[PATCH v3] ocfs2/heartbeat: serialize concurrent writes to a region's dev attribute

[Eklavya Singh <eklavyasingh2106@gmail.com>]

o2hb_region_dev_store() checks reg->hr_bdev_file to see if the region
is already started, but nothing stops two writers from doing that at
the same time. configfs only serializes writes through the same open
file, so two processes that each open the dev attribute can both get
past the check:

  A: hr_bdev_file is NULL, opens file A, sets hr_bdev_file = A
  B: hr_bdev_file is NULL, opens file B, sets hr_bdev_file = B
  A: start fails, fput(hr_bdev_file) drops B, sets it to NULL
  B: start fails, fput(NULL)

Both calls also use the same hr_slots, hr_slot_data and hr_tmp_block,
so the error paths can free the slot data twice.

syzbot hit the fput(NULL) with an unbound loop device, where every
start fails with -EIO. With four threads doing the same thing on a
KASAN kernel I get both crashes within a second:

  BUG: KASAN: double-free in o2hb_unmap_slot_data+0x14d/0x2f0
  BUG: kernel NULL pointer dereference, address: 0000000000000158

Add a mutex to the region and hold it in dev_store(), in the
block_bytes, start_block and blocks stores, and in dev_show(), since
they all look at hr_bdev_file. dev_store() cleans up with goto, so it
takes the mutex in a small wrapper rather than with guard(). The mutex
is dropped while dev_store() waits for the heartbeat to become steady,
which takes seconds, so a read of dev doesn't stall behind a start.
hr_bdev_file is already set at that point, so other writers still get
-EINVAL, and the cleanup after the wait retakes the mutex.

With the patch the same test ran for 10 minutes with lockdep on and
nothing was reported. So did a second one, with a working loop device,
that interrupts starts during the wait while other threads read and
write the attributes.

Fixes: a7f6a5fb4bde ("[PATCH] OCFS2: The Second Oracle Cluster Filesystem")
Reported-by: syzbot+5d4e96acd0febccb5e87@syzkaller.appspotmail.com
Closes: https://syzkaller.appspot.com/bug?extid=5d4e96acd0febccb5e87
Cc: stable@vger.kernel.org
Signed-off-by: Eklavya Singh <eklavyasingh2106@gmail.com>
---
Changes in v3:
- rebase onto linux-next
- drop hr_dev_mutex while waiting for the heartbeat to become steady,
  so a read of dev that races a start no longer blocks (Joseph)
- name the four attributes in the hr_dev_mutex comment and assert the
  mutex is held in __o2hb_region_dev_store() (Joseph)

Changes in v2:
- take the mutex in a wrapper around dev_store() instead of with
  guard(), since dev_store() uses goto for its cleanup

Deepanshu Kartikey [1] and Cen Zhang [2] have posted patches for the
same race, and they conflict with this one. This version also takes
the mutex in dev_show(), which otherwise can read hr_bdev_file while a
failed start is dropping it.

[1] https://lore.kernel.org/all/20261020085100.8819-1-kartikey406@gmail.com/
[2] https://lore.kernel.org/all/pm-ocfs2-objects-candidate-0077-v7-34980443ec254b2a1ffc@gmail.com/

v2: https://lore.kernel.org/all/20261019072936.2279-1-eklavyasingh2106@gmail.com/
v1: https://lore.kernel.org/all/20261018163121.128911-1-eklavyasingh2106@gmail.com/

 fs/ocfs2/cluster/heartbeat.c | 40 ++++++++++++++++++++++++++++++------
 1 file changed, 34 insertions(+), 6 deletions(-)

diff --git a/fs/ocfs2/cluster/heartbeat.c b/fs/ocfs2/cluster/heartbeat.c
index b45adc3..7a18e71 100644
--- a/fs/ocfs2/cluster/heartbeat.c
+++ b/fs/ocfs2/cluster/heartbeat.c
@@ -269,6 +269,8 @@ struct o2hb_region {
 	struct list_head	hr_handler_list;
 	/* Serializes timeout arming against failed-start and teardown. */
 	struct mutex		hr_arming_mutex;
+	/* Serializes the dev, block_bytes, start_block and blocks attributes. */
+	struct mutex		hr_dev_mutex;
 	bool			hr_stopping;
 
 	/* last hb status, 0 for success, other value for error. */
@@ -1681,6 +1683,7 @@ static ssize_t o2hb_region_block_bytes_store(struct config_item *item,
 	unsigned long block_bytes;
 	unsigned int block_bits;
 
+	guard(mutex)(&reg->hr_dev_mutex);
 	if (reg->hr_bdev_file)
 		return -EINVAL;
 
@@ -1710,6 +1713,7 @@ static ssize_t o2hb_region_start_block_store(struct config_item *item,
 	char *p = (char *)page;
 	ssize_t ret;
 
+	guard(mutex)(&reg->hr_dev_mutex);
 	if (reg->hr_bdev_file)
 		return -EINVAL;
 
@@ -1736,6 +1740,7 @@ static ssize_t o2hb_region_blocks_store(struct config_item *item,
 	char *p = (char *)page;
 	int ret;
 
+	guard(mutex)(&reg->hr_dev_mutex);
 	if (reg->hr_bdev_file)
 		return -EINVAL;
 
@@ -1753,10 +1758,12 @@ static ssize_t o2hb_region_blocks_store(struct config_item *item,
 
 static ssize_t o2hb_region_dev_show(struct config_item *item, char *page)
 {
+	struct o2hb_region *reg = to_o2hb_region(item);
 	unsigned int ret = 0;
 
-	if (to_o2hb_region(item)->hr_bdev_file)
-		ret = sprintf(page, "%pg\n", reg_bdev(to_o2hb_region(item)));
+	guard(mutex)(&reg->hr_dev_mutex);
+	if (reg->hr_bdev_file)
+		ret = sprintf(page, "%pg\n", reg_bdev(reg));
 
 	return ret;
 }
@@ -1872,11 +1879,9 @@ out:
  * this is acting as commit; we set up all of hr_bdev_file and hr_task or
  * nothing
  */
-static ssize_t o2hb_region_dev_store(struct config_item *item,
-				     const char *page,
-				     size_t count)
+static ssize_t __o2hb_region_dev_store(struct o2hb_region *reg,
+				       const char *page, size_t count)
 {
-	struct o2hb_region *reg = to_o2hb_region(item);
 	struct task_struct *hb_task;
 	long fd;
 	int sectsize;
@@ -1884,6 +1889,8 @@ static ssize_t o2hb_region_dev_store(struct config_item *item,
 	ssize_t ret = -EINVAL;
 	int live_threshold;
 
+	lockdep_assert_held(&reg->hr_dev_mutex);
+
 	if (reg->hr_bdev_file)
 		return -EINVAL;
 
@@ -1987,9 +1994,15 @@ static ssize_t o2hb_region_dev_store(struct config_item *item,
 
 	wake_up_process(hb_task);
 
+	/*
+	 * Don't hold hr_dev_mutex across the wait, which can take seconds.
+	 * hr_bdev_file is already set, so other writers still get -EINVAL.
+	 */
+	mutex_unlock(&reg->hr_dev_mutex);
 	ret = wait_event_interruptible(o2hb_steady_queue,
 				atomic_read(&reg->hr_steady_iterations) == 0 ||
 				reg->hr_node_deleted);
+	mutex_lock(&reg->hr_dev_mutex);
 	if (ret) {
 		atomic_set(&reg->hr_steady_iterations, 0);
 		reg->hr_aborted_start = 1;
@@ -2041,6 +2054,20 @@ out:
 	return ret;
 }
 
+static ssize_t o2hb_region_dev_store(struct config_item *item,
+				     const char *page,
+				     size_t count)
+{
+	struct o2hb_region *reg = to_o2hb_region(item);
+	ssize_t ret;
+
+	mutex_lock(&reg->hr_dev_mutex);
+	ret = __o2hb_region_dev_store(reg, page, count);
+	mutex_unlock(&reg->hr_dev_mutex);
+
+	return ret;
+}
+
 static ssize_t o2hb_region_pid_show(struct config_item *item, char *page)
 {
 	struct o2hb_region *reg = to_o2hb_region(item);
@@ -2143,6 +2170,7 @@ static struct config_item *o2hb_heartbeat_group_make_item(struct config_group *g
 
 	INIT_LIST_HEAD(&reg->hr_all_item);
 	mutex_init(&reg->hr_arming_mutex);
+	mutex_init(&reg->hr_dev_mutex);
 	reg->hr_stopping = true;
 	INIT_DELAYED_WORK(&reg->hr_write_timeout_work, o2hb_write_timeout);
 	INIT_DELAYED_WORK(&reg->hr_nego_timeout_work, o2hb_nego_timeout);
-- 
2.52.0