On Sat, Aug 24, 2024 at 03:25:04PM +0800, Wen Yang wrote:
> 
> 
> On 2024/8/20 01:45, Greg KH wrote:
> > On Mon, Aug 19, 2024 at 10:59:45PM +0800, Wen Yang wrote:
> > > pstore_dump() is called when both preemption and local IRQ are disabled,
> > > and a spinlock is obtained, which is problematic for the RT kernel because
> > > in this configuration, spinlocks are sleep locks.
> > > 
> > > Replace the spinlock_t with raw_spinlock_t to avoid sleeping in atomic 
> > > context.
> > 
> > This feels odd, is it only an out-of-tree RT thing?  Or does this affect
> > in-kernel code as well?  What prevents any normal spinlock from sleeping
> > in your system configuration as well?
> > 
> 
> Thank you for your comment.
> 
> If we enable PREEMPT_RT, it will also affect in-kernel code, such as in the
> following scenario:
> 
> echo b > /proc/sysrq-trigger
>   - sysrq_handle_reboot
>   - emergency_restart
>   - kmsg_dump
>   - pstore_dump
> Obtained psinfo->buf_lock, if there is a lot of error log in the kernel, it
> will last for a long time
> 
> If the system unexpectedly crashes at this time:
>   - panic()
>   - kmsg_dump
>   - pstore_dump
> Attempting to obtain psinfo->buf_lock while disabling interrupts and
> preemption, but since this lock is already held by the above process, it
> will result in illegal sleep.

Reading Documentation/locking/locktypes.rst seems to suggest pstore does
want the raw version. I'm surprised there aren't many more cases where
this is a problem. :P

-- 
Kees Cook

Reply via email to