Re: [PATCH v2] tty: ttynull: implement console write

[Michael Thalmeier <michael.thalmeier@hale.at>]

Hi Petr,

----- On 15 Feb, 2023, at 15:33, Petr Mladek pmladek@suse.com wrote:
> On Wed 2023-02-15 12:37:36, Vincent Whitchurch wrote:
>> + Cc: John, Petr
>> 
>> On Tue, Feb 14, 2023 at 12:59:21PM +0100, Michael Thalmeier wrote:
>> > Since commit a699449bb13b ("printk: refactor and rework printing logic"),
>> > con->write is called without checking if it is implemented in the console
>> > driver. This does make some sense, because for all "normal" consoles it
>> > is vital to have a write function.
>> > As the ttynull console driver does not need any console output the write
>> > function was not implemented. This caused a "unable to handle kernel NULL
>> > pointer dereference" when the write function is called now.
>> > 
>> > To fix this issue, implement an empty write function.
>> > 
>> > Fixes: a699449bb13b ("printk: refactor and rework printing logic")
>> > Cc: stable@vger.kernel.org
>> > Signed-off-by: Michael Thalmeier <michael.thalmeier@hale.at>
>> 
>> Looking at the referenced commit, the only place I see it calling
>> con->write() is from call_console_driver(), which is in turn only called
>> from console_emit_next_record().  And console_flush_all(), the only
>> caller of console_emit_next_record(), checks that con->write is not NULL
>> using console_is_usable() before calling console_emit_next_record().
> 
> I see the same. The refactoring moved the check of con->write from
> call_console_driver() to console_is_usable(). It detects the NULL
> pointer earlier in console_flush_all()...
> 
>> What am I missing?  Which code path in the referenced commit calls
>> con->write without checking for NULL?
> 
> Vincent, could you please provide log with the backtrace?
> 
> I wonder if the problem is in the RT-patchset where
> console_emit_next_record() is called also from the printk kthread.

You are right. I have encountered this problem with the RT-patchset.
We currently are using the latest v5.15-rt kernel which had this problem.
 
> That said, the current code is error-prone. The check for non-NULL
> con->write is too far from the caller.
> 
> I would prefer to make it more safe. For example, I would prevent
> registration of consoles without con->write callback in the first
> place. It would require, to implement the empty write() callback
> for ttynull console as done by this patch.
> 
> Anyway, I would like to understand if the "unable to handle kernel NULL
> pointer dereference" is a real problem in the current implementation.

Unfortunately I have not yet tested with the current RT or non-RT versions.

> 
> Best Regards,
> Petr

Regards, Michael