Syzbot has reported a potential hang in nilfs_detach_log_writer()
called during nilfs2 unmount.
Analysis revealed that this is because nilfs_segctor_sync(), which
synchronizes with the log writer thread, can be called after
nilfs_segctor_destroy() terminates that thread, as shown in the call
trace below:
nilfs_detach_log_writer
nilfs_segctor_destroy
nilfs_segctor_kill_thread --> Shut down log writer thread
flush_work
nilfs_iput_work_func
nilfs_dispose_list
iput
nilfs_evict_inode
nilfs_transaction_commit
nilfs_construct_segment (if inode needs sync)
nilfs_segctor_sync --> Attempt to synchronize with
log writer thread
*** DEADLOCK ***
Fix this issue by changing nilfs_segctor_sync() so that the log writer
thread returns normally without synchronizing after it terminates, and
by forcing tasks that are already waiting to complete once after the
thread terminates.
The skipped inode metadata flushout will then be processed together in
the subsequent cleanup work in nilfs_segctor_destroy().
Signed-off-by: Ryusuke Konishi <[email protected]>
Reported-by: [email protected]
Closes: https://syzkaller.appspot.com/bug?extid=e3973c409251e136fdd0
Tested-by: Ryusuke Konishi <[email protected]>
Cc: [email protected]
---
fs/nilfs2/segment.c | 21 ++++++++++++++++++---
1 file changed, 18 insertions(+), 3 deletions(-)
diff --git a/fs/nilfs2/segment.c b/fs/nilfs2/segment.c
index 99c78a49e432..c27f0daec9af 100644
--- a/fs/nilfs2/segment.c
+++ b/fs/nilfs2/segment.c
@@ -2190,6 +2190,14 @@ static int nilfs_segctor_sync(struct nilfs_sc_info *sci)
for (;;) {
set_current_state(TASK_INTERRUPTIBLE);
+ /*
+ * Synchronize only while the log writer thread is alive.
+ * Leave flushing out after the log writer thread exits to
+ * the cleanup work in nilfs_segctor_destroy().
+ */
+ if (!sci->sc_task)
+ break;
+
if (atomic_read(&wait_req.done)) {
err = wait_req.err;
break;
@@ -2205,7 +2213,7 @@ static int nilfs_segctor_sync(struct nilfs_sc_info *sci)
return err;
}
-static void nilfs_segctor_wakeup(struct nilfs_sc_info *sci, int err)
+static void nilfs_segctor_wakeup(struct nilfs_sc_info *sci, int err, bool force)
{
struct nilfs_segctor_wait_request *wrq, *n;
unsigned long flags;
@@ -2213,7 +2221,7 @@ static void nilfs_segctor_wakeup(struct nilfs_sc_info *sci, int err)
spin_lock_irqsave(&sci->sc_wait_request.lock, flags);
list_for_each_entry_safe(wrq, n, &sci->sc_wait_request.head, wq.entry) {
if (!atomic_read(&wrq->done) &&
- nilfs_cnt32_ge(sci->sc_seq_done, wrq->seq)) {
+ (force || nilfs_cnt32_ge(sci->sc_seq_done, wrq->seq))) {
wrq->err = err;
atomic_set(&wrq->done, 1);
}
@@ -2362,7 +2370,7 @@ static void nilfs_segctor_notify(struct nilfs_sc_info *sci, int mode, int err)
if (mode == SC_LSEG_SR) {
sci->sc_state &= ~NILFS_SEGCTOR_COMMIT;
sci->sc_seq_done = sci->sc_seq_accepted;
- nilfs_segctor_wakeup(sci, err);
+ nilfs_segctor_wakeup(sci, err, false);
sci->sc_flush_request = 0;
} else {
if (mode == SC_FLUSH_FILE)
@@ -2746,6 +2754,13 @@ static void nilfs_segctor_destroy(struct nilfs_sc_info *sci)
|| sci->sc_seq_request != sci->sc_seq_done);
spin_unlock(&sci->sc_state_lock);
+ /*
+ * Forcibly wake up tasks waiting in nilfs_segctor_sync(), which can
+ * be called from delayed iput() via nilfs_evict_inode() and can race
+ * with the above log writer thread termination.
+ */
+ nilfs_segctor_wakeup(sci, 0, true);
+
if (flush_work(&sci->sc_iput_work))
flag = true;
--
2.34.1