sched/debug: Fix potential deadlock when writing to sched_features

The following lockdep report can be triggered by writing to /sys/kernel/debug/sched_features:

  ======================================================
  WARNING: possible circular locking dependency detected
  4.18.0-rc6-00152-gcd3f77d74ac3-dirty #18 Not tainted
  ------------------------------------------------------
  sh/3358 is trying to acquire lock:
  000000004ad3989d (cpu_hotplug_lock.rw_sem){++++}, at: static_key_enable+0x14/0x30
  but task is already holding lock:
  00000000c1b31a88 (&sb->s_type->i_mutex_key#3){+.+.}, at: sched_feat_write+0x160/0x428
  which lock already depends on the new lock.
  the existing dependency chain (in reverse order) is:
  -> #3 (&sb->s_type->i_mutex_key#3){+.+.}:
         lock_acquire+0xb8/0x148
         down_write+0xac/0x140
         start_creating+0x5c/0x168
         debugfs_create_dir+0x18/0x220
         opp_debug_register+0x8c/0x120
         _add_opp_dev+0x104/0x1f8
         dev_pm_opp_get_opp_table+0x174/0x340
         _of_add_opp_table_v2+0x110/0x760
         dev_pm_opp_of_add_table+0x5c/0x240
         dev_pm_opp_of_cpumask_add_table+0x5c/0x100
         cpufreq_init+0x160/0x430
         cpufreq_online+0x1cc/0xe30
         cpufreq_add_dev+0x78/0x198
         subsys_interface_register+0x168/0x270
         cpufreq_register_driver+0x1c8/0x278
         dt_cpufreq_probe+0xdc/0x1b8
         platform_drv_probe+0xb4/0x168
         driver_probe_device+0x318/0x4b0
         __device_attach_driver+0xfc/0x1f0
         bus_for_each_drv+0xf8/0x180
         __device_attach+0x164/0x200
         device_initial_probe+0x10/0x18
         bus_probe_device+0x110/0x178
         device_add+0x6d8/0x908
         platform_device_add+0x138/0x3d8
         platform_device_register_full+0x1cc/0x1f8
         cpufreq_dt_platdev_init+0x174/0x1bc
         do_one_initcall+0xb8/0x310
         kernel_init_freeable+0x4b8/0x56c
         kernel_init+0x10/0x138
         ret_from_fork+0x10/0x18
  -> #2 (opp_table_lock){+.+.}:
         lock_acquire+0xb8/0x148
         __mutex_lock+0x104/0xf50
         mutex_lock_nested+0x1c/0x28
         _of_add_opp_table_v2+0xb4/0x760
         dev_pm_opp_of_add_table+0x5c/0x240
         dev_pm_opp_of_cpumask_add_table+0x5c/0x100
         cpufreq_init+0x160/0x430
         cpufreq_online+0x1cc/0xe30
         cpufreq_add_dev+0x78/0x198
         subsys_interface_register+0x168/0x270
         cpufreq_register_driver+0x1c8/0x278
         dt_cpufreq_probe+0xdc/0x1b8
         platform_drv_probe+0xb4/0x168
         driver_probe_device+0x318/0x4b0
         __device_attach_driver+0xfc/0x1f0
         bus_for_each_drv+0xf8/0x180
         __device_attach+0x164/0x200
         device_initial_probe+0x10/0x18
         bus_probe_device+0x110/0x178
         device_add+0x6d8/0x908
         platform_device_add+0x138/0x3d8
         platform_device_register_full+0x1cc/0x1f8
         cpufreq_dt_platdev_init+0x174/0x1bc
         do_one_initcall+0xb8/0x310
         kernel_init_freeable+0x4b8/0x56c
         kernel_init+0x10/0x138
         ret_from_fork+0x10/0x18
  -> #1 (subsys mutex#6){+.+.}:
         lock_acquire+0xb8/0x148
         __mutex_lock+0x104/0xf50
         mutex_lock_nested+0x1c/0x28
         subsys_interface_register+0xd8/0x270
         cpufreq_register_driver+0x1c8/0x278
         dt_cpufreq_probe+0xdc/0x1b8
         platform_drv_probe+0xb4/0x168
         driver_probe_device+0x318/0x4b0
         __device_attach_driver+0xfc/0x1f0
         bus_for_each_drv+0xf8/0x180
         __device_attach+0x164/0x200
         device_initial_probe+0x10/0x18
         bus_probe_device+0x110/0x178
         device_add+0x6d8/0x908
         platform_device_add+0x138/0x3d8
         platform_device_register_full+0x1cc/0x1f8
         cpufreq_dt_platdev_init+0x174/0x1bc
         do_one_initcall+0xb8/0x310
         kernel_init_freeable+0x4b8/0x56c
         kernel_init+0x10/0x138
         ret_from_fork+0x10/0x18
  -> #0 (cpu_hotplug_lock.rw_sem){++++}:
         __lock_acquire+0x203c/0x21d0
         lock_acquire+0xb8/0x148
         cpus_read_lock+0x58/0x1c8
         static_key_enable+0x14/0x30
         sched_feat_write+0x314/0x428
         full_proxy_write+0xa0/0x138
         __vfs_write+0xd8/0x388
         vfs_write+0xdc/0x318
         ksys_write+0xb4/0x138
         sys_write+0xc/0x18
         __sys_trace_return+0x0/0x4
  other info that might help us debug this:
  Chain exists of:
    cpu_hotplug_lock.rw_sem --> opp_table_lock --> &sb->s_type->i_mutex_key#3
   Possible unsafe locking scenario:
         CPU0                    CPU1
         ----                    ----
    lock(&sb->s_type->i_mutex_key#3);
                                 lock(opp_table_lock);
                                 lock(&sb->s_type->i_mutex_key#3);
    lock(cpu_hotplug_lock.rw_sem);
   *** DEADLOCK ***
  2 locks held by sh/3358:
   #0: 00000000a8c4b363 (sb_writers#10){.+.+}, at: vfs_write+0x238/0x318
   #1: 00000000c1b31a88 (&sb->s_type->i_mutex_key#3){+.+.}, at: sched_feat_write+0x160/0x428
  stack backtrace:
  CPU: 5 PID: 3358 Comm: sh Not tainted 4.18.0-rc6-00152-gcd3f77d74ac3-dirty #18
  Hardware name: Renesas H3ULCB Kingfisher board based on r8a7795 ES2.0+ (DT)
  Call trace:
   dump_backtrace+0x0/0x288
   show_stack+0x14/0x20
   dump_stack+0x13c/0x1ac
   print_circular_bug.isra.10+0x270/0x438
   check_prev_add.constprop.16+0x4dc/0xb98
   __lock_acquire+0x203c/0x21d0
   lock_acquire+0xb8/0x148
   cpus_read_lock+0x58/0x1c8
   static_key_enable+0x14/0x30
   sched_feat_write+0x314/0x428
   full_proxy_write+0xa0/0x138
   __vfs_write+0xd8/0x388
   vfs_write+0xdc/0x318
   ksys_write+0xb4/0x138
   sys_write+0xc/0x18
   __sys_trace_return+0x0/0x4

This is because when loading the cpufreq_dt module we first acquire
cpu_hotplug_lock.rw_sem lock, then in cpufreq_init(), we are taking
the &sb->s_type->i_mutex_key lock.

But when writing to /sys/kernel/debug/sched_features, the
cpu_hotplug_lock.rw_sem lock depends on the &sb->s_type->i_mutex_key lock.

To fix this bug, reverse the lock acquisition order when writing to
sched_features, this way cpu_hotplug_lock.rw_sem no longer depends on
&sb->s_type->i_mutex_key.

Tested-by: Dietmar Eggemann <[email protected]>
Signed-off-by: Jiada Wang <[email protected]>
Signed-off-by: Peter Zijlstra (Intel) <[email protected]>
Cc: Eugeniu Rosca <[email protected]>
Cc: George G. Davis <[email protected]>
Cc: Linus Torvalds <[email protected]>
Cc: Peter Zijlstra <[email protected]>
Cc: Thomas Gleixner <[email protected]>
Link: http://lkml.kernel.org/r/[email protected]
Signed-off-by: Ingo Molnar <[email protected]>

Author: jiada-wang

kernel/sched/debug.c

@@ -89,12 +89,12 @@ struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
 
 static void sched_feat_disable(int i)
 {
-	static_key_disable(&sched_feat_keys[i]);
+	static_key_disable_cpuslocked(&sched_feat_keys[i]);
 }
 
 static void sched_feat_enable(int i)
 {
-	static_key_enable(&sched_feat_keys[i]);
+	static_key_enable_cpuslocked(&sched_feat_keys[i]);
 }
 #else
 static void sched_feat_disable(int i) { };
@@ -146,9 +146,11 @@ sched_feat_write(struct file *filp, const char __user *ubuf,
 
 	/* Ensure the static_key remains in a consistent state */
 	inode = file_inode(filp);
+	cpus_read_lock();
 	inode_lock(inode);
 	ret = sched_feat_set(cmp);
 	inode_unlock(inode);
+	cpus_read_unlock();
 	if (ret < 0)
 		return ret;