Hi,
first, I am sorry for the late reply. I have been somehow distracted
this week.
On Mon 2021-02-01 18:26:01, Muchun Song wrote:
> We found a deadlock bug on our server when the kernel panic. It can be
> described in the following diagram.
>
> CPU0: CPU1:
> panic rcu_dump_cpu_stacks
> kdump_nmi_shootdown_cpus nmi_trigger_cpumask_backtrace
> register_nmi_handler(crash_nmi_callback) printk_safe_flush
> __printk_safe_flush
>
> raw_spin_lock_irqsave(&read_lock)
> // send NMI to other processors
> apic_send_IPI_allbutself(NMI_VECTOR)
> // NMI interrupt,
> dead loop
> crash_nmi_callback
> printk_safe_flush_on_panic
> printk_safe_flush
> __printk_safe_flush
> // deal lock
> raw_spin_lock_irqsave(&read_lock)
Great catch!
> kernel/printk/printk_safe.c | 58
> ++++++++++++++++++++++++++-------------------
> 1 file changed, 34 insertions(+), 24 deletions(-)
>
> diff --git a/kernel/printk/printk_safe.c b/kernel/printk/printk_safe.c
> index a0e6f746de6c..86d9fa74ac5c 100644
> --- a/kernel/printk/printk_safe.c
> +++ b/kernel/printk/printk_safe.c
> @@ -174,30 +174,13 @@ static void report_message_lost(struct
> printk_safe_seq_buf *s)
> printk_deferred("Lost %d message(s)!\n", lost);
> }
>
> -/*
> - * Flush data from the associated per-CPU buffer. The function
> - * can be called either via IRQ work or independently.
> - */
> -static void __printk_safe_flush(struct irq_work *work)
> +static void __printk_safe_flush_work(struct irq_work *work)
> {
> - static raw_spinlock_t read_lock =
> - __RAW_SPIN_LOCK_INITIALIZER(read_lock);
> struct printk_safe_seq_buf *s =
> container_of(work, struct printk_safe_seq_buf, work);
> - unsigned long flags;
> size_t len;
> - int i;
> + int i = 0;
>
> - /*
> - * The lock has two functions. First, one reader has to flush all
> - * available message to make the lockless synchronization with
> - * writers easier. Second, we do not want to mix messages from
> - * different CPUs. This is especially important when printing
> - * a backtrace.
> - */
> - raw_spin_lock_irqsave(&read_lock, flags);
> -
> - i = 0;
> more:
> len = atomic_read(&s->len);
>
> @@ -232,6 +215,26 @@ static void __printk_safe_flush(struct irq_work *work)
>
> out:
> report_message_lost(s);
> +}
> +
> +/*
> + * Flush data from the associated per-CPU buffer. The function
> + * can be called either via IRQ work or independently.
> + */
> +static void printk_safe_flush_work(struct irq_work *work)
> +{
> + unsigned long flags;
> + static DEFINE_RAW_SPINLOCK(read_lock);
> +
> + /*
> + * The lock has two functions. First, one reader has to flush all
> + * available message to make the lockless synchronization with
> + * writers easier. Second, we do not want to mix messages from
> + * different CPUs. This is especially important when printing
> + * a backtrace.
> + */
> + raw_spin_lock_irqsave(&read_lock, flags);
> + __printk_safe_flush_work(work);
> raw_spin_unlock_irqrestore(&read_lock, flags);
> }
>
> @@ -248,9 +251,9 @@ void printk_safe_flush(void)
>
> for_each_possible_cpu(cpu) {
> #ifdef CONFIG_PRINTK_NMI
> - __printk_safe_flush(&per_cpu(nmi_print_seq, cpu).work);
> + printk_safe_flush_work(&per_cpu(nmi_print_seq, cpu).work);
> #endif
> - __printk_safe_flush(&per_cpu(safe_print_seq, cpu).work);
> + printk_safe_flush_work(&per_cpu(safe_print_seq, cpu).work);
> }
> }
>
> @@ -266,6 +269,8 @@ void printk_safe_flush(void)
> */
> void printk_safe_flush_on_panic(void)
> {
> + int cpu;
> +
> /*
> * Make sure that we could access the main ring buffer.
> * Do not risk a double release when more CPUs are up.
> @@ -278,7 +283,12 @@ void printk_safe_flush_on_panic(void)
> raw_spin_lock_init(&logbuf_lock);
> }
>
> - printk_safe_flush();
> + for_each_possible_cpu(cpu) {
> +#ifdef CONFIG_PRINTK_NMI
> + __printk_safe_flush_work(&per_cpu(nmi_print_seq, cpu).work);
> +#endif
> + __printk_safe_flush_work(&per_cpu(safe_print_seq, cpu).work);
> + }
There might be two CPUs running this code when more CPUs are still running.
The might definitely be the case here:
if (!_crash_kexec_post_notifiers) {
printk_safe_flush_on_panic();
__crash_kexec(NULL);
__printk_safe_flush_work() is relatively safe even in this case.
It accesses only data that are limited by s->len that is updated
atomically.
But it calls:
+ printk_safe_flush_on_panic()
+ __printk_safe_flush_work()
+ printk_safe_flush_buffer()
+ printk_safe_flush_line()
+ printk_deferred()
where printk_deferred() used logbuf_lock() until v5.11-rc1, see
the commit b031a684bfd01d633c ("printk: remove logbuf_lock
writer-protection of ringbuffer").
We need a solution that might be backported back to v4.7-rc1
where logbuf_lock() might cause a deadlock.
Also we really want to call printk_safe_flush_on_panic()
in panic(). It is trying hard but it should not cause
deadlock.
Therefore, I suggest to handle read_lock the same way in
printk_safe_flush_on_panic() as we handle logbuf_lock there.
I mean to make it global in printk_safe.c:
static DEFINE_RAW_SPINLOCK(safe_read_lock);
and do something like this in printk_safe_flush_on_panic()
if (raw_spin_is_locked(&safe_read_lock)) {
if (num_online_cpus() > 1)
return;
debug_locks_off();
raw_spin_lock_init(&safe_read_lock);
}
How does that sound, please?
Best Regards,
Petr
