On 03/06/2017 11:17 AM, Avi Kivity wrote:
>
>
> On 03/06/2017 07:06 PM, Jens Axboe wrote:
>> On 03/06/2017 09:59 AM, Avi Kivity wrote:
>>>
>>> On 03/06/2017 06:08 PM, Jens Axboe wrote:
>>>> On 03/06/2017 08:59 AM, Avi Kivity wrote:
>>>>> On 03/06/2017 05:38 PM, Jens Axboe wrote:
>>>>>> On 03/06/2017 08:29 AM, Avi Kivity wrote:
>>>>>>> On 03/06/2017 05:19 PM, Jens Axboe wrote:
>>>>>>>> On 03/06/2017 01:25 AM, Jan Kara wrote:
>>>>>>>>> On Sun 05-03-17 16:56:21, Avi Kivity wrote:
>>>>>>>>>>> The goal of the patch series is to return -EAGAIN/-EWOULDBLOCK if
>>>>>>>>>>> any of these conditions are met. This way userspace can push most
>>>>>>>>>>> of the write()s to the kernel to the best of its ability to complete
>>>>>>>>>>> and if it returns -EAGAIN, can defer it to another thread.
>>>>>>>>>>>
>>>>>>>>>> Is it not possible to push the iocb to a workqueue? This will allow
>>>>>>>>>> existing userspace to work with the new functionality, unchanged. Any
>>>>>>>>>> userspace implementation would have to do the same thing, so it's not like
>>>>>>>>>> we're saving anything by pushing it there.
>>>>>>>>> That is not easy because until IO is fully submitted, you need some parts
>>>>>>>>> of the context of the process which submits the IO (e.g. memory mappings,
>>>>>>>>> but possibly also other credentials). So you would need to somehow transfer
>>>>>>>>> this information to the workqueue.
>>>>>>>> Outside of technical challenges, the API also needs to return EAGAIN or
>>>>>>>> start blocking at some point. We can't expose a direct connection to
>>>>>>>> queue work like that, and let any user potentially create millions of
>>>>>>>> pending work items (and IOs).
>>>>>>> You wouldn't expect more concurrent events than the maxevents parameter
>>>>>>> that was supplied to io_setup syscall; it should have reserved any
>>>>>>> resources needed.
>>>>>> Doesn't matter what limit you apply, my point still stands - at some
>>>>>> point you have to return EAGAIN, or block. Returning EAGAIN without
>>>>>> the caller having flagged support for that change of behavior would
>>>>>> be problematic.
>>>>> Doesn't it already return EAGAIN (or some other error) if you exceed
>>>>> maxevents?
>>>> It's a setup thing. We check these limits when someone creates an IO
>>>> context, and carve out the specified entries form our global pool. Then
>>>> we free those "resources" when the io context is freed.
>>>>
>>>> Right now I can setup an IO context with 1000 entries on it, yet that
>>>> number has NO bearing on when io_submit() would potentially block or
>>>> return EAGAIN.
>>>>
>>>> We can have a huge gap on the intent signaled by io context setup, and
>>>> the reality imposed by what actually happens on the IO submission side.
>>> Isn't that a bug? Shouldn't that 1001st incomplete io_submit() return
>>> EAGAIN?
>>>
>>> Just tested it, and maxevents is not respected for this:
>>>
>>> io_setup(1, [0x7fc64537f000]) = 0
>>> io_submit(0x7fc64537f000, 10, [{pread, fildes=3, buf=0x1eb4000,
>>> nbytes=4096, offset=0}, {pread, fildes=3, buf=0x1eb4000, nbytes=4096,
>>> offset=0}, {pread, fildes=3, buf=0x1eb4000, nbytes=4096, offset=0},
>>> {pread, fildes=3, buf=0x1eb4000, nbytes=4096, offset=0}, {pread,
>>> fildes=3, buf=0x1eb4000, nbytes=4096, offset=0}, {pread, fildes=3,
>>> buf=0x1eb4000, nbytes=4096, offset=0}, {pread, fildes=3, buf=0x1eb4000,
>>> nbytes=4096, offset=0}, {pread, fildes=3, buf=0x1eb4000, nbytes=4096,
>>> offset=0}, {pread, fildes=3, buf=0x1eb4000, nbytes=4096, offset=0},
>>> {pread, fildes=3, buf=0x1eb4000, nbytes=4096, offset=0}]) = 10
>>>
>>> which is unexpected, to me.
>> ioctx_alloc()
>> {
>> [...]
>>
>> /*
>> * We keep track of the number of available ringbuffer slots, to prevent
>> * overflow (reqs_available), and we also use percpu counters for this.
>> *
>> * So since up to half the slots might be on other cpu's percpu counters
>> * and unavailable, double nr_events so userspace sees what they
>> * expected: additionally, we move req_batch slots to/from percpu
>> * counters at a time, so make sure that isn't 0:
>> */
>> nr_events = max(nr_events, num_possible_cpus() * 4);
>> nr_events *= 2;
>> }
>
> On a 4-lcore desktop:
>
> io_setup(1, [0x7fc210041000]) = 0
> io_submit(0x7fc210041000, 10000, [big array]) = 126
> io_submit(0x7fc210041000, 10000, [big array]) = -1 EAGAIN (Resource
> temporarily unavailable)
>
> so, the user should already expect EAGAIN from io_submit() due to
> resource limits. I'm sure the check could be tightened so that if we do
> have to use a workqueue, we respect the user's limit rather than some
> inflated number.
This is why I previously said that the 1000 requests you potentially
asks for when setting up your IO context has NOTHING to do with when you
will run into EAGAIN. Yes, returning EAGAIN if the app exceeds the
limit that it itself has set is existing behavior and it certainly makes
sense. And it's an easily handled condition, since the app can just
backoff and wait/reap completion events.
But if we allow EAGAIN to bubble up from block request submission, then
that's a change in behavior. This can happen without the app having any
pending IO against that IO context, hence we can return EAGAIN to the
app that then has no reasonable way to handle that condition.
--
Jens Axboe
