeps */
-
set_bh_page(new_bh, new_page, new_offset);
new_jh->b_transaction = NULL;
new_bh->b_size = jh2bh(jh_in)->b_size;
@@ -412,7 +411,11 @@ repeat:
* copying is moved to the transaction's shadow queue.
*/
JBUFFER_TRACE(jh_in, "file as BJ_Shadow");
- jbd2_journal_file_buffer(jh_in, transaction, BJ_Shadow);
+ spin_lock(&journal->j_list_lock);
+ __jbd2_journal_file_buffer(jh_in, transaction, BJ_Shadow);
+ spin_unlock(&journal->j_list_lock);
+ jbd_unlock_bh_state(bh_in);
+
JBUFFER_TRACE(new_jh, "file as BJ_IO");
jbd2_journal_file_buffer(new_jh, transaction, BJ_IO);

--- fs/jbd/journal.c.old 2009-06-21 16:27:37.000000000 +0800
+++ fs/jbd/journal.c 2009-06-21 16:45:06.000000000 +0800
@@ -287,6 +287,7 @@ int journal_write_metadata_buffer(transa
struct page *new_page;
unsigned int new_offset;
struct buffer_head *bh_in = jh2bh(jh_in);
+ journal_t *journal = transaction->t_journal;

/*
* The buffer really shouldn't be locked: only the current committing
@@ -300,6 +301,11 @@ int journal_write_metadata_buffer(transa
J_ASSERT_BH(bh_in, buffer_jbddirty(bh_in));

new_bh = alloc_buffer_head(GFP_NOFS|__GFP_NOFAIL);
+ /* keep subsequent assertions sane */
+ new_bh->b_state = 0;
+ init_buffer(new_bh, NULL, NULL);
+ atomic_set(&new_bh->b_count, 1);
+ new_jh = journal_add_journal_head(new_bh); /* This sleeps */

/*
* If a new transaction has already done a buffer copy-out, then
@@ -361,14 +367,6 @@ repeat:
kunmap_atomic(mapped_data, KM_USER0);
}

- /* keep subsequent assertions sane */
- new_bh->b_state = 0;
- init_buffer(new_bh, NULL, NULL);
- atomic_set(&new_bh->b_count, 1);
- jbd_unlock_bh_state(bh_in);
-
- new_jh = journal_add_journal_head(new_bh); /* This sleeps */