Re: [PATCH 1/1] sched/debug: fix potential deadlock when write to sched_features

From: Peter Zijlstra
Date: Tue Jul 31 2018 - 06:34:09 EST



Leaving a lot of quoted crud for Viresh..

On Tue, Jul 31, 2018 at 06:22:19PM +0900, jiada_wang@xxxxxxxxxx wrote:
> From: Jiada Wang <jiada_wang@xxxxxxxxxx>
>
> Following lockdep report can be triggered by write to
> /sys/kernel/debug/sched_features

Only if you're using this dt_cpufreq thing of course.. I wonder why this
hasn't been reported before.

> [ 43.094116] ======================================================
> [ 43.100345] WARNING: possible circular locking dependency detected
> [ 43.106581] 4.18.0-rc6-00152-gcd3f77d74ac3-dirty #18 Not tainted
> [ 43.112635] ------------------------------------------------------
> [ 43.118864] sh/3358 is trying to acquire lock:
> [ 43.123351] 000000004ad3989d (cpu_hotplug_lock.rw_sem){++++}, at: static_key_enable+0x14/0x30
> [ 43.132008]
> [ 43.132008] but task is already holding lock:
> [ 43.137893] 00000000c1b31a88 (&sb->s_type->i_mutex_key#3){+.+.}, at: sched_feat_write+0x160/0x428
> [ 43.146886]
> [ 43.146886] which lock already depends on the new lock.
> [ 43.146886]
> [ 43.155136]
> [ 43.155136] the existing dependency chain (in reverse order) is:
> [ 43.162681]
> [ 43.162681] -> #3 (&sb->s_type->i_mutex_key#3){+.+.}:
> [ 43.169314] lock_acquire+0xb8/0x148
> [ 43.173462] down_write+0xac/0x140
> [ 43.177430] start_creating+0x5c/0x168
> [ 43.181743] debugfs_create_dir+0x18/0x220
> [ 43.186414] opp_debug_register+0x8c/0x120
> [ 43.191077] _add_opp_dev+0x104/0x1f8
> [ 43.195306] dev_pm_opp_get_opp_table+0x174/0x340
> [ 43.200581] _of_add_opp_table_v2+0x110/0x760
> [ 43.205508] dev_pm_opp_of_add_table+0x5c/0x240
> [ 43.210609] dev_pm_opp_of_cpumask_add_table+0x5c/0x100
> [ 43.216409] cpufreq_init+0x160/0x430
> [ 43.220634] cpufreq_online+0x1cc/0xe30
> [ 43.225033] cpufreq_add_dev+0x78/0x198
> [ 43.229439] subsys_interface_register+0x168/0x270
> [ 43.234803] cpufreq_register_driver+0x1c8/0x278
> [ 43.239990] dt_cpufreq_probe+0xdc/0x1b8
> [ 43.244482] platform_drv_probe+0xb4/0x168
> [ 43.249145] driver_probe_device+0x318/0x4b0
> [ 43.253982] __device_attach_driver+0xfc/0x1f0
> [ 43.258994] bus_for_each_drv+0xf8/0x180
> [ 43.263480] __device_attach+0x164/0x200
> [ 43.267968] device_initial_probe+0x10/0x18
> [ 43.272716] bus_probe_device+0x110/0x178
> [ 43.277298] device_add+0x6d8/0x908
> [ 43.281351] platform_device_add+0x138/0x3d8
> [ 43.286191] platform_device_register_full+0x1cc/0x1f8
> [ 43.291910] cpufreq_dt_platdev_init+0x174/0x1bc
> [ 43.297102] do_one_initcall+0xb8/0x310
> [ 43.301508] kernel_init_freeable+0x4b8/0x56c
> [ 43.306433] kernel_init+0x10/0x138
> [ 43.310485] ret_from_fork+0x10/0x18
> [ 43.314616]
> [ 43.314616] -> #2 (opp_table_lock){+.+.}:
> [ 43.320189] lock_acquire+0xb8/0x148
> [ 43.324329] __mutex_lock+0x104/0xf50
> [ 43.328557] mutex_lock_nested+0x1c/0x28
> [ 43.333047] _of_add_opp_table_v2+0xb4/0x760
> [ 43.337886] dev_pm_opp_of_add_table+0x5c/0x240
> [ 43.342988] dev_pm_opp_of_cpumask_add_table+0x5c/0x100
> [ 43.348786] cpufreq_init+0x160/0x430
> [ 43.353011] cpufreq_online+0x1cc/0xe30
> [ 43.357410] cpufreq_add_dev+0x78/0x198
> [ 43.361811] subsys_interface_register+0x168/0x270
> [ 43.367175] cpufreq_register_driver+0x1c8/0x278
> [ 43.372362] dt_cpufreq_probe+0xdc/0x1b8
> [ 43.376852] platform_drv_probe+0xb4/0x168
> [ 43.381515] driver_probe_device+0x318/0x4b0
> [ 43.386352] __device_attach_driver+0xfc/0x1f0
> [ 43.391363] bus_for_each_drv+0xf8/0x180
> [ 43.395850] __device_attach+0x164/0x200
> [ 43.400338] device_initial_probe+0x10/0x18
> [ 43.405088] bus_probe_device+0x110/0x178
> [ 43.409665] device_add+0x6d8/0x908
> [ 43.413718] platform_device_add+0x138/0x3d8
> [ 43.418559] platform_device_register_full+0x1cc/0x1f8
> [ 43.424270] cpufreq_dt_platdev_init+0x174/0x1bc
> [ 43.429457] do_one_initcall+0xb8/0x310
> [ 43.433857] kernel_init_freeable+0x4b8/0x56c
> [ 43.438782] kernel_init+0x10/0x138
> [ 43.442834] ret_from_fork+0x10/0x18
> [ 43.446965]
> [ 43.446965] -> #1 (subsys mutex#6){+.+.}:
> [ 43.452545] lock_acquire+0xb8/0x148
> [ 43.456684] __mutex_lock+0x104/0xf50
> [ 43.460911] mutex_lock_nested+0x1c/0x28
> [ 43.465400] subsys_interface_register+0xd8/0x270
> [ 43.470677] cpufreq_register_driver+0x1c8/0x278
> [ 43.475864] dt_cpufreq_probe+0xdc/0x1b8
> [ 43.480354] platform_drv_probe+0xb4/0x168
> [ 43.485017] driver_probe_device+0x318/0x4b0
> [ 43.489854] __device_attach_driver+0xfc/0x1f0
> [ 43.494865] bus_for_each_drv+0xf8/0x180
> [ 43.499352] __device_attach+0x164/0x200
> [ 43.503840] device_initial_probe+0x10/0x18
> [ 43.508589] bus_probe_device+0x110/0x178
> [ 43.513166] device_add+0x6d8/0x908
> [ 43.517218] platform_device_add+0x138/0x3d8
> [ 43.522058] platform_device_register_full+0x1cc/0x1f8
> [ 43.527768] cpufreq_dt_platdev_init+0x174/0x1bc
> [ 43.532955] do_one_initcall+0xb8/0x310
> [ 43.537356] kernel_init_freeable+0x4b8/0x56c
> [ 43.542281] kernel_init+0x10/0x138
> [ 43.546333] ret_from_fork+0x10/0x18
> [ 43.550464]
> [ 43.550464] -> #0 (cpu_hotplug_lock.rw_sem){++++}:
> [ 43.556822] __lock_acquire+0x203c/0x21d0
> [ 43.561397] lock_acquire+0xb8/0x148
> [ 43.565543] cpus_read_lock+0x58/0x1c8
> [ 43.569857] static_key_enable+0x14/0x30
> [ 43.574347] sched_feat_write+0x314/0x428
> [ 43.578924] full_proxy_write+0xa0/0x138
> [ 43.583415] __vfs_write+0xd8/0x388
> [ 43.587465] vfs_write+0xdc/0x318
> [ 43.591340] ksys_write+0xb4/0x138
> [ 43.595302] sys_write+0xc/0x18
> [ 43.599004] __sys_trace_return+0x0/0x4
> [ 43.603397]
> [ 43.603397] other info that might help us debug this:
> [ 43.603397]
> [ 43.611473] Chain exists of:
> [ 43.611473] cpu_hotplug_lock.rw_sem --> opp_table_lock --> &sb->s_type->i_mutex_key#3
> [ 43.611473]
> [ 43.623971] Possible unsafe locking scenario:
> [ 43.623971]
> [ 43.629943] CPU0 CPU1
> [ 43.634510] ---- ----
> [ 43.639077] lock(&sb->s_type->i_mutex_key#3);
> [ 43.643667] lock(opp_table_lock);
> [ 43.649733] lock(&sb->s_type->i_mutex_key#3);
> [ 43.656854] lock(cpu_hotplug_lock.rw_sem);
> [ 43.661175]
> [ 43.661175] *** DEADLOCK ***
> [ 43.661175]
> [ 43.667157] 2 locks held by sh/3358:
> [ 43.670765] #0: 00000000a8c4b363 (sb_writers#10){.+.+}, at: vfs_write+0x238/0x318
> [ 43.678444] #1: 00000000c1b31a88 (&sb->s_type->i_mutex_key#3){+.+.}, at: sched_feat_write+0x160/0x428
> [ 43.687872]
> [ 43.687872] stack backtrace:
> [ 43.692284] CPU: 5 PID: 3358 Comm: sh Not tainted 4.18.0-rc6-00152-gcd3f77d74ac3-dirty #18
> [ 43.700616] Hardware name: Renesas H3ULCB Kingfisher board based on r8a7795 ES2.0+ (DT)
> [ 43.708684] Call trace:
> [ 43.711169] dump_backtrace+0x0/0x288
> [ 43.714873] show_stack+0x14/0x20
> [ 43.718226] dump_stack+0x13c/0x1ac
> [ 43.721754] print_circular_bug.isra.10+0x270/0x438
> [ 43.726679] check_prev_add.constprop.16+0x4dc/0xb98
> [ 43.731690] __lock_acquire+0x203c/0x21d0
> [ 43.735740] lock_acquire+0xb8/0x148
> [ 43.739357] cpus_read_lock+0x58/0x1c8
> [ 43.743147] static_key_enable+0x14/0x30
> [ 43.747112] sched_feat_write+0x314/0x428
> [ 43.751163] full_proxy_write+0xa0/0x138
> [ 43.755125] __vfs_write+0xd8/0x388
> [ 43.758651] vfs_write+0xdc/0x318
> [ 43.762001] ksys_write+0xb4/0x138
> [ 43.765438] sys_write+0xc/0x18
> [ 43.768615] __sys_trace_return+0x0/0x4
>
> This is because load of cpufreq_dt module firstly acquires
> cpu_hotplug_lock.rw_sem lock, then in cpufreq_init, it holds
> &sb->s_type->i_mutex_key lock.
>
> But write to /sys/kernel/debug/sched_features, cpu_hotplug_lock.rw_sem lock
> depends on &sb->s_type->i_mutex_key lock.
>
> This patch by reversing the lock acquisition order when write to
> sched_features, this way cpu_hotplug_lock.rw_sem no longer depends on
> &sb->s_type->i_mutex_key.
>
> Cc: George G. Davis <george_davis@xxxxxxxxxx>
> Cc: Eugeniu Rosca <erosca@xxxxxxxxxxxxxx>
> Signed-off-by: Jiada Wang <jiada_wang@xxxxxxxxxx>
> ---
> kernel/sched/debug.c | 22 ++++++++++++----------
> 1 file changed, 12 insertions(+), 10 deletions(-)
>
> diff --git a/kernel/sched/debug.c b/kernel/sched/debug.c
> index e593b4118578..deefa9569bcf 100644
> --- a/kernel/sched/debug.c
> +++ b/kernel/sched/debug.c
> @@ -87,21 +87,21 @@ struct static_key sched_feat_keys[__SCHED_FEAT_NR] = {
>
> #undef SCHED_FEAT
>
> -static void sched_feat_disable(int i)
> +static void sched_feat_disable_cpuslocked(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 void sched_feat_enable_cpuslocked(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) { };
> -static void sched_feat_enable(int i) { };
> +static void sched_feat_disable_cpuslocked(int i) { };
> +static void sched_feat_enable_cpuslocked(int i) { };
> #endif /* HAVE_JUMP_LABEL */
>
> -static int sched_feat_set(char *cmp)
> +static int sched_feat_set_cpuslocked(char *cmp)
> {
> int i;
> int neg = 0;
> @@ -115,10 +115,10 @@ static int sched_feat_set(char *cmp)
> if (strcmp(cmp, sched_feat_names[i]) == 0) {
> if (neg) {
> sysctl_sched_features &= ~(1UL << i);
> - sched_feat_disable(i);
> + sched_feat_disable_cpuslocked(i);
> } else {
> sysctl_sched_features |= (1UL << i);
> - sched_feat_enable(i);
> + sched_feat_enable_cpuslocked(i);
> }
> break;
> }
> @@ -147,9 +147,11 @@ 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);
> - i = sched_feat_set(cmp);
> + i = sched_feat_set_cpuslocked(cmp);
> inode_unlock(inode);
> + cpus_read_unlock();
> if (i == __SCHED_FEAT_NR)
> return -EINVAL;
>

I don't much care for all the _cpuslocked renaming done, but otherwise I
suppose this is the easiest fix.

Note however that this doesn't apply, you're missing that match_string()
patch. Please use tip/sched/core or something.