drm: avoid blocking in drm_clients_info's rcu section
authorDesmond Cheong Zhi Xi <desmondcheongzx@gmail.com>
Mon, 12 Jul 2021 04:35:05 +0000 (12:35 +0800)
committerDaniel Vetter <daniel.vetter@ffwll.ch>
Tue, 20 Jul 2021 18:14:42 +0000 (20:14 +0200)
Inside drm_clients_info, the rcu_read_lock is held to lock
pid_task()->comm. However, within this protected section, a call to
drm_is_current_master is made, which involves a mutex lock in a future
patch. However, this is illegal because the mutex lock might block
while in the RCU read-side critical section.

Since drm_is_current_master isn't protected by rcu_read_lock, we avoid
this by moving it out of the RCU critical section.

The following report came from intel-gfx ci's
igt@debugfs_test@read_all_entries testcase:

=============================
[ BUG: Invalid wait context ]
5.13.0-CI-Patchwork_20515+ #1 Tainted: G        W
-----------------------------
debugfs_test/1101 is trying to lock:
ffff888132d901a8 (&dev->master_mutex){+.+.}-{3:3}, at:
drm_is_current_master+0x1e/0x50
other info that might help us debug this:
context-{4:4}
3 locks held by debugfs_test/1101:
 #0: ffff88810fdffc90 (&p->lock){+.+.}-{3:3}, at:
 seq_read_iter+0x53/0x3b0
 #1: ffff888132d90240 (&dev->filelist_mutex){+.+.}-{3:3}, at:
 drm_clients_info+0x63/0x2a0
 #2: ffffffff82734220 (rcu_read_lock){....}-{1:2}, at:
 drm_clients_info+0x1b1/0x2a0
stack backtrace:
CPU: 8 PID: 1101 Comm: debugfs_test Tainted: G        W
5.13.0-CI-Patchwork_20515+ #1
Hardware name: Intel Corporation CometLake Client Platform/CometLake S
UDIMM (ERB/CRB), BIOS CMLSFWR1.R00.1263.D00.1906260926 06/26/2019
Call Trace:
 dump_stack+0x7f/0xad
 __lock_acquire.cold.78+0x2af/0x2ca
 lock_acquire+0xd3/0x300
 ? drm_is_current_master+0x1e/0x50
 ? __mutex_lock+0x76/0x970
 ? lockdep_hardirqs_on+0xbf/0x130
 __mutex_lock+0xab/0x970
 ? drm_is_current_master+0x1e/0x50
 ? drm_is_current_master+0x1e/0x50
 ? drm_is_current_master+0x1e/0x50
 drm_is_current_master+0x1e/0x50
 drm_clients_info+0x107/0x2a0
 seq_read_iter+0x178/0x3b0
 seq_read+0x104/0x150
 full_proxy_read+0x4e/0x80
 vfs_read+0xa5/0x1b0
 ksys_read+0x5a/0xd0
 do_syscall_64+0x39/0xb0
 entry_SYSCALL_64_after_hwframe+0x44/0xae

Signed-off-by: Desmond Cheong Zhi Xi <desmondcheongzx@gmail.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: https://patchwork.freedesktop.org/patch/msgid/20210712043508.11584-3-desmondcheongzx@gmail.com
drivers/gpu/drm/drm_debugfs.c

index 3d71820..b0a8264 100644 (file)
@@ -91,6 +91,7 @@ static int drm_clients_info(struct seq_file *m, void *data)
        mutex_lock(&dev->filelist_mutex);
        list_for_each_entry_reverse(priv, &dev->filelist, lhead) {
                struct task_struct *task;
+               bool is_current_master = drm_is_current_master(priv);
 
                rcu_read_lock(); /* locks pid_task()->comm */
                task = pid_task(priv->pid, PIDTYPE_PID);
@@ -99,7 +100,7 @@ static int drm_clients_info(struct seq_file *m, void *data)
                           task ? task->comm : "<unknown>",
                           pid_vnr(priv->pid),
                           priv->minor->index,
-                          drm_is_current_master(priv) ? 'y' : 'n',
+                          is_current_master ? 'y' : 'n',
                           priv->authenticated ? 'y' : 'n',
                           from_kuid_munged(seq_user_ns(m), uid),
                           priv->magic);