On Sat 2021-07-03 08:32:02, John Ogness wrote:
> On 2021-07-02, Petr Mladek <pmladek@xxxxxxxx> wrote:
> > The standard printk() tries to flush the message to the console
> > immediately. It tries to take the console lock. If the lock is
> > already taken then the current owner is responsible for flushing
> > even the new message.
> >
> > There is a small race window between checking whether a new message is
> > available and releasing the console lock. It is solved by re-checking
> > the state after releasing the console lock. If the check is positive
> > then console_unlock() tries to take the lock again and process the new
> > message as well.
> >
> > The commit 996e966640ddea7b535c ("printk: remove logbuf_lock") causes that
> > console_seq is not longer read atomically. As a result, the re-check might
> > be done with an inconsistent 64-bit index.
> >
> > Solve it by using the last sequence number that has been checked under
> > the console lock. In the worst case, it will take the lock again only
> > to realized that the new message has already been proceed. But it
> > was possible even before.
> >
> > The variable next_seq is marked as __maybe_unused to call down compiler
> > warning when CONFIG_PRINTK is not defined.
>
> As Sergey already pointed out, this patch is not fixing a real
> problem. An inconsistent value (or an increased consistent value) would
> mean that another printer is actively printing, and thus a retry is not
> necessary anyway.
Ah, I misunderstood that part. You are right. CPU_X might see wrong
console_seq only when CPU_Y incremented console_seq. If CPU_X does not do
retry because of racy console_seq. Then CPU_Y would do retry when
yet another CPU added yet another new message in the meantime.
> But this patch will avoid a KASAN message about an unmarked
> (although safe) data race.
Yup.
OK, I am going to queue the patch for-5.15. There is no need to
rush it for-4.14.
Best Regards,
Petr
