[PATCH v2 1/3] writeback: fix race between cgroup_writeback_umount() and inode_switch_wbs()

[Baokun Li <libaokun@linux.alibaba.com>]

When a container exits, the following BUG_ON() is occasionally triggered:

==================================================================
 VFS: Busy inodes after unmount of sdb (ext4)
 ------------[ cut here ]------------
 kernel BUG at fs/super.c:695!
 CPU: 3 PID: 6 Comm: containerd-shim Tainted: G OE K 6.6 #1
 pstate: 63400009 (nZCv daif +PAN -UAO +TCO +DIT -SSBS BTYPE=--)
 pc : generic_shutdown_super+0xf0/0x100
 lr : generic_shutdown_super+0xf0/0x100
 Call trace:
  generic_shutdown_super+0xf0/0x100
  kill_block_super+0x20/0x48
  ext4_kill_sb+0x28/0x60
  deactivate_locked_super+0x54/0x130
  deactivate_super+0x84/0xa0
  cleanup_mnt+0xa4/0x140
  __cleanup_mnt+0x18/0x28
  task_work_run+0x78/0xe0
  do_notify_resume+0x204/0x240
==================================================================

The root cause is a race between cgroup_writeback_umount() and
inode_switch_wbs()/cleanup_offline_cgwb(). There is a window between
inode_prepare_wbs_switch() returning true and the subsequent
wb_queue_isw() call. Following is the process that triggers the issue:

      CPU A (umount)           |          CPU B (writeback)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
                                 inode_switch_wbs/cleanup_offline_cgwb
                                  atomic_inc(&isw_nr_in_flight)
                                  inode_prepare_wbs_switch
                                   -> passes SB_ACTIVE check
                                   __iget(inode)
 generic_shutdown_super
  sb->s_flags &= ~SB_ACTIVE
  cgroup_writeback_umount(sb)
   smp_mb()
   atomic_read(&isw_nr_in_flight)
   rcu_barrier()
    -> no pending RCU callbacks
   flush_workqueue(isw_wq)
    -> nothing queued, returns
  evict_inodes(sb)
   -> Inode skipped as isw still holds a ref.
  sop->put_super(sb)
   /* destroys percpu counters */
  -> VFS: Busy inodes after unmount!
                                  wb_queue_isw()
                                   queue_work(isw_wq, ...)
                                  /* later in work function */
                                  inode_switch_wbs_work_fn
                                   process_inode_switch_wbs
                                    iput() -> evict
                                     percpu_counter_dec() // UAF!

Fix this by extending the RCU read-side critical section in
inode_switch_wbs() and cleanup_offline_cgwb() to cover from
inode_prepare_wbs_switch() through wb_queue_isw().  Since there is
no sleep in this window, rcu_read_lock() can be used.  Then add a
synchronize_rcu() in cgroup_writeback_umount() before the existing
rcu_barrier(), so that all in-flight switchers that have passed the
SB_ACTIVE check have completed queue_work() before flush_workqueue()
is called.

Fixes: a1a0e23e4903 ("writeback: flush inode cgroup wb switches instead of pinning super_block")
Suggested-by: Jan Kara <jack@suse.cz>
Link: https://lore.kernel.org/all/mxnjq2l6guusfchvauxr3v7c4bwjasybxlleqbbh4efloeqspz@iqylk76ohufz
Signed-off-by: Baokun Li <libaokun@linux.alibaba.com>
---
 fs/fs-writeback.c | 27 ++++++++++++++++++++++++++-
 1 file changed, 26 insertions(+), 1 deletion(-)

diff --git a/fs/fs-writeback.c b/fs/fs-writeback.c
index a65694cbfe68..1f95ddcee363 100644
--- a/fs/fs-writeback.c
+++ b/fs/fs-writeback.c
@@ -665,7 +665,6 @@ static void inode_switch_wbs(struct inode *inode, int new_wb_id)
 	memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys);
 	if (memcg_css && !css_tryget(memcg_css))
 		memcg_css = NULL;
-	rcu_read_unlock();
 	if (!memcg_css)
 		goto out_free;
 
@@ -680,10 +679,18 @@ static void inode_switch_wbs(struct inode *inode, int new_wb_id)
 	isw->inodes[0] = inode;
 
 	trace_inode_switch_wbs_queue(inode->i_wb, new_wb, 1);
+	/*
+	 * Paired with synchronize_rcu() in cgroup_writeback_umount().
+	 * Holding rcu_read_lock across wb_queue_isw() ensures
+	 * synchronize_rcu() cannot return until the work is queued, so
+	 * the subsequent flush_workqueue() will wait for the switch.
+	 */
 	wb_queue_isw(new_wb, isw);
+	rcu_read_unlock();
 	return;
 
 out_free:
+	rcu_read_unlock();
 	atomic_dec(&isw_nr_in_flight);
 	if (new_wb)
 		wb_put(new_wb);
@@ -741,6 +748,14 @@ bool cleanup_offline_cgwb(struct bdi_writeback *wb)
 		new_wb = &wb->bdi->wb; /* wb_get() is noop for bdi's wb */
 
 	nr = 0;
+	/*
+	 * Paired with synchronize_rcu() in cgroup_writeback_umount().
+	 * Holding rcu_read_lock across the SB_ACTIVE check, the inode grab
+	 * and wb_queue_isw() ensures synchronize_rcu() cannot return until
+	 * the work is queued, so the subsequent flush_workqueue() will wait
+	 * for the switch.
+	 */
+	rcu_read_lock();
 	spin_lock(&wb->list_lock);
 	/*
 	 * In addition to the inodes that have completed writeback, also switch
@@ -758,6 +773,7 @@ bool cleanup_offline_cgwb(struct bdi_writeback *wb)
 
 	/* no attached inodes? bail out */
 	if (nr == 0) {
+		rcu_read_unlock();
 		atomic_dec(&isw_nr_in_flight);
 		wb_put(new_wb);
 		kfree(isw);
@@ -766,6 +782,7 @@ bool cleanup_offline_cgwb(struct bdi_writeback *wb)
 
 	trace_inode_switch_wbs_queue(wb, new_wb, nr);
 	wb_queue_isw(new_wb, isw);
+	rcu_read_unlock();
 
 	return restart;
 }
@@ -1221,6 +1238,14 @@ void cgroup_writeback_umount(struct super_block *sb)
 	smp_mb();
 
 	if (atomic_read(&isw_nr_in_flight)) {
+		/*
+		 * Paired with rcu_read_lock() in inode_switch_wbs() and
+		 * cleanup_offline_cgwb().  synchronize_rcu() waits for any
+		 * in-flight switcher that already passed the SB_ACTIVE check
+		 * to finish queueing its work, so flush_workqueue() below
+		 * will then drain it.
+		 */
+		synchronize_rcu();
 		/*
 		 * Use rcu_barrier() to wait for all pending callbacks to
 		 * ensure that all in-flight wb switches are in the workqueue.
-- 
2.43.7