Re: [patch] aio: remove aio-max-nr and instead use the memlock rlimit to limit the number of pages pinned for the aio completion ring

[Eric Dumazet <dada1@cosmosbay.com>]

Eric Dumazet a écrit :
> Eric Dumazet a écrit :
>> Jeff Moyer a écrit :
>>> Avi Kivity <avi@redhat.com> writes:
>>>
>>>> Jeff Moyer wrote:
>>>>> Hi,
>>>>>
>>>>> Believe it or not, I get numerous questions from customers about the
>>>>> suggested tuning value of aio-max-nr.  aio-max-nr limits the total
>>>>> number of io events that can be reserved, system wide, for aio
>>>>> completions.  Each time io_setup is called, a ring buffer is allocated
>>>>> that can hold nr_events I/O completions.  That ring buffer is then
>>>>> mapped into the process' address space, and the pages are pinned in
>>>>> memory.  So, the reason for this upper limit (I believe) is to keep a
>>>>> malicious user from pinning all of kernel memory.  Now, this sounds like
>>>>> a much better job for the memlock rlimit to me, hence the following
>>>>> patch.
>>>>>   
>>>> Is it not possible to get rid of the pinning entirely?  Pinning
>>>> interferes with page migration which is important for NUMA, among
>>>> other issues.
>>> aio_complete is called from interrupt handlers, so can't block faulting
>>> in a page.  Zach mentions there is a possibility of handing completions
>>> off to a kernel thread, with all of the performance worries and extra
>>> bookkeeping that go along with such a scheme (to help frame my concerns,
>>> I often get lambasted over .5% performance regressions).
>> This aio_completion from interrupt handlers keep us from using SLAB_DESTROY_BY_RCU
>> instead of call_rcu() for "struct file" freeing.
>>
>> http://lkml.org/lkml/2008/12/17/364
>>
>> I would love if we could get rid of this mess...
> 
> Speaking of that, I tried to take a look at this aio stuff and have one question.
> 
> Assuming that __fput() cannot be called from interrupt context.
> -> fput() should not be called from interrupt context as well.
> 
> How comes we call fput(req->ki_eventfd) from really_put_req()
> from interrupt context ?
> 
> If user program closes eventfd, then inflight AIO requests can trigger
> a bug.
> 

Path could be :

1) fput() changes so that calling it from interrupt context is possible
   (Using a working queue to make sure __fput() is called from process context)

2) Changes aio to use fput() as is (and zap its internal work_queue and aio_fput_routine() stuff)

3) Once atomic_long_dec_and_test(&filp->f_count) only performed in fput(),
   SLAB_DESTROY_BY_RCU for "struct file" get back :)