Re: [RFC PATCH] serial: core: Use pm_runtime_get_sync() in uart_start()

[John Ogness]

[Added printk maintainers CC.]

On 2023-11-24, Xuewen Yan <xuewen.yan@xxxxxxxxxx> wrote:
> The commit 84a9582fd203("serial: core: Start managing serial
> controllers to enable runtime PM") use the pm_runtime_get() after
> uart_port_lock() which would close the irq and disable preement.  At
> this time, pm_runtime_get may cause the following two problems:
>
> (1) deadlock in try_to_wake_up:
>
> uart_write()
> uart_port_lock() <<< get lock
> __uart_start
> __pm_runtime_resume
> rpm_resume
> queue_work_on
> try_to_wake_up
> _printk
> uart_console_write
> ...
> uart_port_lock() <<< wait forever

I suppose you got this because of the lockdep message generated by
#2. It probably would make sense to call __printk_safe_enter() inside
uart_port_lock(). This would allow printk() to automatically defer the
printing for that CPU until the port lock is released.

> (2) scheduling while atomic:
> uart_write()
> uart_port_lock() <<< get lock
> __uart_start
> __pm_runtime_resume
> rpm_resume
> schedule() << sleep

rpm_resume() is a fascinating function. It requires the caller to hold a
spin_lock (dev->power.lock) with interrupts disabled. But it seems to
believe that this is the *only* spin_lock held so that it can
temporarily spin_unlock and call might_sleep() functions. In the case of
uart_write(), it certainly is not the only spin_lock held.

I do not know enough about the internals of RPM to suggest a proper
solution. But it looks like rpm_resume() cannot assume dev->power.lock
is the only spin_lock held by the caller.

John Ogness