1 Date: Mon, 23 Oct 2006 15:40:48 -0500
2 From: Eric Sandeen <sandeen@redhat.com>
3 Subject: [PATCH RHEL5] handle races w/ truncate in journal_dirty_data()
5 This is for BZ 209647 <https://bugzilla.redhat.com/bugzilla/show_bug.cgi?id=209647>: ext3/jbd panic
7 This patch is now in -mm.
9 When running several fsx's and other filesystem stress tests, we found
10 cases where an unmapped buffer was still being sent to submit_bh by the
11 ext3 dirty data journaling code.
13 I saw this happen in two ways, both related to another thread doing a
14 truncate which would unmap the buffer in question.
16 Either we would get into journal_dirty_data with a bh which was already
17 unmapped (although journal_dirty_data_fn had checked for this earlier, the
18 state was not locked at that point), or it would get unmapped in the middle
19 of journal_dirty_data when we dropped locks to call sync_dirty_buffer.
21 By re-checking for mapped state after we've acquired the bh state lock, we
22 should avoid these races. If we find a buffer which is no longer mapped,
23 we essentially ignore it, because journal_unmap_buffer has already decided
24 that this buffer can go away.
26 I've also added tracepoints in these two cases, and made a couple other
27 tracepoint changes that I found useful in debugging this.
29 Signed-off-by: Eric Sandeen <esandeen@redhat.com>
30 Cc: <linux-ext4@vger.kernel.org>
31 Signed-off-by: Andrew Morton <akpm@osdl.org>
34 fs/jbd/transaction.c | 15 ++++++++++++++-
35 1 file changed, 14 insertions(+), 1 deletion(-)
37 Index: linux-2.6.18-1.2732.el5/fs/jbd/transaction.c
38 ===================================================================
39 --- linux-2.6.18-1.2732.el5.orig/fs/jbd/transaction.c
40 +++ linux-2.6.18-1.2732.el5/fs/jbd/transaction.c
41 @@ -967,6 +967,13 @@ int journal_dirty_data(handle_t *handle,
43 jbd_lock_bh_state(bh);
44 spin_lock(&journal->j_list_lock);
46 + /* Now that we have bh_state locked, are we really still mapped? */
47 + if (!buffer_mapped(bh)) {
48 + JBUFFER_TRACE(jh, "unmapped buffer, bailing out");
52 if (jh->b_transaction) {
53 JBUFFER_TRACE(jh, "has transaction");
54 if (jh->b_transaction != handle->h_transaction) {
55 @@ -1028,6 +1035,11 @@ int journal_dirty_data(handle_t *handle,
56 sync_dirty_buffer(bh);
57 jbd_lock_bh_state(bh);
58 spin_lock(&journal->j_list_lock);
59 + /* Since we dropped the lock... */
60 + if (!buffer_mapped(bh)) {
61 + JBUFFER_TRACE(jh, "buffer got unmapped");
64 /* The buffer may become locked again at any
65 time if it is redirtied */
67 @@ -1823,6 +1835,7 @@ static int journal_unmap_buffer(journal_
70 } else if (transaction == journal->j_committing_transaction) {
71 + JBUFFER_TRACE(jh, "on committing transaction");
72 if (jh->b_jlist == BJ_Locked) {
74 * The buffer is on the committing transaction's locked
75 @@ -1837,7 +1850,6 @@ static int journal_unmap_buffer(journal_
76 * can remove it's next_transaction pointer from the
77 * running transaction if that is set, but nothing
79 - JBUFFER_TRACE(jh, "on committing transaction");
81 if (jh->b_next_transaction) {
82 J_ASSERT(jh->b_next_transaction ==
83 @@ -1857,6 +1869,7 @@ static int journal_unmap_buffer(journal_
84 * i_size already for this truncate so recovery will not
85 * expose the disk blocks we are discarding here.) */
86 J_ASSERT_JH(jh, transaction == journal->j_running_transaction);
87 + JBUFFER_TRACE(jh, "on running transaction");
88 may_free = __dispose_buffer(jh, transaction);