Re: [PATCHSET 0/2] Fix splice from random/urandom

[Jens Axboe]

On 5/20/22 6:43 AM, Jason A. Donenfeld wrote:
> Hi Jens,
> 
> On Thu, May 19, 2022 at 07:10:56PM -0600, Jens Axboe wrote:
>> On 5/19/22 7:00 PM, Jason A. Donenfeld wrote:
>>> Hi Jens,
>>>
>>> On Thu, May 19, 2022 at 06:56:12PM -0600, Jens Axboe wrote:
>>>> On 5/19/22 6:48 PM, Jason A. Donenfeld wrote:
>>>>> sendfile() returns -EINVAL even with your patches. Only splicing to pipes
>>>>> seems to work.
>>>>
>>>> Huh, that really should work. Are you trying to sendfile() to random? If
>>>> so, you need that last write_iter patch too, and add the splice_write as
>>>> I mentioned.
>>>  
>>> No, I've only tried the read side so far. I made a little program:
>>>
>>> #include <sys/sendfile.h>
>>> #include <stdio.h>
>>>
>>> int main(int argc, char *argv[])
>>> {
>>>         ssize_t s = sendfile(1, 0, NULL, 0xffff);
>>>         fprintf(stderr, "ret: %zd\n", s);
>>>         return 0;
>>> }
>>>
>>> Then I ran `./a.out < /dev/urandom > /dev/null`. Fails. OTOH, if I
>>> replace /dev/urandom with an ordinary file, it succeeds.
>>
>> Here's why, it's limited to regular files or block devices:
>>
>> if (unlikely(!S_ISREG(i_mode) && !S_ISBLK(i_mode)))
>> 	return -EINVAL;
>>
>> in splice_direct_to_actor().
> 
> Indeed. Looks like that was your code from long long ago!

Yep I would not be surprised if that is the case!

> I posted
> https://lore.kernel.org/lkml/20220520095747.123748-1-Jason@xxxxxxxxx/ to
> fix it if you'd like to review.

Not in my inbox, but you also used an email that hasn't been valid in 16
years :-)

But looks fine to me, we can open this up to character devices, don't
see an issue with that.

-- 
Jens Axboe