On 1/17/19 1:50 PM, Jeff Moyer wrote:
> Jens Axboe <axboe@xxxxxxxxx> writes:
>
>> On 1/17/19 1:09 PM, Jens Axboe wrote:
>>> On 1/17/19 1:03 PM, Jeff Moyer wrote:
>>>> Jens Axboe <axboe@xxxxxxxxx> writes:
>>>>
>>>>> On 1/17/19 5:48 AM, Roman Penyaev wrote:
>>>>>> On 2019-01-16 18:49, Jens Axboe wrote:
>>>>>>
>>>>>> [...]
>>>>>>
>>>>>>> +static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
>>>>>>> + struct io_uring_params *p)
>>>>>>> +{
>>>>>>> + struct io_sq_ring *sq_ring;
>>>>>>> + struct io_cq_ring *cq_ring;
>>>>>>> + size_t size;
>>>>>>> + int ret;
>>>>>>> +
>>>>>>> + sq_ring = io_mem_alloc(struct_size(sq_ring, array, p->sq_entries));
>>>>>>
>>>>>> It seems that sq_entries, cq_entries are not limited at all. Can nasty
>>>>>> app consume a lot of kernel pages calling io_setup_uring() from a loop
>>>>>> passing random entries number? (or even better: decreasing entries
>>>>>> number,
>>>>>> in order to consume all pages orders with min number of loops).
>>>>>
>>>>> Yes, that's an oversight, we should have a limit in place. I'll add that.
>>>>
>>>> Can we charge the ring memory to the RLIMIT_MEMLOCK as well? I'd prefer
>>>> not to repeat the mistake of fs.aio-max-nr.
>>>
>>> Sure, we can do that. With the ring limited in size (it's now 4k entries
>>> at most), the amount of memory gobbled up by that is much smaller than
>>> the fixed buffers. A max sized ring is about 256k of memory.
>
> Per io_uring. Nothing prevents a user from calling io_uring_setup in a
> loop and continuing to gobble up memory.
>
>> One concern here is that, at least looking at my boxes, the default
>> setting for RLIMIT_MEMLOCK is really low. I'd hate for everyone to run
>> into issues using io_uring just because it seems to require root,
>> because the memlock limit is so low.
>>
>> That's much less of a concern with the fixed buffers, since it's a more
>> esoteric part of it. But everyone should be able to setup a few io_uring
>> queues and use them without having to worry about failing due to an
>> absurdly low RLIMIT_MEMLOCK.
>>
>> Comments?
>
> Yeah, the default is 64k here. We should probably up that. I'd say we
> either tackle the ridiculously low rlimits, or I guess we just go the
> aio route and add a sysctl. :-\ I'll see what's involved in the
> former.
After giving it a bit of thought, let's go the rlimit route. It is cleaner,
and I don't want a sysctl knob for this either. 64k will enable anyone to
set up at least one decently sized ring.
--
Jens Axboe
