On 24/11/2019 02:08, Jens Axboe wrote:
> On 11/23/19 3:49 PM, Pavel Begunkov wrote:
>> There is a bug hunging my system when run fixed-link with /dev/urandom
>> instead of /dev/zero (see patch 1/2).
>>
>> As for me, the easiest way to fix is to grab mm and use userspace
>> address for this specific case (as it's done in patches). The other
>> way is to kmap/vmap, but the first should be short-lived and the
>> second needs mm anyway.
>>
>> Ideas how to do it better way? Suggestions and corrections are welcome.
> 
> OK, took a quick look. kmap() etc doesn't need context, but the copy

Thanks! What copy do you mean? The first and pretty short version was
with kmap.
e.g. while(count) { read(kmap()); ...; knumap(); }

I'll send this shortly. What I don't like here, is that it passes
kmapped virtual address as "void __user *". Is that ok?
	

> does. How about just ensuring we grab the mm for cases that don't have
> ->read_iter() or ->write_iter() and then just map and copy in that
> loop that handles that exact case? I think that's cleaner than what
> you have.
> 

-- 
Pavel Begunkov