Re: New Patch series for the UBD Driver

[Jens Axboe <axboe@kernel.dk>]

On 11/13/18 6:56 AM, Anton Ivanov wrote:
> On 13/11/2018 13:30, Jens Axboe wrote:
>> On 11/13/18 3:15 AM, Anton Ivanov wrote:
>>> I have figured out part of it. It is not caused by the block layer at
>>> all. The offending parts are printks in the io thread. You simply cannot
>>> printk out of there, that causes the spinlock recursion, irqs, on/off,
>>> etc warning.
>>>
>>> This means that any errors MUST be reported only after they have gone
>>> back to the main UML kernel thread.
>> That's a pretty serious issue with UML, there's absolutely nothing wrong
>> with doing printk from IRQ context, or from an IO thread like this one.
>> You don't know what the call stack below you will do for functions you
>> call, this is a ticking time bomb.
> 
> This is not a normal kernel IO thread. It is a host-side helper thread. 
> There are two of those - one for UBD and and one for sigio.
> 
> I am not qualified to judge what is normal and not there, but they 
> operate "outside" normal kernel authority and communicate only via their 
> IPC pipe.
> 
> It is not interrupt context - it is an equivalent of "inside the actual 
> device".

But it's running in the kernel, which means it's eligible to call
any functions. If you have ANY function in your call path that
isn't 100% controller by ubd or UML, then all bets are off.

I don't even see how your new patch 4 fixes anything, whether the
printk happens in the io_thread() or some call off of that should
be the same issue.

-- 
Jens Axboe


_______________________________________________
linux-um mailing list
linux-um@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-um