[v2] ocfs2: fix deadlock on mmapped page in ocfs2_write_begin_nolock()

Commit Message

Zhen Ren Sept. 18, 2016, 4:45 a.m. UTC

The testcase "mmaptruncate" of ocfs2-test deadlocks occasionally.

In this testcase, we create a 2*CLUSTER_SIZE file and mmap() on it;
there are 2 process repeatedly performing the following operations
respectively: one is doing memset(mmaped_addr + 2*CLUSTER_SIZE - 1,
'a', 1), while the another is playing ftruncate(fd, 2*CLUSTER_SIZE)
and then ftruncate(fd, CLUSTER_SIZE) again and again.

This is the backtrace when the deadlock happens:
[<ffffffff817054f0>] __wait_on_bit_lock+0x50/0xa0
[<ffffffff81199bd7>] __lock_page+0xb7/0xc0
[<ffffffff810c4de0>] ? autoremove_wake_function+0x40/0x40
[<ffffffffa0440f4f>] ocfs2_write_begin_nolock+0x163f/0x1790 [ocfs2]
[<ffffffffa0462a50>] ? ocfs2_allocate_extend_trans+0x180/0x180 [ocfs2]
[<ffffffffa0467b47>] ocfs2_page_mkwrite+0x1c7/0x2a0 [ocfs2]
[<ffffffff811cf286>] do_page_mkwrite+0x66/0xc0
[<ffffffff811d3635>] handle_mm_fault+0x685/0x1350
[<ffffffff81039dc0>] ? __fpu__restore_sig+0x70/0x530
[<ffffffff810694c8>] __do_page_fault+0x1d8/0x4d0
[<ffffffff81069827>] trace_do_page_fault+0x37/0xf0
[<ffffffff81061e69>] do_async_page_fault+0x19/0x70
[<ffffffff8170ac98>] async_page_fault+0x28/0x30

In ocfs2_write_begin_nolock(), we first grab the pages and then
allocate disk space for this write; ocfs2_try_to_free_truncate_log()
will be called if -ENOSPC is returned; if we're lucky to get enough clusters,
which is usually the case, we start over again. But in ocfs2_free_write_ctxt()
the target page isn't unlocked, so we will deadlock when trying to grab
the target page again.

Also, -ENOMEM might be returned in ocfs2_grab_pages_for_write(). Another
deadlock will happen in __do_page_mkwrite() if ocfs2_page_mkwrite()
returns non-VM_FAULT_LOCKED, and along with a locked target page.

These two errors fail on the same path, so fix them by unlocking the target
page manually before ocfs2_free_write_ctxt().

Jan Kara helps me clear out the JBD2 part, and suggest the hint for root cause.

Changes since v1:
1. Also put ENOMEM error case into consideration.

Signed-off-by: Eric Ren <[e-mail]>
Reviewed-by: He Gang <[e-mail]>
Acked-by: Joseph Qi <[e-mail]>
---
 fs/ocfs2/aops.c | 10 ++++++++++
 1 file changed, 10 insertions(+)

Patch

diff --git a/fs/ocfs2/aops.c b/fs/ocfs2/aops.c
index 98d3654..bbb4b3e 100644
--- a/fs/ocfs2/aops.c
+++ b/fs/ocfs2/aops.c
@@ -1842,6 +1842,16 @@  int ocfs2_write_begin_nolock(struct address_space *mapping,
 	ocfs2_commit_trans(osb, handle);
 
 out:
+	/*
+	 * The mmapped page won't be unlocked in ocfs2_free_write_ctxt(),
+	 * even in case of error here like ENOSPC and ENOMEM. So, we need
+	 * to unlock the target page manually to prevent deadlocks when
+	 * retrying again on ENOSPC, or when returning non-VM_FAULT_LOCKED
+	 * to VM code.
+	 */
+	if (wc->w_target_locked)
+		unlock_page(mmap_page);
+
 	ocfs2_free_write_ctxt(inode, wc);
 
 	if (data_ac) {