Re: [RFC PATCH v1 20/25] serial: 8250: implement write_atomic

[John Ogness]

On 2019-03-08, John Ogness <john.ogness@xxxxxxxxxxxxx> wrote:
> If the ringbuffer was fully lockless, we should be able to have
> per-console CPU-reentrant spin locks as long as the ordering is
> preserved, which I expect shouldn't be a problem. If any NMI context
> needed a spin lock for its own purposes, it would need to use the
> CPU-reentrant spin lock of the first console so as to preserve the
> ordering in case of a panic.

This point is garbage. Sorry. I do not see how we could safely have
multiple CPU-reentrant spin locks. Example of a deadlock:

CPU0                  CPU1
printk                printk
  console2.lock         console1.lock
NMI                   NMI
  printk                printk
    console1.lock         console2.lock

>> ... it should not be a common lock for the ring buffer and all
>> consoles.
>
> If the ring buffer becomes fully lockless, then we could move to
> per-console CPU-reentrant spin locks.

A fully lockless ring buffer will reduce the scope of the one, global
CPU-reentrant spin lock. But I do not see how we can safely have
multiple of these. If it is part of printk, it is already implicitly on
every line of code.

John Ogness